Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame^] | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
| 2 | #ifndef _LINUX_STATIC_CALL_H |
| 3 | #define _LINUX_STATIC_CALL_H |
| 4 | |
| 5 | /* |
| 6 | * Static call support |
| 7 | * |
| 8 | * Static calls use code patching to hard-code function pointers into direct |
| 9 | * branch instructions. They give the flexibility of function pointers, but |
| 10 | * with improved performance. This is especially important for cases where |
| 11 | * retpolines would otherwise be used, as retpolines can significantly impact |
| 12 | * performance. |
| 13 | * |
| 14 | * |
| 15 | * API overview: |
| 16 | * |
| 17 | * DECLARE_STATIC_CALL(name, func); |
| 18 | * DEFINE_STATIC_CALL(name, func); |
| 19 | * DEFINE_STATIC_CALL_NULL(name, typename); |
| 20 | * static_call(name)(args...); |
| 21 | * static_call_cond(name)(args...); |
| 22 | * static_call_update(name, func); |
| 23 | * |
| 24 | * Usage example: |
| 25 | * |
| 26 | * # Start with the following functions (with identical prototypes): |
| 27 | * int func_a(int arg1, int arg2); |
| 28 | * int func_b(int arg1, int arg2); |
| 29 | * |
| 30 | * # Define a 'my_name' reference, associated with func_a() by default |
| 31 | * DEFINE_STATIC_CALL(my_name, func_a); |
| 32 | * |
| 33 | * # Call func_a() |
| 34 | * static_call(my_name)(arg1, arg2); |
| 35 | * |
| 36 | * # Update 'my_name' to point to func_b() |
| 37 | * static_call_update(my_name, &func_b); |
| 38 | * |
| 39 | * # Call func_b() |
| 40 | * static_call(my_name)(arg1, arg2); |
| 41 | * |
| 42 | * |
| 43 | * Implementation details: |
| 44 | * |
| 45 | * This requires some arch-specific code (CONFIG_HAVE_STATIC_CALL). |
| 46 | * Otherwise basic indirect calls are used (with function pointers). |
| 47 | * |
| 48 | * Each static_call() site calls into a trampoline associated with the name. |
| 49 | * The trampoline has a direct branch to the default function. Updates to a |
| 50 | * name will modify the trampoline's branch destination. |
| 51 | * |
| 52 | * If the arch has CONFIG_HAVE_STATIC_CALL_INLINE, then the call sites |
| 53 | * themselves will be patched at runtime to call the functions directly, |
| 54 | * rather than calling through the trampoline. This requires objtool or a |
| 55 | * compiler plugin to detect all the static_call() sites and annotate them |
| 56 | * in the .static_call_sites section. |
| 57 | * |
| 58 | * |
| 59 | * Notes on NULL function pointers: |
| 60 | * |
| 61 | * Static_call()s support NULL functions, with many of the caveats that |
| 62 | * regular function pointers have. |
| 63 | * |
| 64 | * Clearly calling a NULL function pointer is 'BAD', so too for |
| 65 | * static_call()s (although when HAVE_STATIC_CALL it might not be immediately |
| 66 | * fatal). A NULL static_call can be the result of: |
| 67 | * |
| 68 | * DECLARE_STATIC_CALL_NULL(my_static_call, void (*)(int)); |
| 69 | * |
| 70 | * which is equivalent to declaring a NULL function pointer with just a |
| 71 | * typename: |
| 72 | * |
| 73 | * void (*my_func_ptr)(int arg1) = NULL; |
| 74 | * |
| 75 | * or using static_call_update() with a NULL function. In both cases the |
| 76 | * HAVE_STATIC_CALL implementation will patch the trampoline with a RET |
| 77 | * instruction, instead of an immediate tail-call JMP. HAVE_STATIC_CALL_INLINE |
| 78 | * architectures can patch the trampoline call to a NOP. |
| 79 | * |
| 80 | * In all cases, any argument evaluation is unconditional. Unlike a regular |
| 81 | * conditional function pointer call: |
| 82 | * |
| 83 | * if (my_func_ptr) |
| 84 | * my_func_ptr(arg1) |
| 85 | * |
| 86 | * where the argument evaludation also depends on the pointer value. |
| 87 | * |
| 88 | * When calling a static_call that can be NULL, use: |
