Code

Merge branch 'np/malloc-threading'
authorJunio C Hamano <gitster@pobox.com>
Fri, 21 May 2010 11:02:16 +0000 (04:02 -0700)
committerJunio C Hamano <gitster@pobox.com>
Fri, 21 May 2010 11:02:16 +0000 (04:02 -0700)
* np/malloc-threading:
  Thread-safe xmalloc and xrealloc needs a recursive mutex
  Make xmalloc and xrealloc thread-safe

1  2 
builtin/grep.c
builtin/pack-objects.c
git-compat-util.h
wrapper.c

diff --cc builtin/grep.c
index 8e928e217041a159f4a962f0883d740aa84536d7,0000000000000000000000000000000000000000..b194ea3cea531f3a6c81d6d27b5b3028641ba9ce
mode 100644,000000..100644
--- /dev/null
@@@ -1,1042 -1,0 +1,1042 @@@
- #include "thread-utils.h"
 +/*
 + * Builtin "git grep"
 + *
 + * Copyright (c) 2006 Junio C Hamano
 + */
 +#include "cache.h"
 +#include "blob.h"
 +#include "tree.h"
 +#include "commit.h"
 +#include "tag.h"
 +#include "tree-walk.h"
 +#include "builtin.h"
 +#include "parse-options.h"
 +#include "userdiff.h"
 +#include "grep.h"
 +#include "quote.h"
 +#include "dir.h"
 +
 +#ifndef NO_PTHREADS
 +#include <pthread.h>
++#include "thread-utils.h"
 +#endif
 +
 +static char const * const grep_usage[] = {
 +      "git grep [options] [-e] <pattern> [<rev>...] [[--] path...]",
 +      NULL
 +};
 +
 +static int use_threads = 1;
 +
 +#ifndef NO_PTHREADS
 +#define THREADS 8
 +static pthread_t threads[THREADS];
 +
 +static void *load_sha1(const unsigned char *sha1, unsigned long *size,
 +                     const char *name);
 +static void *load_file(const char *filename, size_t *sz);
 +
 +enum work_type {WORK_SHA1, WORK_FILE};
 +
 +/* We use one producer thread and THREADS consumer
 + * threads. The producer adds struct work_items to 'todo' and the
 + * consumers pick work items from the same array.
 + */
 +struct work_item
 +{
 +      enum work_type type;
 +      char *name;
 +
 +      /* if type == WORK_SHA1, then 'identifier' is a SHA1,
 +       * otherwise type == WORK_FILE, and 'identifier' is a NUL
 +       * terminated filename.
 +       */
 +      void *identifier;
 +      char done;
 +      struct strbuf out;
 +};
 +
 +/* In the range [todo_done, todo_start) in 'todo' we have work_items
 + * that have been or are processed by a consumer thread. We haven't
 + * written the result for these to stdout yet.
 + *
 + * The work_items in [todo_start, todo_end) are waiting to be picked
 + * up by a consumer thread.
 + *
 + * The ranges are modulo TODO_SIZE.
 + */
 +#define TODO_SIZE 128
 +static struct work_item todo[TODO_SIZE];
 +static int todo_start;
 +static int todo_end;
 +static int todo_done;
 +
 +/* Has all work items been added? */
 +static int all_work_added;
 +
 +/* This lock protects all the variables above. */
 +static pthread_mutex_t grep_mutex;
 +
 +/* Used to serialize calls to read_sha1_file. */
 +static pthread_mutex_t read_sha1_mutex;
 +
 +#define grep_lock() pthread_mutex_lock(&grep_mutex)
 +#define grep_unlock() pthread_mutex_unlock(&grep_mutex)
 +#define read_sha1_lock() pthread_mutex_lock(&read_sha1_mutex)
 +#define read_sha1_unlock() pthread_mutex_unlock(&read_sha1_mutex)
 +
 +/* Signalled when a new work_item is added to todo. */
 +static pthread_cond_t cond_add;
 +
 +/* Signalled when the result from one work_item is written to
 + * stdout.
 + */
 +static pthread_cond_t cond_write;
 +
 +/* Signalled when we are finished with everything. */
 +static pthread_cond_t cond_result;
 +
 +static int print_hunk_marks_between_files;
 +static int printed_something;
 +
 +static void add_work(enum work_type type, char *name, void *id)
 +{
 +      grep_lock();
 +
 +      while ((todo_end+1) % ARRAY_SIZE(todo) == todo_done) {
 +              pthread_cond_wait(&cond_write, &grep_mutex);
 +      }
 +
 +      todo[todo_end].type = type;
 +      todo[todo_end].name = name;
 +      todo[todo_end].identifier = id;
 +      todo[todo_end].done = 0;
 +      strbuf_reset(&todo[todo_end].out);
 +      todo_end = (todo_end + 1) % ARRAY_SIZE(todo);
 +
 +      pthread_cond_signal(&cond_add);
 +      grep_unlock();
 +}
 +
 +static struct work_item *get_work(void)
 +{
 +      struct work_item *ret;
 +
 +      grep_lock();
 +      while (todo_start == todo_end && !all_work_added) {
 +              pthread_cond_wait(&cond_add, &grep_mutex);
 +      }
 +
 +      if (todo_start == todo_end && all_work_added) {
 +              ret = NULL;
 +      } else {
 +              ret = &todo[todo_start];
 +              todo_start = (todo_start + 1) % ARRAY_SIZE(todo);
 +      }
 +      grep_unlock();
 +      return ret;
 +}
 +
 +static void grep_sha1_async(struct grep_opt *opt, char *name,
 +                          const unsigned char *sha1)
 +{
 +      unsigned char *s;
 +      s = xmalloc(20);
 +      memcpy(s, sha1, 20);
 +      add_work(WORK_SHA1, name, s);
 +}
 +
 +static void grep_file_async(struct grep_opt *opt, char *name,
 +                          const char *filename)
 +{
 +      add_work(WORK_FILE, name, xstrdup(filename));
 +}
 +
 +static void work_done(struct work_item *w)
 +{
 +      int old_done;
 +
 +      grep_lock();
 +      w->done = 1;
 +      old_done = todo_done;
 +      for(; todo[todo_done].done && todo_done != todo_start;
 +          todo_done = (todo_done+1) % ARRAY_SIZE(todo)) {
 +              w = &todo[todo_done];
 +              if (w->out.len) {
 +                      if (print_hunk_marks_between_files && printed_something)
 +                              write_or_die(1, "--\n", 3);
 +                      write_or_die(1, w->out.buf, w->out.len);
 +                      printed_something = 1;
 +              }
 +              free(w->name);
 +              free(w->identifier);
 +      }
 +
 +      if (old_done != todo_done)
 +              pthread_cond_signal(&cond_write);
 +
 +      if (all_work_added && todo_done == todo_end)
 +              pthread_cond_signal(&cond_result);
 +
 +      grep_unlock();
 +}
 +
 +static void *run(void *arg)
 +{
 +      int hit = 0;
 +      struct grep_opt *opt = arg;
 +
 +      while (1) {
 +              struct work_item *w = get_work();
 +              if (!w)
 +                      break;
 +
 +              opt->output_priv = w;
 +              if (w->type == WORK_SHA1) {
 +                      unsigned long sz;
 +                      void* data = load_sha1(w->identifier, &sz, w->name);
 +
 +                      if (data) {
 +                              hit |= grep_buffer(opt, w->name, data, sz);
 +                              free(data);
 +                      }
 +              } else if (w->type == WORK_FILE) {
 +                      size_t sz;
 +                      void* data = load_file(w->identifier, &sz);
 +                      if (data) {
 +                              hit |= grep_buffer(opt, w->name, data, sz);
 +                              free(data);
 +                      }
 +              } else {
 +                      assert(0);
 +              }
 +
 +              work_done(w);
 +      }
 +      free_grep_patterns(arg);
 +      free(arg);
 +
 +      return (void*) (intptr_t) hit;
 +}
 +
 +static void strbuf_out(struct grep_opt *opt, const void *buf, size_t size)
 +{
 +      struct work_item *w = opt->output_priv;
 +      strbuf_add(&w->out, buf, size);
 +}
 +
 +static void start_threads(struct grep_opt *opt)
 +{
 +      int i;
 +
 +      pthread_mutex_init(&grep_mutex, NULL);
 +      pthread_mutex_init(&read_sha1_mutex, NULL);
 +      pthread_cond_init(&cond_add, NULL);
 +      pthread_cond_init(&cond_write, NULL);
 +      pthread_cond_init(&cond_result, NULL);
 +
 +      for (i = 0; i < ARRAY_SIZE(todo); i++) {
 +              strbuf_init(&todo[i].out, 0);
 +      }
 +
 +      for (i = 0; i < ARRAY_SIZE(threads); i++) {
 +              int err;
 +              struct grep_opt *o = grep_opt_dup(opt);
 +              o->output = strbuf_out;
 +              compile_grep_patterns(o);
 +              err = pthread_create(&threads[i], NULL, run, o);
 +
 +              if (err)
 +                      die("grep: failed to create thread: %s",
 +                          strerror(err));
 +      }
 +}
 +
 +static int wait_all(void)
 +{
 +      int hit = 0;
 +      int i;
 +
 +      grep_lock();
 +      all_work_added = 1;
 +
 +      /* Wait until all work is done. */
 +      while (todo_done != todo_end)
 +              pthread_cond_wait(&cond_result, &grep_mutex);
 +
 +      /* Wake up all the consumer threads so they can see that there
 +       * is no more work to do.
 +       */
 +      pthread_cond_broadcast(&cond_add);
 +      grep_unlock();
 +
 +      for (i = 0; i < ARRAY_SIZE(threads); i++) {
 +              void *h;
 +              pthread_join(threads[i], &h);
 +              hit |= (int) (intptr_t) h;
 +      }
 +
 +      pthread_mutex_destroy(&grep_mutex);
 +      pthread_mutex_destroy(&read_sha1_mutex);
 +      pthread_cond_destroy(&cond_add);
 +      pthread_cond_destroy(&cond_write);
 +      pthread_cond_destroy(&cond_result);
 +
 +      return hit;
 +}
 +#else /* !NO_PTHREADS */
 +#define read_sha1_lock()
 +#define read_sha1_unlock()
 +
 +static int wait_all(void)
 +{
 +      return 0;
 +}
 +#endif
 +
 +static int grep_config(const char *var, const char *value, void *cb)
 +{
 +      struct grep_opt *opt = cb;
 +      char *color = NULL;
 +
 +      switch (userdiff_config(var, value)) {
 +      case 0: break;
 +      case -1: return -1;
 +      default: return 0;
 +      }
 +
 +      if (!strcmp(var, "color.grep"))
 +              opt->color = git_config_colorbool(var, value, -1);
 +      else if (!strcmp(var, "color.grep.context"))
 +              color = opt->color_context;
 +      else if (!strcmp(var, "color.grep.filename"))
 +              color = opt->color_filename;
 +      else if (!strcmp(var, "color.grep.function"))
 +              color = opt->color_function;
 +      else if (!strcmp(var, "color.grep.linenumber"))
 +              color = opt->color_lineno;
 +      else if (!strcmp(var, "color.grep.match"))
 +              color = opt->color_match;
 +      else if (!strcmp(var, "color.grep.selected"))
 +              color = opt->color_selected;
 +      else if (!strcmp(var, "color.grep.separator"))
 +              color = opt->color_sep;
 +      else
 +              return git_color_default_config(var, value, cb);
 +      if (color) {
 +              if (!value)
 +                      return config_error_nonbool(var);
 +              color_parse(value, var, color);
 +      }
 +      return 0;
 +}
 +
 +/*
 + * Return non-zero if max_depth is negative or path has no more then max_depth
 + * slashes.
 + */
 +static int accept_subdir(const char *path, int max_depth)
 +{
 +      if (max_depth < 0)
 +              return 1;
 +
 +      while ((path = strchr(path, '/')) != NULL) {
 +              max_depth--;
 +              if (max_depth < 0)
 +                      return 0;
 +              path++;
 +      }
 +      return 1;
 +}
 +
 +/*
 + * Return non-zero if name is a subdirectory of match and is not too deep.
 + */
 +static int is_subdir(const char *name, int namelen,
 +              const char *match, int matchlen, int max_depth)
 +{
 +      if (matchlen > namelen || strncmp(name, match, matchlen))
 +              return 0;
 +
 +      if (name[matchlen] == '\0') /* exact match */
 +              return 1;
 +
 +      if (!matchlen || match[matchlen-1] == '/' || name[matchlen] == '/')
 +              return accept_subdir(name + matchlen + 1, max_depth);
 +
 +      return 0;
 +}
 +
 +/*
 + * git grep pathspecs are somewhat different from diff-tree pathspecs;
 + * pathname wildcards are allowed.
 + */
 +static int pathspec_matches(const char **paths, const char *name, int max_depth)
 +{
 +      int namelen, i;
 +      if (!paths || !*paths)
 +              return accept_subdir(name, max_depth);
 +      namelen = strlen(name);
 +      for (i = 0; paths[i]; i++) {
 +              const char *match = paths[i];
 +              int matchlen = strlen(match);
 +              const char *cp, *meta;
 +
 +              if (is_subdir(name, namelen, match, matchlen, max_depth))
 +                      return 1;
 +              if (!fnmatch(match, name, 0))
 +                      return 1;
 +              if (name[namelen-1] != '/')
 +                      continue;
 +
 +              /* We are being asked if the directory ("name") is worth
 +               * descending into.
 +               *
 +               * Find the longest leading directory name that does
 +               * not have metacharacter in the pathspec; the name
 +               * we are looking at must overlap with that directory.
 +               */
 +              for (cp = match, meta = NULL; cp - match < matchlen; cp++) {
 +                      char ch = *cp;
 +                      if (ch == '*' || ch == '[' || ch == '?') {
 +                              meta = cp;
 +                              break;
 +                      }
 +              }
 +              if (!meta)
 +                      meta = cp; /* fully literal */
 +
 +              if (namelen <= meta - match) {
 +                      /* Looking at "Documentation/" and
 +                       * the pattern says "Documentation/howto/", or
 +                       * "Documentation/diff*.txt".  The name we
 +                       * have should match prefix.
