[PATCH] read-tree: fix too strong index requirement #5ALT
[git.git] / read-tree.c
index 8c9ffe1..ecd40cc 100644 (file)
@@ -22,14 +22,6 @@ static int unpack_tree(unsigned char *sha1)
        return ret;
 }
 
-static char *lockfile_name;
-
-static void remove_lock_file(void)
-{
-       if (lockfile_name)
-               unlink(lockfile_name);
-}
-
 static int path_matches(struct cache_entry *a, struct cache_entry *b)
 {
        int len = ce_namelen(a);
@@ -47,27 +39,11 @@ static int same(struct cache_entry *a, struct cache_entry *b)
 /*
  * This removes all trivial merges that don't change the tree
  * and collapses them to state 0.
- *
- * _Any_ other merge is left to user policy.  That includes "both
- * created the same file", and "both removed the same file" - which are
- * trivial, but the user might still want to _note_ it. 
  */
 static struct cache_entry *merge_entries(struct cache_entry *a,
                                         struct cache_entry *b,
                                         struct cache_entry *c)
 {
-       int len = ce_namelen(a);
-
-       /*
-        * Are they all the same filename? We won't do
-        * any name merging
-        */
-       if (ce_namelen(b) != len ||
-           ce_namelen(c) != len ||
-           memcmp(a->name, b->name, len) ||
-           memcmp(a->name, c->name, len))
-               return NULL;
-
        /*
         * Ok, all three entries describe the same
         * filename, but maybe the contents or file
@@ -85,12 +61,14 @@ static struct cache_entry *merge_entries(struct cache_entry *a,
         * Here "a" is "original", and "b" and "c" are the two
         * trees we are merging.
         */
-       if (same(b,c))
-               return c;
-       if (same(a,b))
-               return c;
-       if (same(a,c))
-               return b;
+       if (a && b && c) {
+               if (same(b,c))
+                       return c;
+               if (same(a,b))
+                       return c;
+               if (same(a,c))
+                       return b;
+       }
        return NULL;
 }
 
@@ -124,171 +102,202 @@ static void reject_merge(struct cache_entry *ce)
        die("Entry '%s' would be overwritten by merge. Cannot merge.", ce->name);
 }
 
-#define CHECK_OLD(ce) if (old && same(old, ce)) { verify_uptodate(old); old = NULL; }
-
-static void trivially_merge_cache(struct cache_entry **src, int nr)
+static int merged_entry_internal(struct cache_entry *merge, struct cache_entry *old, struct cache_entry **dst, int allow_dirty)
 {
-       struct cache_entry **dst = src;
-       struct cache_entry *old = NULL;
+       merge->ce_flags |= htons(CE_UPDATE);
+       if (old) {
+               /*
+                * See if we can re-use the old CE directly?
+                * That way we get the uptodate stat info.
+                *
+                * This also removes the UPDATE flag on
+                * a match.
+                */
+               if (same(old, merge)) {
+                       *merge = *old;
+               } else if (!allow_dirty) {
+                       verify_uptodate(old);
+               }
+       }
+       merge->ce_flags &= ~htons(CE_STAGEMASK);
+       *dst++ = merge;
+       return 1;
+}
 
-       while (nr--) {
-               struct cache_entry *ce, *result;
+static int merged_entry_allow_dirty(struct cache_entry *merge, struct cache_entry *old, struct cache_entry **dst)
+{
+       return merged_entry_internal(merge, old, dst, 1);
+}
 
-               ce = *src++;
+static int merged_entry(struct cache_entry *merge, struct cache_entry *old, struct cache_entry **dst)
+{
+       return merged_entry_internal(merge, old, dst, 0);
+}
 
