| 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
| 2 | /* |
| 3 | * Variant of atomic_t specialized for reference counts. |
| 4 | * |
| 5 | * The interface matches the atomic_t interface (to aid in porting) but only |
| 6 | * provides the few functions one should use for reference counting. |
| 7 | * |
| 8 | * Saturation semantics |
| 9 | * ==================== |
| 10 | * |
| 11 | * refcount_t differs from atomic_t in that the counter saturates at |
| 12 | * REFCOUNT_SATURATED and will not move once there. This avoids wrapping the |
| 13 | * counter and causing 'spurious' use-after-free issues. In order to avoid the |
| 14 | * cost associated with introducing cmpxchg() loops into all of the saturating |
| 15 | * operations, we temporarily allow the counter to take on an unchecked value |
| 16 | * and then explicitly set it to REFCOUNT_SATURATED on detecting that underflow |
| 17 | * or overflow has occurred. Although this is racy when multiple threads |
| 18 | * access the refcount concurrently, by placing REFCOUNT_SATURATED roughly |
| 19 | * equidistant from 0 and INT_MAX we minimise the scope for error: |
| 20 | * |
| 21 | * INT_MAX REFCOUNT_SATURATED UINT_MAX |
| 22 | * 0 (0x7fff_ffff) (0xc000_0000) (0xffff_ffff) |
| 23 | * +--------------------------------+----------------+----------------+ |
| 24 | * <---------- bad value! ----------> |
| 25 | * |
| 26 | * (in a signed view of the world, the "bad value" range corresponds to |
| 27 | * a negative counter value). |
| 28 | * |
| 29 | * As an example, consider a refcount_inc() operation that causes the counter |
| 30 | * to overflow: |
| 31 | * |
| 32 | * int old = atomic_fetch_add_relaxed(r); |
| 33 | * // old is INT_MAX, refcount now INT_MIN (0x8000_0000) |
| 34 | * if (old < 0) |
| 35 | * atomic_set(r, REFCOUNT_SATURATED); |
| 36 | * |
| 37 | * If another thread also performs a refcount_inc() operation between the two |
| 38 | * atomic operations, then the count will continue to edge closer to 0. If it |
| 39 | * reaches a value of 1 before /any/ of the threads reset it to the saturated |
| 40 | * value, then a concurrent refcount_dec_and_test() may erroneously free the |
| 41 | * underlying object. |
| 42 | * Linux limits the maximum number of tasks to PID_MAX_LIMIT, which is currently |
| 43 | * 0x400000 (and can't easily be raised in the future beyond FUTEX_TID_MASK). |
| 44 | * With the current PID limit, if no batched refcounting operations are used and |
| 45 | * the attacker can't repeatedly trigger kernel oopses in the middle of refcount |
| 46 | * operations, this makes it impossible for a saturated refcount to leave the |
| 47 | * saturation range, even if it is possible for multiple uses of the same |
| 48 | * refcount to nest in the context of a single task: |
| 49 | * |
| 50 | * (UINT_MAX+1-REFCOUNT_SATURATED) / PID_MAX_LIMIT = |
| 51 | * 0x40000000 / 0x400000 = 0x100 = 256 |
| 52 | * |
| 53 | * If hundreds of references are added/removed with a single refcounting |
| 54 | * operation, it may potentially be possible to leave the saturation range; but |
| 55 | * given the precise timing details involved with the round-robin scheduling of |
| 56 | * each thread manipulating the refcount and the need to hit the race multiple |
| 57 | * times in succession, there doesn't appear to be a practical avenue of attack |
| 58 | * even if using refcount_add() operations with larger increments. |
| 59 | * |
| 60 | * Memory ordering |
| 61 | * =============== |
| 62 | * |
| 63 | * Memory ordering rules are slightly relaxed wrt regular atomic_t functions |
| 64 | * and provide only what is strictly required for refcounts. |
| 65 | * |
| 66 | * The increments are fully relaxed; these will not provide ordering. The |
| 67 | * rationale is that whatever is used to obtain the object we're increasing the |
