Code

pack-objects: reuse data from existing packs.
authorJunio C Hamano <junkio@cox.net>
Thu, 16 Feb 2006 01:34:29 +0000 (17:34 -0800)
committerJunio C Hamano <junkio@cox.net>
Fri, 17 Feb 2006 10:11:38 +0000 (02:11 -0800)
When generating a new pack, notice if we have already needed
objects in existing packs.  If an object is stored deltified,
and its base object is also what we are going to pack, then
reuse the existing deltified representation unconditionally,
bypassing all the expensive find_deltas() and try_deltas()
calls.

Also, notice if what we are going to write out exactly match
what is already in an existing pack (either deltified or just
compressed).  In such a case, we can just copy it instead of
going through the usual uncompressing & recompressing cycle.

Without this patch, in linux-2.6 repository with about 1500
loose objects and a single mega pack:

    $ git-rev-list --objects v2.6.16-rc3 >RL
    $ wc -l RL
    184141 RL
    $ time git-pack-objects p <RL
    Generating pack...
    Done counting 184141 objects.
    Packing 184141 objects....................
    a1fc7b3e537fcb9b3c46b7505df859f0a11e79d2

    real    12m4.323s
    user    11m2.560s
    sys     0m55.950s

With this patch, the same input:

    $ time ../git.junio/git-pack-objects q <RL
    Generating pack...
    Done counting 184141 objects.
    Packing 184141 objects.....................
    a1fc7b3e537fcb9b3c46b7505df859f0a11e79d2
    Total 184141, written 184141, reused 182441

    real    1m2.608s
    user    0m55.090s
    sys     0m1.830s

Signed-off-by: Junio C Hamano <junkio@cox.net>
pack-objects.c
pack.h
sha1_file.c

index c5a5e61605b7aefde9fd3cc8c43a085156671bc9..70fb2afeb8cfdef06b20dc7f96cd427f7c50ab60 100644 (file)
@@ -9,15 +9,31 @@ static const char pack_usage[] = "git-pack-objects [-q] [--non-empty] [--local]
 
 struct object_entry {
        unsigned char sha1[20];
-       unsigned long size;
-       unsigned long offset;
-       unsigned int depth;
-       unsigned int hash;
+       unsigned long size;     /* uncompressed size */
+       unsigned long offset;   /* offset into the final pack file (nonzero if already written) */
+       unsigned int depth;     /* delta depth */
+       unsigned int hash;      /* name hint hash */
        enum object_type type;
-       unsigned long delta_size;
-       struct object_entry *delta;
+       unsigned long delta_size;       /* delta data size (uncompressed) */
+       struct object_entry *delta;     /* delta base object */
+       struct packed_git *in_pack;     /* already in pack */
+       enum object_type in_pack_type;  /* could be delta */
+       unsigned int in_pack_offset;
 };
 
+/*
+ * Objects we are going to pack are colected in objects array (dynamically
+ * expanded).  nr_objects & nr_alloc controls this array.  They are stored
+ * in the order we see -- typically rev-list --objects order that gives us
+ * nice "minimum seek" order.
+ *
+ * sorted-by-sha ans sorted-by-type are arrays of pointers that point at
+ * elements in the objects array.  The former is used to build the pack
+ * index (lists object names in the ascending order to help offset lookup),
+ * and the latter is used to group similar things together by try_delta()
+ * heuristics.
+ */
+
 static unsigned char object_list_sha1[20];
 static int non_empty = 0;
 static int local = 0;
@@ -29,6 +45,135 @@ static const char *base_name;
 static unsigned char pack_file_sha1[20];
 static int progress = 1;
 
