summaryrefslogtreecommitdiff
path: root/security/integrity/evm/evm_main.c
diff options
context:
space:
mode:
Diffstat (limited to 'security/integrity/evm/evm_main.c')
-rw-r--r--security/integrity/evm/evm_main.c17
1 files changed, 14 insertions, 3 deletions
diff --git a/security/integrity/evm/evm_main.c b/security/integrity/evm/evm_main.c
index 2e6fb6e2ffd2..23d484e05e6f 100644
--- a/security/integrity/evm/evm_main.c
+++ b/security/integrity/evm/evm_main.c
@@ -457,10 +457,21 @@ static int evm_xattr_acl_change(struct user_namespace *mnt_userns,
int rc;
/*
- * user_ns is not relevant here, ACL_USER/ACL_GROUP don't have impact
- * on the inode mode (see posix_acl_equiv_mode()).
+ * An earlier comment here mentioned that the idmappings for
+ * ACL_{GROUP,USER} don't matter since EVM is only interested in the
+ * mode stored as part of POSIX ACLs. Nonetheless, if it must translate
+ * from the uapi POSIX ACL representation to the VFS internal POSIX ACL
+ * representation it should do so correctly. There's no guarantee that
+ * we won't change POSIX ACLs in a way that ACL_{GROUP,USER} matters
+ * for the mode at some point and it's difficult to keep track of all
+ * the LSM and integrity modules and what they do to POSIX ACLs.
+ *
+ * Frankly, EVM shouldn't try to interpret the uapi struct for POSIX
+ * ACLs it received. It requires knowledge that only the VFS is
+ * guaranteed to have.
*/
- acl = posix_acl_from_xattr(&init_user_ns, xattr_value, xattr_value_len);
+ acl = vfs_set_acl_prepare(mnt_userns, i_user_ns(inode),
+ xattr_value, xattr_value_len);
if (IS_ERR_OR_NULL(acl))
return 1;