Commit | Line | Data |
---|---|---|
90702366 | 1 | /* SPDX-License-Identifier: MIT */ |
21b58a3b | 2 | /* SPDX-FileCopyrightText: 2016-2024 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> */ |
f2d7b530 MJ |
3 | /* SPDX-FileCopyrightText: 2018 Will Deacon <will.deacon@arm.com> */ |
4 | ||
d78a16c2 | 5 | /* |
44ec21eb | 6 | * rseq/arch/aarch64.h |
d78a16c2 MD |
7 | */ |
8 | ||
44ec21eb MJ |
9 | #ifndef _RSEQ_RSEQ_H |
10 | #error "Never use <rseq/arch/aarch64.h> directly; include <rseq/rseq.h> instead." | |
11 | #endif | |
12 | ||
21b58a3b MD |
13 | /* |
14 | * RSEQ_ASM_*() macro helpers are internal to the librseq headers. Those | |
15 | * are not part of the public API. | |
16 | */ | |
17 | ||
8b7c64f7 MD |
18 | /* |
19 | * aarch64 -mbig-endian generates mixed endianness code vs data: | |
20 | * little-endian code and big-endian data. Ensure the RSEQ_SIG signature | |
21 | * matches code endianness. | |
22 | */ | |
23 | #define RSEQ_SIG_CODE 0xd428bc00 /* BRK #0x45E0. */ | |
24 | ||
dbbd006a MD |
25 | #ifdef __AARCH64EB__ /* Big endian */ |
26 | # define RSEQ_SIG_DATA 0x00bc28d4 /* BRK #0x45E0. */ | |
27 | #else /* Little endian */ | |
28 | # define RSEQ_SIG_DATA RSEQ_SIG_CODE | |
8b7c64f7 MD |
29 | #endif |
30 | ||
31 | #define RSEQ_SIG RSEQ_SIG_DATA | |
d78a16c2 | 32 | |
21b58a3b MD |
33 | /* |
34 | * Refer to the Linux kernel memory model (LKMM) for documentation of | |
35 | * the memory barriers. | |
36 | */ | |
37 | ||
38 | /* CPU memory barrier. */ | |
d78a16c2 | 39 | #define rseq_smp_mb() __asm__ __volatile__ ("dmb ish" ::: "memory") |
21b58a3b | 40 | /* CPU read memory barrier */ |
d78a16c2 | 41 | #define rseq_smp_rmb() __asm__ __volatile__ ("dmb ishld" ::: "memory") |
21b58a3b | 42 | /* CPU write memory barrier */ |
d78a16c2 MD |
43 | #define rseq_smp_wmb() __asm__ __volatile__ ("dmb ishst" ::: "memory") |
44 | ||
21b58a3b | 45 | /* Acquire: One-way permeable barrier. */ |
d78a16c2 MD |
46 | #define rseq_smp_load_acquire(p) \ |
47 | __extension__ ({ \ | |
18ca65e0 MD |
48 | union { rseq_unqual_scalar_typeof(*(p)) __val; char __c[sizeof(*(p))]; } __u; \ |
49 | switch (sizeof(*(p))) { \ | |
d78a16c2 | 50 | case 1: \ |
18ca65e0 MD |
51 | __asm__ __volatile__ ("ldarb %w0, %1" \ |
52 | : "=r" (*(__u8 *)__u.__c) \ | |
53 | : "Q" (*(p)) : "memory"); \ | |
d78a16c2 MD |
54 | break; \ |
55 | case 2: \ | |
18ca65e0 MD |
56 | __asm__ __volatile__ ("ldarh %w0, %1" \ |
57 | : "=r" (*(__u16 *)__u.__c) \ | |
58 | : "Q" (*(p)) : "memory"); \ | |
d78a16c2 MD |
59 | break; \ |
60 | case 4: \ | |
18ca65e0 MD |
61 | __asm__ __volatile__ ("ldar %w0, %1" \ |
62 | : "=r" (*(__u32 *)__u.__c) \ | |
63 | : "Q" (*(p)) : "memory"); \ | |
d78a16c2 MD |
64 | break; \ |
65 | case 8: \ | |
18ca65e0 MD |
66 | __asm__ __volatile__ ("ldar %0, %1" \ |
67 | : "=r" (*(__u64 *)__u.__c) \ | |
68 | : "Q" (*(p)) : "memory"); \ | |
d78a16c2 MD |
69 | break; \ |
70 | } \ | |
0aac367c | 71 | (rseq_unqual_scalar_typeof(*(p)))__u.__val; \ |
d78a16c2 MD |
72 | }) |
73 | ||
