1.. SPDX-License-Identifier: GPL-2.0
2
3Spectre Side Channels
4=====================
5
6Spectre is a class of side channel attacks that exploit branch prediction
7and speculative execution on modern CPUs to read memory, possibly
8bypassing access controls. Speculative execution side channel exploits
9do not modify memory but attempt to infer privileged data in the memory.
10
11This document covers Spectre variant 1 and Spectre variant 2.
12
13Affected processors
14-------------------
15
16Speculative execution side channel methods affect a wide range of modern
17high performance processors, since most modern high speed processors
18use branch prediction and speculative execution.
19
20The following CPUs are vulnerable:
21
22    - Intel Core, Atom, Pentium, and Xeon processors
23
24    - AMD Phenom, EPYC, and Zen processors
25
26    - IBM POWER and zSeries processors
27
28    - Higher end ARM processors
29
30    - Apple CPUs
31
32    - Higher end MIPS CPUs
33
34    - Likely most other high performance CPUs. Contact your CPU vendor for details.
35
36Whether a processor is affected or not can be read out from the Spectre
37vulnerability files in sysfs. See :ref:`spectre_sys_info`.
38
39Related CVEs
40------------
41
42The following CVE entries describe Spectre variants:
43
44   =============   =======================  =================
45   CVE-2017-5753   Bounds check bypass      Spectre variant 1
46   CVE-2017-5715   Branch target injection  Spectre variant 2
47   =============   =======================  =================
48
49Problem
50-------
51
52CPUs use speculative operations to improve performance. That may leave
53traces of memory accesses or computations in the processor's caches,
54buffers, and branch predictors. Malicious software may be able to
55influence the speculative execution paths, and then use the side effects
56of the speculative execution in the CPUs' caches and buffers to infer
57privileged data touched during the speculative execution.
58
59Spectre variant 1 attacks take advantage of speculative execution of
60conditional branches, while Spectre variant 2 attacks use speculative
61execution of indirect branches to leak privileged memory.
62See :ref:`[1] <spec_ref1>` :ref:`[5] <spec_ref5>` :ref:`[7] <spec_ref7>`
63:ref:`[10] <spec_ref10>` :ref:`[11] <spec_ref11>`.
64
65Spectre variant 1 (Bounds Check Bypass)
66---------------------------------------
67
68The bounds check bypass attack :ref:`[2] <spec_ref2>` takes advantage
69of speculative execution that bypasses conditional branch instructions
70used for memory access bounds check (e.g. checking if the index of an
71array results in memory access within a valid range). This results in
72memory accesses to invalid memory (with out-of-bound index) that are
73done speculatively before validation checks resolve. Such speculative
74memory accesses can leave side effects, creating side channels which
75leak information to the attacker.
76
77There are some extensions of Spectre variant 1 attacks for reading data
78over the network, see :ref:`[12] <spec_ref12>`. However such attacks
79are difficult, low bandwidth, fragile, and are considered low risk.
80
81Spectre variant 2 (Branch Target Injection)
82-------------------------------------------
83
84The branch target injection attack takes advantage of speculative
85execution of indirect branches :ref:`[3] <spec_ref3>`.  The indirect
86branch predictors inside the processor used to guess the target of
87indirect branches can be influenced by an attacker, causing gadget code
88to be speculatively executed, thus exposing sensitive data touched by
89the victim. The side effects left in the CPU's caches during speculative
90execution can be measured to infer data values.
91
92.. _poison_btb:
93
94In Spectre variant 2 attacks, the attacker can steer speculative indirect
95branches in the victim to gadget code by poisoning the branch target
96buffer of a CPU used for predicting indirect branch addresses. Such
97poisoning could be done by indirect branching into existing code,
98with the address offset of the indirect branch under the attacker's
99control. Since the branch prediction on impacted hardware does not
100fully disambiguate branch address and uses the offset for prediction,
101this could cause privileged code's indirect branch to jump to a gadget
102code with the same offset.
