verify_signed_buffer: prefer close_tempfile() to close()
[gitweb.git] / packfile.c
index 60d9fc3b036a7d7ab62c6adb01953f743ddceb6a..f86fa051c9e67def994b8b487ad34f200967d5e9 100644 (file)
@@ -1,5 +1,13 @@
 #include "cache.h"
 #include "mru.h"
+#include "pack.h"
+#include "dir.h"
+#include "mergesort.h"
+#include "packfile.h"
+#include "delta.h"
+#include "list.h"
+#include "streaming.h"
+#include "sha1-lookup.h"
 
 char *odb_pack_name(struct strbuf *buf,
                    const unsigned char *sha1,
@@ -23,14 +31,14 @@ char *sha1_pack_index_name(const unsigned char *sha1)
        return odb_pack_name(&buf, sha1, "idx");
 }
 
-unsigned int pack_used_ctr;
-unsigned int pack_mmap_calls;
-unsigned int peak_pack_open_windows;
-unsigned int pack_open_windows;
-unsigned int pack_open_fds;
-unsigned int pack_max_fds;
-size_t peak_pack_mapped;
-size_t pack_mapped;
+static unsigned int pack_used_ctr;
+static unsigned int pack_mmap_calls;
+static unsigned int peak_pack_open_windows;
+static unsigned int pack_open_windows;
+static unsigned int pack_open_fds;
+static unsigned int pack_max_fds;
+static size_t peak_pack_mapped;
+static size_t pack_mapped;
 struct packed_git *packed_git;
 
 static struct mru packed_git_mru_storage;
@@ -59,3 +67,1830 @@ void pack_report(void)
                pack_open_windows, peak_pack_open_windows,
                sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 }
+
+/*
+ * Open and mmap the index file at path, perform a couple of
+ * consistency checks, then record its information to p.  Return 0 on
+ * success.
+ */
+static int check_packed_git_idx(const char *path, struct packed_git *p)
+{
+       void *idx_map;
+       struct pack_idx_header *hdr;
+       size_t idx_size;
+       uint32_t version, nr, i, *index;
+       int fd = git_open(path);
+       struct stat st;
+
+       if (fd < 0)
+               return -1;
+       if (fstat(fd, &st)) {
+               close(fd);
+               return -1;
+       }
+       idx_size = xsize_t(st.st_size);
+       if (idx_size < 4 * 256 + 20 + 20) {
+               close(fd);
+               return error("index file %s is too small", path);
+       }
+       idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
+       close(fd);
+
+       hdr = idx_map;
+       if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
+               version = ntohl(hdr->idx_version);
+               if (version < 2 || version > 2) {
+                       munmap(idx_map, idx_size);
+                       return error("index file %s is version %"PRIu32
+                                    " and is not supported by this binary"
+                                    " (try upgrading GIT to a newer version)",
+                                    path, version);
+               }
+       } else
+               version = 1;
+
+       nr = 0;
+       index = idx_map;
+       if (version > 1)
+               index += 2;  /* skip index header */
+       for (i = 0; i < 256; i++) {
+               uint32_t n = ntohl(index[i]);
+               if (n < nr) {
+                       munmap(idx_map, idx_size);
+                       return error("non-monotonic index %s", path);
+               }
+               nr = n;
+       }
+
+       if (version == 1) {
+               /*
+                * Total size:
+                *  - 256 index entries 4 bytes each
+                *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
+                *  - 20-byte SHA1 of the packfile
+                *  - 20-byte SHA1 file checksum
+                */
+               if (idx_size != 4*256 + nr * 24 + 20 + 20) {
+                       munmap(idx_map, idx_size);
+                       return error("wrong index v1 file size in %s", path);
+               }
+       } else if (version == 2) {
+               /*
+                * Minimum size:
+                *  - 8 bytes of header
+                *  - 256 index entries 4 bytes each
+                *  - 20-byte sha1 entry * nr
+                *  - 4-byte crc entry * nr
+                *  - 4-byte offset entry * nr
+                *  - 20-byte SHA1 of the packfile
+                *  - 20-byte SHA1 file checksum
+                * And after the 4-byte offset table might be a
+                * variable sized table containing 8-byte entries
+                * for offsets larger than 2^31.
+                */
+               unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
+               unsigned long max_size = min_size;
+               if (nr)
+                       max_size += (nr - 1)*8;
+               if (idx_size < min_size || idx_size > max_size) {
+                       munmap(idx_map, idx_size);
+                       return error("wrong index v2 file size in %s", path);
+               }
+               if (idx_size != min_size &&
+                   /*
+                    * make sure we can deal with large pack offsets.
+                    * 31-bit signed offset won't be enough, neither
+                    * 32-bit unsigned one will be.
+                    */
+                   (sizeof(off_t) <= 4)) {
+                       munmap(idx_map, idx_size);
+                       return error("pack too large for current definition of off_t in %s", path);
+               }
+       }
+
+       p->index_version = version;
+       p->index_data = idx_map;
+       p->index_size = idx_size;
+       p->num_objects = nr;
+       return 0;
+}
+
+int open_pack_index(struct packed_git *p)
+{
+       char *idx_name;
+       size_t len;
+       int ret;
+
+       if (p->index_data)
+               return 0;
+
+       if (!strip_suffix(p->pack_name, ".pack", &len))
+               die("BUG: pack_name does not end in .pack");
+       idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
+       ret = check_packed_git_idx(idx_name, p);
+       free(idx_name);
+       return ret;
+}
+
+static struct packed_git *alloc_packed_git(int extra)
+{
+       struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
+       memset(p, 0, sizeof(*p));
+       p->pack_fd = -1;
+       return p;
+}
+
+struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
+{
+       const char *path = sha1_pack_name(sha1);
+       size_t alloc = st_add(strlen(path), 1);
+       struct packed_git *p = alloc_packed_git(alloc);
+
+       memcpy(p->pack_name, path, alloc); /* includes NUL */
+       hashcpy(p->sha1, sha1);
+       if (check_packed_git_idx(idx_path, p)) {
+               free(p);
+               return NULL;
+       }
+
+       return p;
+}
+
+static void scan_windows(struct packed_git *p,
+       struct packed_git **lru_p,
+       struct pack_window **lru_w,
+       struct pack_window **lru_l)
+{
+       struct pack_window *w, *w_l;
+
+       for (w_l = NULL, w = p->windows; w; w = w->next) {
+               if (!w->inuse_cnt) {
+                       if (!*lru_w || w->last_used < (*lru_w)->last_used) {
+                               *lru_p = p;
+                               *lru_w = w;
+                               *lru_l = w_l;
+                       }
+               }
+               w_l = w;
+       }
+}
+
+static int unuse_one_window(struct packed_git *current)
+{
+       struct packed_git *p, *lru_p = NULL;
+       struct pack_window *lru_w = NULL, *lru_l = NULL;
+
+       if (current)
+               scan_windows(current, &lru_p, &lru_w, &lru_l);