21b58a3b | 74 | /* Acquire barrier after control dependency. */ |
d78a16c2 MD |
75 | #define rseq_smp_acquire__after_ctrl_dep() rseq_smp_rmb() |
76 | ||
21b58a3b | 77 | /* Release: One-way permeable barrier. */ |
d78a16c2 MD |
78 | #define rseq_smp_store_release(p, v) \ |
79 | do { \ | |
18ca65e0 MD |
80 | union { rseq_unqual_scalar_typeof(*(p)) __val; char __c[sizeof(*(p))]; } __u = \ |
81 | { .__val = (rseq_unqual_scalar_typeof(*(p))) (v) }; \ | |
82 | switch (sizeof(*(p))) { \ | |
d78a16c2 | 83 | case 1: \ |
18ca65e0 MD |
84 | __asm__ __volatile__ ("stlrb %w1, %0" \ |
85 | : "=Q" (*(p)) \ | |
86 | : "r" (*(__u8 *)__u.__c) \ | |
d78a16c2 MD |
87 | : "memory"); \ |
88 | break; \ | |
89 | case 2: \ | |
18ca65e0 MD |
90 | __asm__ __volatile__ ("stlrh %w1, %0" \ |
91 | : "=Q" (*(p)) \ | |
92 | : "r" (*(__u16 *)__u.__c) \ | |
d78a16c2 MD |
93 | : "memory"); \ |
94 | break; \ | |
95 | case 4: \ | |
18ca65e0 MD |
96 | __asm__ __volatile__ ("stlr %w1, %0" \ |
97 | : "=Q" (*(p)) \ | |
98 | : "r" (*(__u32 *)__u.__c) \ | |
d78a16c2 MD |
99 | : "memory"); \ |
100 | break; \ | |
101 | case 8: \ | |
18ca65e0 MD |
102 | __asm__ __volatile__ ("stlr %1, %0" \ |
103 | : "=Q" (*(p)) \ | |
104 | : "r" (*(__u64 *)__u.__c) \ | |
d78a16c2 MD |
105 | : "memory"); \ |
106 | break; \ | |
107 | } \ | |
108 | } while (0) | |
109 | ||
9ed48921 | 110 | #define RSEQ_ASM_U64_PTR(x) ".quad " x |
0f986786 | 111 | #define RSEQ_ASM_U32(x) ".long " x |
9ed48921 | 112 | |
21b58a3b | 113 | /* Temporary scratch registers. */ |
d78a16c2 MD |
114 | #define RSEQ_ASM_TMP_REG32 "w15" |
115 | #define RSEQ_ASM_TMP_REG "x15" | |
116 | #define RSEQ_ASM_TMP_REG_2 "x14" | |
117 | ||
ed21bf6d MD |
118 | /* Common architecture support macros. */ |
119 | #include "rseq/arch/generic/common.h" | |
90d9876e | 120 | |
21b58a3b MD |
121 | /* |
122 | * Define a critical section abort handler. | |
123 | * | |
124 | * @label: | |
125 | * Local label to the abort handler. | |
126 | * @teardown: | |
127 | * Sequence of instructions to run on abort. | |
128 | * @abort_label: | |
129 | * C label to jump to at the end of the sequence. | |
130 | */ | |
ebd27573 | 131 | #define RSEQ_ASM_DEFINE_ABORT(label, teardown, abort_label) \ |
d78a16c2 | 132 | " b 222f\n" \ |
8b7c64f7 | 133 | " .inst " __rseq_str(RSEQ_SIG_CODE) "\n" \ |
d78a16c2 | 134 | __rseq_str(label) ":\n" \ |
ebd27573 | 135 | teardown \ |
d78a16c2 MD |
136 | " b %l[" __rseq_str(abort_label) "]\n" \ |
137 | "222:\n" | |
138 | ||
21b58a3b MD |
139 | /* Jump to local label @label when @cpu_id != @current_cpu_id. */ |
140 | #define RSEQ_ASM_STORE_RSEQ_CS(label, cs_label, rseq_cs) \ | |
141 | RSEQ_INJECT_ASM(1) \ | |
142 | " adrp " RSEQ_ASM_TMP_REG ", " __rseq_str(cs_label) "\n" \ | |
143 | " add " RSEQ_ASM_TMP_REG ", " RSEQ_ASM_TMP_REG \ | |
144 | ", :lo12:" __rseq_str(cs_label) "\n" \ | |
145 | " str " RSEQ_ASM_TMP_REG ", %[" __rseq_str(rseq_cs) "]\n" \ | |
146 | __rseq_str(label) ":\n" | |
147 | ||
148 | /* Store @value to address @var. */ | |
d78a16c2 MD |