-               /* We throw away original cache entries except for the stat information */
-               if (!ce_stage(ce)) {
-                       if (old)
-                               reject_merge(old);
-                       old = ce;
-                       active_nr--;
-                       continue;
-               }
-               if (old && !path_matches(old, ce))
-                       reject_merge(old);
-               if (nr > 2 && (result = merge_entries(ce, src[1], src[2])) != NULL) {
-                       result->ce_flags |= htons(CE_UPDATE);
-                       /*
-                        * See if we can re-use the old CE directly?
-                        * That way we get the uptodate stat info.
-                        *
-                        * This also removes the UPDATE flag on
-                        * a match.
-                        */
-                       if (old && same(old, result)) {
-                               *result = *old;
-                               old = NULL;
-                       }
-                       CHECK_OLD(ce);
-                       CHECK_OLD(src[1]);
-                       CHECK_OLD(src[2]);
-                       ce = result;
-                       ce->ce_flags &= ~htons(CE_STAGEMASK);
-                       src += 2;
-                       nr -= 2;
-                       active_nr -= 2;
-               }
+static int deleted_entry(struct cache_entry *ce, struct cache_entry *old, struct cache_entry **dst)
+{
+       if (old)
+               verify_uptodate(old);
+       ce->ce_mode = 0;
+       *dst++ = ce;
+       return 1;
+}
 
-               /*
-                * If we had an old entry that we now effectively
-                * overwrite, make sure it wasn't dirty.
-                */
-               CHECK_OLD(ce);
-               *dst++ = ce;
+static int threeway_merge(struct cache_entry *stages[4], struct cache_entry **dst)
+{
+       struct cache_entry *old = stages[0];
+       struct cache_entry *a = stages[1], *b = stages[2], *c = stages[3];
+       struct cache_entry *merge;
+       int count;
+
+       /* #5ALT */
+       if (!a && b && c && same(b, c)) {
+               if (old && !same(b, old))
+                       return -1;
+               return merged_entry_allow_dirty(b, old, dst);
        }
+       /*
+        * If we have an entry in the index cache ("old"), then we want
+        * to make sure that it matches any entries in stage 2 ("first
+        * branch", aka "b").
+        */
+       if (old) {
+               if (!b || !same(old, b))
+                       return -1;
+       }
+       merge = merge_entries(a, b, c);
+       if (merge)
+               return merged_entry(merge, old, dst);
        if (old)
-               reject_merge(old);
+               verify_uptodate(old);
+       count = 0;
+       if (a) { *dst++ = a; count++; }
+       if (b) { *dst++ = b; count++; }
+       if (c) { *dst++ = c; count++; }
+       return count;
 }
 
 /*
- * When we find a "stage2" entry in the two-way merge, that's
- * the one that will remain. If we have an exact old match,
- * we don't care whether the file is up-to-date or not, we just
- * re-use the thing directly.
+ * Two-way merge.
+ *
+ * The rule is to "carry forward" what is in the index without losing
+ * information across a "fast forward", favoring a successful merge
+ * over a merge failure when it makes sense.  For details of the
+ * "carry forward" rule, please see <Documentation/git-read-tree.txt>.
  *
- * If we didn't have an exact match, then we want to make sure
- * that we've seen a stage1 that matched the old, and that the
- * old file was up-to-date. Because it will be gone after this
- * merge..
  */
-static void twoway_check(struct cache_entry *old, int seen_stage1, struct cache_entry *ce)
+static int twoway_merge(struct cache_entry **src, struct cache_entry **dst)
 {
-       if (path_matches(old, ce)) {
-               /*
-                * This also removes the UPDATE flag on
-                * a match
-                */
-               if (same(old, ce)) {
-                       *ce = *old;
-                       return;
+       struct cache_entry *current = src[0];
+       struct cache_entry *oldtree = src[1], *newtree = src[2];
+
+       if (src[3])
+               return -1;
+
+       if (current) {
+               if ((!oldtree && !newtree) || /* 4 and 5 */
+                   (!oldtree && newtree &&
+                    same(current, newtree)) || /* 6 and 7 */
+                   (oldtree && newtree &&
+                    same(oldtree, newtree)) || /* 14 and 15 */
+                   (oldtree && newtree &&
+                    !same(oldtree, newtree) && /* 18 and 19*/
+                    same(current, newtree))) {
+                       *dst++ = current;
+                       return 1;
+               }
+               else if (oldtree && !newtree && same(current, oldtree)) {
+                       /* 10 or 11 */
+                       return deleted_entry(oldtree, current, dst);
+               }
+               else if (oldtree && newtree &&
+                        same(current, oldtree) && !same(current, newtree)) {
+                       /* 20 or 21 */
+                       return merged_entry(newtree, current, dst);
                }
-               if (!seen_stage1)
-                       reject_merge(old);
+               else
+                       /* all other failures */
+                       return -1;
        }
-       verify_uptodate(old);
+       else if (newtree)
+               return merged_entry(newtree, current, dst);
+       else
+               return deleted_entry(oldtree, current, dst);
 }
 