 +                       */
 +                      if (!memcmp(match, name, namelen))
 +                              return 1;
 +                      continue;
 +              }
 +
 +              if (meta - match < namelen) {
 +                      /* Looking at "Documentation/howto/" and
 +                       * the pattern says "Documentation/h*";
 +                       * match up to "Do.../h"; this avoids descending
 +                       * into "Documentation/technical/".
 +                       */
 +                      if (!memcmp(match, name, meta - match))
 +                              return 1;
 +                      continue;
 +              }
 +      }
 +      return 0;
 +}
 +
 +static void *lock_and_read_sha1_file(const unsigned char *sha1, enum object_type *type, unsigned long *size)
 +{
 +      void *data;
 +
 +      if (use_threads) {
 +              read_sha1_lock();
 +              data = read_sha1_file(sha1, type, size);
 +              read_sha1_unlock();
 +      } else {
 +              data = read_sha1_file(sha1, type, size);
 +      }
 +      return data;
 +}
 +
 +static void *load_sha1(const unsigned char *sha1, unsigned long *size,
 +                     const char *name)
 +{
 +      enum object_type type;
 +      void *data = lock_and_read_sha1_file(sha1, &type, size);
 +
 +      if (!data)
 +              error("'%s': unable to read %s", name, sha1_to_hex(sha1));
 +
 +      return data;
 +}
 +
 +static int grep_sha1(struct grep_opt *opt, const unsigned char *sha1,
 +                   const char *filename, int tree_name_len)
 +{
 +      struct strbuf pathbuf = STRBUF_INIT;
 +      char *name;
 +
 +      if (opt->relative && opt->prefix_length) {
 +              quote_path_relative(filename + tree_name_len, -1, &pathbuf,
 +                                  opt->prefix);
 +              strbuf_insert(&pathbuf, 0, filename, tree_name_len);
 +      } else {
 +              strbuf_addstr(&pathbuf, filename);
 +      }
 +
 +      name = strbuf_detach(&pathbuf, NULL);
 +
 +#ifndef NO_PTHREADS
 +      if (use_threads) {
 +              grep_sha1_async(opt, name, sha1);
 +              return 0;
 +      } else
 +#endif
 +      {
 +              int hit;
 +              unsigned long sz;
 +              void *data = load_sha1(sha1, &sz, name);
 +              if (!data)
 +                      hit = 0;
 +              else
 +                      hit = grep_buffer(opt, name, data, sz);
 +
 +              free(data);
 +              free(name);
 +              return hit;
 +      }
 +}
 +
 +static void *load_file(const char *filename, size_t *sz)
 +{
 +      struct stat st;
 +      char *data;
 +      int i;
 +
 +      if (lstat(filename, &st) < 0) {
 +      err_ret:
 +              if (errno != ENOENT)
 +                      error("'%s': %s", filename, strerror(errno));
 +              return 0;
 +      }
 +      if (!S_ISREG(st.st_mode))
 +              return 0;
 +      *sz = xsize_t(st.st_size);
 +      i = open(filename, O_RDONLY);
 +      if (i < 0)
 +              goto err_ret;
 +      data = xmalloc(*sz + 1);
 +      if (st.st_size != read_in_full(i, data, *sz)) {
 +              error("'%s': short read %s", filename, strerror(errno));
 +              close(i);
 +              free(data);
 +              return 0;
 +      }
 +      close(i);
 +      data[*sz] = 0;
 +      return data;
 +}
 +
 +static int grep_file(struct grep_opt *opt, const char *filename)
 +{
 +      struct strbuf buf = STRBUF_INIT;
 +      char *name;
 +
 +      if (opt->relative && opt->prefix_length)
 +              quote_path_relative(filename, -1, &buf, opt->prefix);
 +      else
 +              strbuf_addstr(&buf, filename);
 +      name = strbuf_detach(&buf, NULL);
 +
 +#ifndef NO_PTHREADS
 +      if (use_threads) {
 +              grep_file_async(opt, name, filename);
 +              return 0;
 +      } else
 +#endif
 +      {
 +              int hit;
 +              size_t sz;
 +              void *data = load_file(filename, &sz);
 +              if (!data)
 +                      hit = 0;
 +              else
 +                      hit = grep_buffer(opt, name, data, sz);
 +
 +              free(data);
 +              free(name);
 +              return hit;
 +      }
 +}
 +
 +static int grep_cache(struct grep_opt *opt, const char **paths, int cached)
 +{
 +      int hit = 0;
 +      int nr;
 +      read_cache();
 +
 +      for (nr = 0; nr < active_nr; nr++) {
 +              struct cache_entry *ce = active_cache[nr];
 +              if (!S_ISREG(ce->ce_mode))
 +                      continue;
 +              if (!pathspec_matches(paths, ce->name, opt->max_depth))
 +                      continue;
 +              /*
 +               * If CE_VALID is on, we assume worktree file and its cache entry
 +               * are identical, even if worktree file has been modified, so use
 +               * cache version instead
 +               */
 +              if (cached || (ce->ce_flags & CE_VALID) || ce_skip_worktree(ce)) {
 +                      if (ce_stage(ce))
 +                              continue;
 +                      hit |= grep_sha1(opt, ce->sha1, ce->name, 0);
 +              }
 +              else
 +                      hit |= grep_file(opt, ce->name);
 +              if (ce_stage(ce)) {
 +                      do {
 +                              nr++;
 +                      } while (nr < active_nr &&
 +                               !strcmp(ce->name, active_cache[nr]->name));
 +                      nr--; /* compensate for loop control */
 +              }
 +              if (hit && opt->status_only)
 +                      break;
 +      }
 +      free_grep_patterns(opt);
 +      return hit;
 +}
 +
 +static int grep_tree(struct grep_opt *opt, const char **paths,
 +                   struct tree_desc *tree,
 +                   const char *tree_name, const char *base)
 +{
 +      int len;
 +      int hit = 0;
 +      struct name_entry entry;
 +      char *down;
 +      int tn_len = strlen(tree_name);
 +      struct strbuf pathbuf;
 +
 +      strbuf_init(&pathbuf, PATH_MAX + tn_len);
 +
 +      if (tn_len) {
 +              strbuf_add(&pathbuf, tree_name, tn_len);
 +              strbuf_addch(&pathbuf, ':');
 +              tn_len = pathbuf.len;
 +      }
 +      strbuf_addstr(&pathbuf, base);
 +      len = pathbuf.len;
 +
 +      while (tree_entry(tree, &entry)) {
 +              int te_len = tree_entry_len(entry.path, entry.sha1);
 +              pathbuf.len = len;
 +              strbuf_add(&pathbuf, entry.path, te_len);
 +
 +              if (S_ISDIR(entry.mode))
 +                      /* Match "abc/" against pathspec to
 +                       * decide if we want to descend into "abc"
 +                       * directory.
 +                       */
 +                      strbuf_addch(&pathbuf, '/');
 +
 +              down = pathbuf.buf + tn_len;
 +              if (!pathspec_matches(paths, down, opt->max_depth))
 +                      ;
 +              else if (S_ISREG(entry.mode))
 +                      hit |= grep_sha1(opt, entry.sha1, pathbuf.buf, tn_len);
 +              else if (S_ISDIR(entry.mode)) {
 +                      enum object_type type;
 +                      struct tree_desc sub;
 +                      void *data;
 +                      unsigned long size;
 +
 +                      data = lock_and_read_sha1_file(entry.sha1, &type, &size);
 +                      if (!data)
 +                              die("unable to read tree (%s)",
 +                                  sha1_to_hex(entry.sha1));
 +                      init_tree_desc(&sub, data, size);
 +                      hit |= grep_tree(opt, paths, &sub, tree_name, down);
 +                      free(data);
 +              }
 +              if (hit && opt->status_only)
 +                      break;
 +      }
 +      strbuf_release(&pathbuf);
 +      return hit;
 +}
 +
 +static int grep_object(struct grep_opt *opt, const char **paths,
 +                     struct object *obj, const char *name)
 +{
 +      if (obj->type == OBJ_BLOB)
 +              return grep_sha1(opt, obj->sha1, name, 0);
 +      if (obj->type == OBJ_COMMIT || obj->type == OBJ_TREE) {
 +              struct tree_desc tree;
 +              void *data;
 +              unsigned long size;
 +              int hit;
 +              data = read_object_with_reference(obj->sha1, tree_type,
 +                                                &size, NULL);
 +              if (!data)
 +                      die("unable to read tree (%s)", sha1_to_hex(obj->sha1));
 +              init_tree_desc(&tree, data, size);
 +              hit = grep_tree(opt, paths, &tree, name, "");
 +              free(data);
 +              return hit;
 +      }
 +      die("unable to grep from object of type %s", typename(obj->type));
 +}
 +
 +static int grep_directory(struct grep_opt *opt, const char **paths)
 +{
 +      struct dir_struct dir;
 +      int i, hit = 0;
 +
 +      memset(&dir, 0, sizeof(dir));
 +      setup_standard_excludes(&dir);
 +
 +      fill_directory(&dir, paths);
 +      for (i = 0; i < dir.nr; i++) {
 +              hit |= grep_file(opt, dir.entries[i]->name);
 +              if (hit && opt->status_only)
 +                      break;
 +      }
 +      free_grep_patterns(opt);
 +      return hit;
 +}
 +
 +static int context_callback(const struct option *opt, const char *arg,
 +                          int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      int value;
 +      const char *endp;
 +
 +      if (unset) {
 +              grep_opt->pre_context = grep_opt->post_context = 0;
 +              return 0;
 +      }
 +      value = strtol(arg, (char **)&endp, 10);
 +      if (*endp) {
 +              return error("switch `%c' expects a numerical value",
 +                           opt->short_name);
 +      }
 +      grep_opt->pre_context = grep_opt->post_context = value;
 +      return 0;
 +}
 +
 +static int file_callback(const struct option *opt, const char *arg, int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      FILE *patterns;
 +      int lno = 0;
 +      struct strbuf sb = STRBUF_INIT;
 +
 +      patterns = fopen(arg, "r");
 +      if (!patterns)
 +              die_errno("cannot open '%s'", arg);
 +      while (strbuf_getline(&sb, patterns, '\n') == 0) {
 +              /* ignore empty line like grep does */
 +              if (sb.len == 0)
 +                      continue;
 +              append_grep_pattern(grep_opt, strbuf_detach(&sb, NULL), arg,
 +                                  ++lno, GREP_PATTERN);
 +      }
 +      fclose(patterns);
 +      strbuf_release(&sb);
 +      return 0;
 +}
 +
 +static int not_callback(const struct option *opt, const char *arg, int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      append_grep_pattern(grep_opt, "--not", "command line", 0, GREP_NOT);
 +      return 0;
 +}
 +
 +static int and_callback(const struct option *opt, const char *arg, int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      append_grep_pattern(grep_opt, "--and", "command line", 0, GREP_AND);
 +      return 0;
 +}
 +
 +static int open_callback(const struct option *opt, const char *arg, int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      append_grep_pattern(grep_opt, "(", "command line", 0, GREP_OPEN_PAREN);
 +      return 0;
 +}
 +
 +static int close_callback(const struct option *opt, const char *arg, int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      append_grep_pattern(grep_opt, ")", "command line", 0, GREP_CLOSE_PAREN);
 +      return 0;
 +}
 +
 +static int pattern_callback(const struct option *opt, const char *arg,
 +                          int unset)
 +{
 +      struct grep_opt *grep_opt = opt->value;
 +      append_grep_pattern(grep_opt, arg, "-e option", 0, GREP_PATTERN);
 +      return 0;
 +}
 +
 +static int help_callback(const struct option *opt, const char *arg, int unset)
 +{
 +      return -1;
 +}
 +
 +int cmd_grep(int argc, const char **argv, const char *prefix)
 +{
 +      int hit = 0;
 +      int cached = 0;
 +      int seen_dashdash = 0;
 +      int external_grep_allowed__ignored;
 +      struct grep_opt opt;
 +      struct object_array list = { 0, 0, NULL };
 +      const char **paths = NULL;
 +      int i;
 +      int dummy;
 +      int nongit = 0, use_index = 1;
 +      struct option options[] = {
 +              OPT_BOOLEAN(0, "cached", &cached,
 +                      "search in index instead of in the work tree"),
 +              OPT_BOOLEAN(0, "index", &use_index,
 +                      "--no-index finds in contents not managed by git"),
 +              OPT_GROUP(""),
 +              OPT_BOOLEAN('v', "invert-match", &opt.invert,
 +                      "show non-matching lines"),
 +              OPT_BOOLEAN('i', "ignore-case", &opt.ignore_case,
 +                      "case insensitive matching"),
 +              OPT_BOOLEAN('w', "word-regexp", &opt.word_regexp,
 +                      "match patterns only at word boundaries"),
 +              OPT_SET_INT('a', "text", &opt.binary,
 +                      "process binary files as text", GREP_BINARY_TEXT),
 +              OPT_SET_INT('I', NULL, &opt.binary,
 +                      "don't match patterns in binary files",
 +                      GREP_BINARY_NOMATCH),
 +              { OPTION_INTEGER, 0, "max-depth", &opt.max_depth, "depth",
 +                      "descend at most <depth> levels", PARSE_OPT_NONEG,
 +                      NULL, 1 },
 +              OPT_GROUP(""),
 +              OPT_BIT('E', "extended-regexp", &opt.regflags,
 +                      "use extended POSIX regular expressions", REG_EXTENDED),
 +              OPT_NEGBIT('G', "basic-regexp", &opt.regflags,
 +                      "use basic POSIX regular expressions (default)",
 +                      REG_EXTENDED),
 +              OPT_BOOLEAN('F', "fixed-strings", &opt.fixed,
 +                      "interpret patterns as fixed strings"),
 +              OPT_GROUP(""),
 +              OPT_BOOLEAN('n', NULL, &opt.linenum, "show line numbers"),
 +              OPT_NEGBIT('h', NULL, &opt.pathname, "don't show filenames", 1),
 +              OPT_BIT('H', NULL, &opt.pathname, "show filenames", 1),
 +              OPT_NEGBIT(0, "full-name", &opt.relative,
 +                      "show filenames relative to top directory", 1),
 +              OPT_BOOLEAN('l', "files-with-matches", &opt.name_only,
 +                      "show only filenames instead of matching lines"),
 +              OPT_BOOLEAN(0, "name-only", &opt.name_only,
 +                      "synonym for --files-with-matches"),
 +              OPT_BOOLEAN('L', "files-without-match",
 +                      &opt.unmatch_name_only,
 +                      "show only the names of files without match"),
 +              OPT_BOOLEAN('z', "null", &opt.null_following_name,
 +                      "print NUL after filenames"),
 +              OPT_BOOLEAN('c', "count", &opt.count,
 +                      "show the number of matches instead of matching lines"),
 +              OPT__COLOR(&opt.color, "highlight matches"),
 +              OPT_GROUP(""),
 +              OPT_CALLBACK('C', NULL, &opt, "n",
 +                      "show <n> context lines before and after matches",
 +                      context_callback),
 +              OPT_INTEGER('B', NULL, &opt.pre_context,
 +                      "show <n> context lines before matches"),
 +              OPT_INTEGER('A', NULL, &opt.post_context,
 +                      "show <n> context lines after matches"),
 +              OPT_NUMBER_CALLBACK(&opt, "shortcut for -C NUM",
 +                      context_callback),
 +              OPT_BOOLEAN('p', "show-function", &opt.funcname,
 +                      "show a line with the function name before matches"),
 +              OPT_GROUP(""),
 +              OPT_CALLBACK('f', NULL, &opt, "file",
 +                      "read patterns from file", file_callback),
 +              { OPTION_CALLBACK, 'e', NULL, &opt, "pattern",
 +                      "match <pattern>", PARSE_OPT_NONEG, pattern_callback },
 +              { OPTION_CALLBACK, 0, "and", &opt, NULL,
 +                "combine patterns specified with -e",
 +                PARSE_OPT_NOARG | PARSE_OPT_NONEG, and_callback },
 +              OPT_BOOLEAN(0, "or", &dummy, ""),
 +              { OPTION_CALLBACK, 0, "not", &opt, NULL, "",
 +                PARSE_OPT_NOARG | PARSE_OPT_NONEG, not_callback },
 +              { OPTION_CALLBACK, '(', NULL, &opt, NULL, "",
 +                PARSE_OPT_NOARG | PARSE_OPT_NONEG | PARSE_OPT_NODASH,
 +                open_callback },
 +              { OPTION_CALLBACK, ')', NULL, &opt, NULL, "",
 +                PARSE_OPT_NOARG | PARSE_OPT_NONEG | PARSE_OPT_NODASH,
 +                close_callback },
 +              OPT_BOOLEAN('q', "quiet", &opt.status_only,
 +                          "indicate hit with exit status without output"),
 +              OPT_BOOLEAN(0, "all-match", &opt.all_match,
 +                      "show only matches from files that match all patterns"),
 +              OPT_GROUP(""),
 +              OPT_BOOLEAN(0, "ext-grep", &external_grep_allowed__ignored,
 +                          "allow calling of grep(1) (ignored by this build)"),
 +              { OPTION_CALLBACK, 0, "help-all", &options, NULL, "show usage",
 +                PARSE_OPT_HIDDEN | PARSE_OPT_NOARG, help_callback },
 +              OPT_END()
 +      };
 +
 +      prefix = setup_git_directory_gently(&nongit);
 +
 +      /*
 +       * 'git grep -h', unlike 'git grep -h <pattern>', is a request
 +       * to show usage information and exit.
