1=======================================
2Pointer authentication in AArch64 Linux
3=======================================
4
5Author: Mark Rutland <mark.rutland@arm.com>
6
7Date: 2017-07-19
8
9This document briefly describes the provision of pointer authentication
10functionality in AArch64 Linux.
11
12
13Architecture overview
14---------------------
15
16The ARMv8.3 Pointer Authentication extension adds primitives that can be
17used to mitigate certain classes of attack where an attacker can corrupt
18the contents of some memory (e.g. the stack).
19
20The extension uses a Pointer Authentication Code (PAC) to determine
21whether pointers have been modified unexpectedly. A PAC is derived from
22a pointer, another value (such as the stack pointer), and a secret key
23held in system registers.
24
25The extension adds instructions to insert a valid PAC into a pointer,
26and to verify/remove the PAC from a pointer. The PAC occupies a number
27of high-order bits of the pointer, which varies dependent on the
28configured virtual address size and whether pointer tagging is in use.
29
30A subset of these instructions have been allocated from the HINT
31encoding space. In the absence of the extension (or when disabled),
32these instructions behave as NOPs. Applications and libraries using
33these instructions operate correctly regardless of the presence of the
34extension.
35
36The extension provides five separate keys to generate PACs - two for
37instruction addresses (APIAKey, APIBKey), two for data addresses
38(APDAKey, APDBKey), and one for generic authentication (APGAKey).
39
40
41Basic support
42-------------
43
44When CONFIG_ARM64_PTR_AUTH is selected, and relevant HW support is
45present, the kernel will assign random key values to each process at
46exec*() time. The keys are shared by all threads within the process, and
47are preserved across fork().
48
49Presence of address authentication functionality is advertised via
50HWCAP_PACA, and generic authentication functionality via HWCAP_PACG.
51
52The number of bits that the PAC occupies in a pointer is 55 minus the
53virtual address size configured by the kernel. For example, with a
54virtual address size of 48, the PAC is 7 bits wide.
55
56Recent versions of GCC can compile code with APIAKey-based return
57address protection when passed the -msign-return-address option. This
58uses instructions in the HINT space (unless -march=armv8.3-a or higher
59is also passed), and such code can run on systems without the pointer
60authentication extension.
61
62In addition to exec(), keys can also be reinitialized to random values
63using the PR_PAC_RESET_KEYS prctl. A bitmask of PR_PAC_APIAKEY,
64PR_PAC_APIBKEY, PR_PAC_APDAKEY, PR_PAC_APDBKEY and PR_PAC_APGAKEY
65specifies which keys are to be reinitialized; specifying 0 means "all
66keys".
67
68
69Debugging
70---------
71
72When CONFIG_ARM64_PTR_AUTH is selected, and HW support for address
73authentication is present, the kernel will expose the position of TTBR0
74PAC bits in the NT_ARM_PAC_MASK regset (struct user_pac_mask), which
75userspace can acquire via PTRACE_GETREGSET.
76
77The regset is exposed only when HWCAP_PACA is set. Separate masks are
78exposed for data pointers and instruction pointers, as the set of PAC
79bits can vary between the two. Note that the masks apply to TTBR0
80addresses, and are not valid to apply to TTBR1 addresses (e.g. kernel
81pointers).
82
83Additionally, when CONFIG_CHECKPOINT_RESTORE is also set, the kernel
84will expose the NT_ARM_PACA_KEYS and NT_ARM_PACG_KEYS regsets (struct
85user_pac_address_keys and struct user_pac_generic_keys). These can be
86used to get and set the keys for a thread.
87
88
89Virtualization
90--------------
91
92Pointer authentication is enabled in KVM guest when each virtual cpu is
93initialised by passing flags KVM_ARM_VCPU_PTRAUTH_[ADDRESS/GENERIC] and
94requesting these two separate cpu features to be enabled. The current KVM
95guest implementation works by enabling both features together, so both
96these userspace flags are checked before enabling pointer authentication.
97The separate userspace flag will allow to have no userspace ABI changes
98if support is added in the future to allow these two features to be
99enabled independently of one another.
100
101As Arm Architecture specifies that Pointer Authentication feature is
102implemented along with the VHE feature so KVM arm64 ptrauth code relies
103on VHE mode to be present.
104
105Additionally, when these vcpu feature flags are not set then KVM will
106filter out the Pointer Authentication system key registers from
107KVM_GET/SET_REG_* ioctls and mask those features from cpufeature ID
108register. Any attempt to use the Pointer Authentication instructions will
109result in an UNDEFINED exception being injected into the guest.
110
111
112Enabling and disabling keys
113---------------------------
114
115The prctl PR_PAC_SET_ENABLED_KEYS allows the user program to control which
116PAC keys are enabled in a particular task. It takes two arguments, the
117first being a bitmask of PR_PAC_APIAKEY, PR_PAC_APIBKEY, PR_PAC_APDAKEY
118and PR_PAC_APDBKEY specifying which keys shall be affected by this prctl,
119and the second being a bitmask of the same bits specifying whether the key
120should be enabled or disabled. For example::
121
122  prctl(PR_PAC_SET_ENABLED_KEYS,
123        PR_PAC_APIAKEY | PR_PAC_APIBKEY | PR_PAC_APDAKEY | PR_PAC_APDBKEY,
124        PR_PAC_APIBKEY, 0, 0);
125
126disables all keys except the IB key.
127
128The main reason why this is useful is to enable a userspace ABI that uses PAC
129instructions to sign and authenticate function pointers and other pointers
130exposed outside of the function, while still allowing binaries conforming to
131the ABI to interoperate with legacy binaries that do not sign or authenticate
132pointers.
133
134The idea is that a dynamic loader or early startup code would issue this
135prctl very early after establishing that a process may load legacy binaries,
136but before executing any PAC instructions.
137
138For compatibility with previous kernel versions, processes start up with IA,
139IB, DA and DB enabled, and are reset to this state on exec(). Processes created
140via fork() and clone() inherit the key enabled state from the calling process.
141
142It is recommended to avoid disabling the IA key, as this has higher performance
143overhead than disabling any of the other keys.
144