Add check program "git-check-racy"
[gitweb.git] / read-cache.c
index e3d3ddb18ce03c0cbd29242e779ae603653c5913..4d621ef4448b34ce8c46674989191513a950cd35 100644 (file)
  *
  * Copyright (C) Linus Torvalds, 2005
  */
-#include <stdarg.h>
 #include "cache.h"
+#include "cache-tree.h"
 
-const char *sha1_file_directory = NULL;
-struct cache_entry **active_cache = NULL;
-unsigned int active_nr = 0, active_alloc = 0;
-
-void usage(const char *err)
-{
-       fprintf(stderr, "usage: %s\n", err);
-       exit(1);
-}
+/* Index extensions.
+ *
+ * The first letter should be 'A'..'Z' for extensions that are not
+ * necessary for a correct operation (i.e. optimization data).
+ * When new extensions are added that _needs_ to be understood in
+ * order to correctly interpret the index file, pick character that
+ * is outside the range, to cause the reader to abort.
+ */
 
-static void report(const char *prefix, const char *err, va_list params)
-{
-       fputs(prefix, stderr);
-       vfprintf(stderr, err, params);
-       fputs("\n", stderr);
-}
+#define CACHE_EXT(s) ( (s[0]<<24)|(s[1]<<16)|(s[2]<<8)|(s[3]) )
+#define CACHE_EXT_TREE 0x54524545      /* "TREE" */
 
-void die(const char *err, ...)
-{
-       va_list params;
-
-       va_start(params, err);
-       report("fatal: ", err, params);
-       va_end(params);
-       exit(1);
-}
+struct cache_entry **active_cache = NULL;
+static time_t index_file_timestamp;
+unsigned int active_nr = 0, active_alloc = 0, active_cache_changed = 0;
 
-int error(const char *err, ...)
-{
-       va_list params;
+struct cache_tree *active_cache_tree = NULL;
 
-       va_start(params, err);
-       report("error: ", err, params);
-       va_end(params);
-       return -1;
-}
+int cache_errno = 0;
 
+static void *cache_mmap = NULL;
+static size_t cache_mmap_size = 0;
 
-static unsigned hexval(char c)
+/*
+ * This only updates the "non-critical" parts of the directory
+ * cache, ie the parts that aren't tracked by GIT, and only used
+ * to validate the cache.
+ */
+void fill_stat_cache_info(struct cache_entry *ce, struct stat *st)
 {
-       if (c >= '0' && c <= '9')
-               return c - '0';
-       if (c >= 'a' && c <= 'f')
-               return c - 'a' + 10;
-       if (c >= 'A' && c <= 'F')
-               return c - 'A' + 10;
-       return ~0;
-}
+       ce->ce_ctime.sec = htonl(st->st_ctime);
+       ce->ce_mtime.sec = htonl(st->st_mtime);
+#ifdef USE_NSEC
+       ce->ce_ctime.nsec = htonl(st->st_ctim.tv_nsec);
+       ce->ce_mtime.nsec = htonl(st->st_mtim.tv_nsec);
+#endif
+       ce->ce_dev = htonl(st->st_dev);
+       ce->ce_ino = htonl(st->st_ino);
+       ce->ce_uid = htonl(st->st_uid);
+       ce->ce_gid = htonl(st->st_gid);
+       ce->ce_size = htonl(st->st_size);
 
-int get_sha1_hex(const char *hex, unsigned char *sha1)
-{
-       int i;
-       for (i = 0; i < 20; i++) {
-               unsigned int val = (hexval(hex[0]) << 4) | hexval(hex[1]);
-               if (val & ~0xff)
-                       return -1;
-               *sha1++ = val;
-               hex += 2;
-       }
-       return 0;
+       if (assume_unchanged)
+               ce->ce_flags |= htons(CE_VALID);
 }
 
-char * sha1_to_hex(const unsigned char *sha1)
+static int ce_compare_data(struct cache_entry *ce, struct stat *st)
 {
-       static char buffer[50];
-       static const char hex[] = "0123456789abcdef";
-       char *buf = buffer;
-       int i;
+       int match = -1;
+       int fd = open(ce->name, O_RDONLY);
 
-       for (i = 0; i < 20; i++) {
-               unsigned int val = *sha1++;
-               *buf++ = hex[val >> 4];
-               *buf++ = hex[val & 0xf];
+       if (fd >= 0) {
+               unsigned char sha1[20];
+               if (!index_fd(sha1, fd, st, 0, NULL))
+                       match = memcmp(sha1, ce->sha1, 20);
+               /* index_fd() closed the file descriptor already */
        }
-       return buffer;
+       return match;
 }
 
-/*
- * NOTE! This returns a statically allocated buffer, so you have to be
- * careful about using it. Do a "strdup()" if you need to save the
- * filename.
- */
-char *sha1_file_name(const unsigned char *sha1)
+static int ce_compare_link(struct cache_entry *ce, unsigned long expected_size)
 {
-       int i;
-       static char *name, *base;
+       int match = -1;
+       char *target;
+       void *buffer;
+       unsigned long size;
+       char type[10];
+       int len;
 
