grsecurity 2.1.0 and kernel vulnerabilities

grsecurity 2.1.0 release / Linux Kernel advisories
--------------------------------------------------------------------

Table Of Contents:
1) grsecurity 2.1.0 announcement and changelog
2) Linux Kernel advisory introduction
3) 2.4/2.6 random poolsize sysctl handler integer overflow
4) 2.6 scsi ioctl integer overflow and information leak
5) 2.2/2.4/2.6 moxa serial driver bss overflow
6) 2.4/2.6 RLIMIT_MEMLOCK bypass and (2.6) unprivileged user DoS
7) Attachments, including patches for all vulns, a POC for #3, and a
   working exploit for #6

1) grsecurity 2.1.0 announcement and changelog

I'm happy to announce the release of grsecurity 2.1.0.  It is being
released initially for the 2.4.28 and 2.6.10 kernels and will be ported
immediately to the next kernel versions when released.  It can be
downloaded at http://www.grsecurity.net.  We are still actively seeking
sponsorship, so if you benefit from using grsecurity and like the
changes you see in 2.1.0, please consider sponsoring the future
development and maintenance of the project.
Changes in this release include:

* New configuration file for full learning: /etc/grsec/learn_config
* Learning heuristics have been optimized to better detect temporary
  file usage and reduce appropriately.
* Learning heuristics have been modified to weight against reducing
  certain additional important directories.
* User/group ID transitions have been added to the learning system.
  Any subject transitioning to less than 3 different users or 3
  different groups that has CAP_SETUID or CAP_SETGID will have ID
  transitions added.  This is useful to automatically secure
  applications that only transition to one or few users/groups like
  nobody/nogroup.
* /proc/<pid>/* accesses are automatically rewritten as /proc by grlearn
  before being cached and written to disk
* The inherit-based learning usable through the learning configuration
  file is usable through a regular policy as well simply by adding "i"
  instead of "l" to a subject for learning.
* Inheritance is preserved whenever possible across uid/gid changes when
  the role resulting from the uid/gid change is no different from that
  before the change.
* A complete ~95-99% efficient LFU-hash hybrid caching system has been
  added that greatly reduces the number of full object lookups by
  caching the result.  The cache essentially mimics the filesystem
  around where applications are operating: nearly equivalent to having
  an object for every file and directory on the system, but without the
  wasted memory.  The cache is invalidated on creates and deletes that
  cause a change in policy (through policy re-creation) and on renames
  of directories or symlinks.
* Memory usage for non-full learning has been significantly reduced and
  all memory leaks have been plugged.
* A new object mode has been added for hardlinks for more fine-grained
  permissions.  See the sample policy file for information on what
  permissions are required to create a hardlink.  Its corresponding
  audit flag has been added as well.
* Destruction of unused shared memory feature added and included in
  the sysctl framework of grsecurity.  This feature was ported from
  Openwall (http://www.openwall.com/linux).
* A new option was added to the sysctl feature that enables at boot all
  features enabled in the kernel configuration, while allowing them to
  be changed via the sysctl interface until grsec_lock is set.
* Policy statistics have been added to gradm that provide useful,
  security-relevant information on the policy you are loading into the
  kernel.  You can view these statistics when enabling the system by
  running gradm -V -E.
* Interactive performance of full-learning has improved by ~15% by
  reducing the number of context switches caused by grlearn doing small
  disk writes by using a write buffer (writing more once instead of
  less 1000 times) and keeping track of log entry lengths for quicker
  string matching.  A signal handler was added to grlearn so that when
  learning is stopped, the write buffer is flushed to disk.
* Kernel headers are no longer used for gradm
* Updates from the PaX project (http://pax.grsecurity.net)
* Bugfixes for things mentioned on the list, etc

When patching your kernel for the 2.4.28 and 2.6.10 kernels, since they
contain several vulnerabilities, make sure to also apply the secfix
patches located on the website.

