From patchwork Sat Feb 27 15:30:12 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: John Wood X-Patchwork-Id: 12107895 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-13.7 required=3.0 tests=BAYES_00,DKIM_SIGNED, DKIM_VALID,FREEMAIL_FORGED_FROMDOMAIN,FREEMAIL_FROM, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,USER_AGENT_GIT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 05FBEC433E0 for ; Sat, 27 Feb 2021 18:03:03 +0000 (UTC) Received: from mother.openwall.net (mother.openwall.net [195.42.179.200]) by mail.kernel.org (Postfix) with SMTP id DD70064E20 for ; Sat, 27 Feb 2021 18:03:01 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org DD70064E20 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=gmx.com Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=kernel-hardening-return-20856-kernel-hardening=archiver.kernel.org@lists.openwall.com Received: (qmail 24034 invoked by uid 550); 27 Feb 2021 18:02:53 -0000 Mailing-List: contact kernel-hardening-help@lists.openwall.com; run by ezmlm Precedence: bulk List-Post: List-Help: List-Unsubscribe: List-Subscribe: List-ID: Received: (qmail 24002 invoked from network); 27 Feb 2021 18:02:53 -0000 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=gmx.net; s=badeba3b8450; t=1614448958; bh=BUrFkJRdMorIGyPHZ/sn3asqSQHApToumYQiv4PUyts=; h=X-UI-Sender-Class:From:To:Cc:Subject:Date:In-Reply-To:References; b=c0FOR6fPgbSK7rkVNiSmZ1Z5GAnqKXApxUUhTg9gNC/PnTWd9gdZpxJT9NOIEzwji CtwIM7QG6S1SHUQrpfa8zBRDYzakki+jYq8ztI6PElYh3llNXkjJfCeegLLzHsjieh FnaE6SPuRD6akUDTTZmwG1k/gBrKTiZVgoL0Vb6Y= X-UI-Sender-Class: 01bb95c1-4bf8-414a-932a-4f6e2808ef9c From: John Wood To: Kees Cook , Jann Horn , Randy Dunlap , Jonathan Corbet , James Morris , Shuah Khan Cc: John Wood , "Serge E. Hallyn" , Greg Kroah-Hartman , linux-doc@vger.kernel.org, linux-kernel@vger.kernel.org, linux-security-module@vger.kernel.org, linux-kselftest@vger.kernel.org, kernel-hardening@lists.openwall.com Subject: [PATCH v5 7/8] Documentation: Add documentation for the Brute LSM Date: Sat, 27 Feb 2021 16:30:12 +0100 Message-Id: <20210227153013.6747-8-john.wood@gmx.com> X-Mailer: git-send-email 2.25.1 In-Reply-To: <20210227153013.6747-1-john.wood@gmx.com> References: <20210227153013.6747-1-john.wood@gmx.com> MIME-Version: 1.0 X-Provags-ID: V03:K1:lgDBTFtAnDfrXI03PXITKePnX0sTM5BOZYhusvycVNJ9anmZQAq E6XNkOyHnOXbG/h0J1S0fdEDTnagZ455otTbtEb9+V84gaRrWdRcarSJ1TPpLHg47mXeVOm HU/OYKZdwKPVLAuSMpwAl7mlySnDq7/twkbgZeoA+MInl5IcifzHaws+RmBwGRIPbZROz81 9EPxlJ24Wk22cFiYm/8kQ== X-UI-Out-Filterresults: notjunk:1;V03:K0:9QeVEX4n87I=:53y8CTzNX4HW2oPzSES8ZS WNsNyzo8B7ARXI9hxOjDyQeaeUOt29cWYp0jwrlkBlfu0klk/xRhz/uJ9DFr9dSD31uZLdcbO ZmNElXOHeZ08ux+MZoD5Z91OkMMUegMZL5RyDbAoWmyekmI0sCAdj+w/nBJ4ncyd3jHxE9x/y zrglSeXRImn63dU2DiQxbXkQKDszmxUnmL+xcUTohC0FPyO4Kug1HaOUI4Ix9gq9xFXFZh/G8 +kJabTgHhtavMx0UqltdELBJ4TYL6cUSkka5hjjW15FIFDINqETLybfTMajbZxlFB5D/fCwMT JvPhkbrQCuURpNyrSLrUu+94A29Pcg1OL1+sW0PwWO9R507nOP4nA6Ln7vHaafNs0GO2bap5L fozEOR/9yQ+NIH1k9aNp9NEsC4JVbLXDJkqENwZ4B+aJxXud9DfXsay2i3Qnc52oB4VWMQBAX rwA3Dfg8kZ9OHKPJAigivD54HDKncujhpsEMwXXihANG8+t9iW/8sDEQrdZQd9OdY5zTQxnxa sT0/HSAzZFnhKT2dGI/zGSP/Wb0AeaK3Jk9n6Vu8em+073tW1nF+jBiHqDZqVw9yqzyZrRHwU ywWe6j+WIgOXQB6tGG5+IIGHRE/Eh+iCUskq1shc8TwkW8vy6Ve8yv4LOvG7IHy6VSZZHXYaw uWwklSXU/oxixowL4/GGrzzRfniui8qIfSMH7Y6zT14i1vJ/pF0ctFK29rqjPV6Q49xziQljL ak98beAkvqlyi9xlchTW1zeOCPuQ99mp+lNxmi7rxWC3jLUVh5VjIydYevWwjphReEOhlSKn+ x+BNwS/su9TlPh2ArbEOy2s39T+au1pDa2J/zveArlhZC52b+KjrTyAFGuivJJG92pOuxs2Og DslPyd93rQmZ76lDA4Ig== Add some info detailing what is the Brute LSM, its motivation, weak points of existing implementations, proposed solutions, enabling, disabling and self-tests. Signed-off-by: John Wood --- Documentation/admin-guide/LSM/Brute.rst | 224 ++++++++++++++++++++++++ Documentation/admin-guide/LSM/index.rst | 1 + security/brute/Kconfig | 3 +- 3 files changed, 227 insertions(+), 1 deletion(-) create mode 100644 Documentation/admin-guide/LSM/Brute.rst -- 2.25.1 diff --git a/Documentation/admin-guide/LSM/Brute.rst b/Documentation/admin-guide/LSM/Brute.rst new file mode 100644 index 000000000000..485966a610bb --- /dev/null +++ b/Documentation/admin-guide/LSM/Brute.rst @@ -0,0 +1,224 @@ +.. SPDX-License-Identifier: GPL-2.0 +=========================================================== +Brute: Fork brute force attack detection and mitigation LSM +=========================================================== + +Attacks against vulnerable userspace applications with the purpose to break ASLR +or bypass canaries traditionally use some level of brute force with the help of +the fork system call. This is possible since when creating a new process using +fork its memory contents are the same as those of the parent process (the +process that called the fork system call). So, the attacker can test the memory +infinite times to find the correct memory values or the correct memory addresses +without worrying about crashing the application. + +Based on the above scenario it would be nice to have this detected and +mitigated, and this is the goal of this implementation. Specifically the +following attacks are expected to be detected: + +1.- Launching (fork()/exec()) a setuid/setgid process repeatedly until a + desirable memory layout is got (e.g. Stack Clash). +2.- Connecting to an exec()ing network daemon (e.g. xinetd) repeatedly until a + desirable memory layout is got (e.g. what CTFs do for simple network + service). +3.- Launching processes without exec() (e.g. Android Zygote) and exposing state + to attack a sibling. +4.- Connecting to a fork()ing network daemon (e.g. apache) repeatedly until the + previously shared memory layout of all the other children is exposed (e.g. + kind of related to HeartBleed). + +In each case, a privilege boundary has been crossed: + +Case 1: setuid/setgid process +Case 2: network to local +Case 3: privilege changes +Case 4: network to local + +So, what really needs to be detected are fork/exec brute force attacks that +cross any of the commented bounds. + + +Other implementations +===================== + +The public version of grsecurity, as a summary, is based on the idea of delaying +the fork system call if a child died due to some fatal signal (SIGSEGV, SIGBUS, +SIGKILL or SIGILL). This has some issues: + +Bad practices +------------- + +Adding delays to the kernel is, in general, a bad idea. + +Scenarios not detected (false negatives) +---------------------------------------- + +This protection acts only when the fork system call is called after a child has +crashed. So, it would still be possible for an attacker to fork a big amount of +children (in the order of thousands), then probe all of them, and finally wait +the protection time before repeating the steps. + +Moreover, this method is based on the idea that the protection doesn't act if +the parent crashes. So, it would still be possible for an attacker to fork a +process and probe itself. Then, fork the child process and probe itself again. +This way, these steps can be repeated infinite times without any mitigation. + +Scenarios detected (false positives) +------------------------------------ + +Scenarios where an application rarely fails for reasons unrelated to a real +attack. + + +This implementation +=================== + +The main idea behind this implementation is to improve the existing ones +focusing on the weak points annotated before. Basically, the adopted solution is +to detect a fast crash rate instead