-       if (!base) {
-               char *sha1_file_directory = getenv(DB_ENVIRONMENT) ? : DEFAULT_DB_ENVIRONMENT;
-               int len = strlen(sha1_file_directory);
-               base = malloc(len + 60);
-               memcpy(base, sha1_file_directory, len);
-               memset(base+len, 0, 60);
-               base[len] = '/';
-               base[len+3] = '/';
-               name = base + len + 1;
+       target = xmalloc(expected_size);
+       len = readlink(ce->name, target, expected_size);
+       if (len != expected_size) {
+               free(target);
+               return -1;
        }
-       for (i = 0; i < 20; i++) {
-               static char hex[] = "0123456789abcdef";
-               unsigned int val = sha1[i];
-               char *pos = name + i*2 + (i > 0);
-               *pos++ = hex[val >> 4];
-               *pos = hex[val & 0xf];
+       buffer = read_sha1_file(ce->sha1, type, &size);
+       if (!buffer) {
+               free(target);
+               return -1;
        }
-       return base;
+       if (size == expected_size)
+               match = memcmp(buffer, target, size);
+       free(buffer);
+       free(target);
+       return match;
 }
 
-int check_sha1_signature(unsigned char *sha1, void *map, unsigned long size)
+static int ce_modified_check_fs(struct cache_entry *ce, struct stat *st)
 {
-       unsigned char real_sha1[20];
-       SHA_CTX c;
-
-       SHA1_Init(&c);
-       SHA1_Update(&c, map, size);
-       SHA1_Final(real_sha1, &c);
-       return memcmp(sha1, real_sha1, 20) ? -1 : 0;
+       switch (st->st_mode & S_IFMT) {
+       case S_IFREG:
+               if (ce_compare_data(ce, st))
+                       return DATA_CHANGED;
+               break;
+       case S_IFLNK:
+               if (ce_compare_link(ce, st->st_size))
+                       return DATA_CHANGED;
+               break;
+       default:
+               return TYPE_CHANGED;
+       }
+       return 0;
 }
 
-void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
+static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
 {
-       char *filename = sha1_file_name(sha1);
-       int fd = open(filename, O_RDONLY);
-       struct stat st;
-       void *map;
+       unsigned int changed = 0;
 
-       if (fd < 0) {
-               perror(filename);
-               return NULL;
+       switch (ntohl(ce->ce_mode) & S_IFMT) {
+       case S_IFREG:
+               changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
+               /* We consider only the owner x bit to be relevant for
+                * "mode changes"
+                */
+               if (trust_executable_bit &&
+                   (0100 & (ntohl(ce->ce_mode) ^ st->st_mode)))
+                       changed |= MODE_CHANGED;
+               break;
+       case S_IFLNK:
+               changed |= !S_ISLNK(st->st_mode) ? TYPE_CHANGED : 0;
+               break;
+       default:
+               die("internal error: ce_mode is %o", ntohl(ce->ce_mode));
        }
-       if (fstat(fd, &st) < 0) {
-               close(fd);
-               return NULL;
-       }
-       map = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
-       close(fd);
-       if (-1 == (int)(long)map)
-               return NULL;
-       *size = st.st_size;
-       return map;
-}
+       if (ce->ce_mtime.sec != htonl(st->st_mtime))
+               changed |= MTIME_CHANGED;
+       if (ce->ce_ctime.sec != htonl(st->st_ctime))
+               changed |= CTIME_CHANGED;
 
-void * unpack_sha1_file(void *map, unsigned long mapsize, char *type, unsigned long *size)
-{
-       int ret, bytes;
-       z_stream stream;
-       char buffer[8192];
-       char *buf;
+#ifdef USE_NSEC
+       /*
+        * nsec seems unreliable - not all filesystems support it, so
+        * as long as it is in the inode cache you get right nsec
+        * but after it gets flushed, you get zero nsec.
+        */
+       if (ce->ce_mtime.nsec != htonl(st->st_mtim.tv_nsec))
+               changed |= MTIME_CHANGED;
+       if (ce->ce_ctime.nsec != htonl(st->st_ctim.tv_nsec))
+               changed |= CTIME_CHANGED;
+#endif 
 
-       /* Get the data stream */
-       memset(&stream, 0, sizeof(stream));
-       stream.next_in = map;
-       stream.avail_in = mapsize;
-       stream.next_out = buffer;
-       stream.avail_out = sizeof(buffer);
+       if (ce->ce_uid != htonl(st->st_uid) ||
+           ce->ce_gid != htonl(st->st_gid))
+               changed |= OWNER_CHANGED;
+       if (ce->ce_ino != htonl(st->st_ino))
+               changed |= INODE_CHANGED;
 
-       inflateInit(&stream);
-       ret = inflate(&stream, 0);
-       if (sscanf(buffer, "%10s %lu", type, size) != 2)
-               return NULL;
+#ifdef USE_STDEV
+       /*
+        * st_dev breaks on network filesystems where different
+        * clients will have different views of what "device"
+        * the filesystem is on
+        */
+       if (ce->ce_dev != htonl(st->st_dev))
+               changed |= INODE_CHANGED;
+#endif
 
-       bytes = strlen(buffer) + 1;
-       buf = malloc(*size);
-       if (!buf)
-               return NULL;
+       if (ce->ce_size != htonl(st->st_size))
+               changed |= DATA_CHANGED;
 
-       memcpy(buf, buffer + bytes, stream.total_out - bytes);
-       bytes = stream.total_out - bytes;
-       if (bytes < *size && ret == Z_OK) {
-               stream.next_out = buf + bytes;
-               stream.avail_out = *size - bytes;
-               while (inflate(&stream, Z_FINISH) == Z_OK)
-                       /* nothing */;
-       }
-       inflateEnd(&stream);
-       return buf;
+       return changed;
 }
 
