| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | #define CREATE_TRACE_POINTS |
| 3 | #include <trace/events/mmap_lock.h> |
| 4 | |
| 5 | #include <linux/mm.h> |
| 6 | #include <linux/cgroup.h> |
| 7 | #include <linux/memcontrol.h> |
| 8 | #include <linux/mmap_lock.h> |
| 9 | #include <linux/mutex.h> |
| 10 | #include <linux/percpu.h> |
| 11 | #include <linux/rcupdate.h> |
| 12 | #include <linux/smp.h> |
| 13 | #include <linux/trace_events.h> |
| 14 | #include <linux/local_lock.h> |
| 15 | |
| 16 | EXPORT_TRACEPOINT_SYMBOL(mmap_lock_start_locking); |
| 17 | EXPORT_TRACEPOINT_SYMBOL(mmap_lock_acquire_returned); |
| 18 | EXPORT_TRACEPOINT_SYMBOL(mmap_lock_released); |
| 19 | |
| 20 | #ifdef CONFIG_TRACING |
| 21 | /* |
| 22 | * Trace calls must be in a separate file, as otherwise there's a circular |
| 23 | * dependency between linux/mmap_lock.h and trace/events/mmap_lock.h. |
| 24 | */ |
| 25 | |
| 26 | void __mmap_lock_do_trace_start_locking(struct mm_struct *mm, bool write) |
| 27 | { |
| 28 | trace_mmap_lock_start_locking(mm, write); |
| 29 | } |
| 30 | EXPORT_SYMBOL(__mmap_lock_do_trace_start_locking); |
| 31 | |
| 32 | void __mmap_lock_do_trace_acquire_returned(struct mm_struct *mm, bool write, |
| 33 | bool success) |
| 34 | { |
| 35 | trace_mmap_lock_acquire_returned(mm, write, success); |
| 36 | } |
| 37 | EXPORT_SYMBOL(__mmap_lock_do_trace_acquire_returned); |
| 38 | |
| 39 | void __mmap_lock_do_trace_released(struct mm_struct *mm, bool write) |
| 40 | { |
| 41 | trace_mmap_lock_released(mm, write); |
| 42 | } |
| 43 | EXPORT_SYMBOL(__mmap_lock_do_trace_released); |
| 44 | #endif /* CONFIG_TRACING */ |
| 45 | |
| 46 | #ifdef CONFIG_MMU |
| 47 | #ifdef CONFIG_PER_VMA_LOCK |
| 48 | static inline bool __vma_enter_locked(struct vm_area_struct *vma, bool detaching) |
| 49 | { |
| 50 | unsigned int tgt_refcnt = VMA_LOCK_OFFSET; |
| 51 | |
| 52 | /* Additional refcnt if the vma is attached. */ |
| 53 | if (!detaching) |
| 54 | tgt_refcnt++; |
| 55 | |
| 56 | /* |
| 57 | * If vma is detached then only vma_mark_attached() can raise the |
| 58 | * vm_refcnt. mmap_write_lock prevents racing with vma_mark_attached(). |
| 59 | */ |
| 60 | if (!refcount_add_not_zero(VMA_LOCK_OFFSET, r: &vma->vm_refcnt)) |
| 61 | return false; |
| 62 | |
| 63 | rwsem_acquire(&vma->vmlock_dep_map, 0, 0, _RET_IP_); |
| 64 | rcuwait_wait_event(&vma->vm_mm->vma_writer_wait, |
| 65 | refcount_read(&vma->vm_refcnt) == tgt_refcnt, |
| 66 | TASK_UNINTERRUPTIBLE); |
| 67 | lock_acquired(&vma->vmlock_dep_map, _RET_IP_); |
| 68 | |
| 69 | return true; |
| 70 | } |
| 71 | |
| 72 | static inline void __vma_exit_locked(struct vm_area_struct *vma, bool *detached) |
| 73 | { |
| 74 | *detached = refcount_sub_and_test(VMA_LOCK_OFFSET, r: &vma->vm_refcnt); |
| 75 | rwsem_release(&vma->vmlock_dep_map, _RET_IP_); |
| 76 | } |
| 77 | |
| 78 | void __vma_start_write(struct vm_area_struct *vma, unsigned int mm_lock_seq) |
| 79 | { |
| 80 | bool locked; |
| 81 | |
| 82 | /* |
| 83 | * __vma_enter_locked() returns false immediately if the vma is not |
| 84 | * attached, otherwise it waits until refcnt is indicating that vma |
| 85 | * is attached with no readers. |
| 86 | */ |
| 87 | locked = __vma_enter_locked(vma, detaching: false); |
| 88 | |
| 89 | /* |
| 90 | * We should use WRITE_ONCE() here because we can have concurrent reads |
| 91 | * from the early lockless pessimistic check in vma_start_read(). |
| 92 | * We don't really care about the correctness of that early check, but |
| 93 | * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy. |
| 94 | */ |
| 95 | WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq); |