 +       */
 +      if (argc == 2 && !strcmp(argv[1], "-h"))
 +              usage_with_options(grep_usage, options);
 +
 +      memset(&opt, 0, sizeof(opt));
 +      opt.prefix = prefix;
 +      opt.prefix_length = (prefix && *prefix) ? strlen(prefix) : 0;
 +      opt.relative = 1;
 +      opt.pathname = 1;
 +      opt.pattern_tail = &opt.pattern_list;
 +      opt.header_tail = &opt.header_list;
 +      opt.regflags = REG_NEWLINE;
 +      opt.max_depth = -1;
 +
 +      strcpy(opt.color_context, "");
 +      strcpy(opt.color_filename, "");
 +      strcpy(opt.color_function, "");
 +      strcpy(opt.color_lineno, "");
 +      strcpy(opt.color_match, GIT_COLOR_BOLD_RED);
 +      strcpy(opt.color_selected, "");
 +      strcpy(opt.color_sep, GIT_COLOR_CYAN);
 +      opt.color = -1;
 +      git_config(grep_config, &opt);
 +      if (opt.color == -1)
 +              opt.color = git_use_color_default;
 +
 +      /*
 +       * If there is no -- then the paths must exist in the working
 +       * tree.  If there is no explicit pattern specified with -e or
 +       * -f, we take the first unrecognized non option to be the
 +       * pattern, but then what follows it must be zero or more
 +       * valid refs up to the -- (if exists), and then existing
 +       * paths.  If there is an explicit pattern, then the first
 +       * unrecognized non option is the beginning of the refs list
 +       * that continues up to the -- (if exists), and then paths.
 +       */
 +      argc = parse_options(argc, argv, prefix, options, grep_usage,
 +                           PARSE_OPT_KEEP_DASHDASH |
 +                           PARSE_OPT_STOP_AT_NON_OPTION |
 +                           PARSE_OPT_NO_INTERNAL_HELP);
 +
 +      if (use_index && nongit)
 +              /* die the same way as if we did it at the beginning */
 +              setup_git_directory();
 +
 +      /*
 +       * skip a -- separator; we know it cannot be
 +       * separating revisions from pathnames if
 +       * we haven't even had any patterns yet
 +       */
 +      if (argc > 0 && !opt.pattern_list && !strcmp(argv[0], "--")) {
 +              argv++;
 +              argc--;
 +      }
 +
 +      /* First unrecognized non-option token */
 +      if (argc > 0 && !opt.pattern_list) {
 +              append_grep_pattern(&opt, argv[0], "command line", 0,
 +                                  GREP_PATTERN);
 +              argv++;
 +              argc--;
 +      }
 +
 +      if (!opt.pattern_list)
 +              die("no pattern given.");
 +      if (!opt.fixed && opt.ignore_case)
 +              opt.regflags |= REG_ICASE;
 +      if ((opt.regflags != REG_NEWLINE) && opt.fixed)
 +              die("cannot mix --fixed-strings and regexp");
 +
 +#ifndef NO_PTHREADS
 +      if (online_cpus() == 1 || !grep_threads_ok(&opt))
 +              use_threads = 0;
 +
 +      if (use_threads) {
 +              if (opt.pre_context || opt.post_context)
 +                      print_hunk_marks_between_files = 1;
 +              start_threads(&opt);
 +      }
 +#else
 +      use_threads = 0;
 +#endif
 +
 +      compile_grep_patterns(&opt);
 +
 +      /* Check revs and then paths */
 +      for (i = 0; i < argc; i++) {
 +              const char *arg = argv[i];
 +              unsigned char sha1[20];
 +              /* Is it a rev? */
 +              if (!get_sha1(arg, sha1)) {
 +                      struct object *object = parse_object(sha1);
 +                      if (!object)
 +                              die("bad object %s", arg);
 +                      add_object_array(object, arg, &list);
 +                      continue;
 +              }
 +              if (!strcmp(arg, "--")) {
 +                      i++;
 +                      seen_dashdash = 1;
 +              }
 +              break;
 +      }
 +
 +      /* The rest are paths */
 +      if (!seen_dashdash) {
 +              int j;
 +              for (j = i; j < argc; j++)
 +                      verify_filename(prefix, argv[j]);
 +      }
 +
 +      if (i < argc)
 +              paths = get_pathspec(prefix, argv + i);
 +      else if (prefix) {
 +              paths = xcalloc(2, sizeof(const char *));
 +              paths[0] = prefix;
 +              paths[1] = NULL;
 +      }
 +
 +      if (!use_index) {
 +              int hit;
 +              if (cached)
 +                      die("--cached cannot be used with --no-index.");
 +              if (list.nr)
 +                      die("--no-index cannot be used with revs.");
 +              hit = grep_directory(&opt, paths);
 +              if (use_threads)
 +                      hit |= wait_all();
 +              return !hit;
 +      }
 +
 +      if (!list.nr) {
 +              int hit;
 +              if (!cached)
 +                      setup_work_tree();
 +
 +              hit = grep_cache(&opt, paths, cached);
 +              if (use_threads)
 +                      hit |= wait_all();
 +              return !hit;
 +      }
 +
 +      if (cached)
 +              die("both --cached and trees are given.");
 +
 +      for (i = 0; i < list.nr; i++) {
 +              struct object *real_obj;
 +              real_obj = deref_tag(list.objects[i].item, NULL, 0);
 +              if (grep_object(&opt, paths, real_obj, list.objects[i].name)) {
 +                      hit = 1;
 +                      if (opt.status_only)
 +                              break;
 +              }
 +      }
 +
 +      if (use_threads)
 +              hit |= wait_all();
 +      free_grep_patterns(&opt);
 +      return !hit;
 +}
index 97802585ea3ac69ac6ed2e7995605bdcae84558e,0000000000000000000000000000000000000000..214d7ef2b12d41e8d70539fe4a4b16fde71a780a
mode 100644,000000..100644
--- /dev/null
@@@ -1,2334 -1,0 +1,2343 @@@
- #include "thread-utils.h"
 +#include "builtin.h"
 +#include "cache.h"
 +#include "attr.h"
 +#include "object.h"
 +#include "blob.h"
 +#include "commit.h"
 +#include "tag.h"
 +#include "tree.h"
 +#include "delta.h"
 +#include "pack.h"
 +#include "pack-revindex.h"
 +#include "csum-file.h"
 +#include "tree-walk.h"
 +#include "diff.h"
 +#include "revision.h"
 +#include "list-objects.h"
 +#include "progress.h"
 +#include "refs.h"
 +
 +#ifndef NO_PTHREADS
-       pthread_mutex_init(&read_mutex, NULL);
 +#include <pthread.h>
++#include "thread-utils.h"
 +#endif
 +
 +static const char pack_usage[] =
 +  "git pack-objects [{ -q | --progress | --all-progress }]\n"
 +  "        [--all-progress-implied]\n"
 +  "        [--max-pack-size=N] [--local] [--incremental]\n"
 +  "        [--window=N] [--window-memory=N] [--depth=N]\n"
 +  "        [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset]\n"
 +  "        [--threads=N] [--non-empty] [--revs [--unpacked | --all]*]\n"
 +  "        [--reflog] [--stdout | base-name] [--include-tag]\n"
 +  "        [--keep-unreachable | --unpack-unreachable \n"
 +  "        [<ref-list | <object-list]";
 +
 +struct object_entry {
 +      struct pack_idx_entry idx;
 +      unsigned long size;     /* uncompressed size */
 +      struct packed_git *in_pack;     /* already in pack */
 +      off_t in_pack_offset;
 +      struct object_entry *delta;     /* delta base object */
 +      struct object_entry *delta_child; /* deltified objects who bases me */
 +      struct object_entry *delta_sibling; /* other deltified objects who
 +                                           * uses the same base as me
 +                                           */
 +      void *delta_data;       /* cached delta (uncompressed) */
 +      unsigned long delta_size;       /* delta data size (uncompressed) */
 +      unsigned long z_delta_size;     /* delta data size (compressed) */
 +      unsigned int hash;      /* name hint hash */
 +      enum object_type type;
 +      enum object_type in_pack_type;  /* could be delta */
 +      unsigned char in_pack_header_size;
 +      unsigned char preferred_base; /* we do not pack this, but is available
 +                                     * to be used as the base object to delta
 +                                     * objects against.
 +                                     */
 +      unsigned char no_try_delta;
 +};
 +
 +/*
 + * Objects we are going to pack are collected 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.
 + */
 +static struct object_entry *objects;
 +static struct pack_idx_entry **written_list;
 +static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
 +
 +static int non_empty;
 +static int reuse_delta = 1, reuse_object = 1;
 +static int keep_unreachable, unpack_unreachable, include_tag;
 +static int local;
 +static int incremental;
 +static int ignore_packed_keep;
 +static int allow_ofs_delta;
 +static const char *base_name;
 +static int progress = 1;
 +static int window = 10;
 +static unsigned long pack_size_limit, pack_size_limit_cfg;
 +static int depth = 50;
 +static int delta_search_threads;
 +static int pack_to_stdout;
 +static int num_preferred_base;
 +static struct progress *progress_state;
 +static int pack_compression_level = Z_DEFAULT_COMPRESSION;
 +static int pack_compression_seen;
 +
 +static unsigned long delta_cache_size = 0;
 +static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
 +static unsigned long cache_max_small_delta_size = 1000;
 +
 +static unsigned long window_memory_limit = 0;
 +
 +/*
 + * The object names in objects array are hashed with this hashtable,
 + * to help looking up the entry by object name.
 + * This hashtable is built after all the objects are seen.
 + */
 +static int *object_ix;
 +static int object_ix_hashsz;
 +
 +/*
 + * stats
 + */
 +static uint32_t written, written_delta;
 +static uint32_t reused, reused_delta;
 +
 +
 +static void *get_delta(struct object_entry *entry)
 +{
 +      unsigned long size, base_size, delta_size;
 +      void *buf, *base_buf, *delta_buf;
 +      enum object_type type;
 +
 +      buf = read_sha1_file(entry->idx.sha1, &type, &size);
 +      if (!buf)
 +              die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 +      base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 +      if (!base_buf)
 +              die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 +      delta_buf = diff_delta(base_buf, base_size,
 +                             buf, size, &delta_size, 0);
 +      if (!delta_buf || delta_size != entry->delta_size)
 +              die("delta size changed");
 +      free(buf);
 +      free(base_buf);
 +      return delta_buf;
 +}
 +
 +static unsigned long do_compress(void **pptr, unsigned long size)
 +{
 +      z_stream stream;
 +      void *in, *out;
 +      unsigned long maxsize;
 +
 +      memset(&stream, 0, sizeof(stream));
 +      deflateInit(&stream, pack_compression_level);
 +      maxsize = deflateBound(&stream, size);
 +
 +      in = *pptr;
 +      out = xmalloc(maxsize);
 +      *pptr = out;
 +
 +      stream.next_in = in;
 +      stream.avail_in = size;
 +      stream.next_out = out;
 +      stream.avail_out = maxsize;
 +      while (deflate(&stream, Z_FINISH) == Z_OK)
 +              ; /* nothing */
 +      deflateEnd(&stream);
 +
 +      free(in);
 +      return stream.total_out;
 +}
 +
 +/*
 + * we are going to reuse the existing object data as is.  make
 + * sure it is not corrupt.