103
104The most useful gadgets take an attacker-controlled input parameter (such
105as a register value) so that the memory read can be controlled. Gadgets
106without input parameters might be possible, but the attacker would have
107very little control over what memory can be read, reducing the risk of
108the attack revealing useful data.
109
110One other variant 2 attack vector is for the attacker to poison the
111return stack buffer (RSB) :ref:`[13] <spec_ref13>` to cause speculative
112subroutine return instruction execution to go to a gadget.  An attacker's
113imbalanced subroutine call instructions might "poison" entries in the
114return stack buffer which are later consumed by a victim's subroutine
115return instructions.  This attack can be mitigated by flushing the return
116stack buffer on context switch, or virtual machine (VM) exit.
117
118On systems with simultaneous multi-threading (SMT), attacks are possible
119from the sibling thread, as level 1 cache and branch target buffer
120(BTB) may be shared between hardware threads in a CPU core.  A malicious
121program running on the sibling thread may influence its peer's BTB to
122steer its indirect branch speculations to gadget code, and measure the
123speculative execution's side effects left in level 1 cache to infer the
124victim's data.
125
126Attack scenarios
127----------------
128
129The following list of attack scenarios have been anticipated, but may
130not cover all possible attack vectors.
131
1321. A user process attacking the kernel
133^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
134
135   The attacker passes a parameter to the kernel via a register or
136   via a known address in memory during a syscall. Such parameter may
137   be used later by the kernel as an index to an array or to derive
138   a pointer for a Spectre variant 1 attack.  The index or pointer
139   is invalid, but bound checks are bypassed in the code branch taken
140   for speculative execution. This could cause privileged memory to be
141   accessed and leaked.
142
143   For kernel code that has been identified where data pointers could
144   potentially be influenced for Spectre attacks, new "nospec" accessor
145   macros are used to prevent speculative loading of data.
146
147   Spectre variant 2 attacker can :ref:`poison <poison_btb>` the branch
148   target buffer (BTB) before issuing syscall to launch an attack.
149   After entering the kernel, the kernel could use the poisoned branch
150   target buffer on indirect jump and jump to gadget code in speculative
151   execution.
152
153   If an attacker tries to control the memory addresses leaked during
154   speculative execution, he would also need to pass a parameter to the
155   gadget, either through a register or a known address in memory. After
156   the gadget has executed, he can measure the side effect.
157
158   The kernel can protect itself against consuming poisoned branch
159   target buffer entries by using return trampolines (also known as
160   "retpoline") :ref:`[3] <spec_ref3>` :ref:`[9] <spec_ref9>` for all
161   indirect branches. Return trampolines trap speculative execution paths
162   to prevent jumping to gadget code during speculative execution.
163   x86 CPUs with Enhanced Indirect Branch Restricted Speculation
164   (Enhanced IBRS) available in hardware should use the feature to
165   mitigate Spectre variant 2 instead of retpoline. Enhanced IBRS is
166   more efficient than retpoline.
167
168   There may be gadget code in firmware which could be exploited with
169   Spectre variant 2 attack by a rogue user process. To mitigate such
170   attacks on x86, Indirect Branch Restricted Speculation (IBRS) feature
171   is turned on before the kernel invokes any firmware code.
172
1732. A user process attacking another user process
174^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
175
176   A malicious user process can try to attack another user process,
177   either via a context switch on the same hardware thread, or from the
178   sibling hyperthread sharing a physical processor core on simultaneous
179   multi-threading (SMT) system.
180
181   Spectre variant 1 attacks generally require passing parameters
182   between the processes, which needs a data passing relationship, such
183   as remote procedure calls (RPC).  Those parameters are used in gadget
184   code to derive invalid data pointers accessing privileged memory in
185   the attacked process.
186
187   Spectre variant 2 attacks can be launched from a rogue process by
188   :ref:`poisoning <poison_btb>` the branch target buffer.  This can
189   influence the indirect branch targets for a victim process that either
190   runs later on the same hardware thread, or running concurrently on
191   a sibling hardware thread sharing the same physical core.