-void * read_sha1_file(const unsigned char *sha1, char *type, unsigned long *size)
+int ce_match_stat(struct cache_entry *ce, struct stat *st, int options)
 {
-       unsigned long mapsize;
-       void *map, *buf;
+       unsigned int changed;
+       int ignore_valid = options & 01;
+       int assume_racy_is_modified = options & 02;
 
-       map = map_sha1_file(sha1, &mapsize);
-       if (map) {
-               buf = unpack_sha1_file(map, mapsize, type, size);
-               munmap(map, mapsize);
-               return buf;
-       }
-       return NULL;
-}
+       /*
+        * If it's marked as always valid in the index, it's
+        * valid whatever the checked-out copy says.
+        */
+       if (!ignore_valid && (ce->ce_flags & htons(CE_VALID)))
+               return 0;
 
-int write_sha1_file(char *buf, unsigned len, unsigned char *returnsha1)
-{
-       int size;
-       char *compressed;
-       z_stream stream;
-       unsigned char sha1[20];
-       SHA_CTX c;
+       changed = ce_match_stat_basic(ce, st);
 
-       /* Set it up */
-       memset(&stream, 0, sizeof(stream));
-       deflateInit(&stream, Z_BEST_COMPRESSION);
-       size = deflateBound(&stream, len);
-       compressed = malloc(size);
-
-       /* Compress it */
-       stream.next_in = buf;
-       stream.avail_in = len;
-       stream.next_out = compressed;
-       stream.avail_out = size;
-       while (deflate(&stream, Z_FINISH) == Z_OK)
-               /* nothing */;
-       deflateEnd(&stream);
-       size = stream.total_out;
-
-       /* Sha1.. */
-       SHA1_Init(&c);
-       SHA1_Update(&c, compressed, size);
-       SHA1_Final(sha1, &c);
+       /*
+        * Within 1 second of this sequence:
+        *      echo xyzzy >file && git-update-index --add file
+        * running this command:
+        *      echo frotz >file
+        * would give a falsely clean cache entry.  The mtime and
+        * length match the cache, and other stat fields do not change.
+        *
+        * We could detect this at update-index time (the cache entry
+        * being registered/updated records the same time as "now")
+        * and delay the return from git-update-index, but that would
+        * effectively mean we can make at most one commit per second,
+        * which is not acceptable.  Instead, we check cache entries
+        * whose mtime are the same as the index file timestamp more
+        * carefully than others.
+        */
+       if (!changed &&
+           index_file_timestamp &&
+           index_file_timestamp <= ntohl(ce->ce_mtime.sec)) {
+               if (assume_racy_is_modified)
+                       changed |= DATA_CHANGED;
+               else
+                       changed |= ce_modified_check_fs(ce, st);
+       }
 
-       if (write_sha1_buffer(sha1, compressed, size) < 0)
-               return -1;
-       if (returnsha1)
-               memcpy(returnsha1, sha1, 20);
-       return 0;
+       return changed;
 }
 
-int write_sha1_buffer(const unsigned char *sha1, void *buf, unsigned int size)
+int ce_modified(struct cache_entry *ce, struct stat *st, int really)
 {
-       char *filename = sha1_file_name(sha1);
-       int fd;
+       int changed, changed_fs;
+       changed = ce_match_stat(ce, st, really);
+       if (!changed)
+               return 0;
+       /*
+        * If the mode or type has changed, there's no point in trying
+        * to refresh the entry - it's not going to match
+        */
+       if (changed & (MODE_CHANGED | TYPE_CHANGED))
+               return changed;
 
-       fd = open(filename, O_WRONLY | O_CREAT | O_EXCL, 0666);
-       if (fd < 0) {
-               void *map;
+       /* Immediately after read-tree or update-index --cacheinfo,
+        * the length field is zero.  For other cases the ce_size
+        * should match the SHA1 recorded in the index entry.
+        */
+       if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
+               return changed;
 
-               if (errno != EEXIST)
-                       return -1;
-#ifndef COLLISION_CHECK
-               fd = open(filename, O_RDONLY);
-               if (fd < 0)
-                       return -1;
-               map = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
-               if (map == MAP_FAILED)
-                       return -1;
-               if (memcmp(buf, map, size))
-                       return error("SHA1 collision detected!"
-                                       " This is bad, bad, BAD!\a\n");
-#endif
-               return 0;
-       }
-       write(fd, buf, size);
-       close(fd);
+       changed_fs = ce_modified_check_fs(ce, st);
+       if (changed_fs)
+               return changed | changed_fs;
        return 0;
 }
 
-int cache_match_stat(struct cache_entry *ce, struct stat *st)
+int base_name_compare(const char *name1, int len1, int mode1,
+                     const char *name2, int len2, int mode2)
 {
-       unsigned int changed = 0;
+       unsigned char c1, c2;
+       int len = len1 < len2 ? len1 : len2;
+       int cmp;
 