| 96 | |
| 97 | if (locked) { |
| 98 | bool detached; |
| 99 | |
| 100 | __vma_exit_locked(vma, detached: &detached); |
| 101 | WARN_ON_ONCE(detached); /* vma should remain attached */ |
| 102 | } |
| 103 | } |
| 104 | EXPORT_SYMBOL_GPL(__vma_start_write); |
| 105 | |
| 106 | void vma_mark_detached(struct vm_area_struct *vma) |
| 107 | { |
| 108 | vma_assert_write_locked(vma); |
| 109 | vma_assert_attached(vma); |
| 110 | |
| 111 | /* |
| 112 | * We are the only writer, so no need to use vma_refcount_put(). |
| 113 | * The condition below is unlikely because the vma has been already |
| 114 | * write-locked and readers can increment vm_refcnt only temporarily |
| 115 | * before they check vm_lock_seq, realize the vma is locked and drop |
| 116 | * back the vm_refcnt. That is a narrow window for observing a raised |
| 117 | * vm_refcnt. |
| 118 | */ |
| 119 | if (unlikely(!refcount_dec_and_test(&vma->vm_refcnt))) { |
| 120 | /* Wait until vma is detached with no readers. */ |
| 121 | if (__vma_enter_locked(vma, detaching: true)) { |
| 122 | bool detached; |
| 123 | |
| 124 | __vma_exit_locked(vma, detached: &detached); |
| 125 | WARN_ON_ONCE(!detached); |
| 126 | } |
| 127 | } |
| 128 | } |
| 129 | |
| 130 | /* |
| 131 | * Try to read-lock a vma. The function is allowed to occasionally yield false |
| 132 | * locked result to avoid performance overhead, in which case we fall back to |
| 133 | * using mmap_lock. The function should never yield false unlocked result. |
| 134 | * False locked result is possible if mm_lock_seq overflows or if vma gets |
| 135 | * reused and attached to a different mm before we lock it. |
| 136 | * Returns the vma on success, NULL on failure to lock and EAGAIN if vma got |
| 137 | * detached. |
| 138 | * |
| 139 | * IMPORTANT: RCU lock must be held upon entering the function, but upon error |
| 140 | * IT IS RELEASED. The caller must handle this correctly. |
| 141 | */ |
| 142 | static inline struct vm_area_struct *vma_start_read(struct mm_struct *mm, |
| 143 | struct vm_area_struct *vma) |
| 144 | { |
| 145 | struct mm_struct *other_mm; |
| 146 | int oldcnt; |
| 147 | |
| 148 | RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "no rcu lock held" ); |
| 149 | /* |
| 150 | * Check before locking. A race might cause false locked result. |
| 151 | * We can use READ_ONCE() for the mm_lock_seq here, and don't need |
| 152 | * ACQUIRE semantics, because this is just a lockless check whose result |
| 153 | * we don't rely on for anything - the mm_lock_seq read against which we |
| 154 | * need ordering is below. |
| 155 | */ |
| 156 | if (READ_ONCE(vma->vm_lock_seq) == READ_ONCE(mm->mm_lock_seq.sequence)) { |
| 157 | vma = NULL; |
| 158 | goto err; |
| 159 | } |
| 160 | |
| 161 | /* |
| 162 | * If VMA_LOCK_OFFSET is set, __refcount_inc_not_zero_limited_acquire() |
| 163 | * will fail because VMA_REF_LIMIT is less than VMA_LOCK_OFFSET. |
| 164 | * Acquire fence is required here to avoid reordering against later |
| 165 | * vm_lock_seq check and checks inside lock_vma_under_rcu(). |
| 166 | */ |
| 167 | if (unlikely(!__refcount_inc_not_zero_limited_acquire(&vma->vm_refcnt, &oldcnt, |
| 168 | VMA_REF_LIMIT))) { |
| 169 | /* return EAGAIN if vma got detached from under us */ |
| 170 | vma = oldcnt ? NULL : ERR_PTR(error: -EAGAIN); |
| 171 | goto err; |
| 172 | } |
| 173 | |
| 174 | rwsem_acquire_read(&vma->vmlock_dep_map, 0, 1, _RET_IP_); |
| 175 | |
| 176 | if (unlikely(vma->vm_mm != mm)) |
| 177 | goto err_unstable; |
| 178 | |
| 179 | /* |
| 180 | * Overflow of vm_lock_seq/mm_lock_seq might produce false locked result. |