 + */
 +static int check_pack_inflate(struct packed_git *p,
 +              struct pack_window **w_curs,
 +              off_t offset,
 +              off_t len,
 +              unsigned long expect)
 +{
 +      z_stream stream;
 +      unsigned char fakebuf[4096], *in;
 +      int st;
 +
 +      memset(&stream, 0, sizeof(stream));
 +      git_inflate_init(&stream);
 +      do {
 +              in = use_pack(p, w_curs, offset, &stream.avail_in);
 +              stream.next_in = in;
 +              stream.next_out = fakebuf;
 +              stream.avail_out = sizeof(fakebuf);
 +              st = git_inflate(&stream, Z_FINISH);
 +              offset += stream.next_in - in;
 +      } while (st == Z_OK || st == Z_BUF_ERROR);
 +      git_inflate_end(&stream);
 +      return (st == Z_STREAM_END &&
 +              stream.total_out == expect &&
 +              stream.total_in == len) ? 0 : -1;
 +}
 +
 +static void copy_pack_data(struct sha1file *f,
 +              struct packed_git *p,
 +              struct pack_window **w_curs,
 +              off_t offset,
 +              off_t len)
 +{
 +      unsigned char *in;
 +      unsigned int avail;
 +
 +      while (len) {
 +              in = use_pack(p, w_curs, offset, &avail);
 +              if (avail > len)
 +                      avail = (unsigned int)len;
 +              sha1write(f, in, avail);
 +              offset += avail;
 +              len -= avail;
 +      }
 +}
 +
 +static unsigned long write_object(struct sha1file *f,
 +                                struct object_entry *entry,
 +                                off_t write_offset)
 +{
 +      unsigned long size, limit, datalen;
 +      void *buf;
 +      unsigned char header[10], dheader[10];
 +      unsigned hdrlen;
 +      enum object_type type;
 +      int usable_delta, to_reuse;
 +
 +      if (!pack_to_stdout)
 +              crc32_begin(f);
 +
 +      type = entry->type;
 +
 +      /* apply size limit if limited packsize and not first object */
 +      if (!pack_size_limit || !nr_written)
 +              limit = 0;
 +      else if (pack_size_limit <= write_offset)
 +              /*
 +               * the earlier object did not fit the limit; avoid
 +               * mistaking this with unlimited (i.e. limit = 0).
 +               */
 +              limit = 1;
 +      else
 +              limit = pack_size_limit - write_offset;
 +
 +      if (!entry->delta)
 +              usable_delta = 0;       /* no delta */
 +      else if (!pack_size_limit)
 +             usable_delta = 1;        /* unlimited packfile */
 +      else if (entry->delta->idx.offset == (off_t)-1)
 +              usable_delta = 0;       /* base was written to another pack */
 +      else if (entry->delta->idx.offset)
 +              usable_delta = 1;       /* base already exists in this pack */
 +      else
 +              usable_delta = 0;       /* base could end up in another pack */
 +
 +      if (!reuse_object)
 +              to_reuse = 0;   /* explicit */
 +      else if (!entry->in_pack)
 +              to_reuse = 0;   /* can't reuse what we don't have */
 +      else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
 +                              /* check_object() decided it for us ... */
 +              to_reuse = usable_delta;
 +                              /* ... but pack split may override that */
 +      else if (type != entry->in_pack_type)
 +              to_reuse = 0;   /* pack has delta which is unusable */
 +      else if (entry->delta)
 +              to_reuse = 0;   /* we want to pack afresh */
 +      else
 +              to_reuse = 1;   /* we have it in-pack undeltified,
 +                               * and we do not need to deltify it.
 +                               */
 +
 +      if (!to_reuse) {
 +              no_reuse:
 +              if (!usable_delta) {
 +                      buf = read_sha1_file(entry->idx.sha1, &type, &size);
 +                      if (!buf)
 +                              die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 +                      /*
 +                       * make sure no cached delta data remains from a
 +                       * previous attempt before a pack split occurred.
 +                       */
 +                      free(entry->delta_data);
 +                      entry->delta_data = NULL;
 +                      entry->z_delta_size = 0;
 +              } else if (entry->delta_data) {
 +                      size = entry->delta_size;
 +                      buf = entry->delta_data;
 +                      entry->delta_data = NULL;
 +                      type = (allow_ofs_delta && entry->delta->idx.offset) ?
 +                              OBJ_OFS_DELTA : OBJ_REF_DELTA;
 +              } else {
 +                      buf = get_delta(entry);
 +                      size = entry->delta_size;
 +                      type = (allow_ofs_delta && entry->delta->idx.offset) ?
 +                              OBJ_OFS_DELTA : OBJ_REF_DELTA;
 +              }
 +
 +              if (entry->z_delta_size)
 +                      datalen = entry->z_delta_size;
 +              else
 +                      datalen = do_compress(&buf, size);
 +
 +              /*
 +               * The object header is a byte of 'type' followed by zero or
 +               * more bytes of length.
 +               */
 +              hdrlen = encode_in_pack_object_header(type, size, header);
 +
 +              if (type == OBJ_OFS_DELTA) {
 +                      /*
 +                       * Deltas with relative base contain an additional
 +                       * encoding of the relative offset for the delta
 +                       * base from this object's position in the pack.
 +                       */
 +                      off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 +                      unsigned pos = sizeof(dheader) - 1;
 +                      dheader[pos] = ofs & 127;
 +                      while (ofs >>= 7)
 +                              dheader[--pos] = 128 | (--ofs & 127);
 +                      if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 +                              free(buf);
 +                              return 0;
 +                      }
 +                      sha1write(f, header, hdrlen);
 +                      sha1write(f, dheader + pos, sizeof(dheader) - pos);
 +                      hdrlen += sizeof(dheader) - pos;
 +              } else if (type == OBJ_REF_DELTA) {
 +                      /*
 +                       * Deltas with a base reference contain
 +                       * an additional 20 bytes for the base sha1.
 +                       */
 +                      if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 +                              free(buf);
 +                              return 0;
 +                      }
 +                      sha1write(f, header, hdrlen);
 +                      sha1write(f, entry->delta->idx.sha1, 20);
 +                      hdrlen += 20;
 +              } else {
 +                      if (limit && hdrlen + datalen + 20 >= limit) {
 +                              free(buf);
 +                              return 0;
 +                      }
 +                      sha1write(f, header, hdrlen);
 +              }
 +              sha1write(f, buf, datalen);
 +              free(buf);
 +      }
 +      else {
 +              struct packed_git *p = entry->in_pack;
 +              struct pack_window *w_curs = NULL;
 +              struct revindex_entry *revidx;
 +              off_t offset;
 +
 +              if (entry->delta)
 +                      type = (allow_ofs_delta && entry->delta->idx.offset) ?
 +                              OBJ_OFS_DELTA : OBJ_REF_DELTA;
 +              hdrlen = encode_in_pack_object_header(type, entry->size, header);
 +
 +              offset = entry->in_pack_offset;
 +              revidx = find_pack_revindex(p, offset);
 +              datalen = revidx[1].offset - offset;
 +              if (!pack_to_stdout && p->index_version > 1 &&
 +                  check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 +                      error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 +                      unuse_pack(&w_curs);
 +                      goto no_reuse;
 +              }
 +
 +              offset += entry->in_pack_header_size;
 +              datalen -= entry->in_pack_header_size;
 +              if (!pack_to_stdout && p->index_version == 1 &&
 +                  check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 +                      error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 +                      unuse_pack(&w_curs);
 +                      goto no_reuse;
 +              }
 +
 +              if (type == OBJ_OFS_DELTA) {
 +                      off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 +                      unsigned pos = sizeof(dheader) - 1;
 +                      dheader[pos] = ofs & 127;
 +                      while (ofs >>= 7)
 +                              dheader[--pos] = 128 | (--ofs & 127);
 +                      if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 +                              unuse_pack(&w_curs);
 +                              return 0;
 +                      }
 +                      sha1write(f, header, hdrlen);
 +                      sha1write(f, dheader + pos, sizeof(dheader) - pos);
 +                      hdrlen += sizeof(dheader) - pos;
 +                      reused_delta++;
 +              } else if (type == OBJ_REF_DELTA) {
 +                      if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 +                              unuse_pack(&w_curs);
 +                              return 0;
 +                      }
 +                      sha1write(f, header, hdrlen);
 +                      sha1write(f, entry->delta->idx.sha1, 20);
 +                      hdrlen += 20;
 +                      reused_delta++;
 +              } else {
 +                      if (limit && hdrlen + datalen + 20 >= limit) {
 +                              unuse_pack(&w_curs);
 +                              return 0;
 +                      }
 +                      sha1write(f, header, hdrlen);
 +              }
 +              copy_pack_data(f, p, &w_curs, offset, datalen);
 +              unuse_pack(&w_curs);
 +              reused++;
 +      }
 +      if (usable_delta)
 +              written_delta++;
 +      written++;
 +      if (!pack_to_stdout)
 +              entry->idx.crc32 = crc32_end(f);
 +      return hdrlen + datalen;
 +}
 +
 +static int write_one(struct sha1file *f,
 +                             struct object_entry *e,
 +                             off_t *offset)
 +{
 +      unsigned long size;
 +
 +      /* offset is non zero if object is written already. */
 +      if (e->idx.offset || e->preferred_base)
 +              return -1;
 +
 +      /* if we are deltified, write out base object first. */
 +      if (e->delta && !write_one(f, e->delta, offset))
 +              return 0;
 +
 +      e->idx.offset = *offset;
 +      size = write_object(f, e, *offset);
 +      if (!size) {
 +              e->idx.offset = 0;
 +              return 0;
 +      }
 +      written_list[nr_written++] = &e->idx;
 +
 +      /* make sure off_t is sufficiently large not to wrap */
 +      if (*offset > *offset + size)
 +              die("pack too large for current definition of off_t");
 +      *offset += size;
 +      return 1;
 +}
 +
 +static void write_pack_file(void)
 +{
 +      uint32_t i = 0, j;
 +      struct sha1file *f;
 +      off_t offset;
 +      struct pack_header hdr;
 +      uint32_t nr_remaining = nr_result;
 +      time_t last_mtime = 0;
 +
 +      if (progress > pack_to_stdout)
 +              progress_state = start_progress("Writing objects", nr_result);
 +      written_list = xmalloc(nr_objects * sizeof(*written_list));
 +
 +      do {
 +              unsigned char sha1[20];
 +              char *pack_tmp_name = NULL;
 +
 +              if (pack_to_stdout) {
 +                      f = sha1fd_throughput(1, "<stdout>", progress_state);
 +              } else {
 +                      char tmpname[PATH_MAX];
 +                      int fd;
 +                      fd = odb_mkstemp(tmpname, sizeof(tmpname),
 +                                       "pack/tmp_pack_XXXXXX");
 +                      pack_tmp_name = xstrdup(tmpname);
 +                      f = sha1fd(fd, pack_tmp_name);
 +              }
 +
 +              hdr.hdr_signature = htonl(PACK_SIGNATURE);
 +              hdr.hdr_version = htonl(PACK_VERSION);
 +              hdr.hdr_entries = htonl(nr_remaining);
 +              sha1write(f, &hdr, sizeof(hdr));
 +              offset = sizeof(hdr);
 +              nr_written = 0;
 +              for (; i < nr_objects; i++) {
 +                      if (!write_one(f, objects + i, &offset))
 +                              break;
 +                      display_progress(progress_state, written);
 +              }
 +
 +              /*
 +               * Did we write the wrong # entries in the header?
 +               * If so, rewrite it like in fast-import
 +               */
 +              if (pack_to_stdout) {
 +                      sha1close(f, sha1, CSUM_CLOSE);
 +              } else if (nr_written == nr_remaining) {
 +                      sha1close(f, sha1, CSUM_FSYNC);
 +              } else {
 +                      int fd = sha1close(f, sha1, 0);
 +                      fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 +                                               nr_written, sha1, offset);
 +                      close(fd);
 +              }
 +
 +              if (!pack_to_stdout) {
 +                      struct stat st;
 +                      const char *idx_tmp_name;
 +                      char tmpname[PATH_MAX];
 +
 +                      idx_tmp_name = write_idx_file(NULL, written_list,
 +                                                    nr_written, sha1);
 +
 +                      snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
 +                               base_name, sha1_to_hex(sha1));
 +                      free_pack_by_name(tmpname);
 +                      if (adjust_shared_perm(pack_tmp_name))
 +                              die_errno("unable to make temporary pack file readable");
 +                      if (rename(pack_tmp_name, tmpname))
 +                              die_errno("unable to rename temporary pack file");
 +
 +                      /*
 +                       * Packs are runtime accessed in their mtime
 +                       * order since newer packs are more likely to contain
 +                       * younger objects.  So if we are creating multiple
 +                       * packs then we should modify the mtime of later ones
 +                       * to preserve this property.
 +                       */
 +                      if (stat(tmpname, &st) < 0) {
 +                              warning("failed to stat %s: %s",
 +                                      tmpname, strerror(errno));
 +                      } else if (!last_mtime) {
 +                              last_mtime = st.st_mtime;
 +                      } else {
 +                              struct utimbuf utb;
 +                              utb.actime = st.st_atime;
 +                              utb.modtime = --last_mtime;
 +                              if (utime(tmpname, &utb) < 0)
 +                                      warning("failed utime() on %s: %s",
 +                                              tmpname, strerror(errno));
 +                      }
 +
 +                      snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
 +                               base_name, sha1_to_hex(sha1));
 +                      if (adjust_shared_perm(idx_tmp_name))
 +                              die_errno("unable to make temporary index file readable");
 +                      if (rename(idx_tmp_name, tmpname))
 +                              die_errno("unable to rename temporary index file");
 +
 +                      free((void *) idx_tmp_name);
 +                      free(pack_tmp_name);
 +                      puts(sha1_to_hex(sha1));
 +              }
 +
 +              /* mark written objects as written to previous pack */
 +              for (j = 0; j < nr_written; j++) {
 +                      written_list[j]->offset = (off_t)-1;
 +              }
 +              nr_remaining -= nr_written;
 +      } while (nr_remaining && i < nr_objects);
 +
 +      free(written_list);
 +      stop_progress(&progress_state);
 +      if (written != nr_result)
 +              die("wrote %"PRIu32" objects while expecting %"PRIu32,
 +                      written, nr_result);
 +}
 +
 +static int locate_object_entry_hash(const 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 (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
 +                      return i;
 +              if (++i == object_ix_hashsz)
 +                      i = 0;
 +      }
 +      return -1 - i;
 +}
 +
 +static struct object_entry *locate_object_entry(const unsigned char *sha1)
 +{
 +      int i;
 +
 +      if (!object_ix_hashsz)
 +              return NULL;
 +
 +      i = locate_object_entry_hash(sha1);
 +      if (0 <= i)
 +              return &objects[object_ix[i]-1];
 +      return NULL;
 +}
 +
 +static void rehash_objects(void)
 +{
 +      uint32_t i;
 +      struct object_entry *oe;
 +
 +      object_ix_hashsz = nr_objects * 3;
 +      if (object_ix_hashsz < 1024)
 +              object_ix_hashsz = 1024;
 +      object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
 +      memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
 +      for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
 +              int ix = locate_object_entry_hash(oe->idx.sha1);
 +              if (0 <= ix)
 +                      continue;
 +              ix = -1 - ix;
 +              object_ix[ix] = i + 1;
 +      }
 +}
 +
 +static unsigned name_hash(const char *name)
 +{
 +      unsigned c, hash = 0;
 +
 +      if (!name)
 +              return 0;
 +
 +      /*
 +       * This effectively just creates a sortable number from the
 +       * last sixteen non-whitespace characters. Last characters
 +       * count "most", so things that end in ".c" sort together.
