1 | //===-- xray_AArch64.cpp ----------------------------------------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file is a part of XRay, a dynamic runtime instrumentation system. |
10 | // |
11 | // Implementation of AArch64-specific routines (64-bit). |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | #include "sanitizer_common/sanitizer_common.h" |
15 | #include "xray_defs.h" |
16 | #include "xray_interface_internal.h" |
17 | #include <atomic> |
18 | #include <cassert> |
19 | |
20 | extern "C" void __clear_cache(void *start, void *end); |
21 | |
22 | namespace __xray { |
23 | |
24 | // The machine codes for some instructions used in runtime patching. |
25 | enum class PatchOpcodes : uint32_t { |
26 | PO_StpX0X30SP_m16e = 0xA9BF7BE0, // STP X0, X30, [SP, #-16]! |
27 | PO_LdrX16_12 = 0x58000070, // LDR X16, #12 |
28 | PO_BlrX16 = 0xD63F0200, // BLR X16 |
29 | PO_LdpX0X30SP_16 = 0xA8C17BE0, // LDP X0, X30, [SP], #16 |
30 | PO_B32 = 0x14000008 // B #32 |
31 | }; |
32 | |
33 | inline static bool patchSled(const bool Enable, const uint32_t FuncId, |
34 | const XRaySledEntry &Sled, |
35 | void (*TracingHook)()) XRAY_NEVER_INSTRUMENT { |
36 | // When |Enable| == true, |
37 | // We replace the following compile-time stub (sled): |
38 | // |
39 | // xray_sled_n: |
40 | // B #32 |
41 | // 7 NOPs (24 bytes) |
42 | // |
43 | // With the following runtime patch: |
44 | // |
45 | // xray_sled_n: |
46 | // STP X0, X30, [SP, #-16]! ; PUSH {r0, lr} |
47 | // LDR W17, #12 ; W17 := function ID |
48 | // LDR X16,#12 ; X16 := address of the trampoline |
49 | // BLR X16 |
50 | // ;DATA: 32 bits of function ID |
51 | // ;DATA: lower 32 bits of the address of the trampoline |
52 | // ;DATA: higher 32 bits of the address of the trampoline |
53 | // LDP X0, X30, [SP], #16 ; POP {r0, lr} |
54 | // |
55 | // Replacement of the first 4-byte instruction should be the last and atomic |
56 | // operation, so that the user code which reaches the sled concurrently |
57 | // either jumps over the whole sled, or executes the whole sled when the |
58 | // latter is ready. |
59 | // |
60 | // When |Enable|==false, we set back the first instruction in the sled to be |
61 | // B #32 |
62 | |
63 | uint32_t *FirstAddress = reinterpret_cast<uint32_t *>(Sled.address()); |
64 | uint32_t *CurAddress = FirstAddress + 1; |
65 | if (Enable) { |
66 | *CurAddress++ = 0x18000071; // ldr w17, #12 |
67 | *CurAddress = uint32_t(PatchOpcodes::PO_LdrX16_12); |
68 | CurAddress++; |
69 | *CurAddress = uint32_t(PatchOpcodes::PO_BlrX16); |
70 | CurAddress++; |
71 | *CurAddress = FuncId; |
72 | CurAddress++; |
73 | *reinterpret_cast<void (**)()>(CurAddress) = TracingHook; |
74 | CurAddress += 2; |
75 | *CurAddress = uint32_t(PatchOpcodes::PO_LdpX0X30SP_16); |
76 | CurAddress++; |
77 | std::atomic_store_explicit( |
78 | a: reinterpret_cast<std::atomic<uint32_t> *>(FirstAddress), |
79 | i: uint32_t(PatchOpcodes::PO_StpX0X30SP_m16e), m: std::memory_order_release); |
80 | } else { |
81 | std::atomic_store_explicit( |
82 | a: reinterpret_cast<std::atomic<uint32_t> *>(FirstAddress), |
83 | i: uint32_t(PatchOpcodes::PO_B32), m: std::memory_order_release); |
84 | } |
85 | __clear_cache(start: reinterpret_cast<char *>(FirstAddress), |
86 | end: reinterpret_cast<char *>(CurAddress)); |
87 | return true; |
88 | } |
89 | |
90 | bool patchFunctionEntry(const bool Enable, const uint32_t FuncId, |
91 | const XRaySledEntry &Sled, |
92 | void (*Trampoline)()) XRAY_NEVER_INSTRUMENT { |
93 | return patchSled(Enable, FuncId, Sled, TracingHook: Trampoline); |
94 | } |
95 | |
96 | bool patchFunctionExit(const bool Enable, const uint32_t FuncId, |
97 | const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT { |
98 | return patchSled(Enable, FuncId, Sled, TracingHook: __xray_FunctionExit); |
99 | } |
100 | |
101 | bool patchFunctionTailExit(const bool Enable, const uint32_t FuncId, |
102 | const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT { |
103 | return patchSled(Enable, FuncId, Sled, TracingHook: __xray_FunctionTailExit); |
104 | } |
105 | |
106 | // AArch64AsmPrinter::LowerPATCHABLE_EVENT_CALL generates this code sequence: |
107 | // |
108 | // .Lxray_event_sled_N: |
109 | // b 1f |
110 | // save x0 and x1 (and also x2 for TYPED_EVENT_CALL) |
111 | // set up x0 and x1 (and also x2 for TYPED_EVENT_CALL) |
112 | // bl __xray_CustomEvent or __xray_TypedEvent |
113 | // restore x0 and x1 (and also x2 for TYPED_EVENT_CALL) |
114 | // 1f |
115 | // |
116 | // There are 6 instructions for EVENT_CALL and 9 for TYPED_EVENT_CALL. |
117 | // |
118 | // Enable: b .+24 => nop |
119 | // Disable: nop => b .+24 |
120 | bool patchCustomEvent(const bool Enable, const uint32_t FuncId, |
121 | const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT { |
122 | uint32_t Inst = Enable ? 0xd503201f : 0x14000006; |
123 | std::atomic_store_explicit( |
124 | a: reinterpret_cast<std::atomic<uint32_t> *>(Sled.address()), i: Inst, |
125 | m: std::memory_order_release); |
126 | return false; |
127 | } |
128 | |
129 | // Enable: b +36 => nop |
130 | // Disable: nop => b +36 |
131 | bool patchTypedEvent(const bool Enable, const uint32_t FuncId, |
132 | const XRaySledEntry &Sled) XRAY_NEVER_INSTRUMENT { |
133 | uint32_t Inst = Enable ? 0xd503201f : 0x14000009; |
134 | std::atomic_store_explicit( |
135 | a: reinterpret_cast<std::atomic<uint32_t> *>(Sled.address()), i: Inst, |
136 | m: std::memory_order_release); |
137 | return false; |
138 | } |
139 | |
140 | // FIXME: Maybe implement this better? |
141 | bool probeRequiredCPUFeatures() XRAY_NEVER_INSTRUMENT { return true; } |
142 | |
143 | } // namespace __xray |
144 | |