149 | #define RSEQ_ASM_OP_STORE(value, var) \ |
150 | " str %[" __rseq_str(value) "], %[" __rseq_str(var) "]\n" | |
151 | ||
21b58a3b | 152 | /* Store-release @value to address @var. */ |
d78a16c2 MD |
153 | #define RSEQ_ASM_OP_STORE_RELEASE(value, var) \ |
154 | " stlr %[" __rseq_str(value) "], %[" __rseq_str(var) "]\n" | |
155 | ||
21b58a3b MD |
156 | /* |
157 | * End-of-sequence store of @value to address @var. Emit | |
158 | * @post_commit_label label after the store instruction. | |
159 | */ | |
d78a16c2 MD |
160 | #define RSEQ_ASM_OP_FINAL_STORE(value, var, post_commit_label) \ |
161 | RSEQ_ASM_OP_STORE(value, var) \ | |
162 | __rseq_str(post_commit_label) ":\n" | |
163 | ||
21b58a3b MD |
164 | /* |
165 | * End-of-sequence store-release of @value to address @var. Emit | |
166 | * @post_commit_label label after the store instruction. | |
167 | */ | |
d78a16c2 MD |
168 | #define RSEQ_ASM_OP_FINAL_STORE_RELEASE(value, var, post_commit_label) \ |
169 | RSEQ_ASM_OP_STORE_RELEASE(value, var) \ | |
170 | __rseq_str(post_commit_label) ":\n" | |
171 | ||
21b58a3b | 172 | /* Jump to local label @label when @var != @expect. */ |
769ec9a5 | 173 | #define RSEQ_ASM_OP_CBNE(var, expect, label) \ |
d78a16c2 MD |
174 | " ldr " RSEQ_ASM_TMP_REG ", %[" __rseq_str(var) "]\n" \ |
175 | " sub " RSEQ_ASM_TMP_REG ", " RSEQ_ASM_TMP_REG \ | |
176 | ", %[" __rseq_str(expect) "]\n" \ | |
177 | " cbnz " RSEQ_ASM_TMP_REG ", " __rseq_str(label) "\n" | |
178 | ||
21b58a3b MD |
179 | /* |
180 | * Jump to local label @label when @var != @expect (32-bit register | |
181 | * comparison). | |
182 | */ | |
769ec9a5 | 183 | #define RSEQ_ASM_OP_CBNE32(var, expect, label) \ |
d78a16c2 MD |
184 | " ldr " RSEQ_ASM_TMP_REG32 ", %[" __rseq_str(var) "]\n" \ |
185 | " sub " RSEQ_ASM_TMP_REG32 ", " RSEQ_ASM_TMP_REG32 \ | |
186 | ", %w[" __rseq_str(expect) "]\n" \ | |
187 | " cbnz " RSEQ_ASM_TMP_REG32 ", " __rseq_str(label) "\n" | |
188 | ||
21b58a3b | 189 | /* Jump to local label @label when @var == @expect. */ |
769ec9a5 | 190 | #define RSEQ_ASM_OP_CBEQ(var, expect, label) \ |
d78a16c2 MD |
191 | " ldr " RSEQ_ASM_TMP_REG ", %[" __rseq_str(var) "]\n" \ |
192 | " sub " RSEQ_ASM_TMP_REG ", " RSEQ_ASM_TMP_REG \ | |
193 | ", %[" __rseq_str(expect) "]\n" \ | |
194 | " cbz " RSEQ_ASM_TMP_REG ", " __rseq_str(label) "\n" | |
195 | ||
21b58a3b | 196 | /* Jump to local label @label when @cpu_id != @current_cpu_id. */ |
769ec9a5 | 197 | #define RSEQ_ASM_CBNE_CPU_ID(cpu_id, current_cpu_id, label) \ |
d78a16c2 | 198 | RSEQ_INJECT_ASM(2) \ |
769ec9a5 | 199 | RSEQ_ASM_OP_CBNE32(current_cpu_id, cpu_id, label) |
d78a16c2 | 200 | |
21b58a3b | 201 | /* Load @var into temporary register. */ |
d78a16c2 MD |
202 | #define RSEQ_ASM_OP_R_LOAD(var) \ |
203 | " ldr " RSEQ_ASM_TMP_REG ", %[" __rseq_str(var) "]\n" | |
204 | ||
21b58a3b | 205 | /* Store from temporary register into @var. */ |
d78a16c2 MD |
206 | #define RSEQ_ASM_OP_R_STORE(var) \ |
207 | " str " RSEQ_ASM_TMP_REG ", %[" __rseq_str(var) "]\n" | |