| 89 | * |
| 90 | * static_call_cond(name)(arg1); |
| 91 | * |
| 92 | * which will include the required value tests to avoid NULL-pointer |
| 93 | * dereferences. |
| 94 | */ |
| 95 | |
| 96 | #include <linux/types.h> |
| 97 | #include <linux/cpu.h> |
| 98 | #include <linux/static_call_types.h> |
| 99 | |
| 100 | #ifdef CONFIG_HAVE_STATIC_CALL |
| 101 | #include <asm/static_call.h> |
| 102 | |
| 103 | /* |
| 104 | * Either @site or @tramp can be NULL. |
| 105 | */ |
| 106 | extern void arch_static_call_transform(void *site, void *tramp, void *func, bool tail); |
| 107 | |
| 108 | #define STATIC_CALL_TRAMP_ADDR(name) &STATIC_CALL_TRAMP(name) |
| 109 | |
| 110 | #else |
| 111 | #define STATIC_CALL_TRAMP_ADDR(name) NULL |
| 112 | #endif |
| 113 | |
| 114 | #define static_call_update(name, func) \ |
| 115 | ({ \ |
| 116 | BUILD_BUG_ON(!__same_type(*(func), STATIC_CALL_TRAMP(name))); \ |
| 117 | __static_call_update(&STATIC_CALL_KEY(name), \ |
| 118 | STATIC_CALL_TRAMP_ADDR(name), func); \ |
| 119 | }) |
| 120 | |
| 121 | #ifdef CONFIG_HAVE_STATIC_CALL_INLINE |
| 122 | |
| 123 | extern int __init static_call_init(void); |
| 124 | |
| 125 | struct static_call_mod { |
| 126 | struct static_call_mod *next; |
| 127 | struct module *mod; /* for vmlinux, mod == NULL */ |
| 128 | struct static_call_site *sites; |
| 129 | }; |
| 130 | |
| 131 | struct static_call_key { |
| 132 | void *func; |
| 133 | union { |
| 134 | /* bit 0: 0 = mods, 1 = sites */ |
| 135 | unsigned long type; |
| 136 | struct static_call_mod *mods; |
| 137 | struct static_call_site *sites; |
| 138 | }; |
| 139 | }; |
| 140 | |
| 141 | /* For finding the key associated with a trampoline */ |
| 142 | struct static_call_tramp_key { |
| 143 | s32 tramp; |
| 144 | s32 key; |
| 145 | }; |
| 146 | |
| 147 | extern void __static_call_update(struct static_call_key *key, void *tramp, void *func); |
| 148 | extern int static_call_mod_init(struct module *mod); |
| 149 | extern int static_call_text_reserved(void *start, void *end); |
| 150 | |
| 151 | #define DEFINE_STATIC_CALL(name, _func) \ |
| 152 | DECLARE_STATIC_CALL(name, _func); \ |
| 153 | struct static_call_key STATIC_CALL_KEY(name) = { \ |
| 154 | .func = _func, \ |
| 155 | .type = 1, \ |
| 156 | }; \ |
| 157 | ARCH_DEFINE_STATIC_CALL_TRAMP(name, _func) |
| 158 | |
| 159 | #define DEFINE_STATIC_CALL_NULL(name, _func) \ |
| 160 | DECLARE_STATIC_CALL(name, _func); \ |
| 161 | struct static_call_key STATIC_CALL_KEY(name) = { \ |
| 162 | .func = NULL, \ |
| 163 | .type = 1, \ |
| 164 | }; \ |
| 165 | ARCH_DEFINE_STATIC_CALL_NULL_TRAMP(name) |
| 166 | |
| 167 | #define static_call_cond(name) (void)__static_call(name) |
| 168 | |
| 169 | #define EXPORT_STATIC_CALL(name) \ |
| 170 | EXPORT_SYMBOL(STATIC_CALL_KEY(name)); \ |
| 171 | EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)) |
| 172 | #define EXPORT_STATIC_CALL_GPL(name) \ |
| 173 | EXPORT_SYMBOL_GPL(STATIC_CALL_KEY(name)); \ |
| 174 | EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)) |
| 175 | |
| 176 | /* Leave the key unexported, so modules can't change static call targets: */ |
| 177 | #define EXPORT_STATIC_CALL_TRAMP(name) \ |
| 178 | EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)); \ |
| 179 | ARCH_ADD_TRAMP_KEY(name) |
| 180 | #define EXPORT_STATIC_CALL_TRAMP_GPL(name) \ |
| 181 | EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)); \ |
| 182 | ARCH_ADD_TRAMP_KEY(name) |
| 183 | |
| 184 | #elif defined(CONFIG_HAVE_STATIC_CALL) |
| 185 | |
| 186 | static inline int static_call_init(void) { return 0; } |
| 187 | |
| 188 | struct static_call_key { |
| 189 | void *func; |
| 190 | }; |
| 191 | |
| 192 | #define DEFINE_STATIC_CALL(name, _func) \ |
| 193 | DECLARE_STATIC_CALL(name, _func); \ |
