1L1TF - L1 Terminal Fault
2========================
3
4L1 Terminal Fault is a hardware vulnerability which allows unprivileged
5speculative access to data which is available in the Level 1 Data Cache
6when the page table entry controlling the virtual address, which is used
7for the access, has the Present bit cleared or other reserved bits set.
8
9Affected processors
10-------------------
11
12This vulnerability affects a wide range of Intel processors. The
13vulnerability is not present on:
14
15   - Processors from AMD, Centaur and other non Intel vendors
16
17   - Older processor models, where the CPU family is < 6
18
19   - A range of Intel ATOM processors (Cedarview, Cloverview, Lincroft,
20     Penwell, Pineview, Silvermont, Airmont, Merrifield)
21
22   - The Intel XEON PHI family
23
24   - Intel processors which have the ARCH_CAP_RDCL_NO bit set in the
25     IA32_ARCH_CAPABILITIES MSR. If the bit is set the CPU is not affected
26     by the Meltdown vulnerability either. These CPUs should become
27     available by end of 2018.
28
29Whether a processor is affected or not can be read out from the L1TF
30vulnerability file in sysfs. See :ref:`l1tf_sys_info`.
31
32Related CVEs
33------------
34
35The following CVE entries are related to the L1TF vulnerability:
36
37   =============  =================  ==============================
38   CVE-2018-3615  L1 Terminal Fault  SGX related aspects
39   CVE-2018-3620  L1 Terminal Fault  OS, SMM related aspects
40   CVE-2018-3646  L1 Terminal Fault  Virtualization related aspects
41   =============  =================  ==============================
42
43Problem
44-------
45
46If an instruction accesses a virtual address for which the relevant page
47table entry (PTE) has the Present bit cleared or other reserved bits set,
48then speculative execution ignores the invalid PTE and loads the referenced
49data if it is present in the Level 1 Data Cache, as if the page referenced
50by the address bits in the PTE was still present and accessible.
51
52While this is a purely speculative mechanism and the instruction will raise
53a page fault when it is retired eventually, the pure act of loading the
54data and making it available to other speculative instructions opens up the
55opportunity for side channel attacks to unprivileged malicious code,
56similar to the Meltdown attack.
57
58While Meltdown breaks the user space to kernel space protection, L1TF
59allows to attack any physical memory address in the system and the attack
60works across all protection domains. It allows an attack of SGX and also
61works from inside virtual machines because the speculation bypasses the
62extended page table (EPT) protection mechanism.
63
64
65Attack scenarios
66----------------
67
681. Malicious user space
69^^^^^^^^^^^^^^^^^^^^^^^
70
71   Operating Systems store arbitrary information in the address bits of a
72   PTE which is marked non present. This allows a malicious user space
73   application to attack the physical memory to which these PTEs resolve.
74   In some cases user-space can maliciously influence the information
75   encoded in the address bits of the PTE, thus making attacks more
76   deterministic and more practical.
77
78   The Linux kernel contains a mitigation for this attack vector, PTE
79   inversion, which is permanently enabled and has no performance
80   impact. The kernel ensures that the address bits of PTEs, which are not
81   marked present, never point to cacheable physical memory space.
82
83   A system with an up to date kernel is protected against attacks from
84   malicious user space applications.
85
862. Malicious guest in a virtual machine
87^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
88
89   The fact that L1TF breaks all domain protections allows malicious guest
90   OSes, which can control the PTEs directly, and malicious guest user
91   space applications, which run on an unprotected guest kernel lacking the
92   PTE inversion mitigation for L1TF, to attack physical host memory.
93
94   A special aspect of L1TF in the context of virtualization is symmetric
95   multi threading (SMT). The Intel implementation of SMT is called
96   HyperThreading. The fact that Hyperthreads on the affected processors
97   share the L1 Data Cache (L1D) is important for this. As the flaw allows
98   only to attack data which is present in L1D, a malicious guest running
99   on one Hyperthread can attack the data which is brought into the L1D by
100   the context which runs on the sibling Hyperthread of the same physical
101   core. This context can be host OS, host user space or a different guest.