 +       */
 +      while ((c = *name++) != 0) {
 +              if (isspace(c))
 +                      continue;
 +              hash = (hash >> 2) + (c << 24);
 +      }
 +      return hash;
 +}
 +
 +static void setup_delta_attr_check(struct git_attr_check *check)
 +{
 +      static struct git_attr *attr_delta;
 +
 +      if (!attr_delta)
 +              attr_delta = git_attr("delta");
 +
 +      check[0].attr = attr_delta;
 +}
 +
 +static int no_try_delta(const char *path)
 +{
 +      struct git_attr_check check[1];
 +
 +      setup_delta_attr_check(check);
 +      if (git_checkattr(path, ARRAY_SIZE(check), check))
 +              return 0;
 +      if (ATTR_FALSE(check->value))
 +              return 1;
 +      return 0;
 +}
 +
 +static int add_object_entry(const unsigned char *sha1, enum object_type type,
 +                          const char *name, int exclude)
 +{
 +      struct object_entry *entry;
 +      struct packed_git *p, *found_pack = NULL;
 +      off_t found_offset = 0;
 +      int ix;
 +      unsigned hash = name_hash(name);
 +
 +      ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
 +      if (ix >= 0) {
 +              if (exclude) {
 +                      entry = objects + object_ix[ix] - 1;
 +                      if (!entry->preferred_base)
 +                              nr_result--;
 +                      entry->preferred_base = 1;
 +              }
 +              return 0;
 +      }
 +
 +      if (!exclude && local && has_loose_object_nonlocal(sha1))
 +              return 0;
 +
 +      for (p = packed_git; p; p = p->next) {
 +              off_t offset = find_pack_entry_one(sha1, p);
 +              if (offset) {
 +                      if (!found_pack) {
 +                              found_offset = offset;
 +                              found_pack = p;
 +                      }
 +                      if (exclude)
 +                              break;
 +                      if (incremental)
 +                              return 0;
 +                      if (local && !p->pack_local)
 +                              return 0;
 +                      if (ignore_packed_keep && p->pack_local && p->pack_keep)
 +                              return 0;
 +              }
 +      }
 +
 +      if (nr_objects >= nr_alloc) {
 +              nr_alloc = (nr_alloc  + 1024) * 3 / 2;
 +              objects = xrealloc(objects, nr_alloc * sizeof(*entry));
 +      }
 +
 +      entry = objects + nr_objects++;
 +      memset(entry, 0, sizeof(*entry));
 +      hashcpy(entry->idx.sha1, sha1);
 +      entry->hash = hash;
 +      if (type)
 +              entry->type = type;
 +      if (exclude)
 +              entry->preferred_base = 1;
 +      else
 +              nr_result++;
 +      if (found_pack) {
 +              entry->in_pack = found_pack;
 +              entry->in_pack_offset = found_offset;
 +      }
 +
 +      if (object_ix_hashsz * 3 <= nr_objects * 4)
 +              rehash_objects();
 +      else
 +              object_ix[-1 - ix] = nr_objects;
 +
 +      display_progress(progress_state, nr_objects);
 +
 +      if (name && no_try_delta(name))
 +              entry->no_try_delta = 1;
 +
 +      return 1;
 +}
 +
 +struct pbase_tree_cache {
 +      unsigned char sha1[20];
 +      int ref;
 +      int temporary;
 +      void *tree_data;
 +      unsigned long tree_size;
 +};
 +
 +static struct pbase_tree_cache *(pbase_tree_cache[256]);
 +static int pbase_tree_cache_ix(const unsigned char *sha1)
 +{
 +      return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
 +}
 +static int pbase_tree_cache_ix_incr(int ix)
 +{
 +      return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
 +}
 +
 +static struct pbase_tree {
 +      struct pbase_tree *next;
 +      /* This is a phony "cache" entry; we are not
 +       * going to evict it nor find it through _get()
 +       * mechanism -- this is for the toplevel node that
 +       * would almost always change with any commit.
 +       */
 +      struct pbase_tree_cache pcache;
 +} *pbase_tree;
 +
 +static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
 +{
 +      struct pbase_tree_cache *ent, *nent;
 +      void *data;
 +      unsigned long size;
 +      enum object_type type;
 +      int neigh;
 +      int my_ix = pbase_tree_cache_ix(sha1);
 +      int available_ix = -1;
 +
 +      /* pbase-tree-cache acts as a limited hashtable.
 +       * your object will be found at your index or within a few
 +       * slots after that slot if it is cached.
 +       */
 +      for (neigh = 0; neigh < 8; neigh++) {
 +              ent = pbase_tree_cache[my_ix];
 +              if (ent && !hashcmp(ent->sha1, sha1)) {
 +                      ent->ref++;
 +                      return ent;
 +              }
 +              else if (((available_ix < 0) && (!ent || !ent->ref)) ||
 +                       ((0 <= available_ix) &&
 +                        (!ent && pbase_tree_cache[available_ix])))
 +                      available_ix = my_ix;
 +              if (!ent)
 +                      break;
 +              my_ix = pbase_tree_cache_ix_incr(my_ix);
 +      }
 +
 +      /* Did not find one.  Either we got a bogus request or
 +       * we need to read and perhaps cache.
 +       */
 +      data = read_sha1_file(sha1, &type, &size);
 +      if (!data)
 +              return NULL;
 +      if (type != OBJ_TREE) {
 +              free(data);
 +              return NULL;
 +      }
 +
 +      /* We need to either cache or return a throwaway copy */
 +
 +      if (available_ix < 0)
 +              ent = NULL;
 +      else {
 +              ent = pbase_tree_cache[available_ix];
 +              my_ix = available_ix;
 +      }
 +
 +      if (!ent) {
 +              nent = xmalloc(sizeof(*nent));
 +              nent->temporary = (available_ix < 0);
 +      }
 +      else {
 +              /* evict and reuse */
 +              free(ent->tree_data);
 +              nent = ent;
 +      }
 +      hashcpy(nent->sha1, sha1);
 +      nent->tree_data = data;
 +      nent->tree_size = size;
 +      nent->ref = 1;
 +      if (!nent->temporary)
 +              pbase_tree_cache[my_ix] = nent;
 +      return nent;
 +}
 +
 +static void pbase_tree_put(struct pbase_tree_cache *cache)
 +{
 +      if (!cache->temporary) {
 +              cache->ref--;
 +              return;
 +      }
 +      free(cache->tree_data);
 +      free(cache);
 +}
 +
 +static int name_cmp_len(const char *name)
 +{
 +      int i;
 +      for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
 +              ;
 +      return i;
 +}
 +
 +static void add_pbase_object(struct tree_desc *tree,
 +                           const char *name,
 +                           int cmplen,
 +                           const char *fullname)
 +{
 +      struct name_entry entry;
 +      int cmp;
 +
 +      while (tree_entry(tree,&entry)) {
 +              if (S_ISGITLINK(entry.mode))
 +                      continue;
 +              cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
 +                    memcmp(name, entry.path, cmplen);
 +              if (cmp > 0)
 +                      continue;
 +              if (cmp < 0)
 +                      return;
 +              if (name[cmplen] != '/') {
 +                      add_object_entry(entry.sha1,
 +                                       object_type(entry.mode),
 +                                       fullname, 1);
 +                      return;
 +              }
 +              if (S_ISDIR(entry.mode)) {
 +                      struct tree_desc sub;
 +                      struct pbase_tree_cache *tree;
 +                      const char *down = name+cmplen+1;
 +                      int downlen = name_cmp_len(down);
 +
 +                      tree = pbase_tree_get(entry.sha1);
 +                      if (!tree)
 +                              return;
 +                      init_tree_desc(&sub, tree->tree_data, tree->tree_size);
 +
 +                      add_pbase_object(&sub, down, downlen, fullname);
 +                      pbase_tree_put(tree);
 +              }
 +      }
 +}
 +
 +static unsigned *done_pbase_paths;
 +static int done_pbase_paths_num;
 +static int done_pbase_paths_alloc;
 +static int done_pbase_path_pos(unsigned hash)
 +{
 +      int lo = 0;
 +      int hi = done_pbase_paths_num;
 +      while (lo < hi) {
 +              int mi = (hi + lo) / 2;
 +              if (done_pbase_paths[mi] == hash)
 +                      return mi;
 +              if (done_pbase_paths[mi] < hash)
 +                      hi = mi;
 +              else
 +                      lo = mi + 1;
 +      }
 +      return -lo-1;
 +}
 +
 +static int check_pbase_path(unsigned hash)
 +{
 +      int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
 +      if (0 <= pos)
 +              return 1;
 +      pos = -pos - 1;
 +      if (done_pbase_paths_alloc <= done_pbase_paths_num) {
 +              done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
 +              done_pbase_paths = xrealloc(done_pbase_paths,
 +                                          done_pbase_paths_alloc *
 +                                          sizeof(unsigned));
 +      }
 +      done_pbase_paths_num++;
 +      if (pos < done_pbase_paths_num)
 +              memmove(done_pbase_paths + pos + 1,
 +                      done_pbase_paths + pos,
 +                      (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
 +      done_pbase_paths[pos] = hash;
 +      return 0;
 +}
 +
 +static void add_preferred_base_object(const char *name)
 +{
 +      struct pbase_tree *it;
 +      int cmplen;
 +      unsigned hash = name_hash(name);
 +
 +      if (!num_preferred_base || check_pbase_path(hash))
 +              return;
 +
 +      cmplen = name_cmp_len(name);
 +      for (it = pbase_tree; it; it = it->next) {
 +              if (cmplen == 0) {
 +                      add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
 +              }
 +              else {
 +                      struct tree_desc tree;
 +                      init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
 +                      add_pbase_object(&tree, name, cmplen, name);
 +              }
 +      }
 +}
 +
 +static void add_preferred_base(unsigned char *sha1)
 +{
 +      struct pbase_tree *it;
 +      void *data;
 +      unsigned long size;
 +      unsigned char tree_sha1[20];
 +
 +      if (window <= num_preferred_base++)
 +              return;
 +
 +      data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
 +      if (!data)
 +              return;
 +
 +      for (it = pbase_tree; it; it = it->next) {
 +              if (!hashcmp(it->pcache.sha1, tree_sha1)) {
 +                      free(data);
 +                      return;
 +              }
 +      }
 +
 +      it = xcalloc(1, sizeof(*it));
 +      it->next = pbase_tree;
 +      pbase_tree = it;
 +
 +      hashcpy(it->pcache.sha1, tree_sha1);
 +      it->pcache.tree_data = data;
 +      it->pcache.tree_size = size;
 +}
 +
 +static void cleanup_preferred_base(void)
 +{
 +      struct pbase_tree *it;
 +      unsigned i;
 +
 +      it = pbase_tree;
 +      pbase_tree = NULL;
 +      while (it) {
 +              struct pbase_tree *this = it;
 +              it = this->next;
 +              free(this->pcache.tree_data);
 +              free(this);
 +      }
 +
 +      for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
 +              if (!pbase_tree_cache[i])
 +                      continue;
 +              free(pbase_tree_cache[i]->tree_data);
 +              free(pbase_tree_cache[i]);
 +              pbase_tree_cache[i] = NULL;
 +      }
 +
 +      free(done_pbase_paths);
 +      done_pbase_paths = NULL;
 +      done_pbase_paths_num = done_pbase_paths_alloc = 0;
 +}
 +
 +static void check_object(struct object_entry *entry)
 +{
 +      if (entry->in_pack) {
 +              struct packed_git *p = entry->in_pack;
 +              struct pack_window *w_curs = NULL;
 +              const unsigned char *base_ref = NULL;
 +              struct object_entry *base_entry;
 +              unsigned long used, used_0;
 +              unsigned int avail;
 +              off_t ofs;
 +              unsigned char *buf, c;
 +
 +              buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
 +
 +              /*
 +               * We want in_pack_type even if we do not reuse delta
 +               * since non-delta representations could still be reused.
 +               */
 +              used = unpack_object_header_buffer(buf, avail,
 +                                                 &entry->in_pack_type,
 +                                                 &entry->size);
 +              if (used == 0)
 +                      goto give_up;
 +
 +              /*
 +               * Determine if this is a delta and if so whether we can
 +               * reuse it or not.  Otherwise let's find out as cheaply as
 +               * possible what the actual type and size for this object is.