192
193   A user process can protect itself against Spectre variant 2 attacks
194   by using the prctl() syscall to disable indirect branch speculation
195   for itself.  An administrator can also cordon off an unsafe process
196   from polluting the branch target buffer by disabling the process's
197   indirect branch speculation. This comes with a performance cost
198   from not using indirect branch speculation and clearing the branch
199   target buffer.  When SMT is enabled on x86, for a process that has
200   indirect branch speculation disabled, Single Threaded Indirect Branch
201   Predictors (STIBP) :ref:`[4] <spec_ref4>` are turned on to prevent the
202   sibling thread from controlling branch target buffer.  In addition,
203   the Indirect Branch Prediction Barrier (IBPB) is issued to clear the
204   branch target buffer when context switching to and from such process.
205
206   On x86, the return stack buffer is stuffed on context switch.
207   This prevents the branch target buffer from being used for branch
208   prediction when the return stack buffer underflows while switching to
209   a deeper call stack. Any poisoned entries in the return stack buffer
210   left by the previous process will also be cleared.
211
212   User programs should use address space randomization to make attacks
213   more difficult (Set /proc/sys/kernel/randomize_va_space = 1 or 2).
214
2153. A virtualized guest attacking the host
216^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
217
218   The attack mechanism is similar to how user processes attack the
219   kernel.  The kernel is entered via hyper-calls or other virtualization
220   exit paths.
221
222   For Spectre variant 1 attacks, rogue guests can pass parameters
223   (e.g. in registers) via hyper-calls to derive invalid pointers to
224   speculate into privileged memory after entering the kernel.  For places
225   where such kernel code has been identified, nospec accessor macros
226   are used to stop speculative memory access.
227
228   For Spectre variant 2 attacks, rogue guests can :ref:`poison
229   <poison_btb>` the branch target buffer or return stack buffer, causing
230   the kernel to jump to gadget code in the speculative execution paths.
231
232   To mitigate variant 2, the host kernel can use return trampolines
233   for indirect branches to bypass the poisoned branch target buffer,
234   and flushing the return stack buffer on VM exit.  This prevents rogue
235   guests from affecting indirect branching in the host kernel.
236
237   To protect host processes from rogue guests, host processes can have
238   indirect branch speculation disabled via prctl().  The branch target
239   buffer is cleared before context switching to such processes.
240
2414. A virtualized guest attacking other guest
242^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
243
244   A rogue guest may attack another guest to get data accessible by the
245   other guest.
246
247   Spectre variant 1 attacks are possible if parameters can be passed
248   between guests.  This may be done via mechanisms such as shared memory
249   or message passing.  Such parameters could be used to derive data
250   pointers to privileged data in guest.  The privileged data could be
251   accessed by gadget code in the victim's speculation paths.
252
253   Spectre variant 2 attacks can be launched from a rogue guest by
254   :ref:`poisoning <poison_btb>` the branch target buffer or the return
255   stack buffer. Such poisoned entries could be used to influence
256   speculation execution paths in the victim guest.
257
258   Linux kernel mitigates attacks to other guests running in the same
259   CPU hardware thread by flushing the return stack buffer on VM exit,
260   and clearing the branch target buffer before switching to a new guest.
261
262   If SMT is used, Spectre variant 2 attacks from an untrusted guest
263   in the sibling hyperthread can be mitigated by the administrator,
264   by turning off the unsafe guest's indirect branch speculation via
265   prctl().  A guest can also protect itself by turning on microcode
266   based mitigations (such as IBPB or STIBP on x86) within the guest.
267
268.. _spectre_sys_info:
269
270Spectre system information
271--------------------------
272
273The Linux kernel provides a sysfs interface to enumerate the current
274mitigation status of the system for Spectre: whether the system is
275vulnerable, and which mitigations are active.
276
277The sysfs file showing Spectre variant 1 mitigation status is:
278
279   /sys/devices/system/cpu/vulnerabilities/spectre_v1
280
281The possible values in this file are:
282
283  =======================================  =================================
284  'Mitigation: __user pointer sanitation'  Protection in kernel on a case by
285                                           case base with explicit pointer
286                                           sanitation.