102
103   If the processor does not support Extended Page Tables, the attack is
104   only possible, when the hypervisor does not sanitize the content of the
105   effective (shadow) page tables.
106
107   While solutions exist to mitigate these attack vectors fully, these
108   mitigations are not enabled by default in the Linux kernel because they
109   can affect performance significantly. The kernel provides several
110   mechanisms which can be utilized to address the problem depending on the
111   deployment scenario. The mitigations, their protection scope and impact
112   are described in the next sections.
113
114   The default mitigations and the rationale for choosing them are explained
115   at the end of this document. See :ref:`default_mitigations`.
116
117.. _l1tf_sys_info:
118
119L1TF system information
120-----------------------
121
122The Linux kernel provides a sysfs interface to enumerate the current L1TF
123status of the system: whether the system is vulnerable, and which
124mitigations are active. The relevant sysfs file is:
125
126/sys/devices/system/cpu/vulnerabilities/l1tf
127
128The possible values in this file are:
129
130  ===========================   ===============================
131  'Not affected'		The processor is not vulnerable
132  'Mitigation: PTE Inversion'	The host protection is active
133  ===========================   ===============================
134
135If KVM/VMX is enabled and the processor is vulnerable then the following
136information is appended to the 'Mitigation: PTE Inversion' part:
137
138  - SMT status:
139
140    =====================  ================
141    'VMX: SMT vulnerable'  SMT is enabled
142    'VMX: SMT disabled'    SMT is disabled
143    =====================  ================
144
145  - L1D Flush mode:
146
147    ================================  ====================================
148    'L1D vulnerable'		      L1D flushing is disabled
149
150    'L1D conditional cache flushes'   L1D flush is conditionally enabled
151
152    'L1D cache flushes'		      L1D flush is unconditionally enabled
153    ================================  ====================================
154
155The resulting grade of protection is discussed in the following sections.
156
157
158Host mitigation mechanism
159-------------------------
160
161The kernel is unconditionally protected against L1TF attacks from malicious
162user space running on the host.
163
164
165Guest mitigation mechanisms
166---------------------------
167
168.. _l1d_flush:
169
1701. L1D flush on VMENTER
171^^^^^^^^^^^^^^^^^^^^^^^
172
173   To make sure that a guest cannot attack data which is present in the L1D
174   the hypervisor flushes the L1D before entering the guest.
175
176   Flushing the L1D evicts not only the data which should not be accessed
177   by a potentially malicious guest, it also flushes the guest
178   data. Flushing the L1D has a performance impact as the processor has to
179   bring the flushed guest data back into the L1D. Depending on the
180   frequency of VMEXIT/VMENTER and the type of computations in the guest
181   performance degradation in the range of 1% to 50% has been observed. For
182   scenarios where guest VMEXIT/VMENTER are rare the performance impact is
183   minimal. Virtio and mechanisms like posted interrupts are designed to
184   confine the VMEXITs to a bare minimum, but specific configurations and
185   application scenarios might still suffer from a high VMEXIT rate.
186
187   The kernel provides two L1D flush modes:
188    - conditional ('cond')
189    - unconditional ('always')
190
191   The conditional mode avoids L1D flushing after VMEXITs which execute
192   only audited code paths before the corresponding VMENTER. These code
193   paths have been verified that they cannot expose secrets or other
194   interesting data to an attacker, but they can leak information about the
195   address space layout of the hypervisor.
196
197   Unconditional mode flushes L1D on all VMENTER invocations and provides
198   maximum protection. It has a higher overhead than the conditional
199   mode. The overhead cannot be quantified correctly as it depends on the
200   workload scenario and the resulting number of VMEXITs.
201
202   The general recommendation is to enable L1D flush on VMENTER. The kernel
203   defaults to conditional mode on affected processors.