208 | ||
21b58a3b | 209 | /* Load from address in temporary register+@offset into temporary register. */ |
d78a16c2 MD |
210 | #define RSEQ_ASM_OP_R_LOAD_OFF(offset) \ |
211 | " ldr " RSEQ_ASM_TMP_REG ", [" RSEQ_ASM_TMP_REG \ | |
212 | ", %[" __rseq_str(offset) "]]\n" | |
213 | ||
21b58a3b | 214 | /* Add @count to temporary register. */ |
d78a16c2 MD |
215 | #define RSEQ_ASM_OP_R_ADD(count) \ |
216 | " add " RSEQ_ASM_TMP_REG ", " RSEQ_ASM_TMP_REG \ | |
217 | ", %[" __rseq_str(count) "]\n" | |
218 | ||
21b58a3b MD |
219 | /* |
220 | * End-of-sequence store of temporary register to address @var. Emit | |
221 | * @post_commit_label label after the store instruction. | |
222 | */ | |
d78a16c2 MD |
223 | #define RSEQ_ASM_OP_R_FINAL_STORE(var, post_commit_label) \ |
224 | " str " RSEQ_ASM_TMP_REG ", %[" __rseq_str(var) "]\n" \ | |
225 | __rseq_str(post_commit_label) ":\n" | |
226 | ||
21b58a3b MD |
227 | /* |
228 | * Copy @len bytes from @src to @dst. This is an inefficient bytewise | |
229 | * copy and could be improved in the future. | |
230 | */ | |
6020ff66 | 231 | #define RSEQ_ASM_OP_R_BYTEWISE_MEMCPY(dst, src, len) \ |
d78a16c2 MD |
232 | " cbz %[" __rseq_str(len) "], 333f\n" \ |
233 | " mov " RSEQ_ASM_TMP_REG_2 ", %[" __rseq_str(len) "]\n" \ | |
234 | "222: sub " RSEQ_ASM_TMP_REG_2 ", " RSEQ_ASM_TMP_REG_2 ", #1\n" \ | |
235 | " ldrb " RSEQ_ASM_TMP_REG32 ", [%[" __rseq_str(src) "]" \ | |
236 | ", " RSEQ_ASM_TMP_REG_2 "]\n" \ | |
237 | " strb " RSEQ_ASM_TMP_REG32 ", [%[" __rseq_str(dst) "]" \ | |
238 | ", " RSEQ_ASM_TMP_REG_2 "]\n" \ | |
239 | " cbnz " RSEQ_ASM_TMP_REG_2 ", 222b\n" \ | |
240 | "333:\n" | |
241 | ||
a98a3b6d | 242 | /* Per-cpu-id indexing. */ |
d78a16c2 | 243 | |
abf9e855 | 244 | #define RSEQ_TEMPLATE_INDEX_CPU_ID |
a98a3b6d | 245 | #define RSEQ_TEMPLATE_MO_RELAXED |
44ec21eb | 246 | #include "rseq/arch/aarch64/bits.h" |
a98a3b6d | 247 | #undef RSEQ_TEMPLATE_MO_RELAXED |
d78a16c2 | 248 | |
a98a3b6d | 249 | #define RSEQ_TEMPLATE_MO_RELEASE |
44ec21eb | 250 | #include "rseq/arch/aarch64/bits.h" |
a98a3b6d | 251 | #undef RSEQ_TEMPLATE_MO_RELEASE |
abf9e855 | 252 | #undef RSEQ_TEMPLATE_INDEX_CPU_ID |
a98a3b6d MD |
253 | |
254 | /* Per-mm-cid indexing. */ | |
255 | ||
abf9e855 | 256 | #define RSEQ_TEMPLATE_INDEX_MM_CID |
a98a3b6d | 257 | #define RSEQ_TEMPLATE_MO_RELAXED |
44ec21eb | 258 | #include "rseq/arch/aarch64/bits.h" |
a98a3b6d MD |
259 | #undef RSEQ_TEMPLATE_MO_RELAXED |
260 | ||
261 | #define RSEQ_TEMPLATE_MO_RELEASE | |
44ec21eb | 262 | #include "rseq/arch/aarch64/bits.h" |
a98a3b6d | 263 | #undef RSEQ_TEMPLATE_MO_RELEASE |
abf9e855 | 264 | #undef RSEQ_TEMPLATE_INDEX_MM_CID |
a98a3b6d | 265 | |
abf9e855 | 266 | /* APIs which are not indexed. */ |
a98a3b6d | 267 | |
abf9e855 | 268 | #define RSEQ_TEMPLATE_INDEX_NONE |
a98a3b6d | 269 | #define RSEQ_TEMPLATE_MO_RELAXED |
44ec21eb | 270 | #include "rseq/arch/aarch64/bits.h" |
a98a3b6d | 271 | #undef RSEQ_TEMPLATE_MO_RELAXED |
abf9e855 | 272 | #undef RSEQ_TEMPLATE_INDEX_NONE |