| 194 | struct static_call_key STATIC_CALL_KEY(name) = { \ |
| 195 | .func = _func, \ |
| 196 | }; \ |
| 197 | ARCH_DEFINE_STATIC_CALL_TRAMP(name, _func) |
| 198 | |
| 199 | #define DEFINE_STATIC_CALL_NULL(name, _func) \ |
| 200 | DECLARE_STATIC_CALL(name, _func); \ |
| 201 | struct static_call_key STATIC_CALL_KEY(name) = { \ |
| 202 | .func = NULL, \ |
| 203 | }; \ |
| 204 | ARCH_DEFINE_STATIC_CALL_NULL_TRAMP(name) |
| 205 | |
| 206 | #define static_call_cond(name) (void)__static_call(name) |
| 207 | |
| 208 | static inline |
| 209 | void __static_call_update(struct static_call_key *key, void *tramp, void *func) |
| 210 | { |
| 211 | cpus_read_lock(); |
| 212 | WRITE_ONCE(key->func, func); |
| 213 | arch_static_call_transform(NULL, tramp, func, false); |
| 214 | cpus_read_unlock(); |
| 215 | } |
| 216 | |
| 217 | static inline int static_call_text_reserved(void *start, void *end) |
| 218 | { |
| 219 | return 0; |
| 220 | } |
| 221 | |
| 222 | #define EXPORT_STATIC_CALL(name) \ |
| 223 | EXPORT_SYMBOL(STATIC_CALL_KEY(name)); \ |
| 224 | EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)) |
| 225 | #define EXPORT_STATIC_CALL_GPL(name) \ |
| 226 | EXPORT_SYMBOL_GPL(STATIC_CALL_KEY(name)); \ |
| 227 | EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)) |
| 228 | |
| 229 | /* Leave the key unexported, so modules can't change static call targets: */ |
| 230 | #define EXPORT_STATIC_CALL_TRAMP(name) \ |
| 231 | EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)) |
| 232 | #define EXPORT_STATIC_CALL_TRAMP_GPL(name) \ |
| 233 | EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)) |
| 234 | |
| 235 | #else /* Generic implementation */ |
| 236 | |
| 237 | static inline int static_call_init(void) { return 0; } |
| 238 | |
| 239 | struct static_call_key { |
| 240 | void *func; |
| 241 | }; |
| 242 | |
| 243 | #define DEFINE_STATIC_CALL(name, _func) \ |
| 244 | DECLARE_STATIC_CALL(name, _func); \ |
| 245 | struct static_call_key STATIC_CALL_KEY(name) = { \ |
| 246 | .func = _func, \ |
| 247 | } |
| 248 | |
| 249 | #define DEFINE_STATIC_CALL_NULL(name, _func) \ |
| 250 | DECLARE_STATIC_CALL(name, _func); \ |
| 251 | struct static_call_key STATIC_CALL_KEY(name) = { \ |
| 252 | .func = NULL, \ |
| 253 | } |
| 254 | |
| 255 | static inline void __static_call_nop(void) { } |
| 256 | |
| 257 | /* |
| 258 | * This horrific hack takes care of two things: |
| 259 | * |
| 260 | * - it ensures the compiler will only load the function pointer ONCE, |
| 261 | * which avoids a reload race. |
| 262 | * |
| 263 | * - it ensures the argument evaluation is unconditional, similar |
| 264 | * to the HAVE_STATIC_CALL variant. |
| 265 | * |
| 266 | * Sadly current GCC/Clang (10 for both) do not optimize this properly |
| 267 | * and will emit an indirect call for the NULL case :-( |
| 268 | */ |
| 269 | #define __static_call_cond(name) \ |
| 270 | ({ \ |
| 271 | void *func = READ_ONCE(STATIC_CALL_KEY(name).func); \ |
| 272 | if (!func) \ |
| 273 | func = &__static_call_nop; \ |
| 274 | (typeof(STATIC_CALL_TRAMP(name))*)func; \ |
| 275 | }) |
| 276 | |
| 277 | #define static_call_cond(name) (void)__static_call_cond(name) |
| 278 | |
| 279 | static inline |
| 280 | void __static_call_update(struct static_call_key *key, void *tramp, void *func) |
| 281 | { |
| 282 | WRITE_ONCE(key->func, func); |
| 283 | } |
| 284 | |
| 285 | static inline int static_call_text_reserved(void *start, void *end) |
| 286 | { |
| 287 | return 0; |
| 288 | } |
| 289 | |
| 290 | #define EXPORT_STATIC_CALL(name) EXPORT_SYMBOL(STATIC_CALL_KEY(name)) |
| 291 | #define EXPORT_STATIC_CALL_GPL(name) EXPORT_SYMBOL_GPL(STATIC_CALL_KEY(name)) |
| 292 | |
| 293 | #endif /* CONFIG_HAVE_STATIC_CALL */ |
| 294 | |
| 295 | #endif /* _LINUX_STATIC_CALL_H */ |