+       for (p = packed_git; p; p = p->next)
+               scan_windows(p, &lru_p, &lru_w, &lru_l);
+       if (lru_p) {
+               munmap(lru_w->base, lru_w->len);
+               pack_mapped -= lru_w->len;
+               if (lru_l)
+                       lru_l->next = lru_w->next;
+               else
+                       lru_p->windows = lru_w->next;
+               free(lru_w);
+               pack_open_windows--;
+               return 1;
+       }
+       return 0;
+}
+
+void release_pack_memory(size_t need)
+{
+       size_t cur = pack_mapped;
+       while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
+               ; /* nothing */
+}
+
+void close_pack_windows(struct packed_git *p)
+{
+       while (p->windows) {
+               struct pack_window *w = p->windows;
+
+               if (w->inuse_cnt)
+                       die("pack '%s' still has open windows to it",
+                           p->pack_name);
+               munmap(w->base, w->len);
+               pack_mapped -= w->len;
+               pack_open_windows--;
+               p->windows = w->next;
+               free(w);
+       }
+}
+
+static int close_pack_fd(struct packed_git *p)
+{
+       if (p->pack_fd < 0)
+               return 0;
+
+       close(p->pack_fd);
+       pack_open_fds--;
+       p->pack_fd = -1;
+
+       return 1;
+}
+
+void close_pack_index(struct packed_git *p)
+{
+       if (p->index_data) {
+               munmap((void *)p->index_data, p->index_size);
+               p->index_data = NULL;
+       }
+}
+
+static void close_pack(struct packed_git *p)
+{
+       close_pack_windows(p);
+       close_pack_fd(p);
+       close_pack_index(p);
+}
+
+void close_all_packs(void)
+{
+       struct packed_git *p;
+
+       for (p = packed_git; p; p = p->next)
+               if (p->do_not_close)
+                       die("BUG: want to close pack marked 'do-not-close'");
+               else
+                       close_pack(p);
+}
+
+/*
+ * The LRU pack is the one with the oldest MRU window, preferring packs
+ * with no used windows, or the oldest mtime if it has no windows allocated.
+ */
+static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
+{
+       struct pack_window *w, *this_mru_w;
+       int has_windows_inuse = 0;
+
+       /*
+        * Reject this pack if it has windows and the previously selected
+        * one does not.  If this pack does not have windows, reject
+        * it if the pack file is newer than the previously selected one.
+        */
+       if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
+               return;
+
+       for (w = this_mru_w = p->windows; w; w = w->next) {
+               /*
+                * Reject this pack if any of its windows are in use,
+                * but the previously selected pack did not have any
+                * inuse windows.  Otherwise, record that this pack
+                * has windows in use.
+                */
+               if (w->inuse_cnt) {
+                       if (*accept_windows_inuse)
+                               has_windows_inuse = 1;
+                       else
+                               return;
+               }
+
+               if (w->last_used > this_mru_w->last_used)
+                       this_mru_w = w;
+
+               /*
+                * Reject this pack if it has windows that have been
+                * used more recently than the previously selected pack.
+                * If the previously selected pack had windows inuse and
+                * we have not encountered a window in this pack that is
+                * inuse, skip this check since we prefer a pack with no
+                * inuse windows to one that has inuse windows.
+                */
+               if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
+                   this_mru_w->last_used > (*mru_w)->last_used)
+                       return;
+       }
+
+       /*
+        * Select this pack.
+        */
+       *mru_w = this_mru_w;
+       *lru_p = p;
+       *accept_windows_inuse = has_windows_inuse;
+}
+
+static int close_one_pack(void)
+{
+       struct packed_git *p, *lru_p = NULL;
+       struct pack_window *mru_w = NULL;
+       int accept_windows_inuse = 1;
+
+       for (p = packed_git; p; p = p->next) {
+               if (p->pack_fd == -1)
+                       continue;
+               find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
+       }
+
+       if (lru_p)
+               return close_pack_fd(lru_p);
+
+       return 0;
+}
+
+static unsigned int get_max_fd_limit(void)
+{
+#ifdef RLIMIT_NOFILE
+       {
+               struct rlimit lim;
+
+               if (!getrlimit(RLIMIT_NOFILE, &lim))
+                       return lim.rlim_cur;
+       }
+#endif
+
+#ifdef _SC_OPEN_MAX
+       {
+               long open_max = sysconf(_SC_OPEN_MAX);
+               if (0 < open_max)
+                       return open_max;
+               /*
+                * Otherwise, we got -1 for one of the two
+                * reasons:
+                *
+                * (1) sysconf() did not understand _SC_OPEN_MAX
+                *     and signaled an error with -1; or
+                * (2) sysconf() said there is no limit.
+                *
+                * We _could_ clear errno before calling sysconf() to
+                * tell these two cases apart and return a huge number
+                * in the latter case to let the caller cap it to a
+                * value that is not so selfish, but letting the
+                * fallback OPEN_MAX codepath take care of these cases
+                * is a lot simpler.
+                */
+       }
+#endif
+
+#ifdef OPEN_MAX
+       return OPEN_MAX;
+#else
+       return 1; /* see the caller ;-) */
+#endif
+}
+
+/*
+ * Do not call this directly as this leaks p->pack_fd on error return;
+ * call open_packed_git() instead.
+ */
+static int open_packed_git_1(struct packed_git *p)
+{
+       struct stat st;
+       struct pack_header hdr;
+       unsigned char sha1[20];
+       unsigned char *idx_sha1;
+       long fd_flag;
+
+       if (!p->index_data && open_pack_index(p))
+               return error("packfile %s index unavailable", p->pack_name);
+
+       if (!pack_max_fds) {
+               unsigned int max_fds = get_max_fd_limit();
+
+               /* Save 3 for stdin/stdout/stderr, 22 for work */
+               if (25 < max_fds)
+                       pack_max_fds = max_fds - 25;
+               else
+                       pack_max_fds = 1;
+       }
+
+       while (pack_max_fds <= pack_open_fds && close_one_pack())
+               ; /* nothing */
+
+       p->pack_fd = git_open(p->pack_name);
+       if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
+               return -1;
+       pack_open_fds++;
+
+       /* If we created the struct before we had the pack we lack size. */
+       if (!p->pack_size) {
+               if (!S_ISREG(st.st_mode))
+                       return error("packfile %s not a regular file", p->pack_name);
+               p->pack_size = st.st_size;
+       } else if (p->pack_size != st.st_size)
+               return error("packfile %s size changed", p->pack_name);
+
+       /* We leave these file descriptors open with sliding mmap;
+        * there is no point keeping them open across exec(), though.