-       /* nsec seems unreliable - not all filesystems support it, so
-        * as long as it is in the inode cache you get right nsec
-        * but after it gets flushed, you get zero nsec. */
-       if (ce->mtime.sec  != (unsigned int)st->st_mtim.tv_sec
-#ifdef NSEC
-           || ce->mtime.nsec != (unsigned int)st->st_mtim.tv_nsec
-#endif
-           )
-               changed |= MTIME_CHANGED;
-       if (ce->ctime.sec  != (unsigned int)st->st_ctim.tv_sec
-#ifdef NSEC
-           || ce->ctime.nsec != (unsigned int)st->st_ctim.tv_nsec
-#endif
-           )
-               changed |= CTIME_CHANGED;
-       if (ce->st_uid != (unsigned int)st->st_uid ||
-           ce->st_gid != (unsigned int)st->st_gid)
-               changed |= OWNER_CHANGED;
-       if (ce->st_mode != (unsigned int)st->st_mode)
-               changed |= MODE_CHANGED;
-       if (ce->st_dev != (unsigned int)st->st_dev ||
-           ce->st_ino != (unsigned int)st->st_ino)
-               changed |= INODE_CHANGED;
-       if (ce->st_size != (unsigned int)st->st_size)
-               changed |= DATA_CHANGED;
-       return changed;
+       cmp = memcmp(name1, name2, len);
+       if (cmp)
+               return cmp;
+       c1 = name1[len];
+       c2 = name2[len];
+       if (!c1 && S_ISDIR(mode1))
+               c1 = '/';
+       if (!c2 && S_ISDIR(mode2))
+               c2 = '/';
+       return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
 }
 
-int cache_name_compare(const char *name1, int len1, const char *name2, int len2)
+int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
 {
+       int len1 = flags1 & CE_NAMEMASK;
+       int len2 = flags2 & CE_NAMEMASK;
        int len = len1 < len2 ? len1 : len2;
        int cmp;
 
@@ -306,6 +271,15 @@ int cache_name_compare(const char *name1, int len1, const char *name2, int len2)
                return -1;
        if (len1 > len2)
                return 1;
+
+       /* Compare stages  */
+       flags1 &= CE_STAGEMASK;
+       flags2 &= CE_STAGEMASK;
+
+       if (flags1 < flags2)
+               return -1;
+       if (flags1 > flags2)
+               return 1;
        return 0;
 }
 
@@ -318,7 +292,7 @@ int cache_name_pos(const char *name, int namelen)
        while (last > first) {
                int next = (last + first) >> 1;
                struct cache_entry *ce = active_cache[next];
-               int cmp = cache_name_compare(name, namelen, ce->name, ce->namelen);
+               int cmp = cache_name_compare(name, namelen, ce->name, ntohs(ce->ce_flags));
                if (!cmp)
                        return next;
                if (cmp < 0) {
@@ -330,37 +304,318 @@ int cache_name_pos(const char *name, int namelen)
        return -first-1;
 }
 
-int remove_file_from_cache(char *path)
+/* Remove entry, return true if there are more entries to go.. */
+int remove_cache_entry_at(int pos)
+{
+       active_cache_changed = 1;
+       active_nr--;
+       if (pos >= active_nr)
+               return 0;
+       memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
+       return 1;
+}
+
+int remove_file_from_cache(const char *path)
 {
        int pos = cache_name_pos(path, strlen(path));
-       if (pos >= 0) {
-               active_nr--;
-               if (pos < active_nr)
-                       memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
+       if (pos < 0)
+               pos = -pos-1;
+       while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
+               remove_cache_entry_at(pos);
+       return 0;
+}
+
+int add_file_to_index(const char *path, int verbose)
+{
+       int size, namelen;
+       struct stat st;
+       struct cache_entry *ce;
+
+       if (lstat(path, &st))
+               die("%s: unable to stat (%s)", path, strerror(errno));
+
+       if (!S_ISREG(st.st_mode) && !S_ISLNK(st.st_mode))
+               die("%s: can only add regular files or symbolic links", path);
+
+       namelen = strlen(path);
+       size = cache_entry_size(namelen);
+       ce = xcalloc(1, size);
+       memcpy(ce->name, path, namelen);
+       ce->ce_flags = htons(namelen);
+       fill_stat_cache_info(ce, &st);
+
+       ce->ce_mode = create_ce_mode(st.st_mode);
+       if (!trust_executable_bit) {
+               /* If there is an existing entry, pick the mode bits
+                * from it.
+                */
+               int pos = cache_name_pos(path, namelen);
+               if (pos >= 0)
+                       ce->ce_mode = active_cache[pos]->ce_mode;
        }
+
+       if (index_path(ce->sha1, path, &st, 1))
+               die("unable to index file %s", path);
+       if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD))
+               die("unable to add %s to index",path);
+       if (verbose)
+               printf("add '%s'\n", path);
+       cache_tree_invalidate_path(active_cache_tree, path);
        return 0;
 }
 
