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 * DEFINE_STATIC_CALL_RET0(name, typename); 21 * 22 * __static_call_return0; 23 * 24 * static_call(name)(args...); 25 * static_call_cond(name)(args...); 26 * static_call_update(name, func); 27 * static_call_query(name); 28 * 29 * EXPORT_STATIC_CALL{,_TRAMP}{,_GPL}() 30 * 31 * Usage example: 32 * 33 * # Start with the following functions (with identical prototypes): 34 * int func_a(int arg1, int arg2); 35 * int func_b(int arg1, int arg2); 36 * 37 * # Define a 'my_name' reference, associated with func_a() by default 38 * DEFINE_STATIC_CALL(my_name, func_a); 39 * 40 * # Call func_a() 41 * static_call(my_name)(arg1, arg2); 42 * 43 * # Update 'my_name' to point to func_b() 44 * static_call_update(my_name, &func_b); 45 * 46 * # Call func_b() 47 * static_call(my_name)(arg1, arg2); 48 * 49 * 50 * Implementation details: 51 * 52 * This requires some arch-specific code (CONFIG_HAVE_STATIC_CALL). 53 * Otherwise basic indirect calls are used (with function pointers). 54 * 55 * Each static_call() site calls into a trampoline associated with the name. 56 * The trampoline has a direct branch to the default function. Updates to a 57 * name will modify the trampoline's branch destination. 58 * 59 * If the arch has CONFIG_HAVE_STATIC_CALL_INLINE, then the call sites 60 * themselves will be patched at runtime to call the functions directly, 61 * rather than calling through the trampoline. This requires objtool or a 62 * compiler plugin to detect all the static_call() sites and annotate them 63 * in the .static_call_sites section. 64 * 65 * 66 * Notes on NULL function pointers: 67 * 68 * Static_call()s support NULL functions, with many of the caveats that 69 * regular function pointers have. 70 * 71 * Clearly calling a NULL function pointer is 'BAD', so too for 72 * static_call()s (although when HAVE_STATIC_CALL it might not be immediately 73 * fatal). A NULL static_call can be the result of: 74 * 75 * DECLARE_STATIC_CALL_NULL(my_static_call, void (*)(int)); 76 * 77 * which is equivalent to declaring a NULL function pointer with just a 78 * typename: 79 * 80 * void (*my_func_ptr)(int arg1) = NULL; 81 * 82 * or using static_call_update() with a NULL function. In both cases the 83 * HAVE_STATIC_CALL implementation will patch the trampoline with a RET 84 * instruction, instead of an immediate tail-call JMP. HAVE_STATIC_CALL_INLINE 85 * architectures can patch the trampoline call to a NOP. 86 * 87 * In all cases, any argument evaluation is unconditional. Unlike a regular 88 * conditional function pointer call: 89 * 90 * if (my_func_ptr) 91 * my_func_ptr(arg1) 92 * 93 * where the argument evaludation also depends on the pointer value. 94 * 95 * When calling a static_call that can be NULL, use: 96 * 97 * static_call_cond(name)(arg1); 98 * 99 * which will include the required value tests to avoid NULL-pointer 100 * dereferences. 101 * 102 * To query which function is currently set to be called, use: 103 * 104 * func = static_call_query(name); 105 * 106 * 107 * DEFINE_STATIC_CALL_RET0 / __static_call_return0: 108 * 109 * Just like how DEFINE_STATIC_CALL_NULL() / static_call_cond() optimize the 110 * conditional void function call, DEFINE_STATIC_CALL_RET0 / 111 * __static_call_return0 optimize the do nothing return 0 function. 112 * 113 * This feature is strictly UB per the C standard (since it casts a function 114 * pointer to a different signature) and relies on the architecture ABI to 115 * make things work. In particular it relies on Caller Stack-cleanup and the 116 * whole return register being clobbered for short return values. All normal 117 * CDECL style ABIs conform. 