+/*
+ * The object names in objects array are hashed with this hashtable,
+ * to help looking up the entry by object name.  Binary search from
+ * sorted_by_sha is also possible but this was easier to code and faster.
+ * This hashtable is built after all the objects are seen.
+ */
+static int *object_ix = NULL;
+static int object_ix_hashsz = 0;
+
+/*
+ * Pack index for existing packs give us easy access to the offsets into
+ * corresponding pack file where each object's data starts, but the entries
+ * do not store the size of the compressed representation (uncompressed
+ * size is easily available by examining the pack entry header).  We build
+ * a hashtable of existing packs (pack_revindex), and keep reverse index
+ * here -- pack index file is sorted by object name mapping to offset; this
+ * pack_revindex[].revindex array is an ordered list of offsets, so if you
+ * know the offset of an object, next offset is where its packed
+ * representation ends.
+ */
+struct pack_revindex {
+       struct packed_git *p;
+       unsigned long *revindex;
+} *pack_revindex = NULL;
+static int pack_revindex_hashsz = 0;
+
+/*
+ * stats
+ */
+static int written = 0;
+static int reused = 0;
+
+static int pack_revindex_ix(struct packed_git *p)
+{
+       unsigned int ui = (unsigned int) p;
+       int i;
+
+       ui = ui ^ (ui >> 16); /* defeat structure alignment */
+       i = (int)(ui % pack_revindex_hashsz);
+       while (pack_revindex[i].p) {
+               if (pack_revindex[i].p == p)
+                       return i;
+               if (++i == pack_revindex_hashsz)
+                       i = 0;
+       }
+       return -1 - i;
+}
+
+static void prepare_pack_ix(void)
+{
+       int num;
+       struct packed_git *p;
+       for (num = 0, p = packed_git; p; p = p->next)
+               num++;
+       if (!num)
+               return;
+       pack_revindex_hashsz = num * 11;
+       pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
+       for (p = packed_git; p; p = p->next) {
+               num = pack_revindex_ix(p);
+               num = - 1 - num;
+               pack_revindex[num].p = p;
+       }
+       /* revindex elements are lazily initialized */
+}
+
+static int cmp_offset(const void *a_, const void *b_)
+{
+       unsigned long a = *(unsigned long *) a_;
+       unsigned long b = *(unsigned long *) b_;
+       if (a < b)
+               return -1;
+       else if (a == b)
+               return 0;
+       else
+               return 1;
+}
+
+/*
+ * Ordered list of offsets of objects in the pack.
+ */
+static void prepare_pack_revindex(struct pack_revindex *rix)
+{
+       struct packed_git *p = rix->p;
+       int num_ent = num_packed_objects(p);
+       int i;
+       void *index = p->index_base + 256;
+
+       rix->revindex = xmalloc(sizeof(unsigned long) * (num_ent + 1));
+       for (i = 0; i < num_ent; i++) {
+               long hl = *((long *)(index + 24 * i));
+               rix->revindex[i] = ntohl(hl);
+       }
+       /* This knows the pack format -- the 20-byte trailer
+        * follows immediately after the last object data.
+        */
+       rix->revindex[num_ent] = p->pack_size - 20;
+       qsort(rix->revindex, num_ent, sizeof(unsigned long), cmp_offset);
+}
+
+static unsigned long find_packed_object_size(struct packed_git *p,
+                                            unsigned long ofs)
+{
+       int num;
+       int lo, hi;
+       struct pack_revindex *rix;
+       unsigned long *revindex;
+       num = pack_revindex_ix(p);
+       if (num < 0)
+               die("internal error: pack revindex uninitialized");
+       rix = &pack_revindex[num];
+       if (!rix->revindex)
+               prepare_pack_revindex(rix);
+       revindex = rix->revindex;
+       lo = 0;
+       hi = num_packed_objects(p) + 1;
+       do {
+               int mi = (lo + hi) / 2;
+               if (revindex[mi] == ofs) {
+                       return revindex[mi+1] - ofs;
+               }
+               else if (ofs < revindex[mi])
+                       hi = mi;
+               else
+                       lo = mi + 1;
+       } while (lo < hi);
+       die("internal error: pack revindex corrupt");
+}
+
 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
 {
        unsigned long othersize, delta_size;
@@ -78,35 +223,52 @@ static unsigned long write_object(struct sha1file *f, struct object_entry *entry
 {
        unsigned long size;
        char type[10];
-       void *buf = read_sha1_file(entry->sha1, type, &size);
+       void *buf;
        unsigned char header[10];
        unsigned hdrlen, datalen;
        enum object_type obj_type;
 
-       if (!buf)
-               die("unable to read %s", sha1_to_hex(entry->sha1));
-       if (size != entry->size)
-               die("object %s size inconsistency (%lu vs %lu)", sha1_to_hex(entry->sha1), size, entry->size);
-
-       /*
-        * The object header is a byte of 'type' followed by zero or
-        * more bytes of length.  For deltas, the 20 bytes of delta sha1
-        * follows that.
-        */
        obj_type = entry->type;
-       if (entry->delta) {
-               buf = delta_against(buf, size, entry);
-               size = entry->delta_size;
-               obj_type = OBJ_DELTA;
+       if (!entry->in_pack ||
+           (obj_type != entry->in_pack_type)) {
+               buf = read_sha1_file(entry->sha1, type, &size);
+               if (!buf)
+                       die("unable to read %s", sha1_to_hex(entry->sha1));
+               if (size != entry->size)
+                       die("object %s size inconsistency (%lu vs %lu)",
+                           sha1_to_hex(entry->sha1), size, entry->size);
+               if (entry->delta) {
+                       buf = delta_against(buf, size, entry);
+                       size = entry->delta_size;
+                       obj_type = OBJ_DELTA;
+               }
+               /*
+                * The object header is a byte of 'type' followed by zero or
+                * more bytes of length.  For deltas, the 20 bytes of delta
+                * sha1 follows that.
+                */
+               hdrlen = encode_header(obj_type, size, header);
+               sha1write(f, header, hdrlen);
+
+               if (entry->delta) {
+                       sha1write(f, entry->delta, 20);
+                       hdrlen += 20;
+               }
+               datalen = sha1write_compressed(f, buf, size);
+               free(buf);
        }
-       hdrlen = encode_header(obj_type, size, header);
-       sha1write(f, header, hdrlen);
-       if (entry->delta) {
-               sha1write(f, entry->delta, 20);
-               hdrlen += 20;
+       else {
+               struct packed_git *p = entry->in_pack;
+               use_packed_git(p);
+
+               datalen = find_packed_object_size(p, entry->in_pack_offset);
+               buf = p->pack_base + entry->in_pack_offset;
+               sha1write(f, buf, datalen);
+               unuse_packed_git(p);
+               hdrlen = 0; /* not really */
+               reused++;
        }
-       datalen = sha1write_compressed(f, buf, size);
-       free(buf);
+       written++;
        return hdrlen + datalen;
 }
 