204
205   **Note**, that L1D flush does not prevent the SMT problem because the
206   sibling thread will also bring back its data into the L1D which makes it
207   attackable again.
208
209   L1D flush can be controlled by the administrator via the kernel command
210   line and sysfs control files. See :ref:`mitigation_control_command_line`
211   and :ref:`mitigation_control_kvm`.
212
213.. _guest_confinement:
214
2152. Guest VCPU confinement to dedicated physical cores
216^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
217
218   To address the SMT problem, it is possible to make a guest or a group of
219   guests affine to one or more physical cores. The proper mechanism for
220   that is to utilize exclusive cpusets to ensure that no other guest or
221   host tasks can run on these cores.
222
223   If only a single guest or related guests run on sibling SMT threads on
224   the same physical core then they can only attack their own memory and
225   restricted parts of the host memory.
226
227   Host memory is attackable, when one of the sibling SMT threads runs in
228   host OS (hypervisor) context and the other in guest context. The amount
229   of valuable information from the host OS context depends on the context
230   which the host OS executes, i.e. interrupts, soft interrupts and kernel
231   threads. The amount of valuable data from these contexts cannot be
232   declared as non-interesting for an attacker without deep inspection of
233   the code.
234
235   **Note**, that assigning guests to a fixed set of physical cores affects
236   the ability of the scheduler to do load balancing and might have
237   negative effects on CPU utilization depending on the hosting
238   scenario. Disabling SMT might be a viable alternative for particular
239   scenarios.
240
241   For further information about confining guests to a single or to a group
242   of cores consult the cpusets documentation:
243
244   https://www.kernel.org/doc/Documentation/admin-guide/cgroup-v1/cpusets.rst
245
246.. _interrupt_isolation:
247
2483. Interrupt affinity
249^^^^^^^^^^^^^^^^^^^^^
250
251   Interrupts can be made affine to logical CPUs. This is not universally
252   true because there are types of interrupts which are truly per CPU
253   interrupts, e.g. the local timer interrupt. Aside of that multi queue
254   devices affine their interrupts to single CPUs or groups of CPUs per
255   queue without allowing the administrator to control the affinities.
256
257   Moving the interrupts, which can be affinity controlled, away from CPUs
258   which run untrusted guests, reduces the attack vector space.
259
260   Whether the interrupts with are affine to CPUs, which run untrusted
261   guests, provide interesting data for an attacker depends on the system
262   configuration and the scenarios which run on the system. While for some
263   of the interrupts it can be assumed that they won't expose interesting
264   information beyond exposing hints about the host OS memory layout, there
265   is no way to make general assumptions.
266
267   Interrupt affinity can be controlled by the administrator via the
268   /proc/irq/$NR/smp_affinity[_list] files. Limited documentation is
269   available at:
270
271   https://www.kernel.org/doc/Documentation/IRQ-affinity.txt
272
273.. _smt_control:
274
2754. SMT control
276^^^^^^^^^^^^^^
277
278   To prevent the SMT issues of L1TF it might be necessary to disable SMT
279   completely. Disabling SMT can have a significant performance impact, but
280   the impact depends on the hosting scenario and the type of workloads.
281   The impact of disabling SMT needs also to be weighted against the impact
282   of other mitigation solutions like confining guests to dedicated cores.
283
284   The kernel provides a sysfs interface to retrieve the status of SMT and
285   to control it. It also provides a kernel command line interface to
286   control SMT.
287
288   The kernel command line interface consists of the following options:
289
290     =========== ==========================================================
291     nosmt	 Affects the bring up of the secondary CPUs during boot. The
292		 kernel tries to bring all present CPUs online during the
293		 boot process. "nosmt" makes sure that from each physical
294		 core only one - the so called primary (hyper) thread is
295		 activated. Due to a design flaw of Intel processors related
296		 to Machine Check Exceptions the non primary siblings have
297		 to be brought up at least partially and are then shut down
298		 again.  "nosmt" can be undone via the sysfs interface.