 +               */
 +              switch (entry->in_pack_type) {
 +              default:
 +                      /* Not a delta hence we've already got all we need. */
 +                      entry->type = entry->in_pack_type;
 +                      entry->in_pack_header_size = used;
 +                      if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
 +                              goto give_up;
 +                      unuse_pack(&w_curs);
 +                      return;
 +              case OBJ_REF_DELTA:
 +                      if (reuse_delta && !entry->preferred_base)
 +                              base_ref = use_pack(p, &w_curs,
 +                                              entry->in_pack_offset + used, NULL);
 +                      entry->in_pack_header_size = used + 20;
 +                      break;
 +              case OBJ_OFS_DELTA:
 +                      buf = use_pack(p, &w_curs,
 +                                     entry->in_pack_offset + used, NULL);
 +                      used_0 = 0;
 +                      c = buf[used_0++];
 +                      ofs = c & 127;
 +                      while (c & 128) {
 +                              ofs += 1;
 +                              if (!ofs || MSB(ofs, 7)) {
 +                                      error("delta base offset overflow in pack for %s",
 +                                            sha1_to_hex(entry->idx.sha1));
 +                                      goto give_up;
 +                              }
 +                              c = buf[used_0++];
 +                              ofs = (ofs << 7) + (c & 127);
 +                      }
 +                      ofs = entry->in_pack_offset - ofs;
 +                      if (ofs <= 0 || ofs >= entry->in_pack_offset) {
 +                              error("delta base offset out of bound for %s",
 +                                    sha1_to_hex(entry->idx.sha1));
 +                              goto give_up;
 +                      }
 +                      if (reuse_delta && !entry->preferred_base) {
 +                              struct revindex_entry *revidx;
 +                              revidx = find_pack_revindex(p, ofs);
 +                              if (!revidx)
 +                                      goto give_up;
 +                              base_ref = nth_packed_object_sha1(p, revidx->nr);
 +                      }
 +                      entry->in_pack_header_size = used + used_0;
 +                      break;
 +              }
 +
 +              if (base_ref && (base_entry = locate_object_entry(base_ref))) {
 +                      /*
 +                       * If base_ref was set above that means we wish to
 +                       * reuse delta data, and we even found that base
 +                       * in the list of objects we want to pack. Goodie!
 +                       *
 +                       * Depth value does not matter - find_deltas() will
 +                       * never consider reused delta as the base object to
 +                       * deltify other objects against, in order to avoid
 +                       * circular deltas.
 +                       */
 +                      entry->type = entry->in_pack_type;
 +                      entry->delta = base_entry;
 +                      entry->delta_size = entry->size;
 +                      entry->delta_sibling = base_entry->delta_child;
 +                      base_entry->delta_child = entry;
 +                      unuse_pack(&w_curs);
 +                      return;
 +              }
 +
 +              if (entry->type) {
 +                      /*
 +                       * This must be a delta and we already know what the
 +                       * final object type is.  Let's extract the actual
 +                       * object size from the delta header.
 +                       */
 +                      entry->size = get_size_from_delta(p, &w_curs,
 +                                      entry->in_pack_offset + entry->in_pack_header_size);
 +                      if (entry->size == 0)
 +                              goto give_up;
 +                      unuse_pack(&w_curs);
 +                      return;
 +              }
 +
 +              /*
 +               * No choice but to fall back to the recursive delta walk
 +               * with sha1_object_info() to find about the object type
 +               * at this point...
 +               */
 +              give_up:
 +              unuse_pack(&w_curs);
 +      }
 +
 +      entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
 +      /*
 +       * The error condition is checked in prepare_pack().  This is
 +       * to permit a missing preferred base object to be ignored
 +       * as a preferred base.  Doing so can result in a larger
 +       * pack file, but the transfer will still take place.
 +       */
 +}
 +
 +static int pack_offset_sort(const void *_a, const void *_b)
 +{
 +      const struct object_entry *a = *(struct object_entry **)_a;
 +      const struct object_entry *b = *(struct object_entry **)_b;
 +
 +      /* avoid filesystem trashing with loose objects */
 +      if (!a->in_pack && !b->in_pack)
 +              return hashcmp(a->idx.sha1, b->idx.sha1);
 +
 +      if (a->in_pack < b->in_pack)
 +              return -1;
 +      if (a->in_pack > b->in_pack)
 +              return 1;
 +      return a->in_pack_offset < b->in_pack_offset ? -1 :
 +                      (a->in_pack_offset > b->in_pack_offset);
 +}
 +
 +static void get_object_details(void)
 +{
 +      uint32_t i;
 +      struct object_entry **sorted_by_offset;
 +
 +      sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
 +      for (i = 0; i < nr_objects; i++)
 +              sorted_by_offset[i] = objects + i;
 +      qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
 +
 +      for (i = 0; i < nr_objects; i++)
 +              check_object(sorted_by_offset[i]);
 +
 +      free(sorted_by_offset);
 +}
 +
 +/*
 + * We search for deltas in a list sorted by type, by filename hash, and then
 + * by size, so that we see progressively smaller and smaller files.
 + * That's because we prefer deltas to be from the bigger file
 + * to the smaller -- deletes are potentially cheaper, but perhaps
 + * more importantly, the bigger file is likely the more recent
 + * one.  The deepest deltas are therefore the oldest objects which are
 + * less susceptible to be accessed often.
 + */
 +static int type_size_sort(const void *_a, const void *_b)
 +{
 +      const struct object_entry *a = *(struct object_entry **)_a;
 +      const struct object_entry *b = *(struct object_entry **)_b;
 +
 +      if (a->type > b->type)
 +              return -1;
 +      if (a->type < b->type)
 +              return 1;
 +      if (a->hash > b->hash)
 +              return -1;
 +      if (a->hash < b->hash)
 +              return 1;
 +      if (a->preferred_base > b->preferred_base)
 +              return -1;
 +      if (a->preferred_base < b->preferred_base)
 +              return 1;
 +      if (a->size > b->size)
 +              return -1;
 +      if (a->size < b->size)
 +              return 1;
 +      return a < b ? -1 : (a > b);  /* newest first */
 +}
 +
 +struct unpacked {
 +      struct object_entry *entry;
 +      void *data;
 +      struct delta_index *index;
 +      unsigned depth;
 +};
 +
 +static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
 +                         unsigned long delta_size)
 +{
 +      if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
 +              return 0;
 +
 +      if (delta_size < cache_max_small_delta_size)
 +              return 1;
 +
 +      /* cache delta, if objects are large enough compared to delta size */
 +      if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
 +              return 1;
 +
 +      return 0;
 +}
 +
 +#ifndef NO_PTHREADS
 +
 +static pthread_mutex_t read_mutex;
 +#define read_lock()           pthread_mutex_lock(&read_mutex)
 +#define read_unlock()         pthread_mutex_unlock(&read_mutex)
 +
 +static pthread_mutex_t cache_mutex;
 +#define cache_lock()          pthread_mutex_lock(&cache_mutex)
 +#define cache_unlock()                pthread_mutex_unlock(&cache_mutex)
 +
 +static pthread_mutex_t progress_mutex;
 +#define progress_lock()               pthread_mutex_lock(&progress_mutex)
 +#define progress_unlock()     pthread_mutex_unlock(&progress_mutex)
 +
 +#else
 +
 +#define read_lock()           (void)0
 +#define read_unlock()         (void)0
 +#define cache_lock()          (void)0
 +#define cache_unlock()                (void)0
 +#define progress_lock()               (void)0
 +#define progress_unlock()     (void)0
 +
 +#endif
 +
 +static int try_delta(struct unpacked *trg, struct unpacked *src,
 +                   unsigned max_depth, unsigned long *mem_usage)
 +{
 +      struct object_entry *trg_entry = trg->entry;
 +      struct object_entry *src_entry = src->entry;
 +      unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
 +      unsigned ref_depth;
 +      enum object_type type;
 +      void *delta_buf;
 +
 +      /* Don't bother doing diffs between different types */
 +      if (trg_entry->type != src_entry->type)
 +              return -1;
 +
 +      /*
 +       * We do not bother to try a delta that we discarded
 +       * on an earlier try, but only when reusing delta data.
 +       */
 +      if (reuse_delta && trg_entry->in_pack &&
 +          trg_entry->in_pack == src_entry->in_pack &&
 +          trg_entry->in_pack_type != OBJ_REF_DELTA &&
 +          trg_entry->in_pack_type != OBJ_OFS_DELTA)
 +              return 0;
 +
 +      /* Let's not bust the allowed depth. */
 +      if (src->depth >= max_depth)
 +              return 0;
 +
 +      /* Now some size filtering heuristics. */
 +      trg_size = trg_entry->size;
 +      if (!trg_entry->delta) {
 +              max_size = trg_size/2 - 20;
 +              ref_depth = 1;
 +      } else {
 +              max_size = trg_entry->delta_size;
 +              ref_depth = trg->depth;
 +      }
 +      max_size = (uint64_t)max_size * (max_depth - src->depth) /
 +                                              (max_depth - ref_depth + 1);
 +      if (max_size == 0)
 +              return 0;
 +      src_size = src_entry->size;
 +      sizediff = src_size < trg_size ? trg_size - src_size : 0;
 +      if (sizediff >= max_size)
 +              return 0;
 +      if (trg_size < src_size / 32)
 +              return 0;
 +
 +      /* Load data if not already done */
 +      if (!trg->data) {
 +              read_lock();
 +              trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
 +              read_unlock();
 +              if (!trg->data)
 +                      die("object %s cannot be read",
 +                          sha1_to_hex(trg_entry->idx.sha1));
 +              if (sz != trg_size)
 +                      die("object %s inconsistent object length (%lu vs %lu)",
 +                          sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
 +              *mem_usage += sz;
 +      }
 +      if (!src->data) {
 +              read_lock();
 +              src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
 +              read_unlock();
 +              if (!src->data)
 +                      die("object %s cannot be read",
 +                          sha1_to_hex(src_entry->idx.sha1));
 +              if (sz != src_size)
 +                      die("object %s inconsistent object length (%lu vs %lu)",
 +                          sha1_to_hex(src_entry->idx.sha1), sz, src_size);
 +              *mem_usage += sz;
 +      }
 +      if (!src->index) {
 +              src->index = create_delta_index(src->data, src_size);
 +              if (!src->index) {
 +                      static int warned = 0;
 +                      if (!warned++)
 +                              warning("suboptimal pack - out of memory");
 +                      return 0;
 +              }
 +              *mem_usage += sizeof_delta_index(src->index);
 +      }
 +
 +      delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
 +      if (!delta_buf)
 +              return 0;
 +
 +      if (trg_entry->delta) {
 +              /* Prefer only shallower same-sized deltas. */
 +              if (delta_size == trg_entry->delta_size &&
 +                  src->depth + 1 >= trg->depth) {
 +                      free(delta_buf);
 +                      return 0;
 +              }
 +      }
 +
 +      /*
 +       * Handle memory allocation outside of the cache
 +       * accounting lock.  Compiler will optimize the strangeness
 +       * away when NO_PTHREADS is defined.
 +       */
 +      free(trg_entry->delta_data);
 +      cache_lock();
 +      if (trg_entry->delta_data) {
 +              delta_cache_size -= trg_entry->delta_size;
 +              trg_entry->delta_data = NULL;
 +      }
 +      if (delta_cacheable(src_size, trg_size, delta_size)) {
 +              delta_cache_size += delta_size;
 +              cache_unlock();
 +              trg_entry->delta_data = xrealloc(delta_buf, delta_size);
 +      } else {
 +              cache_unlock();
 +              free(delta_buf);
 +      }
 +
 +      trg_entry->delta = src_entry;
 +      trg_entry->delta_size = delta_size;
 +      trg->depth = src->depth + 1;
 +
 +      return 1;
 +}
 +
 +static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
 +{
 +      struct object_entry *child = me->delta_child;
 +      unsigned int m = n;
 +      while (child) {
 +              unsigned int c = check_delta_limit(child, n + 1);
 +              if (m < c)
 +                      m = c;
 +              child = child->delta_sibling;
 +      }
 +      return m;
 +}
 +
 +static unsigned long free_unpacked(struct unpacked *n)
 +{
 +      unsigned long freed_mem = sizeof_delta_index(n->index);
 +      free_delta_index(n->index);
 +      n->index = NULL;
 +      if (n->data) {
 +              freed_mem += n->entry->size;
 +              free(n->data);
 +              n->data = NULL;
 +      }
 +      n->entry = NULL;
 +      n->depth = 0;
 +      return freed_mem;
 +}
 +
 +static void find_deltas(struct object_entry **list, unsigned *list_size,
 +                      int window, int depth, unsigned *processed)
 +{
 +      uint32_t i, idx = 0, count = 0;
 +      struct unpacked *array;
 +      unsigned long mem_usage = 0;
 +
 +      array = xcalloc(window, sizeof(struct unpacked));
 +
 +      for (;;) {
 +              struct object_entry *entry;
 +              struct unpacked *n = array + idx;
 +              int j, max_depth, best_base = -1;
 +
 +              progress_lock();
 +              if (!*list_size) {
 +                      progress_unlock();
 +                      break;
 +              }
 +              entry = *list++;
 +              (*list_size)--;
 +              if (!entry->preferred_base) {
 +                      (*processed)++;
 +                      display_progress(progress_state, *processed);
 +              }
 +              progress_unlock();
 +
 +              mem_usage -= free_unpacked(n);
 +              n->entry = entry;
 +
 +              while (window_memory_limit &&
 +                     mem_usage > window_memory_limit &&
 +                     count > 1) {
 +                      uint32_t tail = (idx + window - count) % window;
 +                      mem_usage -= free_unpacked(array + tail);
 +                      count--;
 +              }
 +
 +              /* We do not compute delta to *create* objects we are not
 +               * going to pack.
 +               */
 +              if (entry->preferred_base)
 +                      goto next;
 +
 +              /*
 +               * If the current object is at pack edge, take the depth the
 +               * objects that depend on the current object into account
 +               * otherwise they would become too deep.
 +               */
 +              max_depth = depth;
 +              if (entry->delta_child) {
 +                      max_depth -= check_delta_limit(entry, 0);
 +                      if (max_depth <= 0)
 +                              goto next;
 +              }
 +
 +              j = window;
 +              while (--j > 0) {
 +                      int ret;
 +                      uint32_t other_idx = idx + j;
 +                      struct unpacked *m;
 +                      if (other_idx >= window)
 +                              other_idx -= window;
 +                      m = array + other_idx;
 +                      if (!m->entry)
 +                              break;
 +                      ret = try_delta(n, m, max_depth, &mem_usage);
 +                      if (ret < 0)
 +                              break;
 +                      else if (ret > 0)
 +                              best_base = other_idx;
 +              }
 +
 +              /*
 +               * If we decided to cache the delta data, then it is best
 +               * to compress it right away.  First because we have to do
 +               * it anyway, and doing it here while we're threaded will
 +               * save a lot of time in the non threaded write phase,
 +               * as well as allow for caching more deltas within
 +               * the same cache size limit.
