Add check program "git-check-racy"
[gitweb.git] / read-cache.c
index 50e094e0536740050dd2ac5a041c4c030ce3cfc3..4d621ef4448b34ce8c46674989191513a950cd35 100644 (file)
@@ -4,11 +4,31 @@
  * Copyright (C) Linus Torvalds, 2005
  */
 #include "cache.h"
+#include "cache-tree.h"
+
+/* 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.
+ */
+
+#define CACHE_EXT(s) ( (s[0]<<24)|(s[1]<<16)|(s[2]<<8)|(s[3]) )
+#define CACHE_EXT_TREE 0x54524545      /* "TREE" */
 
 struct cache_entry **active_cache = NULL;
 static time_t index_file_timestamp;
 unsigned int active_nr = 0, active_alloc = 0, active_cache_changed = 0;
 
+struct cache_tree *active_cache_tree = NULL;
+
+int cache_errno = 0;
+
+static void *cache_mmap = NULL;
+static size_t cache_mmap_size = 0;
+
 /*
  * This only updates the "non-critical" parts of the directory
  * cache, ie the parts that aren't tracked by GIT, and only used
@@ -41,7 +61,7 @@ static int ce_compare_data(struct cache_entry *ce, struct stat *st)
                unsigned char sha1[20];
                if (!index_fd(sha1, fd, st, 0, NULL))
                        match = memcmp(sha1, ce->sha1, 20);
-               close(fd);
+               /* index_fd() closed the file descriptor already */
        }
        return match;
 }
@@ -149,9 +169,11 @@ static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
        return changed;
 }
 
-int ce_match_stat(struct cache_entry *ce, struct stat *st, int ignore_valid)
+int ce_match_stat(struct cache_entry *ce, struct stat *st, int options)
 {
        unsigned int changed;
+       int ignore_valid = options & 01;
+       int assume_racy_is_modified = options & 02;
 
        /*
         * If it's marked as always valid in the index, it's
@@ -180,8 +202,12 @@ int ce_match_stat(struct cache_entry *ce, struct stat *st, int ignore_valid)
         */
        if (!changed &&
            index_file_timestamp &&
-           index_file_timestamp <= ntohl(ce->ce_mtime.sec))
-               changed |= ce_modified_check_fs(ce, st);
+           index_file_timestamp <= ntohl(ce->ce_mtime.sec)) {
+               if (assume_racy_is_modified)
+                       changed |= DATA_CHANGED;
+               else
+                       changed |= ce_modified_check_fs(ce, st);
+       }
 
        return changed;
 }
@@ -299,6 +325,45 @@ int remove_file_from_cache(const char *path)
        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 ce_same_name(struct cache_entry *a, struct cache_entry *b)
 {
        int len = ce_namelen(a);
@@ -331,6 +396,70 @@ int ce_path_match(const struct cache_entry *ce, const char **pathspec)
        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?
@@ -472,6 +601,8 @@ int add_cache_entry(struct cache_entry *ce, int option)
 
        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)) {
@@ -496,12 +627,115 @@ int add_cache_entry(struct cache_entry *ce, int option)
        return 0;
 }
 
-static int verify_hdr(struct cache_header *hdr, unsigned long size, unsigned char *sha1)
+/*
+ * "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_buf[20];
-       if (!sha1)
-               sha1 = sha1_buf;
+       unsigned char sha1[20];
 
        if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
                return error("bad signature");
@@ -510,63 +744,100 @@ static int verify_hdr(struct cache_header *hdr, unsigned long size, unsigned cha
        SHA1_Init(&c);
        SHA1_Update(&c, hdr, size - 20);
        SHA1_Final(sha1, &c);
-       if (memcmp(sha1, (void *)hdr + size - 20, 20))
+       if (memcmp(sha1, (char *) hdr + size - 20, 20))
                return error("bad index file sha1 signature");
        return 0;
 }
 
