Skip to contentSkip to conversation

Filter out erroneous POSIX ACLs

PATCH In review [email protected] 2 participants 4 messages Feb 24, 2017

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Filter out erroneous POSIX ACLs

From: Anton Gerasimov <[email protected]>

The difference between what we see in pseudo and what happens without pseudo can be seen by typing:

mkdir setfattr-test
setfattr -n system.posix_acl_default -v 0x02000000 setfattr-test getfattr -n system.posix_acl_default setfattr-test

Under some kernel configurations this difference leads to annoying errors, e.g. directories copied with 'cp -a' get broken in a fancy way.

---
 ports/linux/xattr/pseudo_wrappers.c | 23 +++++++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)
ports/linux/xattr/pseudo_wrappers.c +21 -2
index 46bc053..31a6baf 100644
@@ -64,7 +64,7 @@ posix_permissions(const acl_header *header, int entries, int *extra, int *mode)
 	if (le32(header->version) != 2) {
 		pseudo_diag("Fatal: ACL support no available for header version %d.\n",
 			le32(header->version));
-		return 1;
+		return -1;
 	}
 	*mode = 0;
 	*extra = 0;
@@ -140,12 +140,27 @@ static int shared_setxattr(const char *path, int fd, const char *name, const voi
 	pseudo_debug(PDBGF_XATTR, "setxattr(%s [fd %d], %s => '%.*s')\n",
 		path ? path : "<no path>", fd, name, (int) size, (char *) value);
 
+	/* Filter out erroneous sizes for POSIX ACL
+	 *  see posix_acl_xattr_count in include/linux/posix_acl_xattr.h of Linux source code */
+	if (!strcmp(name, "system.posix_acl_access") || !strcmp(name, "system.posix_acl_default")) {
+		// ACL is corrupt, issue an error
+		if(size < sizeof(acl_header) || (size - sizeof(acl_header)) % sizeof(acl_entry) != 0) {
+			errno = EINVAL;
+			return -1;
+		}
+
+		// ACL is empty, do nothing
+		if((size - sizeof(acl_header)) / sizeof(acl_entry) == 0) {
+			return 0;
+		}
+	}
 	/* this may be a plain chmod */
 	if (!strcmp(name, "system.posix_acl_access")) {
 		int extra;
 		int mode;
 		int entries = (size - sizeof(acl_header)) / sizeof(acl_entry);
-		if (!posix_permissions(value, entries, &extra, &mode)) {
+		int res = posix_permissions(value, entries, &extra, &mode);
+		if (res == 0) {
 			pseudo_debug(PDBGF_XATTR, "posix_acl_access translated to mode %04o. Remaining attribute(s): %d.\n",
 				mode, extra);
 			buf.st_mode = mode;
@@ -164,8 +179,12 @@ static int shared_setxattr(const char *path, int fd, const char *name, const voi
 			if (!extra) {
 				return 0;
 			}
+		} else if (res == -1) {
+			errno = EOPNOTSUPP;
+			return -1;
 		}
 	}
+
 	if (!strcmp(name, "user.pseudo_data")) {
 		pseudo_debug(PDBGF_XATTR | PDBGF_XATTRDB, "user.pseudo_data xattribute does not get to go in database.\n");
 		return -1;
Signature · 1 line
2.11.1
#1 lore raw
Quoted 5 lines · On Fri, 24 Feb 2017 15:48:15 +0100 [email protected] wrote:

From: Anton Gerasimov <[email protected]>

The difference between what we see in pseudo and what happens without pseudo can be seen by typing:

Quoted 7 lines

mkdir setfattr-test
setfattr -n system.posix_acl_default -v 0x02000000 setfattr-test getfattr -n system.posix_acl_default setfattr-test

Under some kernel configurations this difference leads to annoying errors, e.g. directories copied with 'cp -a' get broken in a fancy way.

Can you expand on this? What is happening to make the "erroneous" ACLs? What kernel configurations?

-s

#2 lore raw

Hi Peter,

to reproduce the difference, you just need a kernel with POSIX ACL enabled for your FS (e.g. CONFIG_FS_POSIX_ACL=y; CONFIG_EXT4_FS_POSIX_ACL=y). The problem is that 'cp -a' when copying a directory without any ACLs attached will for whatever reason try to assign an empty posix_acl_default list to the target directory. Linux kernel will always filter it out, but pseudo just creates empty list. Later in the build process, when bootable image is created, it will instantiate a directory with empty posix_acl_default list on the bootable media. When you will try to perform certain operations on such directory (e.g. creating another directory inside it), kernel, namely ext4_get_acl, if I resemble correctly, will return EINVAL error, so the resulting filesystem will be effectively broken in many ways.

There is an issue [1] in Automotive Grade Linux Jira with an example of erroneous behaviour. The tricky part of it is that the error itself can not be reproduced easily: although I have one kernel config of a machine that makes broken images, I couldn't repeat it on later kernel versions. But the setfattr/getfattr behaviour difference that causes that error can be reproduced quite easily on any kernel with ACLs enabled.

[1] (Starting with Jan-Simon Moeller's comment on 26th of October)
https://jira.automotivelinux.org/browse/SPEC-254?focusedCommentId=11313&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-11313

Quoted 19 lines · Seebs
Quoted 5 lines · On Fri, 24 Feb 2017 15:48:15 +0100 [email protected] wrote:

From: Anton Gerasimov <[email protected]>

The difference between what we see in pseudo and what happens without pseudo can be seen by typing:

Quoted 7 lines

mkdir setfattr-test
setfattr -n system.posix_acl_default -v 0x02000000 setfattr-test getfattr -n system.posix_acl_default setfattr-test

Under some kernel configurations this difference leads to annoying errors, e.g. directories copied with 'cp -a' get broken in a fancy way.

Can you expand on this? What is happening to make the "erroneous" ACLs? What kernel configurations?

-s

Signature · 4 lines
Anton Gerasimov, ATS Advanced Telematic Systems GmbH
Kantstrasse 162, 10623 Berlin
Managing Directors: Dirk Pöschl, Armin G. Schmidt
Register Court: HRB 151501 B, Amtsgericht Charlottenburg
#3 lore raw
Quoted 4 lines · On Fri, 24 Feb 2017 16:18:38 +0100 Anton Gerasimov…

FIG_FS_POSIX_ACL=y;
CONFIG_EXT4_FS_POSIX_ACL=y). The problem is that 'cp -a' when copying a directory without any ACLs attached will for whatever reason try to assign an empty posix_acl_default list to the target directory.

...

It had totally failed to occur to me to check for *empty* lists. I'll want to re-read the code just in case, but that makes sense and the patch looked reasonable.

-s