| Message ID | 1518813234-5874-19-git-send-email-takondra@cisco.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, 2018-02-16 at 20:33 +0000, Taras Kondratiuk wrote: > From: Victor Kamensky <kamensky@cisco.com> > > With initramfs cpio format that supports extended attributes > we need to skip sid population on sys_lsetxattr call from > initramfs for rootfs if security server is not initialized yet. > > Otherwise callback in selinux_inode_post_setxattr will try to > translate give security.selinux label into sid context and since > security server is not available yet inode will receive default > sid (typically kernel_t). Note that in the same time proper > label will be stored in inode xattrs. Later, since inode sid > would be already populated system will never look back at > actual xattrs. But if we skip sid population for rootfs and > we have policy that direct use of xattrs for rootfs, proper > sid will be filled in from extended attributes one node is > accessed and server is initialized. > > Note new DELAYAFTERINIT_MNT super block flag is introduced > to only mark rootfs for such behavior. For other types of > tmpfs original logic is still used. (cc selinux maintainers) Wondering if we shouldn't just do this always, for all filesystem types. Also, I think this should likely also be done in selinux_inode_setsecurity() for consistency. > > Signed-off-by: Victor Kamensky <kamensky@cisco.com> > --- > security/selinux/hooks.c | 9 ++++++++- > security/selinux/include/security.h | 1 + > 2 files changed, 9 insertions(+), 1 deletion(-) > > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c > index f3fe65589f02..bb25268f734e 100644 > --- a/security/selinux/hooks.c > +++ b/security/selinux/hooks.c > @@ -716,7 +716,7 @@ static int selinux_set_mnt_opts(struct > super_block *sb, > */ > if (!strncmp(sb->s_type->name, "rootfs", > sizeof("rootfs"))) > - sbsec->flags |= SBLABEL_MNT; > + sbsec->flags |= > SBLABEL_MNT|DELAYAFTERINIT_MNT; > > /* Defer initialization until > selinux_complete_init, > after the initial policy is loaded and > the security > @@ -3253,6 +3253,7 @@ static void selinux_inode_post_setxattr(struct > dentry *dentry, const char *name, > { > struct inode *inode = d_backing_inode(dentry); > struct inode_security_struct *isec; > + struct superblock_security_struct *sbsec; > u32 newsid; > int rc; > > @@ -3261,6 +3262,12 @@ static void selinux_inode_post_setxattr(struct > dentry *dentry, const char *name, > return; > } > > + if (!ss_initialized) { > + sbsec = inode->i_sb->s_security; > + if (sbsec->flags & DELAYAFTERINIT_MNT) > + return; > + } > + > rc = security_context_to_sid_force(value, size, &newsid); > if (rc) { > printk(KERN_ERR "SELinux: unable to map context to > SID" > diff --git a/security/selinux/include/security.h > b/security/selinux/include/security.h > index 02f0412d42f2..585acfd6cbcf 100644 > --- a/security/selinux/include/security.h > +++ b/security/selinux/include/security.h > @@ -52,6 +52,7 @@ > #define ROOTCONTEXT_MNT 0x04 > #define DEFCONTEXT_MNT 0x08 > #define SBLABEL_MNT 0x10 > +#define DELAYAFTERINIT_MNT 0x20 > /* Non-mount related flags */ > #define SE_SBINITIALIZED 0x0100 > #define SE_SBPROC 0x0200 -- To unsubscribe from this list: send the line "unsubscribe linux-security-module" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 02/20/2018 12:56 PM, Stephen Smalley wrote: > On Fri, 2018-02-16 at 20:33 +0000, Taras Kondratiuk wrote: >> From: Victor Kamensky <kamensky@cisco.com> >> >> With initramfs cpio format that supports extended attributes >> we need to skip sid population on sys_lsetxattr call from >> initramfs for rootfs if security server is not initialized yet. >> >> Otherwise callback in selinux_inode_post_setxattr will try to >> translate give security.selinux label into sid context and since >> security server is not available yet inode will receive default >> sid (typically kernel_t). Note that in the same time proper >> label will be stored in inode xattrs. Later, since inode sid >> would be already populated system will never look back at >> actual xattrs. But if we skip sid population for rootfs and >> we have policy that direct use of xattrs for rootfs, proper >> sid will be filled in from extended attributes one node is >> accessed and server is initialized. >> >> Note