grsecurity 2.1.0 release / 5 Linux kernel advisories

From: Brad Spengler (
Date: 01/07/05

  • Next message: Paul Starzetz: "Linux kernel sys_uselib local root vulnerability"
    Date: Fri, 7 Jan 2005 13:18:53 -0500

    Let's try this again, since web archives don't like multipart

    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 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
    * /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 (
    * 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 (
    * 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

    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
    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 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.

    7) Attachments

    expoits_and_patches.tgz can be downloaded at:


  • Next message: Paul Starzetz: "Linux kernel sys_uselib local root vulnerability"