287  =======================================  =================================
288
289However, the protections are put in place on a case by case basis,
290and there is no guarantee that all possible attack vectors for Spectre
291variant 1 are covered.
292
293The spectre_v2 kernel file reports if the kernel has been compiled with
294retpoline mitigation or if the CPU has hardware mitigation, and if the
295CPU has support for additional process-specific mitigation.
296
297This file also reports CPU features enabled by microcode to mitigate
298attack between user processes:
299
3001. Indirect Branch Prediction Barrier (IBPB) to add additional
301   isolation between processes of different users.
3022. Single Thread Indirect Branch Predictors (STIBP) to add additional
303   isolation between CPU threads running on the same core.
304
305These CPU features may impact performance when used and can be enabled
306per process on a case-by-case base.
307
308The sysfs file showing Spectre variant 2 mitigation status is:
309
310   /sys/devices/system/cpu/vulnerabilities/spectre_v2
311
312The possible values in this file are:
313
314  - Kernel status:
315
316  ====================================  =================================
317  'Not affected'                        The processor is not vulnerable
318  'Vulnerable'                          Vulnerable, no mitigation
319  'Mitigation: Full generic retpoline'  Software-focused mitigation
320  'Mitigation: Full AMD retpoline'      AMD-specific software mitigation
321  'Mitigation: Enhanced IBRS'           Hardware-focused mitigation
322  ====================================  =================================
323
324  - Firmware status: Show if Indirect Branch Restricted Speculation (IBRS) is
325    used to protect against Spectre variant 2 attacks when calling firmware (x86 only).
326
327  ========== =============================================================
328  'IBRS_FW'  Protection against user program attacks when calling firmware
329  ========== =============================================================
330
331  - Indirect branch prediction barrier (IBPB) status for protection between
332    processes of different users. This feature can be controlled through
333    prctl() per process, or through kernel command line options. This is
334    an x86 only feature. For more details see below.
335
336  ===================   ========================================================
337  'IBPB: disabled'      IBPB unused
338  'IBPB: always-on'     Use IBPB on all tasks
339  'IBPB: conditional'   Use IBPB on SECCOMP or indirect branch restricted tasks
340  ===================   ========================================================
341
342  - Single threaded indirect branch prediction (STIBP) status for protection
343    between different hyper threads. This feature can be controlled through
344    prctl per process, or through kernel command line options. This is x86
345    only feature. For more details see below.
346
347  ====================  ========================================================
348  'STIBP: disabled'     STIBP unused
349  'STIBP: forced'       Use STIBP on all tasks
350  'STIBP: conditional'  Use STIBP on SECCOMP or indirect branch restricted tasks
351  ====================  ========================================================
352
353  - Return stack buffer (RSB) protection status:
354
355  =============   ===========================================
356  'RSB filling'   Protection of RSB on context switch enabled
357  =============   ===========================================
358
359Full mitigation might require a microcode update from the CPU
360vendor. When the necessary microcode is not available, the kernel will
361report vulnerability.
362
363Turning on mitigation for Spectre variant 1 and Spectre variant 2
364-----------------------------------------------------------------
365
3661. Kernel mitigation
367^^^^^^^^^^^^^^^^^^^^
368
369   For the Spectre variant 1, vulnerable kernel code (as determined
370   by code audit or scanning tools) is annotated on a case by case
371   basis to use nospec accessor macros for bounds clipping :ref:`[2]
372   <spec_ref2>` to avoid any usable disclosure gadgets. However, it may
373   not cover all attack vectors for Spectre variant 1.
374
375   For Spectre variant 2 mitigation, the compiler turns indirect calls or
376   jumps in the kernel into equivalent return trampolines (retpolines)
377   :ref:`[3] <spec_ref3>` :ref:`[9] <spec_ref9>` to go to the target
378   addresses.  Speculative execution paths under retpolines are trapped
379   in an infinite loop to prevent any speculative execution jumping to
380   a gadget.
381
382   To turn on retpoline mitigation on a vulnerable CPU, the kernel
383   needs to be compiled with a gcc compiler that supports the
384   -mindirect-branch=thunk-extern -mindirect-branch-register options.