 /*
- * Two-way merge.
+ * Two-way merge emulated with three-way merge.
+ *
+ * This treats "read-tree -m H M" by transforming it internally
+ * into "read-tree -m H I+H M", where I+H is a tree that would
+ * contain the contents of the current index file, overlayed on
+ * top of H.  Unlike the traditional two-way merge, this leaves
+ * the stages in the resulting index file and lets the user resolve
+ * the merge conflicts using standard tools for three-way merge.
  *
- * The rule is: 
- *  - every current entry has to match the old tree
- *  - if the current entry matches the new tree, we leave it
- *    as-is. Otherwise we require that it be up-to-date.
+ * This function is just to set-up such an arrangement, and the
+ * actual merge uses threeway_merge() function.
  */
-static void twoway_merge(struct cache_entry **src, int nr)
+static void setup_emu23(void)
 {
-       int seen_stage1 = 0;
-       struct cache_entry *old = NULL;
-       struct cache_entry **dst = src;
-
-       while (nr--) {
-               struct cache_entry *ce = *src++;
-               int stage = ce_stage(ce);
-
-               switch (stage) {
-               case 0:
-                       if (old)
-                               reject_merge(old);
-                       old = ce;
-                       seen_stage1 = 0;
-                       active_nr--;
-                       continue;
-
-               case 1:
-                       active_nr--;
-                       if (!old)
-                               continue;
-                       if (!path_matches(old, ce) || !same(old, ce))
-                               reject_merge(old);
-                       seen_stage1 = 1;
-                       continue;
+       /* stage0 contains I, stage1 H, stage2 M.
+        * move stage2 to stage3, and create stage2 entries
+        * by scanning stage0 and stage1 entries.
+        */
+       int i, namelen, size;
+       struct cache_entry *ce, *stage2;
 
-               case 2:
-                       ce->ce_flags |= htons(CE_UPDATE);
-                       if (old) {
-                               twoway_check(old, seen_stage1, ce);
-                               old = NULL;
-                       }
-                       ce->ce_flags &= ~htons(CE_STAGEMASK);
-                       *dst++ = ce;
+       for (i = 0; i < active_nr; i++) {
+               ce = active_cache[i];
+               if (ce_stage(ce) != 2)
                        continue;
-               }
-               die("impossible two-way stage");
+               /* hoist them up to stage 3 */
+               namelen = ce_namelen(ce);
+               ce->ce_flags = create_ce_flags(namelen, 3);
        }
 
-       /*
-        * Unmatched with a new entry? Make sure it was
-        * at least uptodate in the working directory _and_
-        * the original tree..
-        */
-       if (old) {
-               if (!seen_stage1)
-                       reject_merge(old);
-               verify_uptodate(old);
+       for (i = 0; i < active_nr; i++) {
+               ce = active_cache[i];
+               if (ce_stage(ce) > 1)
+                       continue;
+               namelen = ce_namelen(ce);
+               size = cache_entry_size(namelen);
+               stage2 = xmalloc(size);
+               memcpy(stage2, ce, size);
+               stage2->ce_flags = create_ce_flags(namelen, 2);
+               if (add_cache_entry(stage2, ADD_CACHE_OK_TO_ADD) < 0)
+                       die("cannot merge index and our head tree");
+
+               /* We are done with this name, so skip to next name */
+               while (i < active_nr &&
+                      ce_namelen(active_cache[i]) == namelen &&
+                      !memcmp(active_cache[i]->name, ce->name, namelen))
+                       i++;
+               i--; /* compensate for the loop control */
        }
 }
 