299
300     nosmt=force Has the same effect as "nosmt" but it does not allow to
301		 undo the SMT disable via the sysfs interface.
302     =========== ==========================================================
303
304   The sysfs interface provides two files:
305
306   - /sys/devices/system/cpu/smt/control
307   - /sys/devices/system/cpu/smt/active
308
309   /sys/devices/system/cpu/smt/control:
310
311     This file allows to read out the SMT control state and provides the
312     ability to disable or (re)enable SMT. The possible states are:
313
314	==============  ===================================================
315	on		SMT is supported by the CPU and enabled. All
316			logical CPUs can be onlined and offlined without
317			restrictions.
318
319	off		SMT is supported by the CPU and disabled. Only
320			the so called primary SMT threads can be onlined
321			and offlined without restrictions. An attempt to
322			online a non-primary sibling is rejected
323
324	forceoff	Same as 'off' but the state cannot be controlled.
325			Attempts to write to the control file are rejected.
326
327	notsupported	The processor does not support SMT. It's therefore
328			not affected by the SMT implications of L1TF.
329			Attempts to write to the control file are rejected.
330	==============  ===================================================
331
332     The possible states which can be written into this file to control SMT
333     state are:
334
335     - on
336     - off
337     - forceoff
338
339   /sys/devices/system/cpu/smt/active:
340
341     This file reports whether SMT is enabled and active, i.e. if on any
342     physical core two or more sibling threads are online.
343
344   SMT control is also possible at boot time via the l1tf kernel command
345   line parameter in combination with L1D flush control. See
346   :ref:`mitigation_control_command_line`.
347
3485. Disabling EPT
349^^^^^^^^^^^^^^^^
350
351  Disabling EPT for virtual machines provides full mitigation for L1TF even
352  with SMT enabled, because the effective page tables for guests are
353  managed and sanitized by the hypervisor. Though disabling EPT has a
354  significant performance impact especially when the Meltdown mitigation
355  KPTI is enabled.
356
357  EPT can be disabled in the hypervisor via the 'kvm-intel.ept' parameter.
358
359There is ongoing research and development for new mitigation mechanisms to
360address the performance impact of disabling SMT or EPT.
361
362.. _mitigation_control_command_line:
363
364Mitigation control on the kernel command line
365---------------------------------------------
366
367The kernel command line allows to control the L1TF mitigations at boot
368time with the option "l1tf=". The valid arguments for this option are:
369
370  ============  =============================================================
371  full		Provides all available mitigations for the L1TF
372		vulnerability. Disables SMT and enables all mitigations in
373		the hypervisors, i.e. unconditional L1D flushing
374
375		SMT control and L1D flush control via the sysfs interface
376		is still possible after boot.  Hypervisors will issue a
377		warning when the first VM is started in a potentially
378		insecure configuration, i.e. SMT enabled or L1D flush
379		disabled.
380
381  full,force	Same as 'full', but disables SMT and L1D flush runtime
382		control. Implies the 'nosmt=force' command line option.
383		(i.e. sysfs control of SMT is disabled.)
384
385  flush		Leaves SMT enabled and enables the default hypervisor
386		mitigation, i.e. conditional L1D flushing
387
388		SMT control and L1D flush control via the sysfs interface
389		is still possible after boot.  Hypervisors will issue a
390		warning when the first VM is started in a potentially
391		insecure configuration, i.e. SMT enabled or L1D flush
392		disabled.
393
394  flush,nosmt	Disables SMT and enables the default hypervisor mitigation,
395		i.e. conditional L1D flushing.
396
397		SMT control and L1D flush control via the sysfs interface
398		is still possible after boot.  Hypervisors will issue a
399		warning when the first VM is started in a potentially
400		insecure configuration, i.e. SMT enabled or L1D flush
401		disabled.
402
403  flush,nowarn	Same as 'flush', but hypervisors will not warn when a VM is
404		started in a potentially insecure configuration.
