fast-import: prevent producing bad delta
[gitweb.git] / sha1_file.c
index 1a7e41070efa6327859d3c09ede07317cc91ba7d..697f4a43c5ef0bdc220a15545d94bc31ad1f248e 100644 (file)
@@ -11,6 +11,7 @@
 #include "pack.h"
 #include "blob.h"
 #include "commit.h"
+#include "run-command.h"
 #include "tag.h"
 #include "tree.h"
 #include "tree-walk.h"
@@ -1204,20 +1205,29 @@ static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
        return map;
 }
 
-static int legacy_loose_object(unsigned char *map)
+/*
+ * There used to be a second loose object header format which
+ * was meant to mimic the in-pack format, allowing for direct
+ * copy of the object data.  This format turned up not to be
+ * really worth it and we no longer write loose objects in that
+ * format.
+ */
+static int experimental_loose_object(unsigned char *map)
 {
        unsigned int word;
 
        /*
         * Is it a zlib-compressed buffer? If so, the first byte
         * must be 0x78 (15-bit window size, deflated), and the
-        * first 16-bit word is evenly divisible by 31
+        * first 16-bit word is evenly divisible by 31. If so,
+        * we are looking at the official format, not the experimental
+        * one.
         */
        word = (map[0] << 8) + map[1];
        if (map[0] == 0x78 && !(word % 31))
-               return 1;
-       else
                return 0;
+       else
+               return 1;
 }
 
 unsigned long unpack_object_header_buffer(const unsigned char *buf,
@@ -1261,34 +1271,29 @@ static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned lon
        stream->next_out = buffer;
        stream->avail_out = bufsiz;
 
-       if (legacy_loose_object(map)) {
-               git_inflate_init(stream);
-               return git_inflate(stream, 0);
-       }
-
+       if (experimental_loose_object(map)) {
+               /*
+                * The old experimental format we no longer produce;
+                * we can still read it.
+                */
+               used = unpack_object_header_buffer(map, mapsize, &type, &size);
+               if (!used || !valid_loose_object_type[type])
+                       return -1;
+               map += used;
+               mapsize -= used;
 
-       /*
-        * There used to be a second loose object header format which
-        * was meant to mimic the in-pack format, allowing for direct
-        * copy of the object data.  This format turned up not to be
-        * really worth it and we don't write it any longer.  But we
-        * can still read it.
-        */
-       used = unpack_object_header_buffer(map, mapsize, &type, &size);
-       if (!used || !valid_loose_object_type[type])
-               return -1;
-       map += used;
-       mapsize -= used;
+               /* Set up the stream for the rest.. */
+               stream->next_in = map;
+               stream->avail_in = mapsize;
+               git_inflate_init(stream);
 
-       /* Set up the stream for the rest.. */
-       stream->next_in = map;
-       stream->avail_in = mapsize;
+               /* And generate the fake traditional header */
+               stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
+                                                typename(type), size);
+               return 0;
+       }
        git_inflate_init(stream);
-
-       /* And generate the fake traditional header */
-       stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
-                                        typename(type), size);
-       return 0;
+       return git_inflate(stream, 0);
 }
 
 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
@@ -2205,23 +2210,21 @@ static void *read_object(const unsigned char *sha1, enum object_type *type,
  * deal with them should arrange to call read_object() and give error
  * messages themselves.
  */
-void *read_sha1_file_repl(const unsigned char *sha1,
-                         enum object_type *type,
-                         unsigned long *size,
-                         const unsigned char **replacement)
+void *read_sha1_file_extended(const unsigned char *sha1,
+                             enum object_type *type,
+                             unsigned long *size,
+                             unsigned flag)
 {
-       const unsigned char *repl = lookup_replace_object(sha1);
        void *data;
        char *path;
        const struct packed_git *p;
+       const unsigned char *repl = (flag & READ_SHA1_FILE_REPLACE)
+               ? lookup_replace_object(sha1) : sha1;
 