| 181 | * False unlocked result is impossible because we modify and check |
| 182 | * vma->vm_lock_seq under vma->vm_refcnt protection and mm->mm_lock_seq |
| 183 | * modification invalidates all existing locks. |
| 184 | * |
| 185 | * We must use ACQUIRE semantics for the mm_lock_seq so that if we are |
| 186 | * racing with vma_end_write_all(), we only start reading from the VMA |
| 187 | * after it has been unlocked. |
| 188 | * This pairs with RELEASE semantics in vma_end_write_all(). |
| 189 | */ |
| 190 | if (unlikely(vma->vm_lock_seq == raw_read_seqcount(&mm->mm_lock_seq))) { |
| 191 | vma_refcount_put(vma); |
| 192 | vma = NULL; |
| 193 | goto err; |
| 194 | } |
| 195 | |
| 196 | return vma; |
| 197 | err: |
| 198 | rcu_read_unlock(); |
| 199 | |
| 200 | return vma; |
| 201 | err_unstable: |
| 202 | /* |
| 203 | * If vma got attached to another mm from under us, that mm is not |
| 204 | * stable and can be freed in the narrow window after vma->vm_refcnt |
| 205 | * is dropped and before rcuwait_wake_up(mm) is called. Grab it before |
| 206 | * releasing vma->vm_refcnt. |
| 207 | */ |
| 208 | other_mm = vma->vm_mm; /* use a copy as vma can be freed after we drop vm_refcnt */ |
| 209 | |
| 210 | /* __mmdrop() is a heavy operation, do it after dropping RCU lock. */ |
| 211 | rcu_read_unlock(); |
| 212 | mmgrab(mm: other_mm); |
| 213 | vma_refcount_put(vma); |
| 214 | mmdrop(mm: other_mm); |
| 215 | |
| 216 | return NULL; |
| 217 | } |
| 218 | |
| 219 | /* |
| 220 | * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be |
| 221 | * stable and not isolated. If the VMA is not found or is being modified the |
| 222 | * function returns NULL. |
| 223 | */ |
| 224 | struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm, |
| 225 | unsigned long address) |
| 226 | { |
| 227 | MA_STATE(mas, &mm->mm_mt, address, address); |
| 228 | struct vm_area_struct *vma; |
| 229 | |
| 230 | retry: |
| 231 | rcu_read_lock(); |
| 232 | vma = mas_walk(mas: &mas); |
| 233 | if (!vma) { |
| 234 | rcu_read_unlock(); |
| 235 | goto inval; |
| 236 | } |
| 237 | |
| 238 | vma = vma_start_read(mm, vma); |
| 239 | if (IS_ERR_OR_NULL(ptr: vma)) { |
| 240 | /* Check if the VMA got isolated after we found it */ |
| 241 | if (PTR_ERR(ptr: vma) == -EAGAIN) { |
| 242 | count_vm_vma_lock_event(VMA_LOCK_MISS); |
| 243 | /* The area was replaced with another one */ |
| 244 | goto retry; |
| 245 | } |
| 246 | |
| 247 | /* Failed to lock the VMA */ |
| 248 | goto inval; |
| 249 | } |
| 250 | /* |
| 251 | * At this point, we have a stable reference to a VMA: The VMA is |
| 252 | * locked and we know it hasn't already been isolated. |
| 253 | * From here on, we can access the VMA without worrying about which |
| 254 | * fields are accessible for RCU readers. |
| 255 | */ |
| 256 | rcu_read_unlock(); |
| 257 | |
| 258 | /* Check if the vma we locked is the right one. */ |
| 259 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { |
| 260 | vma_end_read(vma); |
| 261 | goto inval; |
| 262 | } |
| 263 | |
| 264 | return vma; |
| 265 | |
| 266 | inval: |
| 267 | count_vm_vma_lock_event(VMA_LOCK_ABORT); |
| 268 | return NULL; |
| 269 | } |
| 270 | |
| 271 | static struct vm_area_struct *lock_next_vma_under_mmap_lock(struct mm_struct *mm, |
| 272 | struct vma_iterator *vmi, |
| 273 | unsigned long from_addr) |
| 274 | { |
| 275 | struct vm_area_struct *vma; |
| 276 | int ret; |
| 277 | |
| 278 | ret = mmap_read_lock_killable(mm); |
| 279 | if (ret) |
| 280 | return ERR_PTR(error: ret); |
| 281 | |
| 282 | /* Lookup the vma at the last position again under mmap_read_lock */ |
| 283 | vma_iter_set(vmi, addr: from_addr); |