+        */
+       fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
+       if (fd_flag < 0)
+               return error("cannot determine file descriptor flags");
+       fd_flag |= FD_CLOEXEC;
+       if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
+               return error("cannot set FD_CLOEXEC");
+
+       /* Verify we recognize this pack file format. */
+       if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
+               return error("file %s is far too short to be a packfile", p->pack_name);
+       if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
+               return error("file %s is not a GIT packfile", p->pack_name);
+       if (!pack_version_ok(hdr.hdr_version))
+               return error("packfile %s is version %"PRIu32" and not"
+                       " supported (try upgrading GIT to a newer version)",
+                       p->pack_name, ntohl(hdr.hdr_version));
+
+       /* Verify the pack matches its index. */
+       if (p->num_objects != ntohl(hdr.hdr_entries))
+               return error("packfile %s claims to have %"PRIu32" objects"
+                            " while index indicates %"PRIu32" objects",
+                            p->pack_name, ntohl(hdr.hdr_entries),
+                            p->num_objects);
+       if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
+               return error("end of packfile %s is unavailable", p->pack_name);
+       if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
+               return error("packfile %s signature is unavailable", p->pack_name);
+       idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
+       if (hashcmp(sha1, idx_sha1))
+               return error("packfile %s does not match index", p->pack_name);
+       return 0;
+}
+
+static int open_packed_git(struct packed_git *p)
+{
+       if (!open_packed_git_1(p))
+               return 0;
+       close_pack_fd(p);
+       return -1;
+}
+
+static int in_window(struct pack_window *win, off_t offset)
+{
+       /* We must promise at least 20 bytes (one hash) after the
+        * offset is available from this window, otherwise the offset
+        * is not actually in this window and a different window (which
+        * has that one hash excess) must be used.  This is to support
+        * the object header and delta base parsing routines below.
+        */
+       off_t win_off = win->offset;
+       return win_off <= offset
+               && (offset + 20) <= (win_off + win->len);
+}
+
+unsigned char *use_pack(struct packed_git *p,
+               struct pack_window **w_cursor,
+               off_t offset,
+               unsigned long *left)
+{
+       struct pack_window *win = *w_cursor;
+
+       /* Since packfiles end in a hash of their content and it's
+        * pointless to ask for an offset into the middle of that
+        * hash, and the in_window function above wouldn't match
+        * don't allow an offset too close to the end of the file.
+        */
+       if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
+               die("packfile %s cannot be accessed", p->pack_name);
+       if (offset > (p->pack_size - 20))
+               die("offset beyond end of packfile (truncated pack?)");
+       if (offset < 0)
+               die(_("offset before end of packfile (broken .idx?)"));
+
+       if (!win || !in_window(win, offset)) {
+               if (win)
+                       win->inuse_cnt--;
+               for (win = p->windows; win; win = win->next) {
+                       if (in_window(win, offset))
+                               break;
+               }
+               if (!win) {
+                       size_t window_align = packed_git_window_size / 2;
+                       off_t len;
+
+                       if (p->pack_fd == -1 && open_packed_git(p))
+                               die("packfile %s cannot be accessed", p->pack_name);
+
+                       win = xcalloc(1, sizeof(*win));
+                       win->offset = (offset / window_align) * window_align;
+                       len = p->pack_size - win->offset;
+                       if (len > packed_git_window_size)
+                               len = packed_git_window_size;
+                       win->len = (size_t)len;
+                       pack_mapped += win->len;
+                       while (packed_git_limit < pack_mapped
+                               && unuse_one_window(p))
+                               ; /* nothing */
+                       win->base = xmmap(NULL, win->len,
+                               PROT_READ, MAP_PRIVATE,
+                               p->pack_fd, win->offset);
+                       if (win->base == MAP_FAILED)
+                               die_errno("packfile %s cannot be mapped",
+                                         p->pack_name);
+                       if (!win->offset && win->len == p->pack_size
+                               && !p->do_not_close)
+                               close_pack_fd(p);
+                       pack_mmap_calls++;
+                       pack_open_windows++;
+                       if (pack_mapped > peak_pack_mapped)
+                               peak_pack_mapped = pack_mapped;
+                       if (pack_open_windows > peak_pack_open_windows)
+                               peak_pack_open_windows = pack_open_windows;
+                       win->next = p->windows;
+                       p->windows = win;
+               }
+       }
+       if (win != *w_cursor) {
+               win->last_used = pack_used_ctr++;
+               win->inuse_cnt++;
+               *w_cursor = win;
+       }
+       offset -= win->offset;
+       if (left)
+               *left = win->len - xsize_t(offset);
+       return win->base + offset;
+}
+
+void unuse_pack(struct pack_window **w_cursor)
+{
+       struct pack_window *w = *w_cursor;
+       if (w) {
+               w->inuse_cnt--;
+               *w_cursor = NULL;
+       }
+}
+
+static void try_to_free_pack_memory(size_t size)
+{
+       release_pack_memory(size);
+}
+
+struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
+{
+       static int have_set_try_to_free_routine;
+       struct stat st;
+       size_t alloc;
+       struct packed_git *p;
+
+       if (!have_set_try_to_free_routine) {
+               have_set_try_to_free_routine = 1;
+               set_try_to_free_routine(try_to_free_pack_memory);
+       }
+
+       /*
+        * Make sure a corresponding .pack file exists and that
+        * the index looks sane.
+        */
+       if (!strip_suffix_mem(path, &path_len, ".idx"))
+               return NULL;
+
+       /*
+        * ".pack" is long enough to hold any suffix we're adding (and
+        * the use xsnprintf double-checks that)
+        */
+       alloc = st_add3(path_len, strlen(".pack"), 1);
+       p = alloc_packed_git(alloc);
+       memcpy(p->pack_name, path, path_len);
+
+       xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
+       if (!access(p->pack_name, F_OK))
+               p->pack_keep = 1;
+
+       xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
+       if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
+               free(p);
+               return NULL;
+       }
+
+       /* ok, it looks sane as far as we can check without
+        * actually mapping the pack file.