385   If the kernel is compiled with a Clang compiler, the compiler needs
386   to support -mretpoline-external-thunk option.  The kernel config
387   CONFIG_RETPOLINE needs to be turned on, and the CPU needs to run with
388   the latest updated microcode.
389
390   On Intel Skylake-era systems the mitigation covers most, but not all,
391   cases. See :ref:`[3] <spec_ref3>` for more details.
392
393   On CPUs with hardware mitigation for Spectre variant 2 (e.g. Enhanced
394   IBRS on x86), retpoline is automatically disabled at run time.
395
396   The retpoline mitigation is turned on by default on vulnerable
397   CPUs. It can be forced on or off by the administrator
398   via the kernel command line and sysfs control files. See
399   :ref:`spectre_mitigation_control_command_line`.
400
401   On x86, indirect branch restricted speculation is turned on by default
402   before invoking any firmware code to prevent Spectre variant 2 exploits
403   using the firmware.
404
405   Using kernel address space randomization (CONFIG_RANDOMIZE_SLAB=y
406   and CONFIG_SLAB_FREELIST_RANDOM=y in the kernel configuration) makes
407   attacks on the kernel generally more difficult.
408
4092. User program mitigation
410^^^^^^^^^^^^^^^^^^^^^^^^^^
411
412   User programs can mitigate Spectre variant 1 using LFENCE or "bounds
413   clipping". For more details see :ref:`[2] <spec_ref2>`.
414
415   For Spectre variant 2 mitigation, individual user programs
416   can be compiled with return trampolines for indirect branches.
417   This protects them from consuming poisoned entries in the branch
418   target buffer left by malicious software.  Alternatively, the
419   programs can disable their indirect branch speculation via prctl()
420   (See :ref:`Documentation/userspace-api/spec_ctrl.rst <set_spec_ctrl>`).
421   On x86, this will turn on STIBP to guard against attacks from the
422   sibling thread when the user program is running, and use IBPB to
423   flush the branch target buffer when switching to/from the program.
424
425   Restricting indirect branch speculation on a user program will
426   also prevent the program from launching a variant 2 attack
427   on x86.  All sand-boxed SECCOMP programs have indirect branch
428   speculation restricted by default.  Administrators can change
429   that behavior via the kernel command line and sysfs control files.
430   See :ref:`spectre_mitigation_control_command_line`.
431
432   Programs that disable their indirect branch speculation will have
433   more overhead and run slower.
434
435   User programs should use address space randomization
436   (/proc/sys/kernel/randomize_va_space = 1 or 2) to make attacks more
437   difficult.
438
4393. VM mitigation
440^^^^^^^^^^^^^^^^
441
442   Within the kernel, Spectre variant 1 attacks from rogue guests are
443   mitigated on a case by case basis in VM exit paths. Vulnerable code
444   uses nospec accessor macros for "bounds clipping", to avoid any
445   usable disclosure gadgets.  However, this may not cover all variant
446   1 attack vectors.
447
448   For Spectre variant 2 attacks from rogue guests to the kernel, the
449   Linux kernel uses retpoline or Enhanced IBRS to prevent consumption of
450   poisoned entries in branch target buffer left by rogue guests.  It also
451   flushes the return stack buffer on every VM exit to prevent a return
452   stack buffer underflow so poisoned branch target buffer could be used,
453   or attacker guests leaving poisoned entries in the return stack buffer.
454
455   To mitigate guest-to-guest attacks in the same CPU hardware thread,
456   the branch target buffer is sanitized by flushing before switching
457   to a new guest on a CPU.
458
459   The above mitigations are turned on by default on vulnerable CPUs.
460
461   To mitigate guest-to-guest attacks from sibling thread when SMT is
462   in use, an untrusted guest running in the sibling thread can have
463   its indirect branch speculation disabled by administrator via prctl().
464
465   The kernel also allows guests to use any microcode based mitigation
466   they choose to use (such as IBPB or STIBP on x86) to protect themselves.