of only one simple crash and to detect both +the crash of parent and child processes. Also, fine tune the detection focusing +on privilege boundary crossing. And finally, as a mitigation method, kill all +the offending tasks involved in the attack instead of using delays. + +To achieve this goal, and going into more details, this implementation is based +on the use of some statistical data shared across all the processes that can +have the same memory contents. Or in other words, a statistical data shared +between all the fork hierarchy processes after an execve system call. + +The purpose of these statistics is, basically, collect all the necessary info +to compute the application crash period in order to detect an attack. This crash +period is the time between the execve system call and the first fault or the +time between two consecutive faults, but this has a drawback. If an application +crashes twice in a short period of time for some reason unrelated to a real +attack, a false positive will be triggered. To avoid this scenario the +exponential moving average (EMA) is used. This way, the application crash period +will be a value that is not prone to change due to spurious data and follows the +real crash period. + +To detect a brute force attack it is necessary that the statistics shared by all +the fork hierarchy processes be updated in every fatal crash and the most +important data to update is the application crash period. + +There are two types of brute force attacks that need to be detected. The first +one is an attack that happens through the fork system call and the second one is +an attack that happens through the execve system call. The first type uses the +statistics shared by all the fork hierarchy processes, but the second type +cannot use this statistical data due to these statistics dissapear when the +involved tasks finished. In this last scenario the attack info should be tracked +by the statistics of a higher fork hierarchy (the hierarchy that contains the +process that forks before the execve system call). + +Moreover, these two attack types have two variants. A slow brute force attack +that is detected if a maximum number of faults per fork hierarchy is reached and +a fast brute force attack that is detected if the application crash period falls +below a certain threshold. + +Exponential moving average (EMA) +-------------------------------- + +This kind of average defines a weight (between 0 and 1) for the new value to add +and applies the remainder of the weight to the current average value. This way, +some spurious data will not excessively modify the average and only if the new +values are persistent, the moving average will tend towards them. + +Mathematically the application crash period's EMA can be expressed as follows: + +period_ema = period * weight + period_ema * (1 - weight) + +Related to the attack detection, the EMA must guarantee that not many crashes +are needed. To demonstrate this, the scenario where an application has been +running without any crashes for a month will be used. + +The period's EMA can be written now as: + +period_ema[i] = period[i] * weight + period_ema[i - 1] * (1 - weight) + +If the new crash periods have insignificant values related to the first crash +period (a month in this case), the formula can be rewritten as: + +period_ema[i] = period_ema[i - 1] * (1 - weight) + +And by extension: + +period_ema[i - 1] = period_ema[i - 2] * (1 - weight) +period_ema[i - 2] = period_ema[i - 3] * (1 - weight) +period_ema[i - 3] = period_ema[i - 4] * (1 - weight) + +So, if the substitution is made: + +period_ema[i] = period_ema[i - 1] * (1 - weight) +period_ema[i] = period_ema[i - 2] * pow((1 - weight) , 