+        */
+       p->pack_size = st.st_size;
+       p->pack_local = local;
+       p->mtime = st.st_mtime;
+       if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
+               hashclr(p->sha1);
+       return p;
+}
+
+void install_packed_git(struct packed_git *pack)
+{
+       if (pack->pack_fd != -1)
+               pack_open_fds++;
+
+       pack->next = packed_git;
+       packed_git = pack;
+}
+
+void (*report_garbage)(unsigned seen_bits, const char *path);
+
+static void report_helper(const struct string_list *list,
+                         int seen_bits, int first, int last)
+{
+       if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
+               return;
+
+       for (; first < last; first++)
+               report_garbage(seen_bits, list->items[first].string);
+}
+
+static void report_pack_garbage(struct string_list *list)
+{
+       int i, baselen = -1, first = 0, seen_bits = 0;
+
+       if (!report_garbage)
+               return;
+
+       string_list_sort(list);
+
+       for (i = 0; i < list->nr; i++) {
+               const char *path = list->items[i].string;
+               if (baselen != -1 &&
+                   strncmp(path, list->items[first].string, baselen)) {
+                       report_helper(list, seen_bits, first, i);
+                       baselen = -1;
+                       seen_bits = 0;
+               }
+               if (baselen == -1) {
+                       const char *dot = strrchr(path, '.');
+                       if (!dot) {
+                               report_garbage(PACKDIR_FILE_GARBAGE, path);
+                               continue;
+                       }
+                       baselen = dot - path + 1;
+                       first = i;
+               }
+               if (!strcmp(path + baselen, "pack"))
+                       seen_bits |= 1;
+               else if (!strcmp(path + baselen, "idx"))
+                       seen_bits |= 2;
+       }
+       report_helper(list, seen_bits, first, list->nr);
+}
+
+static void prepare_packed_git_one(char *objdir, int local)
+{
+       struct strbuf path = STRBUF_INIT;
+       size_t dirnamelen;
+       DIR *dir;
+       struct dirent *de;
+       struct string_list garbage = STRING_LIST_INIT_DUP;
+
+       strbuf_addstr(&path, objdir);
+       strbuf_addstr(&path, "/pack");
+       dir = opendir(path.buf);
+       if (!dir) {
+               if (errno != ENOENT)
+                       error_errno("unable to open object pack directory: %s",
+                                   path.buf);
+               strbuf_release(&path);
+               return;
+       }
+       strbuf_addch(&path, '/');
+       dirnamelen = path.len;
+       while ((de = readdir(dir)) != NULL) {
+               struct packed_git *p;
+               size_t base_len;
+
+               if (is_dot_or_dotdot(de->d_name))
+                       continue;
+
+               strbuf_setlen(&path, dirnamelen);
+               strbuf_addstr(&path, de->d_name);
+
+               base_len = path.len;
+               if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
+                       /* Don't reopen a pack we already have. */
+                       for (p = packed_git; p; p = p->next) {
+                               size_t len;
+                               if (strip_suffix(p->pack_name, ".pack", &len) &&
+                                   len == base_len &&
+                                   !memcmp(p->pack_name, path.buf, len))
+                                       break;
+                       }
+                       if (p == NULL &&
+                           /*
+                            * See if it really is a valid .idx file with
+                            * corresponding .pack file that we can map.
+                            */
+                           (p = add_packed_git(path.buf, path.len, local)) != NULL)
+                               install_packed_git(p);
+               }
+
+               if (!report_garbage)
+                       continue;
+
+               if (ends_with(de->d_name, ".idx") ||
+                   ends_with(de->d_name, ".pack") ||
+                   ends_with(de->d_name, ".bitmap") ||
+                   ends_with(de->d_name, ".keep"))
+                       string_list_append(&garbage, path.buf);
+               else
+                       report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
+       }
+       closedir(dir);
+       report_pack_garbage(&garbage);
+       string_list_clear(&garbage, 0);
+       strbuf_release(&path);
+}
+
+static int approximate_object_count_valid;
+
+/*
+ * Give a fast, rough count of the number of objects in the repository. This
+ * ignores loose objects completely. If you have a lot of them, then either
+ * you should repack because your performance will be awful, or they are
+ * all unreachable objects about to be pruned, in which case they're not really
+ * interesting as a measure of repo size in the first place.
+ */
+unsigned long approximate_object_count(void)
+{
+       static unsigned long count;
+       if (!approximate_object_count_valid) {
+               struct packed_git *p;
+
+               prepare_packed_git();
+               count = 0;
+               for (p = packed_git; p; p = p->next) {
+                       if (open_pack_index(p))
+                               continue;
+                       count += p->num_objects;
+               }
+       }
+       return count;
+}
+
+static void *get_next_packed_git(const void *p)
+{
+       return ((const struct packed_git *)p)->next;
+}
+
+static void set_next_packed_git(void *p, void *next)
+{
+       ((struct packed_git *)p)->next = next;
+}
+
+static int sort_pack(const void *a_, const void *b_)
+{
+       const struct packed_git *a = a_;
+       const struct packed_git *b = b_;
+       int st;
+
+       /*
+        * Local packs tend to contain objects specific to our
+        * variant of the project than remote ones.  In addition,
+        * remote ones could be on a network mounted filesystem.
+        * Favor local ones for these reasons.
+        */
+       st = a->pack_local - b->pack_local;
+       if (st)
+               return -st;
+
+       /*
+        * Younger packs tend to contain more recent objects,
+        * and more recent objects tend to get accessed more
+        * often.
+        */
+       if (a->mtime < b->mtime)
+               return 1;
+       else if (a->mtime == b->mtime)
+               return 0;
+       return -1;
+}
+
+static void rearrange_packed_git(void)
+{
+       packed_git = llist_mergesort(packed_git, get_next_packed_git,
+                                    set_next_packed_git, sort_pack);
+}
+
+static void prepare_packed_git_mru(void)
+{
+       struct packed_git *p;
+
+       mru_clear(packed_git_mru);
+       for (p = packed_git; p; p = p->next)
+               mru_append(packed_git_mru, p);
+}
+
+static int prepare_packed_git_run_once = 0;
+void prepare_packed_git(void)
+{
+       struct alternate_object_database *alt;
+
+       if (prepare_packed_git_run_once)
+               return;
+       prepare_packed_git_one(get_object_directory(), 1);
+       prepare_alt_odb();
+       for (alt = alt_odb_list; alt; alt = alt->next)
+               prepare_packed_git_one(alt->path, 0);
+       rearrange_packed_git();
+       prepare_packed_git_mru();
+       prepare_packed_git_run_once = 1;
+}
+
+void reprepare_packed_git(void)
+{
+       approximate_object_count_valid = 0;
+       prepare_packed_git_run_once = 0;
+       prepare_packed_git();
+}
+
+unsigned long unpack_object_header_buffer(const unsigned char *buf,
+               unsigned long len, enum object_type *type, unsigned long *sizep)
+{
+       unsigned shift;
+       unsigned long size, c;
+       unsigned long used = 0;
+
+       c = buf[used++];
+       *type = (c >> 4) & 7;
+       size = c & 15;
+       shift = 4;
+       while (c & 0x80) {
+               if (len <= used || bitsizeof(long) <= shift) {
+                       error("bad object header");
+                       size = used = 0;
+                       break;
+               }
+               c = buf[used++];
+               size += (c & 0x7f) << shift;
+               shift += 7;
+       }
+       *sizep = size;
+       return used;
+}
+
+unsigned long get_size_from_delta(struct packed_git *p,
+                                 struct pack_window **w_curs,
+                                 off_t curpos)
+{
+       const unsigned char *data;
+       unsigned char delta_head[20], *in;
+       git_zstream stream;
+       int st;
+
+       memset(&stream, 0, sizeof(stream));
+       stream.next_out = delta_head;
+       stream.avail_out = sizeof(delta_head);
+
+       git_inflate_init(&stream);
+       do {
+               in = use_pack(p, w_curs, curpos, &stream.avail_in);
+               stream.next_in = in;
+               st = git_inflate(&stream, Z_FINISH);
+               curpos += stream.next_in - in;
+       } while ((st == Z_OK || st == Z_BUF_ERROR) &&
+                stream.total_out < sizeof(delta_head));
+       git_inflate_end(&stream);
+       if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
+               error("delta data unpack-initial failed");
+               return 0;
+       }
+
+       /* Examine the initial part of the delta to figure out
+        * the result size.