-int add_cache_entry(struct cache_entry *ce, int ok_to_add)
+int ce_same_name(struct cache_entry *a, struct cache_entry *b)
+{
+       int len = ce_namelen(a);
+       return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
+}
+
+int ce_path_match(const struct cache_entry *ce, const char **pathspec)
+{
+       const char *match, *name;
+       int len;
+
+       if (!pathspec)
+               return 1;
+
+       len = ce_namelen(ce);
+       name = ce->name;
+       while ((match = *pathspec++) != NULL) {
+               int matchlen = strlen(match);
+               if (matchlen > len)
+                       continue;
+               if (memcmp(name, match, matchlen))
+                       continue;
+               if (matchlen && name[matchlen-1] == '/')
+                       return 1;
+               if (name[matchlen] == '/' || !name[matchlen])
+                       return 1;
+               if (!matchlen)
+                       return 1;
+       }
+       return 0;
+}
+
+/*
+ * We fundamentally don't like some paths: we don't want
+ * dot or dot-dot anywhere, and for obvious reasons don't
+ * want to recurse into ".git" either.
+ *
+ * Also, we don't want double slashes or slashes at the
+ * end that can make pathnames ambiguous.
+ */
+static int verify_dotfile(const char *rest)
+{
+       /*
+        * The first character was '.', but that
+        * has already been discarded, we now test
+        * the rest.
+        */
+       switch (*rest) {
+       /* "." is not allowed */
+       case '\0': case '/':
+               return 0;
+
+       /*
+        * ".git" followed by  NUL or slash is bad. This
+        * shares the path end test with the ".." case.
+        */
+       case 'g':
+               if (rest[1] != 'i')
+                       break;
+               if (rest[2] != 't')
+                       break;
+               rest += 2;
+       /* fallthrough */
+       case '.':
+               if (rest[1] == '\0' || rest[1] == '/')
+                       return 0;
+       }
+       return 1;
+}
+
+int verify_path(const char *path)
+{
+       char c;
+
+       goto inside;
+       for (;;) {
+               if (!c)
+                       return 1;
+               if (c == '/') {
+inside:
+                       c = *path++;
+                       switch (c) {
+                       default:
+                               continue;
+                       case '/': case '\0':
+                               break;
+                       case '.':
+                               if (verify_dotfile(path))
+                                       continue;
+                       }
+                       return 0;
+               }
+               c = *path++;
+       }
+}
+
+/*
+ * Do we have another file that has the beginning components being a
+ * proper superset of the name we're trying to add?
+ */
+static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
+{
+       int retval = 0;
+       int len = ce_namelen(ce);
+       int stage = ce_stage(ce);
+       const char *name = ce->name;
+
+       while (pos < active_nr) {
+               struct cache_entry *p = active_cache[pos++];
+
+               if (len >= ce_namelen(p))
+                       break;
+               if (memcmp(name, p->name, len))
+                       break;
+               if (ce_stage(p) != stage)
+                       continue;
+               if (p->name[len] != '/')
+                       continue;
+               retval = -1;
+               if (!ok_to_replace)
+                       break;
+               remove_cache_entry_at(--pos);
+       }
+       return retval;
+}
+
+/*
+ * Do we have another file with a pathname that is a proper
+ * subset of the name we're trying to add?
+ */
+static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
+{
+       int retval = 0;
+       int stage = ce_stage(ce);
+       const char *name = ce->name;
+       const char *slash = name + ce_namelen(ce);
+
+       for (;;) {
+               int len;
+
+               for (;;) {
+                       if (*--slash == '/')
+                               break;
+                       if (slash <= ce->name)
+                               return retval;
+               }
+               len = slash - name;
+
+               pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
+               if (pos >= 0) {
+                       retval = -1;
+                       if (ok_to_replace)
+                               break;
+                       remove_cache_entry_at(pos);
+                       continue;
+               }
+
+               /*
+                * Trivial optimization: if we find an entry that
+                * already matches the sub-directory, then we know
+                * we're ok, and we can exit.
+                */
+               pos = -pos-1;
+               while (pos < active_nr) {
+                       struct cache_entry *p = active_cache[pos];
+                       if ((ce_namelen(p) <= len) ||
+                           (p->name[len] != '/') ||
+                           memcmp(p->name, name, len))
+                               break; /* not our subdirectory */
+                       if (ce_stage(p) == stage)
+                               /* p is at the same stage as our entry, and
+                                * is a subdirectory of what we are looking
+                                * at, so we cannot have conflicts at our
+                                * level or anything shorter.
+                                */
+                               return retval;
+                       pos++;
+               }
+       }
+       return retval;
+}
+
+/* We may be in a situation where we already have path/file and path
+ * is being added, or we already have path and path/file is being
+ * added.  Either one would result in a nonsense tree that has path
+ * twice when git-write-tree tries to write it out.  Prevent it.
+ * 
+ * If ok-to-replace is specified, we remove the conflicting entries
+ * from the cache so the caller should recompute the insert position.
+ * When this happens, we return non-zero.
+ */
+static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
+{
+       /*
+        * We check if the path is a sub-path of a subsequent pathname
+        * first, since removing those will not change the position
+        * in the array
+        */
+       int retval = has_file_name(ce, pos, ok_to_replace);
+       /*
+        * Then check if the path might have a clashing sub-directory
+        * before it.
+        */
+       return retval + has_dir_name(ce, pos, ok_to_replace);
+}
+
+int add_cache_entry(struct cache_entry *ce, int option)
 {
        int pos;
+       int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
+       int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
+       int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
 
-       pos = cache_name_pos(ce->name, ce->namelen);
+       pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 
-       /* existing match? Just replace it */
+       /* existing match? Just replace it. */
        if (pos >= 0) {
+               active_cache_changed = 1;
                active_cache[pos] = ce;
                return 0;
        }
        pos = -pos-1;
 
+       /*
+        * Inserting a merged entry ("stage 0") into the index
+        * will always replace all non-merged entries..
+        */
+       if (pos < active_nr && ce_stage(ce) == 0) {
+               while (ce_same_name(active_cache[pos], ce)) {
+                       ok_to_add = 1;
+                       if (!remove_cache_entry_at(pos))
+                               break;
+               }
+       }
+
        if (!ok_to_add)
                return -1;
+       if (!verify_path(ce->name))
+               return -1;
+
+       if (!skip_df_check &&
+           check_file_directory_conflict(ce, pos, ok_to_replace)) {
+               if (!ok_to_replace)
+                       return -1;
+               pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
+               pos = -pos-1;
+       }
 