| 68 | * reference count on will provide the ordering. For locked data structures, |
| 69 | * its the lock acquire, for RCU/lockless data structures its the dependent |
| 70 | * load. |
| 71 | * |
| 72 | * Do note that inc_not_zero() provides a control dependency which will order |
| 73 | * future stores against the inc, this ensures we'll never modify the object |
| 74 | * if we did not in fact acquire a reference. |
| 75 | * |
| 76 | * The decrements will provide release order, such that all the prior loads and |
| 77 | * stores will be issued before, it also provides a control dependency, which |
| 78 | * will order us against the subsequent free(). |
| 79 | * |
| 80 | * The control dependency is against the load of the cmpxchg (ll/sc) that |
| 81 | * succeeded. This means the stores aren't fully ordered, but this is fine |
| 82 | * because the 1->0 transition indicates no concurrency. |
| 83 | * |
| 84 | * Note that the allocator is responsible for ordering things between free() |
| 85 | * and alloc(). |
| 86 | * |
| 87 | * The decrements dec_and_test() and sub_and_test() also provide acquire |
| 88 | * ordering on success. |
| 89 | * |
| 90 | */ |
| 91 | |
| 92 | #ifndef _LINUX_REFCOUNT_H |
| 93 | #define _LINUX_REFCOUNT_H |
| 94 | |
| 95 | #include <linux/atomic.h> |
| 96 | #include <linux/bug.h> |
| 97 | #include <linux/compiler.h> |
| 98 | #include <linux/limits.h> |
| 99 | #include <linux/spinlock_types.h> |
| 100 | |
| 101 | struct mutex; |
| 102 | |
| 103 | /** |
| 104 | * typedef refcount_t - variant of atomic_t specialized for reference counts |
| 105 | * @refs: atomic_t counter field |
| 106 | * |
| 107 | * The counter saturates at REFCOUNT_SATURATED and will not move once |
| 108 | * there. This avoids wrapping the counter and causing 'spurious' |
| 109 | * use-after-free bugs. |
| 110 | */ |
| 111 | typedef struct refcount_struct { |
| 112 | atomic_t refs; |
| 113 | } refcount_t; |
| 114 | |
| 115 | #define REFCOUNT_INIT(n) { .refs = ATOMIC_INIT(n), } |
| 116 | #define REFCOUNT_MAX INT_MAX |
| 117 | #define REFCOUNT_SATURATED (INT_MIN / 2) |
| 118 | |
| 119 | enum refcount_saturation_type { |
| 120 | REFCOUNT_ADD_NOT_ZERO_OVF, |
| 121 | REFCOUNT_ADD_OVF, |
| 122 | REFCOUNT_ADD_UAF, |
| 123 | REFCOUNT_SUB_UAF, |
| 124 | REFCOUNT_DEC_LEAK, |
| 125 | }; |
| 126 | |
| 127 | void refcount_warn_saturate(refcount_t *r, enum refcount_saturation_type t); |
| 128 | |
| 129 | /** |
| 130 | * refcount_set - set a refcount's value |
| 131 | * @r: the refcount |
| 132 | * @n: value to which the refcount will be set |
| 133 | */ |
| 134 | static inline void refcount_set(refcount_t *r, int n) |
| 135 | { |
| 136 | atomic_set(&r->refs, n); |
| 137 | } |
| 138 | |
| 139 | /** |
| 140 | * refcount_read - get a refcount's value |
| 141 | * @r: the refcount |
| 142 | * |
| 143 | * Return: the refcount's value |
| 144 | */ |
| 145 | static inline unsigned int refcount_read(const refcount_t *r) |
| 146 | { |
| 147 | return atomic_read(&r->refs); |
| 148 | } |
| 149 | |
| 150 | static inline __must_check bool __refcount_add_not_zero(int i, refcount_t *r, int *oldp) |
| 151 | { |
| 152 | int old = refcount_read(r); |
| 153 | |
| 154 | do { |
| 155 | if (!old) |
| 156 | break; |
| 157 | } while (!atomic_try_cmpxchg_relaxed(&r->refs, &old, old + i)); |
| 158 | |
| 159 | if (oldp) |
| 160 | *oldp = old; |
| 161 | |
| 162 | if (unlikely(old < 0 || old + i < 0)) |
| 163 | refcount_warn_saturate(r, REFCOUNT_ADD_NOT_ZERO_OVF); |
| 164 | |
| 165 | return old; |
| 166 | } |
| 167 | |
| 168 | /** |
| 169 | * refcount_add_not_zero - add a value to a refcount unless it is 0 |
| 170 | * @i: the value to add to the refcount |
| 171 | * @r: the refcount |
| 172 | * |
| 173 | * Will saturate at REFCOUNT_SATURATED and WARN. |
| 174 | * |
| 175 | * Provides no memory ordering, it is assumed the caller has guaranteed the |