-int read_cache_1(unsigned char *cache_sha1)
+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)
+       if (cache_mmap)
                return active_nr;
 
        errno = ENOENT;
        index_file_timestamp = 0;
-       fd = open(get_index_file(), O_RDONLY);
+       fd = open(path, O_RDONLY);
        if (fd < 0) {
                if (errno == ENOENT)
                        return 0;
                die("index file open failed (%s)", strerror(errno));
        }
 
-       size = 0; // avoid gcc warning
-       map = MAP_FAILED;
+       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) + 20)
-                       map = mmap(NULL, size, PROT_READ | PROT_WRITE, 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 (map == MAP_FAILED)
+       if (cache_mmap == MAP_FAILED)
                die("index file mmap failed (%s)", strerror(errno));
 
-       hdr = map;
-       if (verify_hdr(hdr, size, cache_sha1) < 0)
+       hdr = cache_mmap;
+       if (verify_hdr(hdr, cache_mmap_size) < 0)
                goto unmap;
 
        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 < active_nr; i++) {
-               struct cache_entry *ce = map + offset;
+               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;
        die("index file corrupt");
 }
@@ -575,6 +846,18 @@ int read_cache_1(unsigned char *cache_sha1)
 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) {
@@ -585,19 +868,28 @@ static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
                memcpy(write_buffer + buffered, data, partial);
                buffered += partial;
                if (buffered == WRITE_BUFFER_SIZE) {
-                       SHA1_Update(context, write_buffer, WRITE_BUFFER_SIZE);
-                       if (write(fd, write_buffer, WRITE_BUFFER_SIZE) != WRITE_BUFFER_SIZE)
+                       write_buffer_len = buffered;
+                       if (ce_write_flush(context, fd))
                                return -1;
                        buffered = 0;
                }
                write_buffer_len = buffered;
                len -= partial;
-               data += partial;
+               data = (char *) data + partial;
        }
        return 0;
 }
 
-static int ce_flush(SHA_CTX *context, int fd, unsigned char *sha1)
+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;
 
@@ -614,12 +906,9 @@ static int ce_flush(SHA_CTX *context, int fd, unsigned char *sha1)
        }
 
        /* Append the SHA1 signature at the end */
-       SHA1_Final(sha1, context);
-       memcpy(write_buffer + left, sha1, 20);
+       SHA1_Final(write_buffer + left, context);
        left += 20;
-       if (write(fd, write_buffer, left) != left)
-               return -1;
-       return 0;
+       return (write(fd, write_buffer, left) != left) ? -1 : 0;
 }
 
 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
@@ -648,7 +937,7 @@ static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
                 * $ echo filfre >nitfol
                 * $ git-update-index --add nitfol
                 *
-                * but it does not.  Whe the second update-index runs,
+                * 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
@@ -666,14 +955,13 @@ static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
        }
 }
 
-int write_cache_1(int newfd, struct cache_entry **cache, int entries,
-                 unsigned char *cache_sha1)
+int write_cache(int newfd, struct cache_entry **cache, int entries)
 {
        SHA_CTX c;
        struct cache_header hdr;
-       int i, removed;
-       int status;
-       unsigned char sha1[20];
+       int i, removed, recent;
+       struct stat st;
+       time_t now;
 
        for (i = removed = 0; i < entries; i++)
                if (!cache[i]->ce_mode)
@@ -697,18 +985,68 @@ int write_cache_1(int newfd, struct cache_entry **cache, int entries,
                if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
                        return -1;
        }
-       status = ce_flush(&c, newfd, sha1);
-       if (cache_sha1)
-               memcpy(cache_sha1, sha1, 20);
-       return status;
-}
 
-int read_cache(void)
-{
-       return read_cache_1(NULL);
-}
+       /* 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;
+               }
+       }
 
-int write_cache(int newfd, struct cache_entry **cache, int entries)
-{
-       return write_cache_1(newfd, cache, entries, NULL);
+       /*
+        * 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);
 }