2) +period_ema[i] = period_ema[i - 3] * pow((1 - weight) , 3) +period_ema[i] = period_ema[i - 4] * pow((1 - weight) , 4) + +And in a more generic form: + +period_ema[i] = period_ema[i - n] * pow((1 - weight) , n) + +Where n represents the number of iterations to obtain an EMA value. Or in other +words, the number of crashes to detect an attack. + +So, if we isolate the number of crashes: + +period_ema[i] / period_ema[i - n] = pow((1 - weight), n) +log(period_ema[i] / period_ema[i - n]) = log(pow((1 - weight), n)) +log(period_ema[i] / period_ema[i - n]) = n * log(1 - weight) +n = log(period_ema[i] / period_ema[i - n]) / log(1 - weight) + +Then, in the commented scenario (an application has been running without any +crashes for a month), the approximate number of crashes to detect an attack +(using the implementation values for the weight and the crash period threshold) +is: + +weight = 7 / 10 +crash_period_threshold = 30 seconds + +n = log(crash_period_threshold / seconds_per_month) / log(1 - weight) +n = log(30 / (30 * 24 * 3600)) / log(1 - 0.7) +n = 9.44 + +So, with 10 crashes for this scenario an attack will be detected. If these steps +are repeated for different scenarios and the results are collected: + +1 month without any crashes ----> 9.44 crashes to detect an attack +1 year without any crashes -----> 11.50 crashes to detect an attack +10 years without any crashes ---> 13.42 crashes to detect an attack + +However, this computation has a drawback. The first data added to the EMA not +obtains a real average showing a trend. So the solution is simple, the EMA needs +a minimum number of data to be able to be interpreted. This way, the case where +a few first faults are fast enough followed by no crashes is avoided. + +Per system enabling/disabling +----------------------------- + +This feature can be enabled at build time using the CONFIG_SECURITY_FORK_BRUTE +option or using the visual config application under the following menu: + +Security options ---> Fork brute force attack detection and mitigation + +Also, at boot time, this feature can be disable too, by changing the "lsm=" boot +parameter. + +Kernel selftests +---------------- + +To validate all the expectations about this implementation, there is a set of +selftests. This tests cover fork/exec brute force attacks crossing the following +privilege boundaries: + +1.- setuid process +2.- privilege changes +3.- network to local + +Also, there are some tests to check that fork/exec brute force attacks without +crossing any privilege boundariy already commented doesn't trigger the detection +and mitigation stage. + +To build the tests: +make -C tools/testing/selftests/ TARGETS=brute + +To run the tests: +make -C tools/testing/selftests TARGETS=brute run_tests + +To package the tests: +make -C tools/testing/selftests TARGETS=brute gen_tar diff --git a/Documentation/admin-guide/LSM/index.rst b/Documentation/admin-guide/LSM/index.rst index a6ba95fbaa9f..1f68982bb330 100644 --- a/Documentation/admin-guide/LSM/index.rst +++ b/Documentation/admin-guide/LSM/index.rst @@ -41,6 +41,7 @@ subdirectories. :maxdepth: 1 apparmor + Brute LoadPin SELinux Smack diff --git a/security/brute/Kconfig b/security/brute/Kconfig index 1bd2df1e2dec..334d7e88d27f 100644 --- a/security/brute/Kconfig +++ b/security/brute/Kconfig @@ -7,6 +7,7 @@ config SECURITY_FORK_BRUTE vulnerable userspace processes. The detection method is based on the application crash period and as a mitigation procedure all the offending tasks are killed. Like capabilities, this security module - stacks with other LSMs. + stacks with other LSMs. Further information can be found in + Documentation/admin-guide/LSM/Brute.rst. If you are unsure how to answer this question, answer N.