467
468.. _spectre_mitigation_control_command_line:
469
470Mitigation control on the kernel command line
471---------------------------------------------
472
473Spectre variant 2 mitigation can be disabled or force enabled at the
474kernel command line.
475
476	nospectre_v2
477
478		[X86] Disable all mitigations for the Spectre variant 2
479		(indirect branch prediction) vulnerability. System may
480		allow data leaks with this option, which is equivalent
481		to spectre_v2=off.
482
483
484        spectre_v2=
485
486		[X86] Control mitigation of Spectre variant 2
487		(indirect branch speculation) vulnerability.
488		The default operation protects the kernel from
489		user space attacks.
490
491		on
492			unconditionally enable, implies
493			spectre_v2_user=on
494		off
495			unconditionally disable, implies
496		        spectre_v2_user=off
497		auto
498			kernel detects whether your CPU model is
499		        vulnerable
500
501		Selecting 'on' will, and 'auto' may, choose a
502		mitigation method at run time according to the
503		CPU, the available microcode, the setting of the
504		CONFIG_RETPOLINE configuration option, and the
505		compiler with which the kernel was built.
506
507		Selecting 'on' will also enable the mitigation
508		against user space to user space task attacks.
509
510		Selecting 'off' will disable both the kernel and
511		the user space protections.
512
513		Specific mitigations can also be selected manually:
514
515		retpoline
516					replace indirect branches
517		retpoline,generic
518					google's original retpoline
519		retpoline,amd
520					AMD-specific minimal thunk
521
522		Not specifying this option is equivalent to
523		spectre_v2=auto.
524
525For user space mitigation:
526
527        spectre_v2_user=
528
529		[X86] Control mitigation of Spectre variant 2
530		(indirect branch speculation) vulnerability between
531		user space tasks
532
533		on
534			Unconditionally enable mitigations. Is
535			enforced by spectre_v2=on
536
537		off
538			Unconditionally disable mitigations. Is
539			enforced by spectre_v2=off
540
541		prctl
542			Indirect branch speculation is enabled,
543			but mitigation can be enabled via prctl
544			per thread. The mitigation control state
545			is inherited on fork.
546
547		prctl,ibpb
548			Like "prctl" above, but only STIBP is
549			controlled per thread. IBPB is issued
550			always when switching between different user
551			space processes.
552
553		seccomp
554			Same as "prctl" above, but all seccomp
555			threads will enable the mitigation unless
556			they explicitly opt out.
557
558		seccomp,ibpb
559			Like "seccomp" above, but only STIBP is
560			controlled per thread. IBPB is issued
561			always when switching between different
562			user space processes.
563
564		auto
565			Kernel selects the mitigation depending on
566			the available CPU features and vulnerability.
567
568		Default mitigation:
569		If CONFIG_SECCOMP=y then "seccomp", otherwise "prctl"
570
571		Not specifying this option is equivalent to
572		spectre_v2_user=auto.
573
574		In general the kernel by default selects
575		reasonable mitigations for the current CPU. To
576		disable Spectre variant 2 mitigations, boot with
577		spectre_v2=off. Spectre variant 1 mitigations
578		cannot be disabled.
579
580Mitigation selection guide
581--------------------------
582
5831. Trusted userspace
584^^^^^^^^^^^^^^^^^^^^
585
586   If all userspace applications are from trusted sources and do not
587   execute externally supplied untrusted code, then the mitigations can
588   be disabled.
589
5902. Protect sensitive programs
591^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
592
593   For security-sensitive programs that have secrets (e.g. crypto
594   keys), protection against Spectre variant 2 can be put in place by
595   disabling indirect branch speculation when the program is running
596   (See :ref:`Documentation/userspace-api/spec_ctrl.rst <set_spec_ctrl>`).
597
5983. Sandbox untrusted programs
599^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
600
601   Untrusted programs that could be a source of attacks can be cordoned
602   off by disabling their indirect branch speculation when they are run
603   (See :ref:`Documentation/userspace-api/spec_ctrl.rst <set_spec_ctrl>`).