 +               * ...
 +               * But only if not writing to stdout, since in that case
 +               * the network is most likely throttling writes anyway,
 +               * and therefore it is best to go to the write phase ASAP
 +               * instead, as we can afford spending more time compressing
 +               * between writes at that moment.
 +               */
 +              if (entry->delta_data && !pack_to_stdout) {
 +                      entry->z_delta_size = do_compress(&entry->delta_data,
 +                                                        entry->delta_size);
 +                      cache_lock();
 +                      delta_cache_size -= entry->delta_size;
 +                      delta_cache_size += entry->z_delta_size;
 +                      cache_unlock();
 +              }
 +
 +              /* if we made n a delta, and if n is already at max
 +               * depth, leaving it in the window is pointless.  we
 +               * should evict it first.
 +               */
 +              if (entry->delta && max_depth <= n->depth)
 +                      continue;
 +
 +              /*
 +               * Move the best delta base up in the window, after the
 +               * currently deltified object, to keep it longer.  It will
 +               * be the first base object to be attempted next.
 +               */
 +              if (entry->delta) {
 +                      struct unpacked swap = array[best_base];
 +                      int dist = (window + idx - best_base) % window;
 +                      int dst = best_base;
 +                      while (dist--) {
 +                              int src = (dst + 1) % window;
 +                              array[dst] = array[src];
 +                              dst = src;
 +                      }
 +                      array[dst] = swap;
 +              }
 +
 +              next:
 +              idx++;
 +              if (count + 1 < window)
 +                      count++;
 +              if (idx >= window)
 +                      idx = 0;
 +      }
 +
 +      for (i = 0; i < window; ++i) {
 +              free_delta_index(array[i].index);
 +              free(array[i].data);
 +      }
 +      free(array);
 +}
 +
 +#ifndef NO_PTHREADS
 +
++static void try_to_free_from_threads(size_t size)
++{
++      read_lock();
++      release_pack_memory(size, -1);
++      read_unlock();
++}
++
 +/*
 + * The main thread waits on the condition that (at least) one of the workers
 + * has stopped working (which is indicated in the .working member of
 + * struct thread_params).
 + * When a work thread has completed its work, it sets .working to 0 and
 + * signals the main thread and waits on the condition that .data_ready
 + * becomes 1.
 + */
 +
 +struct thread_params {
 +      pthread_t thread;
 +      struct object_entry **list;
 +      unsigned list_size;
 +      unsigned remaining;
 +      int window;
 +      int depth;
 +      int working;
 +      int data_ready;
 +      pthread_mutex_t mutex;
 +      pthread_cond_t cond;
 +      unsigned *processed;
 +};
 +
 +static pthread_cond_t progress_cond;
 +
 +/*
 + * Mutex and conditional variable can't be statically-initialized on Windows.
 + */
 +static void init_threaded_search(void)
 +{
++      init_recursive_mutex(&read_mutex);
 +      pthread_mutex_init(&cache_mutex, NULL);
 +      pthread_mutex_init(&progress_mutex, NULL);
 +      pthread_cond_init(&progress_cond, NULL);
++      set_try_to_free_routine(try_to_free_from_threads);
 +}
 +
 +static void cleanup_threaded_search(void)
 +{
++      set_try_to_free_routine(NULL);
 +      pthread_cond_destroy(&progress_cond);
 +      pthread_mutex_destroy(&read_mutex);
 +      pthread_mutex_destroy(&cache_mutex);
 +      pthread_mutex_destroy(&progress_mutex);
 +}
 +
 +static void *threaded_find_deltas(void *arg)
 +{
 +      struct thread_params *me = arg;
 +
 +      while (me->remaining) {
 +              find_deltas(me->list, &me->remaining,
 +                          me->window, me->depth, me->processed);
 +
 +              progress_lock();
 +              me->working = 0;
 +              pthread_cond_signal(&progress_cond);
 +              progress_unlock();
 +
 +              /*
 +               * We must not set ->data_ready before we wait on the
 +               * condition because the main thread may have set it to 1
 +               * before we get here. In order to be sure that new
 +               * work is available if we see 1 in ->data_ready, it
 +               * was initialized to 0 before this thread was spawned
 +               * and we reset it to 0 right away.
 +               */
 +              pthread_mutex_lock(&me->mutex);
 +              while (!me->data_ready)
 +                      pthread_cond_wait(&me->cond, &me->mutex);
 +              me->data_ready = 0;
 +              pthread_mutex_unlock(&me->mutex);
 +      }
 +      /* leave ->working 1 so that this doesn't get more work assigned */
 +      return NULL;
 +}
 +
 +static void ll_find_deltas(struct object_entry **list, unsigned list_size,
 +                         int window, int depth, unsigned *processed)
 +{
 +      struct thread_params *p;
 +      int i, ret, active_threads = 0;
 +
 +      init_threaded_search();
 +
 +      if (!delta_search_threads)      /* --threads=0 means autodetect */
 +              delta_search_threads = online_cpus();
 +      if (delta_search_threads <= 1) {
 +              find_deltas(list, &list_size, window, depth, processed);
 +              cleanup_threaded_search();
 +              return;
 +      }
 +      if (progress > pack_to_stdout)
 +              fprintf(stderr, "Delta compression using up to %d threads.\n",
 +                              delta_search_threads);
 +      p = xcalloc(delta_search_threads, sizeof(*p));
 +
 +      /* Partition the work amongst work threads. */
 +      for (i = 0; i < delta_search_threads; i++) {
 +              unsigned sub_size = list_size / (delta_search_threads - i);
 +
 +              /* don't use too small segments or no deltas will be found */
 +              if (sub_size < 2*window && i+1 < delta_search_threads)
 +                      sub_size = 0;
 +
 +              p[i].window = window;
 +              p[i].depth = depth;
 +              p[i].processed = processed;
 +              p[i].working = 1;
 +              p[i].data_ready = 0;
 +
 +              /* try to split chunks on "path" boundaries */
 +              while (sub_size && sub_size < list_size &&
 +                     list[sub_size]->hash &&
 +                     list[sub_size]->hash == list[sub_size-1]->hash)
 +                      sub_size++;
 +
 +              p[i].list = list;
 +              p[i].list_size = sub_size;
 +              p[i].remaining = sub_size;
 +
 +              list += sub_size;
 +              list_size -= sub_size;
 +      }
 +
 +      /* Start work threads. */
 +      for (i = 0; i < delta_search_threads; i++) {
 +              if (!p[i].list_size)
 +                      continue;
 +              pthread_mutex_init(&p[i].mutex, NULL);
 +              pthread_cond_init(&p[i].cond, NULL);
 +              ret = pthread_create(&p[i].thread, NULL,
 +                                   threaded_find_deltas, &p[i]);
 +              if (ret)
 +                      die("unable to create thread: %s", strerror(ret));
 +              active_threads++;
 +      }
 +
 +      /*
 +       * Now let's wait for work completion.  Each time a thread is done
 +       * with its work, we steal half of the remaining work from the
 +       * thread with the largest number of unprocessed objects and give
 +       * it to that newly idle thread.  This ensure good load balancing
 +       * until the remaining object list segments are simply too short
 +       * to be worth splitting anymore.
 +       */
 +      while (active_threads) {
 +              struct thread_params *target = NULL;
 +              struct thread_params *victim = NULL;
 +              unsigned sub_size = 0;
 +
 +              progress_lock();
 +              for (;;) {
 +                      for (i = 0; !target && i < delta_search_threads; i++)
 +                              if (!p[i].working)
 +                                      target = &p[i];
 +                      if (target)
 +                              break;
 +                      pthread_cond_wait(&progress_cond, &progress_mutex);
 +              }
 +
 +              for (i = 0; i < delta_search_threads; i++)
 +                      if (p[i].remaining > 2*window &&
 +                          (!victim || victim->remaining < p[i].remaining))
 +                              victim = &p[i];
 +              if (victim) {
 +                      sub_size = victim->remaining / 2;
 +                      list = victim->list + victim->list_size - sub_size;
 +                      while (sub_size && list[0]->hash &&
 +                             list[0]->hash == list[-1]->hash) {
 +                              list++;
 +                              sub_size--;
 +                      }
 +                      if (!sub_size) {
 +                              /*
 +                               * It is possible for some "paths" to have
 +                               * so many objects that no hash boundary
 +                               * might be found.  Let's just steal the
 +                               * exact half in that case.
 +                               */
 +                              sub_size = victim->remaining / 2;
 +                              list -= sub_size;
 +                      }
 +                      target->list = list;
 +                      victim->list_size -= sub_size;
 +                      victim->remaining -= sub_size;
 +              }
 +              target->list_size = sub_size;
 +              target->remaining = sub_size;
 +              target->working = 1;
 +              progress_unlock();
 +
 +              pthread_mutex_lock(&target->mutex);
 +              target->data_ready = 1;
 +              pthread_cond_signal(&target->cond);
 +              pthread_mutex_unlock(&target->mutex);
 +
 +              if (!sub_size) {
 +                      pthread_join(target->thread, NULL);
 +                      pthread_cond_destroy(&target->cond);
 +                      pthread_mutex_destroy(&target->mutex);
 +                      active_threads--;
 +              }
 +      }
 +      cleanup_threaded_search();
 +      free(p);
 +}
 +
 +#else
 +#define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
 +#endif
 +
 +static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
 +{
 +      unsigned char peeled[20];
 +
 +      if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
 +          !peel_ref(path, peeled)        && /* peelable? */
 +          !is_null_sha1(peeled)          && /* annotated tag? */
 +          locate_object_entry(peeled))      /* object packed? */
 +              add_object_entry(sha1, OBJ_TAG, NULL, 0);
 +      return 0;
 +}
 +
 +static void prepare_pack(int window, int depth)
 +{
 +      struct object_entry **delta_list;
 +      uint32_t i, nr_deltas;
 +      unsigned n;
 +
 +      get_object_details();
 +
 +      /*
 +       * If we're locally repacking then we need to be doubly careful
 +       * from now on in order to make sure no stealth corruption gets
 +       * propagated to the new pack.  Clients receiving streamed packs
 +       * should validate everything they get anyway so no need to incur
 +       * the additional cost here in that case.
 +       */
 +      if (!pack_to_stdout)
 +              do_check_packed_object_crc = 1;
 +
 +      if (!nr_objects || !window || !depth)
 +              return;
 +
 +      delta_list = xmalloc(nr_objects * sizeof(*delta_list));
 +      nr_deltas = n = 0;
 +
 +      for (i = 0; i < nr_objects; i++) {
 +              struct object_entry *entry = objects + i;
 +
 +              if (entry->delta)
 +                      /* This happens if we decided to reuse existing
 +                       * delta from a pack.  "reuse_delta &&" is implied.
 +                       */
 +                      continue;
 +
 +              if (entry->size < 50)
 +                      continue;
 +
 +              if (entry->no_try_delta)
 +                      continue;
 +
 +              if (!entry->preferred_base) {
 +                      nr_deltas++;
 +                      if (entry->type < 0)
 +                              die("unable to get type of object %s",
 +                                  sha1_to_hex(entry->idx.sha1));
 +              } else {
 +                      if (entry->type < 0) {
 +                              /*
 +                               * This object is not found, but we
 +                               * don't have to include it anyway.
 +                               */
 +                              continue;
 +                      }
 +              }
 +
 +              delta_list[n++] = entry;
 +      }
 +
 +      if (nr_deltas && n > 1) {
 +              unsigned nr_done = 0;
 +              if (progress)
 +                      progress_state = start_progress("Compressing objects",
 +                                                      nr_deltas);
 +              qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
 +              ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
 +              stop_progress(&progress_state);
 +              if (nr_done != nr_deltas)
 +                      die("inconsistency with delta count");
 +      }
 +      free(delta_list);
 +}
 +
 +static int git_pack_config(const char *k, const char *v, void *cb)
 +{
 +      if (!strcmp(k, "pack.window")) {
 +              window = git_config_int(k, v);
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.windowmemory")) {
 +              window_memory_limit = git_config_ulong(k, v);
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.depth")) {
 +              depth = git_config_int(k, v);
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.compression")) {
 +              int level = git_config_int(k, v);
 +              if (level == -1)
 +                      level = Z_DEFAULT_COMPRESSION;
 +              else if (level < 0 || level > Z_BEST_COMPRESSION)
 +                      die("bad pack compression level %d", level);
 +              pack_compression_level = level;
 +              pack_compression_seen = 1;
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.deltacachesize")) {
 +              max_delta_cache_size = git_config_int(k, v);
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.deltacachelimit")) {
 +              cache_max_small_delta_size = git_config_int(k, v);
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.threads")) {
 +              delta_search_threads = git_config_int(k, v);
 +              if (delta_search_threads < 0)
 +                      die("invalid number of threads specified (%d)",
 +                          delta_search_threads);
 +#ifdef NO_PTHREADS
 +              if (delta_search_threads != 1)
 +                      warning("no threads support, ignoring %s", k);
 +#endif
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.indexversion")) {
 +              pack_idx_default_version = git_config_int(k, v);
 +              if (pack_idx_default_version > 2)
 +                      die("bad pack.indexversion=%"PRIu32,
 +                              pack_idx_default_version);
 +              return 0;
 +      }
 +      if (!strcmp(k, "pack.packsizelimit")) {
 +              pack_size_limit_cfg = git_config_ulong(k, v);
 +              return 0;
 +      }
 +      return git_default_config(k, v, cb);
 +}
 +
 +static void read_object_list_from_stdin(void)
 +{
 +      char line[40 + 1 + PATH_MAX + 2];
 +      unsigned char sha1[20];
 +
 +      for (;;) {
 +              if (!fgets(line, sizeof(line), stdin)) {
 +                      if (feof(stdin))
 +                              break;
 +                      if (!ferror(stdin))
 +                              die("fgets returned NULL, not EOF, not error!");
 +                      if (errno != EINTR)
 +                              die_errno("fgets");
 +                      clearerr(stdin);
 +                      continue;
 +              }
 +              if (line[0] == '-') {
 +                      if (get_sha1_hex(line+1, sha1))
 +                              die("expected edge sha1, got garbage:\n %s",
 +                                  line);
 +                      add_preferred_base(sha1);
 +                      continue;
 +              }
 +              if (get_sha1_hex(line, sha1))
 +                      die("expected sha1, got garbage:\n %s", line);
 +
 +              add_preferred_base_object(line+41);
 +              add_object_entry(sha1, 0, line+41, 0);
 +      }
 +}
 +
 +#define OBJECT_ADDED (1u<<20)
 +
 +static void show_commit(struct commit *commit, void *data)
 +{
 +      add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
 +      commit->object.flags |= OBJECT_ADDED;
 +}
 +
 +static void show_object(struct object *obj, const struct name_path *path, const char *last)
 +{
 +      char *name = path_name(path, last);
 +
 +      add_preferred_base_object(name);
 +      add_object_entry(obj->sha1, obj->type, name, 0);
 +      obj->flags |= OBJECT_ADDED;
 +
 +      /*
 +       * We will have generated the hash from the name,
 +       * but not saved a pointer to it - we can free it
 +       */
 +      free((char *)name);
 +}
 +
 +static void show_edge(struct commit *commit)
 +{
 +      add_preferred_base(commit->object.sha1);
 +}
 +
 +struct in_pack_object {
 +      off_t offset;
 +      struct object *object;
 +};
 +
 +struct in_pack {
 +      int alloc;
 +      int nr;
 +      struct in_pack_object *array;
 +};
 +
 +static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
 +{
 +      in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
 +      in_pack->array[in_pack->nr].object = object;
 +      in_pack->nr++;
 +}
 +
 +/*
 + * Compare the objects in the offset order, in order to emulate the
 + * "git rev-list --objects" output that produced the pack originally.