405
406  off		Disables hypervisor mitigations and doesn't emit any
407		warnings.
408		It also drops the swap size and available RAM limit restrictions
409		on both hypervisor and bare metal.
410
411  ============  =============================================================
412
413The default is 'flush'. For details about L1D flushing see :ref:`l1d_flush`.
414
415
416.. _mitigation_control_kvm:
417
418Mitigation control for KVM - module parameter
419-------------------------------------------------------------
420
421The KVM hypervisor mitigation mechanism, flushing the L1D cache when
422entering a guest, can be controlled with a module parameter.
423
424The option/parameter is "kvm-intel.vmentry_l1d_flush=". It takes the
425following arguments:
426
427  ============  ==============================================================
428  always	L1D cache flush on every VMENTER.
429
430  cond		Flush L1D on VMENTER only when the code between VMEXIT and
431		VMENTER can leak host memory which is considered
432		interesting for an attacker. This still can leak host memory
433		which allows e.g. to determine the hosts address space layout.
434
435  never		Disables the mitigation
436  ============  ==============================================================
437
438The parameter can be provided on the kernel command line, as a module
439parameter when loading the modules and at runtime modified via the sysfs
440file:
441
442/sys/module/kvm_intel/parameters/vmentry_l1d_flush
443
444The default is 'cond'. If 'l1tf=full,force' is given on the kernel command
445line, then 'always' is enforced and the kvm-intel.vmentry_l1d_flush
446module parameter is ignored and writes to the sysfs file are rejected.
447
448.. _mitigation_selection:
449
450Mitigation selection guide
451--------------------------
452
4531. No virtualization in use
454^^^^^^^^^^^^^^^^^^^^^^^^^^^
455
456   The system is protected by the kernel unconditionally and no further
457   action is required.
458
4592. Virtualization with trusted guests
460^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
461
462   If the guest comes from a trusted source and the guest OS kernel is
463   guaranteed to have the L1TF mitigations in place the system is fully
464   protected against L1TF and no further action is required.
465
466   To avoid the overhead of the default L1D flushing on VMENTER the
467   administrator can disable the flushing via the kernel command line and
468   sysfs control files. See :ref:`mitigation_control_command_line` and
469   :ref:`mitigation_control_kvm`.
470
471
4723. Virtualization with untrusted guests
473^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
474
4753.1. SMT not supported or disabled
476""""""""""""""""""""""""""""""""""
477
478  If SMT is not supported by the processor or disabled in the BIOS or by
479  the kernel, it's only required to enforce L1D flushing on VMENTER.
480
481  Conditional L1D flushing is the default behaviour and can be tuned. See
482  :ref:`mitigation_control_command_line` and :ref:`mitigation_control_kvm`.
483
4843.2. EPT not supported or disabled
485""""""""""""""""""""""""""""""""""
486
487  If EPT is not supported by the processor or disabled in the hypervisor,
488  the system is fully protected. SMT can stay enabled and L1D flushing on
489  VMENTER is not required.
490
491  EPT can be disabled in the hypervisor via the 'kvm-intel.ept' parameter.
492
4933.3. SMT and EPT supported and active
494"""""""""""""""""""""""""""""""""""""
495
496  If SMT and EPT are supported and active then various degrees of
497  mitigations can be employed:
498
499  - L1D flushing on VMENTER:
500
501    L1D flushing on VMENTER is the minimal protection requirement, but it
502    is only potent in combination with other mitigation methods.
503
504    Conditional L1D flushing is the default behaviour and can be tuned. See
505    :ref:`mitigation_control_command_line` and :ref:`mitigation_control_kvm`.
506
507  - Guest confinement:
508
509    Confinement of guests to a single or a group of physical cores which
510    are not running any other processes, can reduce the attack surface
511    significantly, but interrupts, soft interrupts and kernel threads can
512    still expose valuable data to a potential attacker. See
513    :ref:`guest_confinement`.