        /* Make sure the array is big enough .. */
        if (active_nr == active_alloc) {
                active_alloc = alloc_nr(active_alloc);
-               active_cache = realloc(active_cache, active_alloc * sizeof(struct cache_entry *));
+               active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
        }
 
        /* Add it in.. */
@@ -368,109 +623,430 @@ int add_cache_entry(struct cache_entry *ce, int ok_to_add)
        if (active_nr > pos)
                memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
        active_cache[pos] = ce;
+       active_cache_changed = 1;
        return 0;
 }
 
+/*
+ * "refresh" does not calculate a new sha1 file or bring the
+ * cache up-to-date for mode/content changes. But what it
+ * _does_ do is to "re-match" the stat information of a file
+ * with the cache, so that you can refresh the cache for a
+ * file that hasn't been changed but where the stat entry is
+ * out of date.
+ *
+ * For example, you'd want to do this after doing a "git-read-tree",
+ * to link up the stat cache details with the proper files.
+ */
+struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
+{
+       struct stat st;
+       struct cache_entry *updated;
+       int changed, size;
+
+       if (lstat(ce->name, &st) < 0) {
+               cache_errno = errno;
+               return NULL;
+       }
+
+       changed = ce_match_stat(ce, &st, really);
+       if (!changed) {
+               if (really && assume_unchanged &&
+                   !(ce->ce_flags & htons(CE_VALID)))
+                       ; /* mark this one VALID again */
+               else
+                       return ce;
+       }
+
+       if (ce_modified(ce, &st, really)) {
+               cache_errno = EINVAL;
+               return NULL;
+       }
+
+       size = ce_size(ce);
+       updated = xmalloc(size);
+       memcpy(updated, ce, size);
+       fill_stat_cache_info(updated, &st);
+
+       /* In this case, if really is not set, we should leave
+        * CE_VALID bit alone.  Otherwise, paths marked with
+        * --no-assume-unchanged (i.e. things to be edited) will
+        * reacquire CE_VALID bit automatically, which is not
+        * really what we want.
+        */
+       if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
+               updated->ce_flags &= ~htons(CE_VALID);
+
+       return updated;
+}
+
+int refresh_cache(unsigned int flags)
+{
+       int i;
+       int has_errors = 0;
+       int really = (flags & REFRESH_REALLY) != 0;
+       int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
+       int quiet = (flags & REFRESH_QUIET) != 0;
+       int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
+
+       for (i = 0; i < active_nr; i++) {
+               struct cache_entry *ce, *new;
+               ce = active_cache[i];
+               if (ce_stage(ce)) {
+                       while ((i < active_nr) &&
+                              ! strcmp(active_cache[i]->name, ce->name))
+                               i++;
+                       i--;
+                       if (allow_unmerged)
+                               continue;
+                       printf("%s: needs merge\n", ce->name);
+                       has_errors = 1;
+                       continue;
+               }
+
+               new = refresh_cache_entry(ce, really);
+               if (new == ce)
+                       continue;
+               if (!new) {
+                       if (not_new && cache_errno == ENOENT)
+                               continue;
+                       if (really && cache_errno == EINVAL) {
+                               /* If we are doing --really-refresh that
+                                * means the index is not valid anymore.
+                                */
+                               ce->ce_flags &= ~htons(CE_VALID);
+                               active_cache_changed = 1;
+                       }
+                       if (quiet)
+                               continue;
+                       printf("%s: needs update\n", ce->name);
+                       has_errors = 1;
+                       continue;
+               }
+               active_cache_changed = 1;
+               /* You can NOT just free active_cache[i] here, since it
+                * might not be necessarily malloc()ed but can also come
+                * from mmap(). */
+               active_cache[i] = new;
+       }
+       return has_errors;
+}
+
 static int verify_hdr(struct cache_header *hdr, unsigned long size)
 {
        SHA_CTX c;
        unsigned char sha1[20];
 
-       if (hdr->signature != CACHE_SIGNATURE)
+       if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
                return error("bad signature");
-       if (hdr->version != 1)
-               return error("bad version");
+       if (hdr->hdr_version != htonl(2))
+               return error("bad index version");
        SHA1_Init(&c);
-       SHA1_Update(&c, hdr, offsetof(struct cache_header, sha1));
-       SHA1_Update(&c, hdr+1, size - sizeof(*hdr));
+       SHA1_Update(&c, hdr, size - 20);
        SHA1_Final(sha1, &c);
-       if (memcmp(sha1, hdr->sha1, 20))
-               return error("bad header sha1");
+       if (memcmp(sha1, (char *) hdr + size - 20, 20))
+               return error("bad index file sha1 signature");
+       return 0;
+}
+
+static int read_index_extension(const char *ext, void *data, unsigned long sz)
+{
+       switch (CACHE_EXT(ext)) {
+       case CACHE_EXT_TREE:
+               active_cache_tree = cache_tree_read(data, sz);
+               break;
+       default:
+               if (*ext < 'A' || 'Z' < *ext)
+                       return error("index uses %.4s extension, which we do not understand",
+                                    ext);
+               fprintf(stderr, "ignoring %.4s extension\n", ext);
+               break;
+       }
        return 0;
 }
 
 int read_cache(void)
+{
+       return read_cache_from(get_index_file());
+}
+
+/* remember to discard_cache() before reading a different cache! */
+int read_cache_from(const char *path)
 {
        int fd, i;
        struct stat st;
-       unsigned long size, offset;
-       void *map;
+       unsigned long offset;
        struct cache_header *hdr;
 