118 * 119 * In particular the x86_64 implementation replaces the 5 byte CALL 120 * instruction at the callsite with a 5 byte clear of the RAX register, 121 * completely eliding any function call overhead. 122 * 123 * Notably argument setup is unconditional. 124 * 125 * 126 * EXPORT_STATIC_CALL() vs EXPORT_STATIC_CALL_TRAMP(): 127 * 128 * The difference is that the _TRAMP variant tries to only export the 129 * trampoline with the result that a module can use static_call{,_cond}() but 130 * not static_call_update(). 131 * 132 */ 133 134#include <linux/types.h> 135#include <linux/cpu.h> 136#include <linux/static_call_types.h> 137 138#ifdef CONFIG_HAVE_STATIC_CALL 139#include <asm/static_call.h> 140 141/* 142 * Either @site or @tramp can be NULL. 143 */ 144extern void arch_static_call_transform(void *site, void *tramp, void *func, bool tail); 145 146#define STATIC_CALL_TRAMP_ADDR(name) &STATIC_CALL_TRAMP(name) 147 148#else 149#define STATIC_CALL_TRAMP_ADDR(name) NULL 150#endif 151 152#define static_call_update(name, func) \ 153({ \ 154 typeof(&STATIC_CALL_TRAMP(name)) __F = (func); \ 155 __static_call_update(&STATIC_CALL_KEY(name), \ 156 STATIC_CALL_TRAMP_ADDR(name), __F); \ 157}) 158 159#define static_call_query(name) (READ_ONCE(STATIC_CALL_KEY(name).func)) 160 161#ifdef CONFIG_HAVE_STATIC_CALL_INLINE 162 163extern int __init static_call_init(void); 164 165extern void static_call_force_reinit(void); 166 167struct static_call_mod { 168 struct static_call_mod *next; 169 struct module *mod; /* for vmlinux, mod == NULL */ 170 struct static_call_site *sites; 171}; 172 173/* For finding the key associated with a trampoline */ 174struct static_call_tramp_key { 175 s32 tramp; 176 s32 key; 177}; 178 179extern void __static_call_update(struct static_call_key *key, void *tramp, void *func); 180extern int static_call_mod_init(struct module *mod); 181extern int static_call_text_reserved(void *start, void *end); 182 183extern long __static_call_return0(void); 184 185#define DEFINE_STATIC_CALL(name, _func) \ 186 DECLARE_STATIC_CALL(name, _func); \ 187 struct static_call_key STATIC_CALL_KEY(name) = { \ 188 .func = _func, \ 189 .type = 1, \ 190 }; \ 191 ARCH_DEFINE_STATIC_CALL_TRAMP(name, _func) 192 193#define DEFINE_STATIC_CALL_NULL(name, _func) \ 194 DECLARE_STATIC_CALL(name, _func); \ 195 struct static_call_key STATIC_CALL_KEY(name) = { \ 196 .func = NULL, \ 197 .type = 1, \ 198 }; \ 199 ARCH_DEFINE_STATIC_CALL_NULL_TRAMP(name) 200 201#define DEFINE_STATIC_CALL_RET0(name, _func) \ 202 DECLARE_STATIC_CALL(name, _func); \ 203 struct static_call_key STATIC_CALL_KEY(name) = { \ 204 .func = __static_call_return0, \ 205 .type = 1, \ 206 }; \ 207 ARCH_DEFINE_STATIC_CALL_RET0_TRAMP(name) 208 209#define static_call_cond(name) (void)__static_call(name) 210 211#define EXPORT_STATIC_CALL(name) \ 212 EXPORT_SYMBOL(STATIC_CALL_KEY(name)); \ 213 EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)) 214#define EXPORT_STATIC_CALL_GPL(name) \ 215 EXPORT_SYMBOL_GPL(STATIC_CALL_KEY(name)); \ 216 EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)) 217 218/* Leave the key unexported, so modules can't change static call targets: */ 219#define EXPORT_STATIC_CALL_TRAMP(name) \ 220 EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)); \ 221 ARCH_ADD_TRAMP_KEY(name) 222#define EXPORT_STATIC_CALL_TRAMP_GPL(name) \ 223 EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)); \ 224 ARCH_ADD_TRAMP_KEY(name) 225 226#elif defined(CONFIG_HAVE_STATIC_CALL) 227 228static inline int static_call_init(void) { return 0; } 229 230#define