2) Linux Kernel advisory introduction

Let me begin by giving you a timeline:

December 15th: I send Linus a mail with a subject line of
"RLIMIT_MEMLOCK bypass with locked stack"
December 27th: The PaX team sends Linus a mail with a subject line of
"2.6.9+ mlockall/expand_down DoS by unprivileged users"
January 2nd: The PaX team resends the previous mail to Linux and Andrew
Morton

Between December 15th and today, Linus has committed many changes to
the kernel.  Between January 2nd and today, Andrew Morton has committed
several changes to the kernel.  3 weeks is a sufficient amount of time
to be able to expect even a reply about a given vulnerability.  A patch
for the vulnerability was attached to the mails, and in the PaX team's
mails, a working exploit as well.  Private notification of
vulnerabilities is a privilege, and when that privilege is abused by not
responding promptly, it deserves to be revoked.

Using 'advanced static analysis': "cd drivers; grep copy_from_user -r ./* |
grep -v sizeof", I discovered 4 exploitable vulnerabilities in a matter
of 15 minutes.  More vulnerabilities were found in 2.6 than in 2.4.
It's a pretty sad state of affairs for Linux security when someone can
find 4 exploitable vulnerabilities in a matter of minutes.  Since there
was no point in sending more vulnerability reports when the first hadn't
even been responded to, I'm including all four of them in this mail, as
well as a POC for the poolsize bug.  The other bugs can have POCs 
written
for just as trivially.  The poolsize bug requires uid 0, but not any
root capabilities.  The scsi and serial bugs depend on the permissions
of their respective devices, and thus can possibly be exploited as
non-root.  The scsi bug in particular has a couple different attack
vectors that I haven't even bothered to investigate.  Some of these bugs
have gone unfixed for several years.

The PaX team discovered the mlockall DoS. It has been fixed in PaX for
2 years.  I have attached their mail and exploit code.

I'd really like to know what's being done about this pitiful trend of
Linux security, where it's 10x as easy to find a vulnerability in the
kernel than it is in any app on the system, where isec releases at
least one critical vulnerability for each kernel version.  I don't see
that the 2.6 development model is doing anything to help this (as the
spectrum of these vulnerabilities demonstrate), by throwing
experimental code into the kernel and claiming it to be "stable".
Hopefully now these vulnerabilities will be fixed in a timely manner.

3) 2.4/2.6 random poolsize sysctl handler integer overflow

In drivers/char/random.c:

at poolsize_strategy():
>        int     len;
        ^ signed integer
>
>        sysctl_poolsize = random_state->poolinfo.POOLBYTES;
>
>        /*
>
>        /*
>         * We only handle the write case, since the read case gets
>         * handled by the default handler (and we don't care if the
>         * write case happens twice; it's harmless).
>         */
>        if (newval && newlen) {
>                len = newlen;
                ^ unsigned int converted to signed
>                if (len > table->maxlen)
                ^ comparison of two signed integers
>                        len = table->maxlen;
>                if (copy_from_user(table->data, newval, len))
                ^ copy_from_user with len possibly > table->maxlen

4) 2.6 scsi ioctl integer overflow and information leak

In drivers/block/scsi_ioctl.c:

at sg_scsi_ioctl():
>        struct request *rq;
>        int err, in_len, out_len, bytes, opcode, cmdlen;
        ^ in_len, out_len are signed int
>        char *buffer = NULL, sense[SCSI_SENSE_BUFFERSIZE];
>
>        /*
>         * get in an out lengths, verify they don't exceed a page worth of
data
>         */
>        if (get_user(in_len, &sic->inlen))
        ^ in_len is user-controlled
>                return -EFAULT;
>        if (get_user(out_len, &sic->outlen))
        ^ out_len is user-controlled
>                return -EFAULT;
>        if (in_len > PAGE_SIZE || out_len > PAGE_SIZE)
        ^ signed int only has upper bound checked
>                return -EINVAL;
>        if (get_user(opcode, sic->data))
>                return -EFAULT;
>        bytes = max(in_len, out_len);
...
>        rq->cmd_len = cmdlen;
>        if (copy_from_user(rq->cmd, sic->data, cmdlen))
>                goto error;
>         
>        if (copy_from_user(buffer, sic->data + cmdlen, in_len))
                ^ copy_from_user with size possibly > PAGE_SIZE
>                goto error;
...
>                if (copy_to_user(sic->data, buffer, out_len))
                ^ copy_to_user with size possibly > PAGE_SIZE
>                        err = -EFAULT;