new DELAYAFTERINIT_MNT super block flag is introduced >> to only mark rootfs for such behavior. For other types of >> tmpfs original logic is still used. > > (cc selinux maintainers) > > Wondering if we shouldn't just do this always, for all filesystem > types. Also, I think this should likely also be done in > selinux_inode_setsecurity() for consistency. I don't understand what selinux thinks it's doing here. Initramfs is special because it's populated early, ideally early enough drivers can load their firmware out of it. This is guaranteed to be before any processes have launched, before any other filesystems have been mounted. I'm surprised selinux is trying to do anything this early because A) what is there for it to do, B) where did it get a ruleset? This isn't really a mount flag, this is a "the selinux subsystem isn't functionally initialized yet" flag. We haven't launched init. In a modular system the module probably isn't loaded. There are no processes, and the only files anywhere are the ones we're in the process of extracting. What's there fore selinux to do? When a filesystem is mounted, none of these cached selinux "we already looked at the xattrs" inode fields are populated yet, correct? It can figure that out when something accesses the file and do it then, so the point is _not_ doing this now and thus not cacheing the wrong info. That's what the mount flag is doing, telling selinux "not yet". So why does selinux not already _know_ "not yet"? Why doesn't load_policy flush the cache of the old default contexts? What happens if you mount an ext2 root and then init reads a dozen files before it gets to the load_policy? Do those doesn't files have bad default contexts forever now? Where does the selinux ruleset come from during the cpio extract? Was it hardwired into the driver? It certainly didn't come out of a file, and it wasn't a process that loaded it. Why is selinux trying to evaluate and cache the security context of files before it has any rules? (It has xattr annotations, but they have no _meaning_ without rules...? Confused, Rob -- To unsubscribe from this list: send the line "unsubscribe linux-security-module" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, 7 Mar 2018, Rob Landley wrote: > On 02/20/2018 12:56 PM, Stephen Smalley wrote: >> On Fri, 2018-02-16 at 20:33 +0000, Taras Kondratiuk wrote: >>> From: Victor Kamensky <kamensky@cisco.com> >>> >>> With initramfs cpio format that supports extended attributes >>> we need to skip sid population on sys_lsetxattr call from >>> initramfs for rootfs if security server is not initialized yet. >>> >>> Otherwise callback in selinux_inode_post_setxattr will try to >>> translate give security.selinux label into sid context and since >>> security server is not available yet inode will receive default >>> sid (typically kernel_t). Note that in the same time proper >>> label will be stored in inode xattrs. Later, since inode sid >>> would be already populated system will never look back at >>> actual xattrs. But if we skip sid population for rootfs and >>> we have policy that direct use of xattrs for rootfs, proper >>> sid will be filled in from extended attributes one node is >>> accessed and server is initialized. >>> >>> Note new DELAYAFTERINIT_MNT super block flag is introduced >>> to only mark rootfs for such behavior. For other types of >>> tmpfs original logic is still used. >> >> (cc selinux maintainers) >> >> Wondering if we shouldn't just do this always, for all filesystem >> types. Also, I think this should likely also be done in >> selinux_inode_setsecurity() for consistency. Sorry, I did not have time to try out Stephen's suggestion, especially given that core initramfs xattrs acceptance and dicussion looks a bit stalled, and for my use case it is dependency before SELinux changes. I will look for both suggestion this week. Hope to see initramfs xattrs patch series review going again. > I don't understand what selinux thinks it's doing here. > > Initramfs is special because it's populated early, ideally early enough drivers > can load their firmware out of it. This is guaranteed to be before any processes > have launched, before any other filesystems have been mounted. I'm surprised > selinux is trying to do anything this early