DEFINE_STATIC_CALL(name, _func) \ 231 DECLARE_STATIC_CALL(name, _func); \ 232 struct static_call_key STATIC_CALL_KEY(name) = { \ 233 .func = _func, \ 234 }; \ 235 ARCH_DEFINE_STATIC_CALL_TRAMP(name, _func) 236 237#define DEFINE_STATIC_CALL_NULL(name, _func) \ 238 DECLARE_STATIC_CALL(name, _func); \ 239 struct static_call_key STATIC_CALL_KEY(name) = { \ 240 .func = NULL, \ 241 }; \ 242 ARCH_DEFINE_STATIC_CALL_NULL_TRAMP(name) 243 244#define DEFINE_STATIC_CALL_RET0(name, _func) \ 245 DECLARE_STATIC_CALL(name, _func); \ 246 struct static_call_key STATIC_CALL_KEY(name) = { \ 247 .func = __static_call_return0, \ 248 }; \ 249 ARCH_DEFINE_STATIC_CALL_RET0_TRAMP(name) 250 251#define static_call_cond(name) (void)__static_call(name) 252 253static inline 254void __static_call_update(struct static_call_key *key, void *tramp, void *func) 255{ 256 cpus_read_lock(); 257 WRITE_ONCE(key->func, func); 258 arch_static_call_transform(NULL, tramp, func, false); 259 cpus_read_unlock(); 260} 261 262static inline int static_call_text_reserved(void *start, void *end) 263{ 264 return 0; 265} 266 267extern long __static_call_return0(void); 268 269#define EXPORT_STATIC_CALL(name) \ 270 EXPORT_SYMBOL(STATIC_CALL_KEY(name)); \ 271 EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)) 272#define EXPORT_STATIC_CALL_GPL(name) \ 273 EXPORT_SYMBOL_GPL(STATIC_CALL_KEY(name)); \ 274 EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)) 275 276/* Leave the key unexported, so modules can't change static call targets: */ 277#define EXPORT_STATIC_CALL_TRAMP(name) \ 278 EXPORT_SYMBOL(STATIC_CALL_TRAMP(name)) 279#define EXPORT_STATIC_CALL_TRAMP_GPL(name) \ 280 EXPORT_SYMBOL_GPL(STATIC_CALL_TRAMP(name)) 281 282#else /* Generic implementation */ 283 284static inline int static_call_init(void) { return 0; } 285 286static inline long __static_call_return0(void) 287{ 288 return 0; 289} 290 291#define __DEFINE_STATIC_CALL(name, _func, _func_init) \ 292 DECLARE_STATIC_CALL(name, _func); \ 293 struct static_call_key STATIC_CALL_KEY(name) = { \ 294 .func = _func_init, \ 295 } 296 297#define DEFINE_STATIC_CALL(name, _func) \ 298 __DEFINE_STATIC_CALL(name, _func, _func) 299 300#define DEFINE_STATIC_CALL_NULL(name, _func) \ 301 __DEFINE_STATIC_CALL(name, _func, NULL) 302 303#define DEFINE_STATIC_CALL_RET0(name, _func) \ 304 __DEFINE_STATIC_CALL(name, _func, __static_call_return0) 305 306static inline void __static_call_nop(void) { } 307 308/* 309 * This horrific hack takes care of two things: 310 * 311 * - it ensures the compiler will only load the function pointer ONCE, 312 * which avoids a reload race. 313 * 314 * - it ensures the argument evaluation is unconditional, similar 315 * to the HAVE_STATIC_CALL variant. 316 * 317 * Sadly current GCC/Clang (10 for both) do not optimize this properly 318 * and will emit an indirect call for the NULL case :-( 319 */ 320#define __static_call_cond(name) \ 321({ \ 322 void *func = READ_ONCE(STATIC_CALL_KEY(name).func); \ 323 if (!func) \ 324 func = &__static_call_nop; \ 325 (typeof(STATIC_CALL_TRAMP(name))*)func; \ 326}) 327 328#define static_call_cond(name) (void)__static_call_cond(name) 329 330static inline 331void __static_call_update(struct static_call_key *key, void *tramp, void *func) 332{ 333 WRITE_ONCE(key->func, func); 334} 335 336static inline int static_call_text_reserved(void *start, void *end) 337{ 338 return 0; 339} 340 341#define EXPORT_STATIC_CALL(name) EXPORT_SYMBOL(STATIC_CALL_KEY(name)) 342#define EXPORT_STATIC_CALL_GPL(name) EXPORT_SYMBOL_GPL(STATIC_CALL_KEY(name)) 343 344#endif /* CONFIG_HAVE_STATIC_CALL */ 345 346#endif /* _LINUX_STATIC_CALL_H */ 347