        errno = EBUSY;
-       if (active_cache)
-               return error("more than one cachefile");
+       if (cache_mmap)
+               return active_nr;
+
        errno = ENOENT;
-       sha1_file_directory = getenv(DB_ENVIRONMENT);
-       if (!sha1_file_directory)
-               sha1_file_directory = DEFAULT_DB_ENVIRONMENT;
-       if (access(sha1_file_directory, X_OK) < 0)
-               return error("no access to SHA1 file directory");
-       fd = open(".git/index", O_RDONLY);
-       if (fd < 0)
-               return (errno == ENOENT) ? 0 : error("open failed");
-
-       size = 0; // avoid gcc warning
-       map = (void *)-1;
+       index_file_timestamp = 0;
+       fd = open(path, O_RDONLY);
+       if (fd < 0) {
+               if (errno == ENOENT)
+                       return 0;
+               die("index file open failed (%s)", strerror(errno));
+       }
+
+       cache_mmap = MAP_FAILED;
        if (!fstat(fd, &st)) {
-               size = st.st_size;
+               cache_mmap_size = st.st_size;
                errno = EINVAL;
-               if (size >= sizeof(struct cache_header))
-                       map = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
+               if (cache_mmap_size >= sizeof(struct cache_header) + 20)
+                       cache_mmap = mmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
        }
        close(fd);
-       if (-1 == (int)(long)map)
-               return error("mmap failed");
+       if (cache_mmap == MAP_FAILED)
+               die("index file mmap failed (%s)", strerror(errno));
 
-       hdr = map;
-       if (verify_hdr(hdr, size) < 0)
+       hdr = cache_mmap;
+       if (verify_hdr(hdr, cache_mmap_size) < 0)
                goto unmap;
 
-       active_nr = hdr->entries;
+       active_nr = ntohl(hdr->hdr_entries);
        active_alloc = alloc_nr(active_nr);
-       active_cache = calloc(active_alloc, sizeof(struct cache_entry *));
+       active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
 
        offset = sizeof(*hdr);
-       for (i = 0; i < hdr->entries; i++) {
-               struct cache_entry *ce = map + offset;
+       for (i = 0; i < active_nr; i++) {
+               struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
                offset = offset + ce_size(ce);
                active_cache[i] = ce;
        }
+       index_file_timestamp = st.st_mtime;
+       while (offset <= cache_mmap_size - 20 - 8) {
+               /* After an array of active_nr index entries,
+                * there can be arbitrary number of extended
+                * sections, each of which is prefixed with
+                * extension name (4-byte) and section length
+                * in 4-byte network byte order.
+                */
+               unsigned long extsize;
+               memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
+               extsize = ntohl(extsize);
+               if (read_index_extension(((const char *) cache_mmap) + offset,
+                                        (char *) cache_mmap + offset + 8,
+                                        extsize) < 0)
+                       goto unmap;
+               offset += 8;
+               offset += extsize;
+       }
        return active_nr;
 
 unmap:
-       munmap(map, size);
+       munmap(cache_mmap, cache_mmap_size);
        errno = EINVAL;
-       return error("verify header failed");
+       die("index file corrupt");
+}
+
+#define WRITE_BUFFER_SIZE 8192
+static unsigned char write_buffer[WRITE_BUFFER_SIZE];
+static unsigned long write_buffer_len;
+
+static int ce_write_flush(SHA_CTX *context, int fd)
+{
+       unsigned int buffered = write_buffer_len;
+       if (buffered) {
+               SHA1_Update(context, write_buffer, buffered);
+               if (write(fd, write_buffer, buffered) != buffered)
+                       return -1;
+               write_buffer_len = 0;
+       }
+       return 0;
+}
+
+static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
+{
+       while (len) {
+               unsigned int buffered = write_buffer_len;
+               unsigned int partial = WRITE_BUFFER_SIZE - buffered;
+               if (partial > len)
+                       partial = len;
+               memcpy(write_buffer + buffered, data, partial);
+               buffered += partial;
+               if (buffered == WRITE_BUFFER_SIZE) {
+                       write_buffer_len = buffered;
+                       if (ce_write_flush(context, fd))
+                               return -1;
+                       buffered = 0;
+               }
+               write_buffer_len = buffered;
+               len -= partial;
+               data = (char *) data + partial;
+       }
+       return 0;
+}
+
+static int write_index_ext_header(SHA_CTX *context, int fd,
+                                 unsigned int ext, unsigned int sz)
+{
+       ext = htonl(ext);
+       sz = htonl(sz);
+       return ((ce_write(context, fd, &ext, 4) < 0) ||
+               (ce_write(context, fd, &sz, 4) < 0)) ? -1 : 0;
+}
+
+static int ce_flush(SHA_CTX *context, int fd)
+{
+       unsigned int left = write_buffer_len;
+
+       if (left) {
+               write_buffer_len = 0;
+               SHA1_Update(context, write_buffer, left);
+       }
+
+       /* Flush first if not enough space for SHA1 signature */
+       if (left + 20 > WRITE_BUFFER_SIZE) {
+               if (write(fd, write_buffer, left) != left)
+                       return -1;
+               left = 0;
+       }
+
+       /* Append the SHA1 signature at the end */
+       SHA1_Final(write_buffer + left, context);
+       left += 20;
+       return (write(fd, write_buffer, left) != left) ? -1 : 0;
+}
+
+static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
+{
+       /*
+        * The only thing we care about in this function is to smudge the
+        * falsely clean entry due to touch-update-touch race, so we leave
+        * everything else as they are.  We are called for entries whose
+        * ce_mtime match the index file mtime.
+        */
+       struct stat st;
+
+       if (lstat(ce->name, &st) < 0)
+               return;
+       if (ce_match_stat_basic(ce, &st))
+               return;
+       if (ce_modified_check_fs(ce, &st)) {
+               /* This is "racily clean"; smudge it.  Note that this
+                * is a tricky code.  At first glance, it may appear
+                * that it can break with this sequence:
+                *
+                * $ echo xyzzy >frotz
+                * $ git-update-index --add frotz
+                * $ : >frotz
+                * $ sleep 3
+                * $ echo filfre >nitfol
+                * $ git-update-index --add nitfol
+                *
+                * but it does not.  When the second update-index runs,
+                * it notices that the entry "frotz" has the same timestamp
+                * as index, and if we were to smudge it by resetting its
+                * size to zero here, then the object name recorded
+                * in index is the 6-byte file but the cached stat information
+                * becomes zero --- which would then match what we would
+                * obtain from the filesystem next time we stat("frotz"). 
+                *
+                * However, the second update-index, before calling
+                * this function, notices that the cached size is 6
+                * bytes and what is on the filesystem is an empty
+                * file, and never calls us, so the cached size information
+                * for "frotz" stays 6 which does not match the filesystem.
+                */
+               ce->ce_size = htonl(0);
+       }
 }
 