514
515  - Interrupt isolation:
516
517    Isolating the guest CPUs from interrupts can reduce the attack surface
518    further, but still allows a malicious guest to explore a limited amount
519    of host physical memory. This can at least be used to gain knowledge
520    about the host address space layout. The interrupts which have a fixed
521    affinity to the CPUs which run the untrusted guests can depending on
522    the scenario still trigger soft interrupts and schedule kernel threads
523    which might expose valuable information. See
524    :ref:`interrupt_isolation`.
525
526The above three mitigation methods combined can provide protection to a
527certain degree, but the risk of the remaining attack surface has to be
528carefully analyzed. For full protection the following methods are
529available:
530
531  - Disabling SMT:
532
533    Disabling SMT and enforcing the L1D flushing provides the maximum
534    amount of protection. This mitigation is not depending on any of the
535    above mitigation methods.
536
537    SMT control and L1D flushing can be tuned by the command line
538    parameters 'nosmt', 'l1tf', 'kvm-intel.vmentry_l1d_flush' and at run
539    time with the matching sysfs control files. See :ref:`smt_control`,
540    :ref:`mitigation_control_command_line` and
541    :ref:`mitigation_control_kvm`.
542
543  - Disabling EPT:
544
545    Disabling EPT provides the maximum amount of protection as well. It is
546    not depending on any of the above mitigation methods. SMT can stay
547    enabled and L1D flushing is not required, but the performance impact is
548    significant.
549
550    EPT can be disabled in the hypervisor via the 'kvm-intel.ept'
551    parameter.
552
5533.4. Nested virtual machines
554""""""""""""""""""""""""""""
555
556When nested virtualization is in use, three operating systems are involved:
557the bare metal hypervisor, the nested hypervisor and the nested virtual
558machine.  VMENTER operations from the nested hypervisor into the nested
559guest will always be processed by the bare metal hypervisor. If KVM is the
560bare metal hypervisor it will:
561
562 - Flush the L1D cache on every switch from the nested hypervisor to the
563   nested virtual machine, so that the nested hypervisor's secrets are not
564   exposed to the nested virtual machine;
565
566 - Flush the L1D cache on every switch from the nested virtual machine to
567   the nested hypervisor; this is a complex operation, and flushing the L1D
568   cache avoids that the bare metal hypervisor's secrets are exposed to the
569   nested virtual machine;
570
571 - Instruct the nested hypervisor to not perform any L1D cache flush. This
572   is an optimization to avoid double L1D flushing.
573
574
575.. _default_mitigations:
576
577Default mitigations
578-------------------
579
580  The kernel default mitigations for vulnerable processors are:
581
582  - PTE inversion to protect against malicious user space. This is done
583    unconditionally and cannot be controlled. The swap storage is limited
584    to ~16TB.
585
586  - L1D conditional flushing on VMENTER when EPT is enabled for
587    a guest.
588
589  The kernel does not by default enforce the disabling of SMT, which leaves
590  SMT systems vulnerable when running untrusted guests with EPT enabled.
591
592  The rationale for this choice is:
593
594  - Force disabling SMT can break existing setups, especially with
595    unattended updates.
596
597  - If regular users run untrusted guests on their machine, then L1TF is
598    just an add on to other malware which might be embedded in an untrusted
599    guest, e.g. spam-bots or attacks on the local network.
600
601    There is no technical way to prevent a user from running untrusted code
602    on their machines blindly.
603
604  - It's technically extremely unlikely and from today's knowledge even
605    impossible that L1TF can be exploited via the most popular attack
606    mechanisms like JavaScript because these mechanisms have no way to
607    control PTEs. If this would be possible and not other mitigation would
608    be possible, then the default might be different.
609
610  - The administrators of cloud and hosting setups have to carefully
611    analyze the risk for their scenarios and make the appropriate
612    mitigation choices, which might even vary across their deployed
613    machines and also result in other changes of their overall setup.
614    There is no way for the kernel to provide a sensible default for this
615    kind of scenarios.
616