| 284 | vma = vma_next(vmi); |
| 285 | if (vma) { |
| 286 | /* Very unlikely vma->vm_refcnt overflow case */ |
| 287 | if (unlikely(!vma_start_read_locked(vma))) |
| 288 | vma = ERR_PTR(error: -EAGAIN); |
| 289 | } |
| 290 | |
| 291 | mmap_read_unlock(mm); |
| 292 | |
| 293 | return vma; |
| 294 | } |
| 295 | |
| 296 | struct vm_area_struct *lock_next_vma(struct mm_struct *mm, |
| 297 | struct vma_iterator *vmi, |
| 298 | unsigned long from_addr) |
| 299 | { |
| 300 | struct vm_area_struct *vma; |
| 301 | unsigned int mm_wr_seq; |
| 302 | bool mmap_unlocked; |
| 303 | |
| 304 | RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "no rcu read lock held" ); |
| 305 | retry: |
| 306 | /* Start mmap_lock speculation in case we need to verify the vma later */ |
| 307 | mmap_unlocked = mmap_lock_speculate_try_begin(mm, seq: &mm_wr_seq); |
| 308 | vma = vma_next(vmi); |
| 309 | if (!vma) |
| 310 | return NULL; |
| 311 | |
| 312 | vma = vma_start_read(mm, vma); |
| 313 | if (IS_ERR_OR_NULL(ptr: vma)) { |
| 314 | /* |
| 315 | * Retry immediately if the vma gets detached from under us. |
| 316 | * Infinite loop should not happen because the vma we find will |
| 317 | * have to be constantly knocked out from under us. |
| 318 | */ |
| 319 | if (PTR_ERR(ptr: vma) == -EAGAIN) { |
| 320 | /* reset to search from the last address */ |
| 321 | rcu_read_lock(); |
| 322 | vma_iter_set(vmi, addr: from_addr); |
| 323 | goto retry; |
| 324 | } |
| 325 | |
| 326 | goto fallback; |
| 327 | } |
| 328 | |
| 329 | /* Verify the vma is not behind the last search position. */ |
| 330 | if (unlikely(from_addr >= vma->vm_end)) |
| 331 | goto fallback_unlock; |
| 332 | |
| 333 | /* |
| 334 | * vma can be ahead of the last search position but we need to verify |
| 335 | * it was not shrunk after we found it and another vma has not been |
| 336 | * installed ahead of it. Otherwise we might observe a gap that should |
| 337 | * not be there. |
| 338 | */ |
| 339 | if (from_addr < vma->vm_start) { |
| 340 | /* Verify only if the address space might have changed since vma lookup. */ |
| 341 | if (!mmap_unlocked || mmap_lock_speculate_retry(mm, seq: mm_wr_seq)) { |
| 342 | vma_iter_set(vmi, addr: from_addr); |
| 343 | if (vma != vma_next(vmi)) |
| 344 | goto fallback_unlock; |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | return vma; |
| 349 | |
| 350 | fallback_unlock: |
| 351 | rcu_read_unlock(); |
| 352 | vma_end_read(vma); |
| 353 | fallback: |
| 354 | vma = lock_next_vma_under_mmap_lock(mm, vmi, from_addr); |
| 355 | rcu_read_lock(); |
| 356 | /* Reinitialize the iterator after re-entering rcu read section */ |
| 357 | vma_iter_set(vmi, addr: IS_ERR_OR_NULL(ptr: vma) ? from_addr : vma->vm_end); |
| 358 | |
| 359 | return vma; |
| 360 | } |
| 361 | #endif /* CONFIG_PER_VMA_LOCK */ |
| 362 | |
| 363 | #ifdef CONFIG_LOCK_MM_AND_FIND_VMA |
| 364 | #include <linux/extable.h> |
| 365 | |
| 366 | static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs) |
| 367 | { |
| 368 | if (likely(mmap_read_trylock(mm))) |
| 369 | return true; |
| 370 | |
| 371 | if (regs && !user_mode(regs)) { |
| 372 | unsigned long ip = exception_ip(regs); |
| 373 | if (!search_exception_tables(add: ip)) |
| 374 | return false; |
| 375 | } |
| 376 | |
| 377 | return !mmap_read_lock_killable(mm); |
| 378 | } |
| 379 | |
| 380 | static inline bool mmap_upgrade_trylock(struct mm_struct *mm) |
| 381 | { |
| 382 | /* |
| 383 | * We don't have this operation yet. |
| 384 | * |
| 385 | * It should be easy enough to do: it's basically a |
| 386 | * atomic_long_try_cmpxchg_acquire() |
| 387 | * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but |