5) 2.2/2.4/2.6 moxa serial driver bss overflow

In drivers/char/moxa.c:

>static unsigned char moxaBuff[10240];

In MoxaDriverIoctl():

>        if(copy_from_user(&dltmp, argp, sizeof(struct dl_str)))
>                return -EFAULT;
                ^ dltmp.len is user-controlled
>        if(dltmp.cardno < 0 || dltmp.cardno >= MAX_BOARDS)
>                return -EINVAL;
>                
>        switch(cmd)
>        {
>        case MOXA_LOAD_BIOS:
>                i = moxaloadbios(dltmp.cardno, dltmp.buf, dltmp.len);
                ^ called with no length checking
>                return (i);
>        case MOXA_FIND_BOARD:
>                return moxafindcard(dltmp.cardno);
>        case MOXA_LOAD_C320B:
>                moxaload320b(dltmp.cardno, dltmp.buf, dltmp.len);
                ^ called with no length checking
>        default: /* to keep gcc happy */
>                return (0);
>        case MOXA_LOAD_CODE:
>                i = moxaloadcode(dltmp.cardno, dltmp.buf, dltmp.len);
                ^ called with no length checking

In moxaloadbios():

>static int moxaloadbios(int cardno, unsigned char __user *tmp, int len)
>{
>        void __iomem *baseAddr;
>        int i;
>
>        if(copy_from_user(moxaBuff, tmp, len))
                ^ copy_from_user with no length checking
>                return -EFAULT;

In moxaloadcode():

> static int moxaloadcode(int cardno, unsigned char __user *tmp, int len)
> {
>        void __iomem *baseAddr, *ofsAddr;
>        int retval, port, i;
>
>        if(copy_from_user(moxaBuff, tmp, len))
                ^ copy_from_user with no length checking
>                return -EFAULT;

In moxaload320b():

>static int moxaload320b(int cardno, unsigned char __user *tmp, int len)
>{
>        void __iomem *baseAddr;
>        int i;
>
>        if(len > sizeof(moxaBuff))
                ^ signed int has only upper-bound checked
>                return -EINVAL;
>        if(copy_from_user(moxaBuff, tmp, len))
                ^ copy_from_user with len possibly > sizeof(moxaBuff)
>                return -EFAULT;

6) 2.4/2.6 RLIMIT_MEMLOCK bypass and (2.6) unprivileged user DoS

Taken from the mail from the PaX team to Linus and Andrew Morton:

the 'culprit' patch is how the default RLIM_MEMLOCK and the privilege
to call mlockall have changed in 2.6.9. namely, the former has been
reduced to 32 pages while the latter has been relaxed to allow it for
otherwise unprivileged users if their RLIM_MEMLOCK is bigger than the
currently allocated vm. which is normally good enough, except as you
now know there's a path that can increase the allocated vm without
checking for RLIM_MEMLOCK.

i'm attaching a small i386-specific demonstration, use the makefile to
create the small self-contained executable, e.g., 'make alloc=0x100000'
to have it allocate 1MB of stack and lock all of it. for demonstrating
the full effect of locking down arbitrary amounts of memory, you'll have
to set your stack rlimit to infinity (ulimit -s unlimited) and allocate
as much memory as your memory overcommit policy allows (this may mean
that you'll have to run multiple instances, if you have lots of memory).

surprisingly, in my tests the kernel survived pretty well, it just crawled
to a snail's speed as every mapped page access required disk i/o ;-). i
didn't play with overcommit policies nor any special workloads, so there
may very well be worse effects with that much locked memory. in any case,
this may warrant 2.6.10.1 because as soon as the fix goes into -bk, anyone
reading the logs can easily figure it out and reproduce the 'exploit'.

the attached patch is the excerpt from PaX that survives the exploit, so
i think it's good to go.