| 176 | * object memory to be stable (RCU, etc.). It does provide a control dependency |
| 177 | * and thereby orders future stores. See the comment on top. |
| 178 | * |
| 179 | * Use of this function is not recommended for the normal reference counting |
| 180 | * use case in which references are taken and released one at a time. In these |
| 181 | * cases, refcount_inc(), or one of its variants, should instead be used to |
| 182 | * increment a reference count. |
| 183 | * |
| 184 | * Return: false if the passed refcount is 0, true otherwise |
| 185 | */ |
| 186 | static inline __must_check bool refcount_add_not_zero(int i, refcount_t *r) |
| 187 | { |
| 188 | return __refcount_add_not_zero(i, r, NULL); |
| 189 | } |
| 190 | |
| 191 | static inline void __refcount_add(int i, refcount_t *r, int *oldp) |
| 192 | { |
| 193 | int old = atomic_fetch_add_relaxed(i, &r->refs); |
| 194 | |
| 195 | if (oldp) |
| 196 | *oldp = old; |
| 197 | |
| 198 | if (unlikely(!old)) |
| 199 | refcount_warn_saturate(r, REFCOUNT_ADD_UAF); |
| 200 | else if (unlikely(old < 0 || old + i < 0)) |
| 201 | refcount_warn_saturate(r, REFCOUNT_ADD_OVF); |
| 202 | } |
| 203 | |
| 204 | /** |
| 205 | * refcount_add - add a value to a refcount |
| 206 | * @i: the value to add to the refcount |
| 207 | * @r: the refcount |
| 208 | * |
| 209 | * Similar to atomic_add(), but will saturate at REFCOUNT_SATURATED and WARN. |
| 210 | * |
| 211 | * Provides no memory ordering, it is assumed the caller has guaranteed the |
| 212 | * object memory to be stable (RCU, etc.). It does provide a control dependency |
| 213 | * and thereby orders future stores. See the comment on top. |
| 214 | * |
| 215 | * Use of this function is not recommended for the normal reference counting |
| 216 | * use case in which references are taken and released one at a time. In these |
| 217 | * cases, refcount_inc(), or one of its variants, should instead be used to |
| 218 | * increment a reference count. |
| 219 | */ |
| 220 | static inline void refcount_add(int i, refcount_t *r) |
| 221 | { |
| 222 | __refcount_add(i, r, NULL); |
| 223 | } |
| 224 | |
| 225 | static inline __must_check bool __refcount_inc_not_zero(refcount_t *r, int *oldp) |
| 226 | { |
| 227 | return __refcount_add_not_zero(1, r, oldp); |
| 228 | } |
| 229 | |
| 230 | /** |
| 231 | * refcount_inc_not_zero - increment a refcount unless it is 0 |
| 232 | * @r: the refcount to increment |
| 233 | * |
| 234 | * Similar to atomic_inc_not_zero(), but will saturate at REFCOUNT_SATURATED |
| 235 | * and WARN. |
| 236 | * |
| 237 | * Provides no memory ordering, it is assumed the caller has guaranteed the |
| 238 | * object memory to be stable (RCU, etc.). It does provide a control dependency |
| 239 | * and thereby orders future stores. See the comment on top. |
| 240 | * |
| 241 | * Return: true if the increment was successful, false otherwise |
| 242 | */ |
| 243 | static inline __must_check bool refcount_inc_not_zero(refcount_t *r) |
| 244 | { |
| 245 | return __refcount_inc_not_zero(r, NULL); |
| 246 | } |
| 247 | |
| 248 | static inline void __refcount_inc(refcount_t *r, int *oldp) |
| 249 | { |
| 250 | __refcount_add(1, r, oldp); |
| 251 | } |
| 252 | |
| 253 | /** |
| 254 | * refcount_inc - increment a refcount |
| 255 | * @r: the refcount to increment |
| 256 | * |
| 257 | * Similar to atomic_inc(), but will saturate at REFCOUNT_SATURATED and WARN. |
| 258 | * |
| 259 | * Provides no memory ordering, it is assumed the caller already has a |
| 260 | * reference on the object. |
| 261 | * |
| 262 | * Will WARN if the refcount is 0, as this represents a possible use-after-free |
| 263 | * condition. |
| 264 | */ |
| 265 | static inline void refcount_inc(refcount_t *r) |
| 266 | { |
| 267 | __refcount_inc(r, NULL); |
| 268 | } |
| 269 | |
| 270 | static inline __must_check bool __refcount_sub_and_test(int i, refcount_t *r, int *oldp) |