| 388 | * it also needs the proper lockdep magic etc. |
| 389 | */ |
| 390 | return false; |
| 391 | } |
| 392 | |
| 393 | static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs) |
| 394 | { |
| 395 | mmap_read_unlock(mm); |
| 396 | if (regs && !user_mode(regs)) { |
| 397 | unsigned long ip = exception_ip(regs); |
| 398 | if (!search_exception_tables(add: ip)) |
| 399 | return false; |
| 400 | } |
| 401 | return !mmap_write_lock_killable(mm); |
| 402 | } |
| 403 | |
| 404 | /* |
| 405 | * Helper for page fault handling. |
| 406 | * |
| 407 | * This is kind of equivalent to "mmap_read_lock()" followed |
| 408 | * by "find_extend_vma()", except it's a lot more careful about |
| 409 | * the locking (and will drop the lock on failure). |
| 410 | * |
| 411 | * For example, if we have a kernel bug that causes a page |
| 412 | * fault, we don't want to just use mmap_read_lock() to get |
| 413 | * the mm lock, because that would deadlock if the bug were |
| 414 | * to happen while we're holding the mm lock for writing. |
| 415 | * |
| 416 | * So this checks the exception tables on kernel faults in |
| 417 | * order to only do this all for instructions that are actually |
| 418 | * expected to fault. |
| 419 | * |
| 420 | * We can also actually take the mm lock for writing if we |
| 421 | * need to extend the vma, which helps the VM layer a lot. |
| 422 | */ |
| 423 | struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm, |
| 424 | unsigned long addr, struct pt_regs *regs) |
| 425 | { |
| 426 | struct vm_area_struct *vma; |
| 427 | |
| 428 | if (!get_mmap_lock_carefully(mm, regs)) |
| 429 | return NULL; |
| 430 | |
| 431 | vma = find_vma(mm, addr); |
| 432 | if (likely(vma && (vma->vm_start <= addr))) |
| 433 | return vma; |
| 434 | |
| 435 | /* |
| 436 | * Well, dang. We might still be successful, but only |
| 437 | * if we can extend a vma to do so. |
| 438 | */ |
| 439 | if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) { |
| 440 | mmap_read_unlock(mm); |
| 441 | return NULL; |
| 442 | } |
| 443 | |
| 444 | /* |
| 445 | * We can try to upgrade the mmap lock atomically, |
| 446 | * in which case we can continue to use the vma |
| 447 | * we already looked up. |
| 448 | * |
| 449 | * Otherwise we'll have to drop the mmap lock and |
| 450 | * re-take it, and also look up the vma again, |
| 451 | * re-checking it. |
| 452 | */ |
| 453 | if (!mmap_upgrade_trylock(mm)) { |
| 454 | if (!upgrade_mmap_lock_carefully(mm, regs)) |
| 455 | return NULL; |
| 456 | |
| 457 | vma = find_vma(mm, addr); |
| 458 | if (!vma) |
| 459 | goto fail; |
| 460 | if (vma->vm_start <= addr) |
| 461 | goto success; |
| 462 | if (!(vma->vm_flags & VM_GROWSDOWN)) |
| 463 | goto fail; |
| 464 | } |
| 465 | |
| 466 | if (expand_stack_locked(vma, address: addr)) |
| 467 | goto fail; |
| 468 | |
| 469 | success: |
| 470 | mmap_write_downgrade(mm); |
| 471 | return vma; |
| 472 | |
| 473 | fail: |
| 474 | mmap_write_unlock(mm); |
| 475 | return NULL; |
| 476 | } |
| 477 | #endif /* CONFIG_LOCK_MM_AND_FIND_VMA */ |
| 478 | |
| 479 | #else /* CONFIG_MMU */ |
| 480 | |
| 481 | /* |
| 482 | * At least xtensa ends up having protection faults even with no |
| 483 | * MMU.. No stack expansion, at least. |
| 484 | */ |
| 485 | struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm, |
| 486 | unsigned long addr, struct pt_regs *regs) |
| 487 | { |
| 488 | struct vm_area_struct *vma; |
| 489 | |
| 490 | mmap_read_lock(mm); |
| 491 | vma = vma_lookup(mm, addr); |
| 492 | if (!vma) |
| 493 | mmap_read_unlock(mm); |
| 494 | return vma; |
| 495 | } |
| 496 | |
| 497 | #endif /* CONFIG_MMU */ |
| 498 | |