 int write_cache(int newfd, struct cache_entry **cache, int entries)
 {
        SHA_CTX c;
        struct cache_header hdr;
-       int i;
+       int i, removed, recent;
+       struct stat st;
+       time_t now;
 
-       hdr.signature = CACHE_SIGNATURE;
-       hdr.version = 1;
-       hdr.entries = entries;
+       for (i = removed = 0; i < entries; i++)
+               if (!cache[i]->ce_mode)
+                       removed++;
 
-       SHA1_Init(&c);
-       SHA1_Update(&c, &hdr, offsetof(struct cache_header, sha1));
-       for (i = 0; i < entries; i++) {
-               struct cache_entry *ce = cache[i];
-               int size = ce_size(ce);
-               SHA1_Update(&c, ce, size);
-       }
-       SHA1_Final(hdr.sha1, &c);
+       hdr.hdr_signature = htonl(CACHE_SIGNATURE);
+       hdr.hdr_version = htonl(2);
+       hdr.hdr_entries = htonl(entries - removed);
 
-       if (write(newfd, &hdr, sizeof(hdr)) != sizeof(hdr))
+       SHA1_Init(&c);
+       if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
                return -1;
 
        for (i = 0; i < entries; i++) {
                struct cache_entry *ce = cache[i];
-               int size = ce_size(ce);
-               if (write(newfd, ce, size) != size)
+               if (!ce->ce_mode)
+                       continue;
+               if (index_file_timestamp &&
+                   index_file_timestamp <= ntohl(ce->ce_mtime.sec))
+                       ce_smudge_racily_clean_entry(ce);
+               if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
                        return -1;
        }
-       return 0;
+
+       /* Write extension data here */
+       if (active_cache_tree) {
+               unsigned long sz;
+               void *data = cache_tree_write(active_cache_tree, &sz);
+               if (data &&
+                   !write_index_ext_header(&c, newfd, CACHE_EXT_TREE, sz) &&
+                   !ce_write(&c, newfd, data, sz))
+                       ;
+               else {
+                       free(data);
+                       return -1;
+               }
+       }
+
+       /*
+        * To prevent later ce_match_stat() from always falling into
+        * check_fs(), if we have too many entries that can trigger
+        * racily clean check, we are better off delaying the return.
+        * We arbitrarily say if more than 20 paths or 25% of total
+        * paths are very new, we delay the return until the index
+        * file gets a new timestamp.
+        *
+        * NOTE! NOTE! NOTE!
+        *
+        * This assumes that nobody is touching the working tree while
+        * we are updating the index.
+        */
+
+       /* Make sure that the new index file has st_mtime
+        * that is current enough -- ce_write() batches the data
+        * so it might not have written anything yet.
+        */
+       ce_write_flush(&c, newfd);
+
+       now = fstat(newfd, &st) ? 0 : st.st_mtime;
+       if (now) {
+               recent = 0;
+               for (i = 0; i < entries; i++) {
+                       struct cache_entry *ce = cache[i];
+                       time_t entry_time = (time_t) ntohl(ce->ce_mtime.sec);
+                       if (!ce->ce_mode)
+                               continue;
+                       if (now && now <= entry_time)
+                               recent++;
+               }
+               if (20 < recent && entries <= recent * 4) {
+#if 0
+                       fprintf(stderr, "entries    %d\n", entries);
+                       fprintf(stderr, "recent     %d\n", recent);
+                       fprintf(stderr, "now        %lu\n", now);
+#endif
+                       while (!fstat(newfd, &st) && st.st_mtime <= now) {
+                               off_t where = lseek(newfd, 0, SEEK_CUR);
+                               sleep(1);
+                               if ((where == (off_t) -1) ||
+                                   (write(newfd, "", 1) != 1) ||
+                                   (lseek(newfd, -1, SEEK_CUR) != where) ||
+                                   ftruncate(newfd, where))
+                                       break;
+                       }
+               }
+       }
+       return ce_flush(&c, newfd);
 }