because A) what is there for it to > do, B) where did it get a ruleset? > > This isn't really a mount flag, this is a "the selinux subsystem isn't > functionally initialized yet" flag. We haven't launched init. In a modular > system the module probably isn't loaded. There are no processes, and the only > files anywhere are the ones we're in the process of extracting. What's there > fore selinux to do? > > When a filesystem is mounted, none of these cached selinux "we already looked at > the xattrs" inode fields are populated yet, correct? It can figure that out when > something accesses the file and do it then, so the point is _not_ doing this now > and thus not cacheing the wrong info. That's what the mount flag is doing, > telling selinux "not yet". So why does selinux not already _know_ "not yet"? > > Why doesn't load_policy flush the cache of the old default contexts? What > happens if you mount an ext2 root and then init reads a dozen files before it > gets to the load_policy? I need to check whether security context caching happens on all file operations, or when setxattr is executed. If latter, setxattr operation before policy load may not be very common use case. Also note there is a second SELinux related patch and corresponding Stephen's comment: if SELinux is enabled in kernel, but policy is not loaded yet, setxattr for security.selinux extended attribute will go for check to SELinux LSM callback, it will be denied. My other patch was relaxing above for "rootfs" only, i.e covering initramfs xattrs case. Stephen's point was that maybe it needs to be relaxed for all cases if policy not loaded yet. I need some time to look at the code and think about what can go wrong, if rule relaxed for all cases. > Do those doesn't files have bad default contexts forever now? > > Where does the selinux ruleset come from during the cpio extract? Yes, in our use case SELinux policy file (/etc/selinux/*/policy/policy.*) comes from cpio initramfs itself. So it is chicken and egg problem. > Was it > hardwired into the driver? It certainly didn't come out of a file, and it wasn't > a process that loaded it. Why is selinux trying to evaluate and cache the > security context of files before it has any rules? Note for ext2 case there is no setxattr first as we have in initramfs xattrs case, so extended attributes values are read from backing persitent storage as they were put there before, and there would not be discrepency what is cached in security context inode data scructure and real "security.selinux" extended attribute value in file system. Thanks, Victor > (It has xattr annotations, > but they have no _meaning_ without rules...? > > Confused, > > Rob > -- To unsubscribe from this list: send the line "unsubscribe linux-security-module" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c index f3fe65589f02..bb25268f734e 100644 --- a/security/selinux/hooks.c +++ b/security/selinux/hooks.c @@ -716,7 +716,7 @@ static int selinux_set_mnt_opts(struct super_block *sb, */ if (!strncmp(sb->s_type->name, "rootfs", sizeof("rootfs"))) - sbsec->flags |= SBLABEL_MNT; + sbsec->flags |= SBLABEL_MNT|DELAYAFTERINIT_MNT; /* Defer initialization until selinux_complete_init, after the initial policy is loaded and the security @@ -3253,6 +3253,7 @@ static void selinux_inode_post_setxattr(struct dentry *dentry, const char *name, { struct inode *inode = d_backing_inode(dentry); struct inode_security_struct *isec; + struct superblock_security_struct *sbsec; u32 newsid; int rc; @@ -3261,6 +3262,12 @@ static void selinux_inode_post_setxattr(struct dentry *dentry, const char *name, return; } + if (!ss_initialized) { + sbsec = inode->i_sb->s_security; + if (sbsec->flags & DELAYAFTERINIT_MNT) + return; + } + rc = security_context_to_sid_force(value, size, &newsid); if (rc) { printk(KERN_ERR "SELinux: unable to map context to SID" diff --git a/security/selinux/include/security.h b/security/selinux/include/security.h index 02f0412d42f2..585acfd6cbcf 100644 --- a/security/selinux/include/security.h +++ b/security/selinux/include/security.h @@ -52,6 +52,7 @@ #define ROOTCONTEXT_MNT 0x04 #define DEFCONTEXT_MNT 0x08 #define SBLABEL_MNT 0x10 +#define DELAYAFTERINIT_MNT 0x20 /* Non-mount related flags */ #define SE_SBINITIALIZED 0x0100 #define SE_SBPROC 0x0200