-static void merge_stat_info(struct cache_entry **src, int nr)
+/*
+ * One-way merge.
+ *
+ * The rule is:
+ * - take the stat information from stage0, take the data from stage1
+ */
+static int oneway_merge(struct cache_entry **src, struct cache_entry **dst)
 {
-       static struct cache_entry null_entry;
-       struct cache_entry **dst = src;
-       struct cache_entry *stat = &null_entry;
+       struct cache_entry *old = src[0];
+       struct cache_entry *a = src[1];
 
-       while (nr--) {
-               struct cache_entry *ce = *src++;
+       if (src[2] || src[3])
+               return -1;
 
-               /* We throw away original cache entries except for the stat information */
-               if (!ce_stage(ce)) {
-                       stat = ce;
-                       active_nr--;
-                       continue;
-               }
-               if (path_matches(ce, stat) && same(ce, stat))
-                       *ce = *stat;
-               ce->ce_flags &= ~htons(CE_STAGEMASK);
-               *dst++ = ce;
+       if (!a)
+               return 0;
+       if (old && same(old, a)) {
+               *dst++ = old;
+               return 1;
        }
+       return merged_entry(a, NULL, dst);
 }
 
 static void check_updates(struct cache_entry **src, int nr)
@@ -302,6 +311,11 @@ static void check_updates(struct cache_entry **src, int nr)
        unsigned short mask = htons(CE_UPDATE);
        while (nr--) {
                struct cache_entry *ce = *src++;
+               if (!ce->ce_mode) {
+                       if (update)
+                               unlink(ce->name);
+                       continue;
+               }
                if (ce->ce_flags & mask) {
                        ce->ce_flags &= ~mask;
                        if (update)
@@ -310,24 +324,75 @@ static void check_updates(struct cache_entry **src, int nr)
        }
 }
 
-static char *read_tree_usage = "git-read-tree (<sha> | -m <sha1> [<sha2> [<sha3>]])";
+typedef int (*merge_fn_t)(struct cache_entry **, struct cache_entry **);
+
+static void merge_cache(struct cache_entry **src, int nr, merge_fn_t fn)
+{
+       struct cache_entry **dst = src;
+
+       while (nr) {
+               int entries;
+               struct cache_entry *name, *ce, *stages[4] = { NULL, };
+
+               name = ce = *src;
+               for (;;) {
+                       int stage = ce_stage(ce);
+                       stages[stage] = ce;
+                       ce = *++src;
+                       active_nr--;
+                       if (!--nr)
+                               break;
+                       if (!path_matches(ce, name))
+                               break;
+               }
+
+               entries = fn(stages, dst);
+               if (entries < 0)
+                       reject_merge(name);
+               dst += entries;
+               active_nr += entries;
+       }
+       check_updates(active_cache, active_nr);
+}
+
+static int read_cache_unmerged(void)
+{
+       int i, deleted;
+       struct cache_entry **dst;
+
+       read_cache();
+       dst = active_cache;
+       deleted = 0;
+       for (i = 0; i < active_nr; i++) {
+               struct cache_entry *ce = active_cache[i];
+               if (ce_stage(ce)) {
+                       deleted++;
+                       continue;
+               }
+               if (deleted)
+                       *dst = ce;
+               dst++;
+       }
+       active_nr -= deleted;
+       return deleted;
+}
+
+static char *read_tree_usage = "git-read-tree (<sha> | -m [-u] <sha1> [<sha2> [<sha3>]])";
+
+static struct cache_file cache_file;
 
 int main(int argc, char **argv)
 {
-       int i, newfd, merge;
+       int i, newfd, merge, reset, emu23;
        unsigned char sha1[20];
-       static char lockfile[MAXPATHLEN+1];
-       const char *indexfile = get_index_file();
 