+        */
+       data = delta_head;
+
+       /* ignore base size */
+       get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
+
+       /* Read the result size */
+       return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
+}
+
+int unpack_object_header(struct packed_git *p,
+                        struct pack_window **w_curs,
+                        off_t *curpos,
+                        unsigned long *sizep)
+{
+       unsigned char *base;
+       unsigned long left;
+       unsigned long used;
+       enum object_type type;
+
+       /* use_pack() assures us we have [base, base + 20) available
+        * as a range that we can look at.  (Its actually the hash
+        * size that is assured.)  With our object header encoding
+        * the maximum deflated object size is 2^137, which is just
+        * insane, so we know won't exceed what we have been given.
+        */
+       base = use_pack(p, w_curs, *curpos, &left);
+       used = unpack_object_header_buffer(base, left, &type, sizep);
+       if (!used) {
+               type = OBJ_BAD;
+       } else
+               *curpos += used;
+
+       return type;
+}
+
+void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
+{
+       unsigned i;
+       for (i = 0; i < p->num_bad_objects; i++)
+               if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
+                       return;
+       p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
+                                     st_mult(GIT_MAX_RAWSZ,
+                                             st_add(p->num_bad_objects, 1)));
+       hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
+       p->num_bad_objects++;
+}
+
+const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
+{
+       struct packed_git *p;
+       unsigned i;
+
+       for (p = packed_git; p; p = p->next)
+               for (i = 0; i < p->num_bad_objects; i++)
+                       if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
+                               return p;
+       return NULL;
+}
+
+static off_t get_delta_base(struct packed_git *p,
+                                   struct pack_window **w_curs,
+                                   off_t *curpos,
+                                   enum object_type type,
+                                   off_t delta_obj_offset)
+{
+       unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
+       off_t base_offset;
+
+       /* use_pack() assured us we have [base_info, base_info + 20)
+        * as a range that we can look at without walking off the
+        * end of the mapped window.  Its actually the hash size
+        * that is assured.  An OFS_DELTA longer than the hash size
+        * is stupid, as then a REF_DELTA would be smaller to store.
+        */
+       if (type == OBJ_OFS_DELTA) {
+               unsigned used = 0;
+               unsigned char c = base_info[used++];
+               base_offset = c & 127;
+               while (c & 128) {
+                       base_offset += 1;
+                       if (!base_offset || MSB(base_offset, 7))
+                               return 0;  /* overflow */
+                       c = base_info[used++];
+                       base_offset = (base_offset << 7) + (c & 127);
+               }
+               base_offset = delta_obj_offset - base_offset;
+               if (base_offset <= 0 || base_offset >= delta_obj_offset)
+                       return 0;  /* out of bound */
+               *curpos += used;
+       } else if (type == OBJ_REF_DELTA) {
+               /* The base entry _must_ be in the same pack */
+               base_offset = find_pack_entry_one(base_info, p);
+               *curpos += 20;
+       } else
+               die("I am totally screwed");
+       return base_offset;
+}
+
+/*
+ * Like get_delta_base above, but we return the sha1 instead of the pack
+ * offset. This means it is cheaper for REF deltas (we do not have to do
+ * the final object lookup), but more expensive for OFS deltas (we
+ * have to load the revidx to convert the offset back into a sha1).
+ */
+static const unsigned char *get_delta_base_sha1(struct packed_git *p,
+                                               struct pack_window **w_curs,
+                                               off_t curpos,
+                                               enum object_type type,
+                                               off_t delta_obj_offset)
+{
+       if (type == OBJ_REF_DELTA) {
+               unsigned char *base = use_pack(p, w_curs, curpos, NULL);
+               return base;
+       } else if (type == OBJ_OFS_DELTA) {
+               struct revindex_entry *revidx;
+               off_t base_offset = get_delta_base(p, w_curs, &curpos,
+                                                  type, delta_obj_offset);
+
+               if (!base_offset)
+                       return NULL;
+
+               revidx = find_pack_revindex(p, base_offset);
+               if (!revidx)
+                       return NULL;
+
+               return nth_packed_object_sha1(p, revidx->nr);
+       } else
+               return NULL;
+}
+
+static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
+{
+       int type;
+       struct revindex_entry *revidx;
+       const unsigned char *sha1;
+       revidx = find_pack_revindex(p, obj_offset);
+       if (!revidx)
+               return OBJ_BAD;
+       sha1 = nth_packed_object_sha1(p, revidx->nr);
+       mark_bad_packed_object(p, sha1);
+       type = sha1_object_info(sha1, NULL);
+       if (type <= OBJ_NONE)
+               return OBJ_BAD;
+       return type;
+}
+
+#define POI_STACK_PREALLOC 64
+
+static enum object_type packed_to_object_type(struct packed_git *p,
+                                             off_t obj_offset,
+                                             enum object_type type,
+                                             struct pack_window **w_curs,
+                                             off_t curpos)
+{
+       off_t small_poi_stack[POI_STACK_PREALLOC];
+       off_t *poi_stack = small_poi_stack;
+       int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
+
+       while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
+               off_t base_offset;
+               unsigned long size;
+               /* Push the object we're going to leave behind */
+               if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
+                       poi_stack_alloc = alloc_nr(poi_stack_nr);
+                       ALLOC_ARRAY(poi_stack, poi_stack_alloc);
+                       memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
+               } else {
+                       ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
+               }
+               poi_stack[poi_stack_nr++] = obj_offset;
+               /* If parsing the base offset fails, just unwind */
+               base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
+               if (!base_offset)
+                       goto unwind;
+               curpos = obj_offset = base_offset;
+               type = unpack_object_header(p, w_curs, &curpos, &size);
+               if (type <= OBJ_NONE) {
+                       /* If getting the base itself fails, we first
+                        * retry the base, otherwise unwind */
+                       type = retry_bad_packed_offset(p, base_offset);
+                       if (type > OBJ_NONE)
+                               goto out;
+                       goto unwind;
+               }
+       }
+
+       switch (type) {
+       case OBJ_BAD:
+       case OBJ_COMMIT:
+       case OBJ_TREE:
+       case OBJ_BLOB:
+       case OBJ_TAG:
+               break;
+       default:
+               error("unknown object type %i at offset %"PRIuMAX" in %s",
+                     type, (uintmax_t)obj_offset, p->pack_name);
+               type = OBJ_BAD;
+       }
+
+out:
+       if (poi_stack != small_poi_stack)
+               free(poi_stack);
+       return type;
+
+unwind:
+       while (poi_stack_nr) {
+               obj_offset = poi_stack[--poi_stack_nr];
+               type = retry_bad_packed_offset(p, obj_offset);
+               if (type > OBJ_NONE)
+                       goto out;
+       }
+       type = OBJ_BAD;
+       goto out;
+}
+
+static struct hashmap delta_base_cache;
+static size_t delta_base_cached;
+
+static LIST_HEAD(delta_base_cache_lru);
+
+struct delta_base_cache_key {
+       struct packed_git *p;
+       off_t base_offset;
+};
+
+struct delta_base_cache_entry {
+       struct hashmap hash;
+       struct delta_base_cache_key key;
+       struct list_head lru;
+       void *data;
+       unsigned long size;
+       enum object_type type;
+};
+
+static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
+{
+       unsigned int hash;
+
+       hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
+       hash += (hash >> 8) + (hash >> 16);
+       return hash;
+}
+
+static struct delta_base_cache_entry *
+get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
+{
+       struct hashmap_entry entry;
+       struct delta_base_cache_key key;
+
+       if (!delta_base_cache.cmpfn)
+               return NULL;
+
+       hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
+       key.p = p;
+       key.base_offset = base_offset;
+       return hashmap_get(&delta_base_cache, &entry, &key);
+}
+
+static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
+                                  const struct delta_base_cache_key *b)
+{
+       return a->p == b->p && a->base_offset == b->base_offset;
+}
+
+static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
+                                    const void *va, const void *vb,
+                                    const void *vkey)
+{
+       const struct delta_base_cache_entry *a = va, *b = vb;
+       const struct delta_base_cache_key *key = vkey;
+       if (key)
+               return !delta_base_cache_key_eq(&a->key, key);
+       else
+               return !delta_base_cache_key_eq(&a->key, &b->key);
+}
+
+static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
+{
+       return !!get_delta_base_cache_entry(p, base_offset);
+}
+
+/*
+ * Remove the entry from the cache, but do _not_ free the associated
+ * entry data. The caller takes ownership of the "data" buffer, and
+ * should copy out any fields it wants before detaching.