        errno = 0;
        data = read_object(repl, type, size);
-       if (data) {
-               if (replacement)
-                       *replacement = repl;
+       if (data)
                return data;
-       }
 
        if (errno && errno != ENOENT)
                die_errno("failed to read object %s", sha1_to_hex(sha1));
@@ -2580,10 +2583,11 @@ static void check_tag(const void *buf, size_t size)
 }
 
 static int index_mem(unsigned char *sha1, void *buf, size_t size,
-                    int write_object, enum object_type type,
-                    const char *path, int format_check)
+                    enum object_type type,
+                    const char *path, unsigned flags)
 {
        int ret, re_allocated = 0;
+       int write_object = flags & HASH_WRITE_OBJECT;
 
        if (!type)
                type = OBJ_BLOB;
@@ -2599,7 +2603,7 @@ static int index_mem(unsigned char *sha1, void *buf, size_t size,
                        re_allocated = 1;
                }
        }
-       if (format_check) {
+       if (flags & HASH_FORMAT_CHECK) {
                if (type == OBJ_TREE)
                        check_tree(buf, size);
                if (type == OBJ_COMMIT)
@@ -2617,44 +2621,141 @@ static int index_mem(unsigned char *sha1, void *buf, size_t size,
        return ret;
 }
 
+static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
+                     const char *path, unsigned flags)
+{
+       struct strbuf sbuf = STRBUF_INIT;
+       int ret;
+
+       if (strbuf_read(&sbuf, fd, 4096) >= 0)
+               ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
+       else
+               ret = -1;
+       strbuf_release(&sbuf);
+       return ret;
+}
+
 #define SMALL_FILE_SIZE (32*1024)
 
-int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
-            enum object_type type, const char *path, int format_check)
+static int index_core(unsigned char *sha1, int fd, size_t size,
+                     enum object_type type, const char *path,
+                     unsigned flags)
 {
        int ret;
-       size_t size = xsize_t(st->st_size);
 