604   This prevents untrusted programs from polluting the branch target
605   buffer.  All programs running in SECCOMP sandboxes have indirect
606   branch speculation restricted by default. This behavior can be
607   changed via the kernel command line and sysfs control files. See
608   :ref:`spectre_mitigation_control_command_line`.
609
6103. High security mode
611^^^^^^^^^^^^^^^^^^^^^
612
613   All Spectre variant 2 mitigations can be forced on
614   at boot time for all programs (See the "on" option in
615   :ref:`spectre_mitigation_control_command_line`).  This will add
616   overhead as indirect branch speculations for all programs will be
617   restricted.
618
619   On x86, branch target buffer will be flushed with IBPB when switching
620   to a new program. STIBP is left on all the time to protect programs
621   against variant 2 attacks originating from programs running on
622   sibling threads.
623
624   Alternatively, STIBP can be used only when running programs
625   whose indirect branch speculation is explicitly disabled,
626   while IBPB is still used all the time when switching to a new
627   program to clear the branch target buffer (See "ibpb" option in
628   :ref:`spectre_mitigation_control_command_line`).  This "ibpb" option
629   has less performance cost than the "on" option, which leaves STIBP
630   on all the time.
631
632References on Spectre
633---------------------
634
635Intel white papers:
636
637.. _spec_ref1:
638
639[1] `Intel analysis of speculative execution side channels <https://newsroom.intel.com/wp-content/uploads/sites/11/2018/01/Intel-Analysis-of-Speculative-Execution-Side-Channels.pdf>`_.
640
641.. _spec_ref2:
642
643[2] `Bounds check bypass <https://software.intel.com/security-software-guidance/software-guidance/bounds-check-bypass>`_.
644
645.. _spec_ref3:
646
647[3] `Deep dive: Retpoline: A branch target injection mitigation <https://software.intel.com/security-software-guidance/insights/deep-dive-retpoline-branch-target-injection-mitigation>`_.
648
649.. _spec_ref4:
650
651[4] `Deep Dive: Single Thread Indirect Branch Predictors <https://software.intel.com/security-software-guidance/insights/deep-dive-single-thread-indirect-branch-predictors>`_.
652
653AMD white papers:
654
655.. _spec_ref5:
656
657[5] `AMD64 technology indirect branch control extension <https://developer.amd.com/wp-content/resources/Architecture_Guidelines_Update_Indirect_Branch_Control.pdf>`_.
658
659.. _spec_ref6:
660
661[6] `Software techniques for managing speculation on AMD processors <https://developer.amd.com/wp-content/resources/90343-B_SoftwareTechniquesforManagingSpeculation_WP_7-18Update_FNL.pdf>`_.
662
663ARM white papers:
664
665.. _spec_ref7:
666
667[7] `Cache speculation side-channels <https://developer.arm.com/support/arm-security-updates/speculative-processor-vulnerability/download-the-whitepaper>`_.
668
669.. _spec_ref8:
670
671[8] `Cache speculation issues update <https://developer.arm.com/support/arm-security-updates/speculative-processor-vulnerability/latest-updates/cache-speculation-issues-update>`_.
672
673Google white paper:
674
675.. _spec_ref9:
676
677[9] `Retpoline: a software construct for preventing branch-target-injection <https://support.google.com/faqs/answer/7625886>`_.
678
679MIPS white paper:
680
681.. _spec_ref10:
682
683[10] `MIPS: response on speculative execution and side channel vulnerabilities <https://www.mips.com/blog/mips-response-on-speculative-execution-and-side-channel-vulnerabilities/>`_.
684
685Academic papers:
686
687.. _spec_ref11:
688
689[11] `Spectre Attacks: Exploiting Speculative Execution <https://spectreattack.com/spectre.pdf>`_.
690
691.. _spec_ref12:
692
693[12] `NetSpectre: Read Arbitrary Memory over Network <https://arxiv.org/abs/1807.10535>`_.
694
695.. _spec_ref13:
696
697[13] `Spectre Returns! Speculation Attacks using the Return Stack Buffer <https://www.usenix.org/system/files/conference/woot18/woot18-paper-koruyeh.pdf>`_.
698