| 271 | { |
| 272 | int old = atomic_fetch_sub_release(i, &r->refs); |
| 273 | |
| 274 | if (oldp) |
| 275 | *oldp = old; |
| 276 | |
| 277 | if (old == i) { |
| 278 | smp_acquire__after_ctrl_dep(); |
| 279 | return true; |
| 280 | } |
| 281 | |
| 282 | if (unlikely(old < 0 || old - i < 0)) |
| 283 | refcount_warn_saturate(r, REFCOUNT_SUB_UAF); |
| 284 | |
| 285 | return false; |
| 286 | } |
| 287 | |
| 288 | /** |
| 289 | * refcount_sub_and_test - subtract from a refcount and test if it is 0 |
| 290 | * @i: amount to subtract from the refcount |
| 291 | * @r: the refcount |
| 292 | * |
| 293 | * Similar to atomic_dec_and_test(), but it will WARN, return false and |
| 294 | * ultimately leak on underflow and will fail to decrement when saturated |
| 295 | * at REFCOUNT_SATURATED. |
| 296 | * |
| 297 | * Provides release memory ordering, such that prior loads and stores are done |
| 298 | * before, and provides an acquire ordering on success such that free() |
| 299 | * must come after. |
| 300 | * |
| 301 | * Use of this function is not recommended for the normal reference counting |
| 302 | * use case in which references are taken and released one at a time. In these |
| 303 | * cases, refcount_dec(), or one of its variants, should instead be used to |
| 304 | * decrement a reference count. |
| 305 | * |
| 306 | * Return: true if the resulting refcount is 0, false otherwise |
| 307 | */ |
| 308 | static inline __must_check bool refcount_sub_and_test(int i, refcount_t *r) |
| 309 | { |
| 310 | return __refcount_sub_and_test(i, r, NULL); |
| 311 | } |
| 312 | |
| 313 | static inline __must_check bool __refcount_dec_and_test(refcount_t *r, int *oldp) |
| 314 | { |
| 315 | return __refcount_sub_and_test(1, r, oldp); |
| 316 | } |
| 317 | |
| 318 | /** |
| 319 | * refcount_dec_and_test - decrement a refcount and test if it is 0 |
| 320 | * @r: the refcount |
| 321 | * |
| 322 | * Similar to atomic_dec_and_test(), it will WARN on underflow and fail to |
| 323 | * decrement when saturated at REFCOUNT_SATURATED. |
| 324 | * |
| 325 | * Provides release memory ordering, such that prior loads and stores are done |
| 326 | * before, and provides an acquire ordering on success such that free() |
| 327 | * must come after. |
| 328 | * |
| 329 | * Return: true if the resulting refcount is 0, false otherwise |
| 330 | */ |
| 331 | static inline __must_check bool refcount_dec_and_test(refcount_t *r) |
| 332 | { |
| 333 | return __refcount_dec_and_test(r, NULL); |
| 334 | } |
| 335 | |
| 336 | static inline void __refcount_dec(refcount_t *r, int *oldp) |
| 337 | { |
| 338 | int old = atomic_fetch_sub_release(1, &r->refs); |
| 339 | |
| 340 | if (oldp) |
| 341 | *oldp = old; |
| 342 | |
| 343 | if (unlikely(old <= 1)) |
| 344 | refcount_warn_saturate(r, REFCOUNT_DEC_LEAK); |
| 345 | } |
| 346 | |
| 347 | /** |
| 348 | * refcount_dec - decrement a refcount |
| 349 | * @r: the refcount |
| 350 | * |
| 351 | * Similar to atomic_dec(), it will WARN on underflow and fail to decrement |
| 352 | * when saturated at REFCOUNT_SATURATED. |
| 353 | * |
| 354 | * Provides release memory ordering, such that prior loads and stores are done |
| 355 | * before. |
| 356 | */ |
| 357 | static inline void refcount_dec(refcount_t *r) |
| 358 | { |
| 359 | __refcount_dec(r, NULL); |
| 360 | } |
| 361 | |
| 362 | extern __must_check bool refcount_dec_if_one(refcount_t *r); |
| 363 | extern __must_check bool refcount_dec_not_one(refcount_t *r); |
| 364 | extern __must_check bool refcount_dec_and_mutex_lock(refcount_t *r, struct mutex *lock); |
| 365 | extern __must_check bool refcount_dec_and_lock(refcount_t *r, spinlock_t *lock); |
| 366 | extern __must_check bool refcount_dec_and_lock_irqsave(refcount_t *r, |
| 367 | spinlock_t *lock, |
| 368 | unsigned long *flags); |
| 369 | #endif /* _LINUX_REFCOUNT_H */ |