@@ -148,8 +310,6 @@ static void write_pack_file(void)
                offset = write_one(f, objects + i, offset);
 
        sha1close(f, pack_file_sha1, 1);
-       mb = offset >> 20;
-       offset &= 0xfffff;
 }
 
 static void write_index_file(void)
@@ -196,18 +356,21 @@ static int add_object_entry(unsigned char *sha1, unsigned int hash)
 {
        unsigned int idx = nr_objects;
        struct object_entry *entry;
-
-       if (incremental || local) {
-               struct packed_git *p;
-
-               for (p = packed_git; p; p = p->next) {
-                       struct pack_entry e;
-
-                       if (find_pack_entry_one(sha1, &e, p)) {
-                               if (incremental)
-                                       return 0;
-                               if (local && !p->pack_local)
-                                       return 0;
+       struct packed_git *p;
+       unsigned int found_offset;
+       struct packed_git *found_pack;
+
+       found_pack = NULL;
+       for (p = packed_git; p; p = p->next) {
+               struct pack_entry e;
+               if (find_pack_entry_one(sha1, &e, p)) {
+                       if (incremental)
+                               return 0;
+                       if (local && !p->pack_local)
+                               return 0;
+                       if (!found_pack) {
+                               found_offset = e.offset;
+                               found_pack = e.p;
                        }
                }
        }
@@ -221,30 +384,107 @@ static int add_object_entry(unsigned char *sha1, unsigned int hash)
        memset(entry, 0, sizeof(*entry));
        memcpy(entry->sha1, sha1, 20);
        entry->hash = hash;
+       if (found_pack) {
+               entry->in_pack = found_pack;
+               entry->in_pack_offset = found_offset;
+       }
        nr_objects = idx+1;
        return 1;
 }
 
+static int locate_object_entry_hash(unsigned char *sha1)
+{
+       int i;
+       unsigned int ui;
+       memcpy(&ui, sha1, sizeof(unsigned int));
+       i = ui % object_ix_hashsz;
+       while (0 < object_ix[i]) {
+               if (!memcmp(sha1, objects[object_ix[i]-1].sha1, 20))
+                       return i;
+               if (++i == object_ix_hashsz)
+                       i = 0;
+       }
+       return -1 - i;
+}
+
+static struct object_entry *locate_object_entry(unsigned char *sha1)
+{
+       int i = locate_object_entry_hash(sha1);
+       if (0 <= i)
+               return &objects[object_ix[i]-1];
+       return NULL;
+}
+
 static void check_object(struct object_entry *entry)
 {
        char type[20];
 