 + */
 +static int ofscmp(const void *a_, const void *b_)
 +{
 +      struct in_pack_object *a = (struct in_pack_object *)a_;
 +      struct in_pack_object *b = (struct in_pack_object *)b_;
 +
 +      if (a->offset < b->offset)
 +              return -1;
 +      else if (a->offset > b->offset)
 +              return 1;
 +      else
 +              return hashcmp(a->object->sha1, b->object->sha1);
 +}
 +
 +static void add_objects_in_unpacked_packs(struct rev_info *revs)
 +{
 +      struct packed_git *p;
 +      struct in_pack in_pack;
 +      uint32_t i;
 +
 +      memset(&in_pack, 0, sizeof(in_pack));
 +
 +      for (p = packed_git; p; p = p->next) {
 +              const unsigned char *sha1;
 +              struct object *o;
 +
 +              if (!p->pack_local || p->pack_keep)
 +                      continue;
 +              if (open_pack_index(p))
 +                      die("cannot open pack index");
 +
 +              ALLOC_GROW(in_pack.array,
 +                         in_pack.nr + p->num_objects,
 +                         in_pack.alloc);
 +
 +              for (i = 0; i < p->num_objects; i++) {
 +                      sha1 = nth_packed_object_sha1(p, i);
 +                      o = lookup_unknown_object(sha1);
 +                      if (!(o->flags & OBJECT_ADDED))
 +                              mark_in_pack_object(o, p, &in_pack);
 +                      o->flags |= OBJECT_ADDED;
 +              }
 +      }
 +
 +      if (in_pack.nr) {
 +              qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
 +                    ofscmp);
 +              for (i = 0; i < in_pack.nr; i++) {
 +                      struct object *o = in_pack.array[i].object;
 +                      add_object_entry(o->sha1, o->type, "", 0);
 +              }
 +      }
 +      free(in_pack.array);
 +}
 +
 +static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
 +{
 +      static struct packed_git *last_found = (void *)1;
 +      struct packed_git *p;
 +
 +      p = (last_found != (void *)1) ? last_found : packed_git;
 +
 +      while (p) {
 +              if ((!p->pack_local || p->pack_keep) &&
 +                      find_pack_entry_one(sha1, p)) {
 +                      last_found = p;
 +                      return 1;
 +              }
 +              if (p == last_found)
 +                      p = packed_git;
 +              else
 +                      p = p->next;
 +              if (p == last_found)
 +                      p = p->next;
 +      }
 +      return 0;
 +}
 +
 +static void loosen_unused_packed_objects(struct rev_info *revs)
 +{
 +      struct packed_git *p;
 +      uint32_t i;
 +      const unsigned char *sha1;
 +
 +      for (p = packed_git; p; p = p->next) {
 +              if (!p->pack_local || p->pack_keep)
 +                      continue;
 +
 +              if (open_pack_index(p))
 +                      die("cannot open pack index");
 +
 +              for (i = 0; i < p->num_objects; i++) {
 +                      sha1 = nth_packed_object_sha1(p, i);
 +                      if (!locate_object_entry(sha1) &&
 +                              !has_sha1_pack_kept_or_nonlocal(sha1))
 +                              if (force_object_loose(sha1, p->mtime))
 +                                      die("unable to force loose object");
 +              }
 +      }
 +}
 +
 +static void get_object_list(int ac, const char **av)
 +{
 +      struct rev_info revs;
 +      char line[1000];
 +      int flags = 0;
 +
 +      init_revisions(&revs, NULL);
 +      save_commit_buffer = 0;
 +      setup_revisions(ac, av, &revs, NULL);
 +
 +      while (fgets(line, sizeof(line), stdin) != NULL) {
 +              int len = strlen(line);
 +              if (len && line[len - 1] == '\n')
 +                      line[--len] = 0;
 +              if (!len)
 +                      break;
 +              if (*line == '-') {
 +                      if (!strcmp(line, "--not")) {
 +                              flags ^= UNINTERESTING;
 +                              continue;
 +                      }
 +                      die("not a rev '%s'", line);
 +              }
 +              if (handle_revision_arg(line, &revs, flags, 1))
 +                      die("bad revision '%s'", line);
 +      }
 +
 +      if (prepare_revision_walk(&revs))
 +              die("revision walk setup failed");
 +      mark_edges_uninteresting(revs.commits, &revs, show_edge);
 +      traverse_commit_list(&revs, show_commit, show_object, NULL);
 +
 +      if (keep_unreachable)
 +              add_objects_in_unpacked_packs(&revs);
 +      if (unpack_unreachable)
 +              loosen_unused_packed_objects(&revs);
 +}
 +
 +int cmd_pack_objects(int argc, const char **argv, const char *prefix)
 +{
 +      int use_internal_rev_list = 0;
 +      int thin = 0;
 +      int all_progress_implied = 0;
 +      uint32_t i;
 +      const char **rp_av;
 +      int rp_ac_alloc = 64;
 +      int rp_ac;
 +
 +      read_replace_refs = 0;
 +
 +      rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
 +
 +      rp_av[0] = "pack-objects";
 +      rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
 +      rp_ac = 2;
 +
 +      git_config(git_pack_config, NULL);
 +      if (!pack_compression_seen && core_compression_seen)
 +              pack_compression_level = core_compression_level;
 +
 +      progress = isatty(2);
 +      for (i = 1; i < argc; i++) {
 +              const char *arg = argv[i];
 +
 +              if (*arg != '-')
 +                      break;
 +
 +              if (!strcmp("--non-empty", arg)) {
 +                      non_empty = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--local", arg)) {
 +                      local = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--incremental", arg)) {
 +                      incremental = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--honor-pack-keep", arg)) {
 +                      ignore_packed_keep = 1;
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--compression=")) {
 +                      char *end;
 +                      int level = strtoul(arg+14, &end, 0);
 +                      if (!arg[14] || *end)
 +                              usage(pack_usage);
 +                      if (level == -1)
 +                              level = Z_DEFAULT_COMPRESSION;
 +                      else if (level < 0 || level > Z_BEST_COMPRESSION)
 +                              die("bad pack compression level %d", level);
 +                      pack_compression_level = level;
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--max-pack-size=")) {
 +                      pack_size_limit_cfg = 0;
 +                      if (!git_parse_ulong(arg+16, &pack_size_limit))
 +                              usage(pack_usage);
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--window=")) {
 +                      char *end;
 +                      window = strtoul(arg+9, &end, 0);
 +                      if (!arg[9] || *end)
 +                              usage(pack_usage);
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--window-memory=")) {
 +                      if (!git_parse_ulong(arg+16, &window_memory_limit))
 +                              usage(pack_usage);
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--threads=")) {
 +                      char *end;
 +                      delta_search_threads = strtoul(arg+10, &end, 0);
 +                      if (!arg[10] || *end || delta_search_threads < 0)
 +                              usage(pack_usage);
 +#ifdef NO_PTHREADS
 +                      if (delta_search_threads != 1)
 +                              warning("no threads support, "
 +                                      "ignoring %s", arg);
 +#endif
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--depth=")) {
 +                      char *end;
 +                      depth = strtoul(arg+8, &end, 0);
 +                      if (!arg[8] || *end)
 +                              usage(pack_usage);
 +                      continue;
 +              }
 +              if (!strcmp("--progress", arg)) {
 +                      progress = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--all-progress", arg)) {
 +                      progress = 2;
 +                      continue;
 +              }
 +              if (!strcmp("--all-progress-implied", arg)) {
 +                      all_progress_implied = 1;
 +                      continue;
 +              }
 +              if (!strcmp("-q", arg)) {
 +                      progress = 0;
 +                      continue;
 +              }
 +              if (!strcmp("--no-reuse-delta", arg)) {
 +                      reuse_delta = 0;
 +                      continue;
 +              }
 +              if (!strcmp("--no-reuse-object", arg)) {
 +                      reuse_object = reuse_delta = 0;
 +                      continue;
 +              }
 +              if (!strcmp("--delta-base-offset", arg)) {
 +                      allow_ofs_delta = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--stdout", arg)) {
 +                      pack_to_stdout = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--revs", arg)) {
 +                      use_internal_rev_list = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--keep-unreachable", arg)) {
 +                      keep_unreachable = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--unpack-unreachable", arg)) {
 +                      unpack_unreachable = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--include-tag", arg)) {
 +                      include_tag = 1;
 +                      continue;
 +              }
 +              if (!strcmp("--unpacked", arg) ||
 +                  !strcmp("--reflog", arg) ||
 +                  !strcmp("--all", arg)) {
 +                      use_internal_rev_list = 1;
 +                      if (rp_ac >= rp_ac_alloc - 1) {
 +                              rp_ac_alloc = alloc_nr(rp_ac_alloc);
 +                              rp_av = xrealloc(rp_av,
 +                                               rp_ac_alloc * sizeof(*rp_av));
 +                      }
 +                      rp_av[rp_ac++] = arg;
 +                      continue;
 +              }
 +              if (!strcmp("--thin", arg)) {
 +                      use_internal_rev_list = 1;
 +                      thin = 1;
 +                      rp_av[1] = "--objects-edge";
 +                      continue;
 +              }
 +              if (!prefixcmp(arg, "--index-version=")) {
 +                      char *c;
 +                      pack_idx_default_version = strtoul(arg + 16, &c, 10);
 +                      if (pack_idx_default_version > 2)
 +                              die("bad %s", arg);
 +                      if (*c == ',')
 +                              pack_idx_off32_limit = strtoul(c+1, &c, 0);
 +                      if (*c || pack_idx_off32_limit & 0x80000000)
 +                              die("bad %s", arg);
 +                      continue;
 +              }
 +              if (!strcmp(arg, "--keep-true-parents")) {
 +                      grafts_replace_parents = 0;
 +                      continue;
 +              }
 +              usage(pack_usage);
 +      }
 +
 +      /* Traditionally "pack-objects [options] base extra" failed;
 +       * we would however want to take refs parameter that would
 +       * have been given to upstream rev-list ourselves, which means
 +       * we somehow want to say what the base name is.  So the
 +       * syntax would be:
 +       *
 +       * pack-objects [options] base <refs...>
 +       *
 +       * in other words, we would treat the first non-option as the
 +       * base_name and send everything else to the internal revision
 +       * walker.
 +       */
 +
 +      if (!pack_to_stdout)
 +              base_name = argv[i++];
 +
 +      if (pack_to_stdout != !base_name)
 +              usage(pack_usage);
 +
 +      if (!pack_to_stdout && !pack_size_limit)
 +              pack_size_limit = pack_size_limit_cfg;
 +      if (pack_to_stdout && pack_size_limit)
 +              die("--max-pack-size cannot be used to build a pack for transfer.");
 +      if (pack_size_limit && pack_size_limit < 1024*1024) {
 +              warning("minimum pack size limit is 1 MiB");
 +              pack_size_limit = 1024*1024;
 +      }
 +
 +      if (!pack_to_stdout && thin)
 +              die("--thin cannot be used to build an indexable pack.");
 +
 +      if (keep_unreachable && unpack_unreachable)
 +              die("--keep-unreachable and --unpack-unreachable are incompatible.");
 +
 +      if (progress && all_progress_implied)
 +              progress = 2;
 +
 +      prepare_packed_git();
 +
 +      if (progress)
 +              progress_state = start_progress("Counting objects", 0);
 +      if (!use_internal_rev_list)
 +              read_object_list_from_stdin();
 +      else {
 +              rp_av[rp_ac] = NULL;
 +              get_object_list(rp_ac, rp_av);
 +      }
 +      cleanup_preferred_base();
 +      if (include_tag && nr_result)
 +              for_each_ref(add_ref_tag, NULL);
 +      stop_progress(&progress_state);
 +
 +      if (non_empty && !nr_result)
 +              return 0;
 +      if (nr_result)
 +              prepare_pack(window, depth);
 +      write_pack_file();
 +      if (progress)
 +              fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
 +                      " reused %"PRIu32" (delta %"PRIu32")\n",
 +                      written, written_delta, reused, reused_delta);
 +      return 0;
 +}
index 6e259772895599b35d13564e62077b882f0de3a6,1c171db8b1fb46c5230280b0ebd872c8f23cf7c0..001e5db4bbfb55c954168dd9b35cf3861b7a0c29
@@@ -346,16 -344,10 +346,18 @@@ static inline char *gitstrchrnul(const 
  }
  #endif
  
 +#ifndef HAVE_MEMPCPY
 +#define mempcpy gitmempcpy
 +static inline void *gitmempcpy(void *dest, const void *src, size_t n)
 +{
 +      return (char *)memcpy(dest, src, n) + n;
 +}
 +#endif
 +
  extern void release_pack_memory(size_t, int);
  
+ extern void set_try_to_free_routine(void (*routine)(size_t));
  extern char *xstrdup(const char *str);
  extern void *xmalloc(size_t size);
  extern void *xmallocz(size_t size);
diff --cc wrapper.c
Simple merge