-       snprintf(lockfile, sizeof(lockfile), "%s.lock", indexfile);
-
-       newfd = open(lockfile, O_RDWR | O_CREAT | O_EXCL, 0600);
+       newfd = hold_index_file_for_update(&cache_file, get_index_file());
        if (newfd < 0)
                die("unable to create new cachefile");
-       atexit(remove_lock_file);
-       lockfile_name = lockfile;
 
        merge = 0;
+       reset = 0;
+       emu23 = 0;
        for (i = 1; i < argc; i++) {
                const char *arg = argv[i];
 
@@ -337,20 +402,37 @@ int main(int argc, char **argv)
                        continue;
                }
 
+               /* This differs from "-m" in that we'll silently ignore unmerged entries */
+               if (!strcmp(arg, "--reset")) {
+                       if (stage || merge || emu23)
+                               usage(read_tree_usage);
+                       reset = 1;
+                       merge = 1;
+                       stage = 1;
+                       read_cache_unmerged();
+               }
+
                /* "-m" stands for "merge", meaning we start in stage 1 */
                if (!strcmp(arg, "-m")) {
-                       int i;
-                       if (stage)
-                               die("-m needs to come first");
-                       read_cache();
-                       for (i = 0; i < active_nr; i++) {
-                               if (ce_stage(active_cache[i]))
-                                       die("you need to resolve your current index first");
-                       }
+                       if (stage || merge || emu23)
+                               usage(read_tree_usage);
+                       if (read_cache_unmerged())
+                               die("you need to resolve your current index first");
                        stage = 1;
                        merge = 1;
                        continue;
                }
+
+               /* "-emu23" uses 3-way merge logic to perform fast-forward */
+               if (!strcmp(arg, "--emu23")) {
+                       if (stage || merge || emu23)
+                               usage(read_tree_usage);
+                       if (read_cache_unmerged())
+                               die("you need to resolve your current index first");
+                       merge = emu23 = stage = 1;
+                       continue;
+               }
+
                if (get_sha1(arg, sha1) < 0)
                        usage(read_tree_usage);
                if (stage > 3)
@@ -359,25 +441,29 @@ int main(int argc, char **argv)
                        die("failed to unpack tree object %s", arg);
                stage++;
        }
+       if (update && !merge)
+               usage(read_tree_usage);
        if (merge) {
-               switch (stage) {
-               case 4: /* Three-way merge */
-                       trivially_merge_cache(active_cache, active_nr);
-                       check_updates(active_cache, active_nr);
-                       break;
-               case 3: /* Update from one tree to another */
-                       twoway_merge(active_cache, active_nr);
-                       check_updates(active_cache, active_nr);
-                       break;
-               case 2: /* Just read a tree, merge with old cache contents */
-                       merge_stat_info(active_cache, active_nr);
-                       break;
-               default:
+               static const merge_fn_t merge_function[] = {
+                       [1] = oneway_merge,
+                       [2] = twoway_merge,
+                       [3] = threeway_merge,
+               };
+               merge_fn_t fn;
+
+               if (stage < 2 || stage > 4)
                        die("just how do you expect me to merge %d trees?", stage-1);
+               if (emu23 && stage != 3)
+                       die("--emu23 takes only two trees");
+               fn = merge_function[stage-1];
+               if (stage == 3 && emu23) { 
+                       setup_emu23();
+                       fn = merge_function[3];
                }
+               merge_cache(active_cache, active_nr, fn);
        }
-       if (write_cache(newfd, active_cache, active_nr) || rename(lockfile, indexfile))
+       if (write_cache(newfd, active_cache, active_nr) ||
+           commit_index_file(&cache_file))
                die("unable to write new index file");
-       lockfile_name = NULL;
        return 0;
 }