+ */
+static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
+{
+       hashmap_remove(&delta_base_cache, ent, &ent->key);
+       list_del(&ent->lru);
+       delta_base_cached -= ent->size;
+       free(ent);
+}
+
+static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
+       unsigned long *base_size, enum object_type *type)
+{
+       struct delta_base_cache_entry *ent;
+
+       ent = get_delta_base_cache_entry(p, base_offset);
+       if (!ent)
+               return unpack_entry(p, base_offset, type, base_size);
+
+       if (type)
+               *type = ent->type;
+       if (base_size)
+               *base_size = ent->size;
+       return xmemdupz(ent->data, ent->size);
+}
+
+static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
+{
+       free(ent->data);
+       detach_delta_base_cache_entry(ent);
+}
+
+void clear_delta_base_cache(void)
+{
+       struct list_head *lru, *tmp;
+       list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
+               struct delta_base_cache_entry *entry =
+                       list_entry(lru, struct delta_base_cache_entry, lru);
+               release_delta_base_cache(entry);
+       }
+}
+
+static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
+       void *base, unsigned long base_size, enum object_type type)
+{
+       struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
+       struct list_head *lru, *tmp;
+
+       delta_base_cached += base_size;
+
+       list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
+               struct delta_base_cache_entry *f =
+                       list_entry(lru, struct delta_base_cache_entry, lru);
+               if (delta_base_cached <= delta_base_cache_limit)
+                       break;
+               release_delta_base_cache(f);
+       }
+
+       ent->key.p = p;
+       ent->key.base_offset = base_offset;
+       ent->type = type;
+       ent->data = base;
+       ent->size = base_size;
+       list_add_tail(&ent->lru, &delta_base_cache_lru);
+
+       if (!delta_base_cache.cmpfn)
+               hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
+       hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
+       hashmap_add(&delta_base_cache, ent);
+}
+
+int packed_object_info(struct packed_git *p, off_t obj_offset,
+                      struct object_info *oi)
+{
+       struct pack_window *w_curs = NULL;
+       unsigned long size;
+       off_t curpos = obj_offset;
+       enum object_type type;
+
+       /*
+        * We always get the representation type, but only convert it to
+        * a "real" type later if the caller is interested.
+        */
+       if (oi->contentp) {
+               *oi->contentp = cache_or_unpack_entry(p, obj_offset, oi->sizep,
+                                                     &type);
+               if (!*oi->contentp)
+                       type = OBJ_BAD;
+       } else {
+               type = unpack_object_header(p, &w_curs, &curpos, &size);
+       }
+
+       if (!oi->contentp && oi->sizep) {
+               if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
+                       off_t tmp_pos = curpos;
+                       off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
+                                                          type, obj_offset);
+                       if (!base_offset) {
+                               type = OBJ_BAD;
+                               goto out;
+                       }
+                       *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
+                       if (*oi->sizep == 0) {
+                               type = OBJ_BAD;
+                               goto out;
+                       }
+               } else {
+                       *oi->sizep = size;
+               }
+       }
+
+       if (oi->disk_sizep) {
+               struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
+               *oi->disk_sizep = revidx[1].offset - obj_offset;
+       }
+
+       if (oi->typep || oi->typename) {
+               enum object_type ptot;
+               ptot = packed_to_object_type(p, obj_offset, type, &w_curs,
+                                            curpos);
+               if (oi->typep)
+                       *oi->typep = ptot;
+               if (oi->typename) {
+                       const char *tn = typename(ptot);
+                       if (tn)
+                               strbuf_addstr(oi->typename, tn);
+               }
+               if (ptot < 0) {
+                       type = OBJ_BAD;
+                       goto out;
+               }
+       }
+
+       if (oi->delta_base_sha1) {
+               if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
+                       const unsigned char *base;
+
+                       base = get_delta_base_sha1(p, &w_curs, curpos,
+                                                  type, obj_offset);
+                       if (!base) {
+                               type = OBJ_BAD;
+                               goto out;
+                       }
+
+                       hashcpy(oi->delta_base_sha1, base);
+               } else
+                       hashclr(oi->delta_base_sha1);
+       }
+
+       oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
+                                                         OI_PACKED;
+
+out:
+       unuse_pack(&w_curs);
+       return type;
+}
+
+static void *unpack_compressed_entry(struct packed_git *p,
+                                   struct pack_window **w_curs,
+                                   off_t curpos,
+                                   unsigned long size)
+{
+       int st;
+       git_zstream stream;
+       unsigned char *buffer, *in;
+
+       buffer = xmallocz_gently(size);
+       if (!buffer)
+               return NULL;
+       memset(&stream, 0, sizeof(stream));
+       stream.next_out = buffer;
+       stream.avail_out = size + 1;
+
+       git_inflate_init(&stream);
+       do {
+               in = use_pack(p, w_curs, curpos, &stream.avail_in);
+               stream.next_in = in;
+               st = git_inflate(&stream, Z_FINISH);
+               if (!stream.avail_out)
+                       break; /* the payload is larger than it should be */
+               curpos += stream.next_in - in;
+       } while (st == Z_OK || st == Z_BUF_ERROR);
+       git_inflate_end(&stream);
+       if ((st != Z_STREAM_END) || stream.total_out != size) {
+               free(buffer);
+               return NULL;
+       }
+
+       return buffer;
+}
+
+static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
+{
+       static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
+       trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
+                        p->pack_name, (uintmax_t)obj_offset);
+}
+
+int do_check_packed_object_crc;
+
+#define UNPACK_ENTRY_STACK_PREALLOC 64
+struct unpack_entry_stack_ent {
+       off_t obj_offset;
+       off_t curpos;
+       unsigned long size;
+};
+
+static void *read_object(const unsigned char *sha1, enum object_type *type,
+                        unsigned long *size)
+{
+       struct object_info oi = OBJECT_INFO_INIT;
+       void *content;
+       oi.typep = type;
+       oi.sizep = size;