-       if (!sha1_object_info(entry->sha1, type, &entry->size)) {
-               if (!strcmp(type, "commit")) {
-                       entry->type = OBJ_COMMIT;
-               } else if (!strcmp(type, "tree")) {
-                       entry->type = OBJ_TREE;
-               } else if (!strcmp(type, "blob")) {
-                       entry->type = OBJ_BLOB;
-               } else if (!strcmp(type, "tag")) {
-                       entry->type = OBJ_TAG;
-               } else
-                       die("unable to pack object %s of type %s",
-                           sha1_to_hex(entry->sha1), type);
+       if (entry->in_pack) {
+               /* Check if it is delta, and the base is also an object
+                * we are going to pack.  If so we will reuse the existing
+                * delta.
+                */
+               unsigned char base[20];
+               unsigned long size;
+               struct object_entry *base_entry;
+               if (!check_reuse_pack_delta(entry->in_pack,
+                                           entry->in_pack_offset,
+                                           base, &size,
+                                           &entry->in_pack_type) &&
+                   (base_entry = locate_object_entry(base))) {
+                       /* We do not know depth at this point, but it
+                        * does not matter.  Getting delta_chain_length
+                        * with packed_object_info_detail() is not so
+                        * expensive, so we could do that later if we
+                        * wanted to.  Calling sha1_object_info to get
+                        * the true size (and later an uncompressed
+                        * representation) of deeply deltified object
+                        * is quite expensive.
+                        */
+                       entry->depth = 1;
+                       /* uncompressed size */
+                       entry->size = entry->delta_size = size;
+                       entry->delta = base_entry;
+                       entry->type = OBJ_DELTA;
+                       return;
+               }
+               /* Otherwise we would do the usual */
        }
-       else
+
+       if (sha1_object_info(entry->sha1, type, &entry->size))
                die("unable to get type of object %s",
                    sha1_to_hex(entry->sha1));
+
+       if (!strcmp(type, "commit")) {
+               entry->type = OBJ_COMMIT;
+       } else if (!strcmp(type, "tree")) {
+               entry->type = OBJ_TREE;
+       } else if (!strcmp(type, "blob")) {
+               entry->type = OBJ_BLOB;
+       } else if (!strcmp(type, "tag")) {
+               entry->type = OBJ_TAG;
+       } else
+               die("unable to pack object %s of type %s",
+                   sha1_to_hex(entry->sha1), type);
+}
+
+static void hash_objects(void)
+{
+       int i;
+       struct object_entry *oe;
+
+       object_ix_hashsz = nr_objects * 2;
+       object_ix = xcalloc(sizeof(int), object_ix_hashsz);
+       for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
+               int ix = locate_object_entry_hash(oe->sha1);
+               if (0 <= ix) {
+                       error("the same object '%s' added twice",
+                             sha1_to_hex(oe->sha1));
+                       continue;
+               }
+               ix = -1 - ix;
+               object_ix[ix] = i + 1;
+       }
 }
 
 static void get_object_details(void)
@@ -252,6 +492,8 @@ static void get_object_details(void)
        int i;
        struct object_entry *entry = objects;
 
+       hash_objects();
+       prepare_pack_ix();
        for (i = 0; i < nr_objects; i++)
                check_object(entry++);
 }
@@ -382,6 +624,13 @@ static void find_deltas(struct object_entry **list, int window, int depth)
                        eye_candy -= nr_objects / 20;
                        fputc('.', stderr);
                }
+
+               if (entry->delta)
+                       /* This happens if we decided to reuse existing
+                        * delta from a pack.
+                        */
+                       continue;
+
                free(n->data);
                n->entry = entry;
                n->data = read_sha1_file(entry->sha1, type, &size);
@@ -411,10 +660,12 @@ static void find_deltas(struct object_entry **list, int window, int depth)
 
 static void prepare_pack(int window, int depth)
 {
-       get_object_details();
-
        if (progress)
                fprintf(stderr, "Packing %d objects", nr_objects);
+       get_object_details();
+       if (progress)
+               fprintf(stderr, ".");
+
        sorted_by_type = create_sorted_list(type_size_sort);
        if (window && depth)
                find_deltas(sorted_by_type, window+1, depth);
@@ -599,5 +850,7 @@ int main(int argc, char **argv)
                        puts(sha1_to_hex(object_list_sha1));
                }
        }
+       fprintf(stderr, "Total %d, written %d, reused %d\n",
+               nr_objects, written, reused);
        return 0;
 }
diff --git a/pack.h b/pack.h
index 9dafa2b6d2c76919594218b12a2a36bd9067f2b7..694e0c56f0b67c1ae9b7a94a243aca7b8859e50c 100644 (file)
--- a/pack.h
+++ b/pack.h
@@ -29,5 +29,7 @@ struct pack_header {
 };
 
 extern int verify_pack(struct packed_git *, int);
-
+extern int check_reuse_pack_delta(struct packed_git *, unsigned long,
+                                 unsigned char *, unsigned long *,
+                                 enum object_type *);
 #endif
index 64cf2454182dbf4cbf1a2ee70dea4823cf8cfaa0..0a3a721e97c89f1dd494831a9321de4668f214c8 100644 (file)
@@ -826,6 +826,25 @@ static unsigned long unpack_object_header(struct packed_git *p, unsigned long of
        return offset;
 }
 
+int check_reuse_pack_delta(struct packed_git *p, unsigned long offset,
+                          unsigned char *base, unsigned long *sizep,
+                          enum object_type *kindp)
+{
+       unsigned long ptr;
+       int status = -1;
+
+       use_packed_git(p);
+       ptr = offset;
+       ptr = unpack_object_header(p, ptr, kindp, sizep);
+       if (*kindp != OBJ_DELTA)
+               goto done;
+       memcpy(base, p->pack_base + ptr, 20);
+       status = 0;
+ done:
+       unuse_packed_git(p);
+       return status;
+}
+
 void packed_object_info_detail(struct pack_entry *e,
                               char *type,
                               unsigned long *size,