From: Tim Kientzle <kientzle@gmail.com>
Date: Sat, 12 Dec 2009 23:40:49 +0000 (-0500)
Subject: Cpio "odc" format limits the ino value to 18 bits.  Previously, we
X-Git-Tag: v2.8.0~97
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=f8cd5b5aab2e28229d352c3e48306346052c2ef4;p=thirdparty%2Flibarchive.git

Cpio "odc" format limits the ino value to 18 bits.  Previously, we
simply truncated larger values, which can lead to false collisions and
false hardlink detection, especially on filesystems such as NTFS that
use 64-bit ino values.  Because of the risk, we must issue a warning
in this case, and the warnings have been creating problems for the
bsdcpio test suite.

In order to avoid this problem, map the incoming ino values to new ino
values.  This allows us to store up to 256k distinct files without
risk of false hardlinks.  The only drawback to this approach is the risk
of losing hardlinks when appending to existing archives.

SVN-Revision: 1727
---

diff --git a/libarchive/archive_write_set_format_cpio.c b/libarchive/archive_write_set_format_cpio.c
index 4becb21cf..bddb529ee 100644
--- a/libarchive/archive_write_set_format_cpio.c
+++ b/libarchive/archive_write_set_format_cpio.c
@@ -54,6 +54,12 @@ static int64_t	format_octal_recursive(int64_t, char *, int);
 
 struct cpio {
 	uint64_t	  entry_bytes_remaining;
+
+	int64_t		  ino_next;
+
+	struct		 { int64_t old; int new;} *ino_list;
+	size_t		  ino_list_size;
+	size_t		  ino_list_next;
 };
 
 struct cpio_header {
@@ -103,6 +109,71 @@ archive_write_set_format_cpio(struct archive *_a)
 	return (ARCHIVE_OK);
 }
 
+/*
+ * Ino values are as long as 64 bits on some systems; cpio format
+ * only allows 18 bits and relies on the ino values to identify hardlinked
+ * files.  So, we can't merely "hash" the ino numbers since collisions
+ * would corrupt the archive.  Instead, we generate synthetic ino values
+ * to store in the archive and maintain a map of original ino values to
+ * synthetic ones so we can preserve hardlink information.
+ *
+ * TODO: Make this more efficient.  It's not as bad as it looks (most
+ * files don't have any hardlinks and we don't do any work here for those),
+ * but it wouldn't be hard to do better.
+ *
+ * TODO: Work with dev/ino pairs here instead of just ino values.
+ */
+static int
+synthesize_ino_value(struct cpio *cpio, struct archive_entry *entry)
+{
+	int64_t ino = archive_entry_ino64(entry);
+	int ino_new;
+	size_t i;
+
+	/*
+	 * If no index number was given, don't assign one.  In
+	 * particular, this handles the end-of-archive marker
+	 * correctly by giving it a zero index value.  (This is also
+	 * why we start our synthetic index numbers with one below.)
+	 */
+	if (ino == 0)
+		return (0);
+
+	/* Don't store a mapping if we don't need to. */
+	if (archive_entry_nlink(entry) < 2) {
+		return ++cpio->ino_next;
+	}
+
+	/* Look up old ino; if we have it, this is a hardlink
+	 * and we reuse the same value. */
+	for (i = 0; i < cpio->ino_list_next; ++i) {
+		if (cpio->ino_list[i].old == ino)
+			return (cpio->ino_list[i].new);
+	}
+
+	/* Assign a new index number. */
+	ino_new = ++cpio->ino_next;
+
+	/* Ensure space for the new mapping. */
+	if (cpio->ino_list_size <= cpio->ino_list_next) {
+		size_t newsize = cpio->ino_list_size < 512
+		    ? 512 : cpio->ino_list_size * 2;
+		void *newlist = realloc(cpio->ino_list,
+		    sizeof(cpio->ino_list[0]) * newsize);
+		if (newlist == NULL)
+			return (-1);
+
+		cpio->ino_list_size = newsize;
+		cpio->ino_list = newlist;
+	}
+
+	/* Record and return the new value. */
+	cpio->ino_list[cpio->ino_list_next].old = ino;
+	cpio->ino_list[cpio->ino_list_next].new = ino_new;
+	++cpio->ino_list_next;
+	return (ino_new);
+}
+
 static int
 archive_write_cpio_header(struct archive_write *a, struct archive_entry *entry)
 {
@@ -121,19 +192,19 @@ archive_write_cpio_header(struct archive_write *a, struct archive_entry *entry)
 	memset(&h, 0, sizeof(h));
 	format_octal(070707, &h.c_magic, sizeof(h.c_magic));
 	format_octal(archive_entry_dev(entry), &h.c_dev, sizeof(h.c_dev));
-	/*
-	 * TODO: Generate artificial inode numbers rather than just
-	 * re-using the ones off the disk.  That way, the 18-bit c_ino
-	 * field only limits the number of files in the archive.
-	 */
-	ino = archive_entry_ino64(entry);
-	if (ino < 0 || ino > 0777777) {
+
+	ino = synthesize_ino_value(cpio, entry);
+	if (ino < 0) {
+		archive_set_error(&a->archive, ENOMEM,
+		    "No memory for ino translation table");
+		return (ARCHIVE_FATAL);
+	} else if (ino > 0777777) {
 		archive_set_error(&a->archive, ERANGE,
-		    "large inode number truncated");
-		ret2 = ARCHIVE_WARN;
+		    "Too many files for this cpio format");
+		return (ARCHIVE_FATAL);
 	}
+	format_octal(ino & 0777777, &h.c_ino, sizeof(h.c_ino));
 
-	format_octal(archive_entry_ino64(entry) & 0777777, &h.c_ino, sizeof(h.c_ino));
 	format_octal(archive_entry_mode(entry), &h.c_mode, sizeof(h.c_mode));
 	format_octal(archive_entry_uid(entry), &h.c_uid, sizeof(h.c_uid));
 	format_octal(archive_entry_gid(entry), &h.c_gid, sizeof(h.c_gid));
@@ -246,6 +317,7 @@ archive_write_cpio_destroy(struct archive_write *a)
 	struct cpio *cpio;
 
 	cpio = (struct cpio *)a->format_data;
+	free(cpio->ino_list);
 	free(cpio);
 	a->format_data = NULL;
 	return (ARCHIVE_OK);