+       oi.contentp = &content;
+
+       if (sha1_object_info_extended(sha1, &oi, 0) < 0)
+               return NULL;
+       return content;
+}
+
+void *unpack_entry(struct packed_git *p, off_t obj_offset,
+                  enum object_type *final_type, unsigned long *final_size)
+{
+       struct pack_window *w_curs = NULL;
+       off_t curpos = obj_offset;
+       void *data = NULL;
+       unsigned long size;
+       enum object_type type;
+       struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
+       struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
+       int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
+       int base_from_cache = 0;
+
+       write_pack_access_log(p, obj_offset);
+
+       /* PHASE 1: drill down to the innermost base object */
+       for (;;) {
+               off_t base_offset;
+               int i;
+               struct delta_base_cache_entry *ent;
+
+               ent = get_delta_base_cache_entry(p, curpos);
+               if (ent) {
+                       type = ent->type;
+                       data = ent->data;
+                       size = ent->size;
+                       detach_delta_base_cache_entry(ent);
+                       base_from_cache = 1;
+                       break;
+               }
+
+               if (do_check_packed_object_crc && p->index_version > 1) {
+                       struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
+                       off_t len = revidx[1].offset - obj_offset;
+                       if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
+                               const unsigned char *sha1 =
+                                       nth_packed_object_sha1(p, revidx->nr);
+                               error("bad packed object CRC for %s",
+                                     sha1_to_hex(sha1));
+                               mark_bad_packed_object(p, sha1);
+                               data = NULL;
+                               goto out;
+                       }
+               }
+
+               type = unpack_object_header(p, &w_curs, &curpos, &size);
+               if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
+                       break;
+
+               base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
+               if (!base_offset) {
+                       error("failed to validate delta base reference "
+                             "at offset %"PRIuMAX" from %s",
+                             (uintmax_t)curpos, p->pack_name);
+                       /* bail to phase 2, in hopes of recovery */
+                       data = NULL;
+                       break;
+               }
+
+               /* push object, proceed to base */
+               if (delta_stack_nr >= delta_stack_alloc
+                   && delta_stack == small_delta_stack) {
+                       delta_stack_alloc = alloc_nr(delta_stack_nr);
+                       ALLOC_ARRAY(delta_stack, delta_stack_alloc);
+                       memcpy(delta_stack, small_delta_stack,
+                              sizeof(*delta_stack)*delta_stack_nr);
+               } else {
+                       ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
+               }
+               i = delta_stack_nr++;
+               delta_stack[i].obj_offset = obj_offset;
+               delta_stack[i].curpos = curpos;
+               delta_stack[i].size = size;
+
+               curpos = obj_offset = base_offset;
+       }
+
+       /* PHASE 2: handle the base */
+       switch (type) {
+       case OBJ_OFS_DELTA:
+       case OBJ_REF_DELTA:
+               if (data)
+                       die("BUG: unpack_entry: left loop at a valid delta");
+               break;
+       case OBJ_COMMIT:
+       case OBJ_TREE:
+       case OBJ_BLOB:
+       case OBJ_TAG:
+               if (!base_from_cache)
+                       data = unpack_compressed_entry(p, &w_curs, curpos, size);
+               break;
+       default:
+               data = NULL;
+               error("unknown object type %i at offset %"PRIuMAX" in %s",
+                     type, (uintmax_t)obj_offset, p->pack_name);
+       }
+
+       /* PHASE 3: apply deltas in order */
+
+       /* invariants:
+        *   'data' holds the base data, or NULL if there was corruption
+        */
+       while (delta_stack_nr) {
+               void *delta_data;
+               void *base = data;
+               void *external_base = NULL;
+               unsigned long delta_size, base_size = size;
+               int i;
+
+               data = NULL;
+
+               if (base)
+                       add_delta_base_cache(p, obj_offset, base, base_size, type);
+
+               if (!base) {
+                       /*
+                        * We're probably in deep shit, but let's try to fetch
+                        * the required base anyway from another pack or loose.
+                        * This is costly but should happen only in the presence
+                        * of a corrupted pack, and is better than failing outright.
+                        */
+                       struct revindex_entry *revidx;
+                       const unsigned char *base_sha1;
+                       revidx = find_pack_revindex(p, obj_offset);
+                       if (revidx) {
+                               base_sha1 = nth_packed_object_sha1(p, revidx->nr);
+                               error("failed to read delta base object %s"
+                                     " at offset %"PRIuMAX" from %s",
+                                     sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
+                                     p->pack_name);
+                               mark_bad_packed_object(p, base_sha1);
+                               base = read_object(base_sha1, &type, &base_size);
+                               external_base = base;
+                       }
+               }
+
+               i = --delta_stack_nr;
+               obj_offset = delta_stack[i].obj_offset;
+               curpos = delta_stack[i].curpos;
+               delta_size = delta_stack[i].size;
+
+               if (!base)
+                       continue;
+
+               delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
+
+               if (!delta_data) {
+                       error("failed to unpack compressed delta "
+                             "at offset %"PRIuMAX" from %s",
+                             (uintmax_t)curpos, p->pack_name);
+                       data = NULL;
+                       free(external_base);
+                       continue;
+               }
+
+               data = patch_delta(base, base_size,
+                                  delta_data, delta_size,
+                                  &size);
+
+               /*
+                * We could not apply the delta; warn the user, but keep going.
+                * Our failure will be noticed either in the next iteration of
+                * the loop, or if this is the final delta, in the caller when
+                * we return NULL. Those code paths will take care of making
+                * a more explicit warning and retrying with another copy of
+                * the object.