-       if (!S_ISREG(st->st_mode)) {
-               struct strbuf sbuf = STRBUF_INIT;
-               if (strbuf_read(&sbuf, fd, 4096) >= 0)
-                       ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
-                                       type, path, format_check);
-               else
-                       ret = -1;
-               strbuf_release(&sbuf);
-       } else if (!size) {
-               ret = index_mem(sha1, NULL, size, write_object, type, path,
-                               format_check);
+       if (!size) {
+               ret = index_mem(sha1, NULL, size, type, path, flags);
        } else if (size <= SMALL_FILE_SIZE) {
                char *buf = xmalloc(size);
                if (size == read_in_full(fd, buf, size))
-                       ret = index_mem(sha1, buf, size, write_object, type,
-                                       path, format_check);
+                       ret = index_mem(sha1, buf, size, type, path, flags);
                else
                        ret = error("short read %s", strerror(errno));
                free(buf);
        } else {
                void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
-               ret = index_mem(sha1, buf, size, write_object, type, path,
-                               format_check);
+               ret = index_mem(sha1, buf, size, type, path, flags);
                munmap(buf, size);
        }
+       return ret;
+}
+
+/*
+ * This creates one packfile per large blob, because the caller
+ * immediately wants the result sha1, and fast-import can report the
+ * object name via marks mechanism only by closing the created
+ * packfile.
+ *
+ * This also bypasses the usual "convert-to-git" dance, and that is on
+ * purpose. We could write a streaming version of the converting
+ * functions and insert that before feeding the data to fast-import
+ * (or equivalent in-core API described above), but the primary
+ * motivation for trying to stream from the working tree file and to
+ * avoid mmaping it in core is to deal with large binary blobs, and
+ * by definition they do _not_ want to get any conversion.
+ */
+static int index_stream(unsigned char *sha1, int fd, size_t size,
+                       enum object_type type, const char *path,
+                       unsigned flags)
+{
+       struct child_process fast_import;
+       char export_marks[512];
+       const char *argv[] = { "fast-import", "--quiet", export_marks, NULL };
+       char tmpfile[512];
+       char fast_import_cmd[512];
+       char buf[512];
+       int len, tmpfd;
+
+       strcpy(tmpfile, git_path("hashstream_XXXXXX"));
+       tmpfd = git_mkstemp_mode(tmpfile, 0600);
+       if (tmpfd < 0)
+               die_errno("cannot create tempfile: %s", tmpfile);
+       if (close(tmpfd))
+               die_errno("cannot close tempfile: %s", tmpfile);
+       sprintf(export_marks, "--export-marks=%s", tmpfile);
+
+       memset(&fast_import, 0, sizeof(fast_import));
+       fast_import.in = -1;
+       fast_import.argv = argv;
+       fast_import.git_cmd = 1;
+       if (start_command(&fast_import))
+               die_errno("index-stream: git fast-import failed");
+
+       len = sprintf(fast_import_cmd, "blob\nmark :1\ndata %lu\n",
+                     (unsigned long) size);
+       write_or_whine(fast_import.in, fast_import_cmd, len,
+                      "index-stream: feeding fast-import");
+       while (size) {
+               char buf[10240];
+               size_t sz = size < sizeof(buf) ? size : sizeof(buf);
+               size_t actual;
+
+               actual = read_in_full(fd, buf, sz);
+               if (actual < 0)
+                       die_errno("index-stream: reading input");
+               if (write_in_full(fast_import.in, buf, actual) != actual)
+                       die_errno("index-stream: feeding fast-import");
+               size -= actual;
+       }
+       if (close(fast_import.in))
+               die_errno("index-stream: closing fast-import");
+       if (finish_command(&fast_import))
+               die_errno("index-stream: finishing fast-import");
+
+       tmpfd = open(tmpfile, O_RDONLY);
+       if (tmpfd < 0)
+               die_errno("index-stream: cannot open fast-import mark");
+       len = read(tmpfd, buf, sizeof(buf));
+       if (len < 0)
+               die_errno("index-stream: reading fast-import mark");
+       if (close(tmpfd) < 0)
+               die_errno("index-stream: closing fast-import mark");
+       if (unlink(tmpfile))
+               die_errno("index-stream: unlinking fast-import mark");
+       if (len != 44 ||
+           memcmp(":1 ", buf, 3) ||
+           get_sha1_hex(buf + 3, sha1))
+               die_errno("index-stream: unexpected fast-import mark: <%s>", buf);
+       return 0;
+}
+
+int index_fd(unsigned char *sha1, int fd, struct stat *st,
+            enum object_type type, const char *path, unsigned flags)
+{
+       int ret;
+       size_t size = xsize_t(st->st_size);
+
+       if (!S_ISREG(st->st_mode))
+               ret = index_pipe(sha1, fd, type, path, flags);
+       else if (size <= big_file_threshold || type != OBJ_BLOB)
+               ret = index_core(sha1, fd, size, type, path, flags);
+       else
+               ret = index_stream(sha1, fd, size, type, path, flags);
        close(fd);
        return ret;
 }
 
-int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
+int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
 {
        int fd;
        struct strbuf sb = STRBUF_INIT;
@@ -2665,7 +2766,7 @@ int index_path(unsigned char *sha1, const char *path, struct stat *st, int write
                if (fd < 0)
                        return error("open(\"%s\"): %s", path,
                                     strerror(errno));
-               if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path, 0) < 0)
+               if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
                        return error("%s: failed to insert into database",
                                     path);
                break;
@@ -2675,7 +2776,7 @@ int index_path(unsigned char *sha1, const char *path, struct stat *st, int write
                        return error("readlink(\"%s\"): %s", path,
                                     errstr);
                }
-               if (!write_object)
+               if (!(flags & HASH_WRITE_OBJECT))
                        hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
                else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
                        return error("%s: failed to insert into database",