+                */
+               if (!data)
+                       error("failed to apply delta");
+
+               free(delta_data);
+               free(external_base);
+       }
+
+       if (final_type)
+               *final_type = type;
+       if (final_size)
+               *final_size = size;
+
+out:
+       unuse_pack(&w_curs);
+
+       if (delta_stack != small_delta_stack)
+               free(delta_stack);
+
+       return data;
+}
+
+const unsigned char *nth_packed_object_sha1(struct packed_git *p,
+                                           uint32_t n)
+{
+       const unsigned char *index = p->index_data;
+       if (!index) {
+               if (open_pack_index(p))
+                       return NULL;
+               index = p->index_data;
+       }
+       if (n >= p->num_objects)
+               return NULL;
+       index += 4 * 256;
+       if (p->index_version == 1) {
+               return index + 24 * n + 4;
+       } else {
+               index += 8;
+               return index + 20 * n;
+       }
+}
+
+const struct object_id *nth_packed_object_oid(struct object_id *oid,
+                                             struct packed_git *p,
+                                             uint32_t n)
+{
+       const unsigned char *hash = nth_packed_object_sha1(p, n);
+       if (!hash)
+               return NULL;
+       hashcpy(oid->hash, hash);
+       return oid;
+}
+
+void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
+{
+       const unsigned char *ptr = vptr;
+       const unsigned char *start = p->index_data;
+       const unsigned char *end = start + p->index_size;
+       if (ptr < start)
+               die(_("offset before start of pack index for %s (corrupt index?)"),
+                   p->pack_name);
+       /* No need to check for underflow; .idx files must be at least 8 bytes */
+       if (ptr >= end - 8)
+               die(_("offset beyond end of pack index for %s (truncated index?)"),
+                   p->pack_name);
+}
+
+off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
+{
+       const unsigned char *index = p->index_data;
+       index += 4 * 256;
+       if (p->index_version == 1) {
+               return ntohl(*((uint32_t *)(index + 24 * n)));
+       } else {
+               uint32_t off;
+               index += 8 + p->num_objects * (20 + 4);
+               off = ntohl(*((uint32_t *)(index + 4 * n)));
+               if (!(off & 0x80000000))
+                       return off;
+               index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
+               check_pack_index_ptr(p, index);
+               return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
+                                  ntohl(*((uint32_t *)(index + 4)));
+       }
+}
+
+off_t find_pack_entry_one(const unsigned char *sha1,
+                                 struct packed_git *p)
+{
+       const uint32_t *level1_ofs = p->index_data;
+       const unsigned char *index = p->index_data;
+       unsigned hi, lo, stride;
+       static int debug_lookup = -1;
+
+       if (debug_lookup < 0)
+               debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
+
+       if (!index) {
+               if (open_pack_index(p))
+                       return 0;
+               level1_ofs = p->index_data;
+               index = p->index_data;
+       }
+       if (p->index_version > 1) {
+               level1_ofs += 2;
+               index += 8;
+       }
+       index += 4 * 256;
+       hi = ntohl(level1_ofs[*sha1]);
+       lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
+       if (p->index_version > 1) {
+               stride = 20;
+       } else {
+               stride = 24;
+               index += 4;
+       }
+
+       if (debug_lookup)
+               printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
+                      sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
+
+       while (lo < hi) {
+               unsigned mi = (lo + hi) / 2;
+               int cmp = hashcmp(index + mi * stride, sha1);
+
+               if (debug_lookup)
+                       printf("lo %u hi %u rg %u mi %u\n",
+                              lo, hi, hi - lo, mi);
+               if (!cmp)
+                       return nth_packed_object_offset(p, mi);
+               if (cmp > 0)
+                       hi = mi;
+               else
+                       lo = mi+1;
+       }
+       return 0;
+}
+
+int is_pack_valid(struct packed_git *p)
+{
+       /* An already open pack is known to be valid. */
+       if (p->pack_fd != -1)
+               return 1;
+
+       /* If the pack has one window completely covering the
+        * file size, the pack is known to be valid even if
+        * the descriptor is not currently open.
+        */
+       if (p->windows) {
+               struct pack_window *w = p->windows;
+
+               if (!w->offset && w->len == p->pack_size)
+                       return 1;
+       }
+
+       /* Force the pack to open to prove its valid. */
+       return !open_packed_git(p);
+}
+
+struct packed_git *find_sha1_pack(const unsigned char *sha1,
+                                 struct packed_git *packs)
+{
+       struct packed_git *p;
+
+       for (p = packs; p; p = p->next) {
+               if (find_pack_entry_one(sha1, p))
+                       return p;
+       }
+       return NULL;
+
+}
+
+static int fill_pack_entry(const unsigned char *sha1,
+                          struct pack_entry *e,
+                          struct packed_git *p)
+{
+       off_t offset;
+
+       if (p->num_bad_objects) {
+               unsigned i;
+               for (i = 0; i < p->num_bad_objects; i++)
+                       if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
+                               return 0;
+       }
+
+       offset = find_pack_entry_one(sha1, p);
+       if (!offset)
+               return 0;
+
+       /*
+        * We are about to tell the caller where they can locate the
+        * requested object.  We better make sure the packfile is
+        * still here and can be accessed before supplying that
+        * answer, as it may have been deleted since the index was
+        * loaded!
+        */
+       if (!is_pack_valid(p))
+               return 0;
+       e->offset = offset;
+       e->p = p;
+       hashcpy(e->sha1, sha1);
+       return 1;
+}
+
+/*
+ * Iff a pack file contains the object named by sha1, return true and
+ * store its location to e.
+ */
+int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
+{
+       struct mru_entry *p;
+
+       prepare_packed_git();
+       if (!packed_git)
+               return 0;
+
+       for (p = packed_git_mru->head; p; p = p->next) {
+               if (fill_pack_entry(sha1, e, p->item)) {
+                       mru_mark(packed_git_mru, p);
+                       return 1;
+               }
+       }
+       return 0;
+}
+
+int has_sha1_pack(const unsigned char *sha1)
+{
+       struct pack_entry e;
+       return find_pack_entry(sha1, &e);
+}
+
+int has_pack_index(const unsigned char *sha1)
+{
+       struct stat st;
+       if (stat(sha1_pack_index_name(sha1), &st))
+               return 0;
+       return 1;
+}
+
+static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
+{
+       uint32_t i;
+       int r = 0;
+
+       for (i = 0; i < p->num_objects; i++) {
+               struct object_id oid;
+
+               if (!nth_packed_object_oid(&oid, p, i))
+                       return error("unable to get sha1 of object %u in %s",
+                                    i, p->pack_name);
+
+               r = cb(&oid, p, i, data);
+               if (r)
+                       break;
+       }
+       return r;
+}
+
+int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
+{
+       struct packed_git *p;
+       int r = 0;
+       int pack_errors = 0;
+
+       prepare_packed_git();
+       for (p = packed_git; p; p = p->next) {
+               if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
+                       continue;
+               if (open_pack_index(p)) {
+                       pack_errors = 1;
+                       continue;
+               }
+               r = for_each_object_in_pack(p, cb, data);
+               if (r)
+                       break;
+       }
+       return r ? r : pack_errors;
+}