| 1 | /* |
| 2 | * An async IO implementation for Linux |
| 3 | * Written by Benjamin LaHaise <bcrl@kvack.org> |
| 4 | * |
| 5 | * Implements an efficient asynchronous io interface. |
| 6 | * |
| 7 | * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved. |
| 8 | * |
| 9 | * See ../COPYING for licensing terms. |
| 10 | */ |
| 11 | #define pr_fmt(fmt) "%s: " fmt, __func__ |
| 12 | |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/init.h> |
| 15 | #include <linux/errno.h> |
| 16 | #include <linux/time.h> |
| 17 | #include <linux/aio_abi.h> |
| 18 | #include <linux/export.h> |
| 19 | #include <linux/syscalls.h> |
| 20 | #include <linux/backing-dev.h> |
| 21 | #include <linux/uio.h> |
| 22 | |
| 23 | #include <linux/sched/signal.h> |
| 24 | #include <linux/fs.h> |
| 25 | #include <linux/file.h> |
| 26 | #include <linux/mm.h> |
| 27 | #include <linux/mman.h> |
| 28 | #include <linux/mmu_context.h> |
| 29 | #include <linux/percpu.h> |
| 30 | #include <linux/slab.h> |
| 31 | #include <linux/timer.h> |
| 32 | #include <linux/aio.h> |
| 33 | #include <linux/highmem.h> |
| 34 | #include <linux/workqueue.h> |
| 35 | #include <linux/security.h> |
| 36 | #include <linux/eventfd.h> |
| 37 | #include <linux/blkdev.h> |
| 38 | #include <linux/compat.h> |
| 39 | #include <linux/migrate.h> |
| 40 | #include <linux/ramfs.h> |
| 41 | #include <linux/percpu-refcount.h> |
| 42 | #include <linux/mount.h> |
| 43 | |
| 44 | #include <asm/kmap_types.h> |
| 45 | #include <linux/uaccess.h> |
| 46 | #include <linux/nospec.h> |
| 47 | |
| 48 | #include "internal.h" |
| 49 | |
| 50 | #define AIO_RING_MAGIC 0xa10a10a1 |
| 51 | #define AIO_RING_COMPAT_FEATURES 1 |
| 52 | #define AIO_RING_INCOMPAT_FEATURES 0 |
| 53 | struct aio_ring { |
| 54 | unsigned id; /* kernel internal index number */ |
| 55 | unsigned nr; /* number of io_events */ |
| 56 | unsigned head; /* Written to by userland or under ring_lock |
| 57 | * mutex by aio_read_events_ring(). */ |
| 58 | unsigned tail; |
| 59 | |
| 60 | unsigned magic; |
| 61 | unsigned compat_features; |
| 62 | unsigned incompat_features; |
| 63 | unsigned header_length; /* size of aio_ring */ |
| 64 | |
| 65 | |
| 66 | struct io_event io_events[0]; |
| 67 | }; /* 128 bytes + ring size */ |
| 68 | |
| 69 | #define AIO_RING_PAGES 8 |
| 70 | |
| 71 | struct kioctx_table { |
| 72 | struct rcu_head rcu; |
| 73 | unsigned nr; |
| 74 | struct kioctx __rcu *table[]; |
| 75 | }; |
| 76 | |
| 77 | struct kioctx_cpu { |
| 78 | unsigned reqs_available; |
| 79 | }; |
| 80 | |
| 81 | struct ctx_rq_wait { |
| 82 | struct completion comp; |
| 83 | atomic_t count; |
| 84 | }; |
| 85 | |
| 86 | struct kioctx { |
| 87 | struct percpu_ref users; |
| 88 | atomic_t dead; |
| 89 | |
| 90 | struct percpu_ref reqs; |
| 91 | |
| 92 | unsigned long user_id; |
| 93 | |
| 94 | struct __percpu kioctx_cpu *cpu; |
| 95 | |
| 96 | /* |
| 97 | * For percpu reqs_available, number of slots we move to/from global |
| 98 | * counter at a time: |
| 99 | */ |
| 100 | unsigned req_batch; |
| 101 | /* |
| 102 | * This is what userspace passed to io_setup(), it's not used for |
| 103 | * anything but counting against the global max_reqs quota. |
| 104 | * |
| 105 | * The real limit is nr_events - 1, which will be larger (see |
| 106 | * aio_setup_ring()) |
| 107 | */ |
| 108 | unsigned max_reqs; |
| 109 | |
| 110 | /* Size of ringbuffer, in units of struct io_event */ |
| 111 | unsigned nr_events; |
| 112 | |
| 113 | unsigned long mmap_base; |
| 114 | unsigned long mmap_size; |
| 115 | |
| 116 | struct page **ring_pages; |
| 117 | long nr_pages; |
| 118 | |
| 119 | struct rcu_head free_rcu; |
| 120 | struct work_struct free_work; /* see free_ioctx() */ |
| 121 | |
| 122 | /* |
| 123 | * signals when all in-flight requests are done |
| 124 | */ |
| 125 | struct ctx_rq_wait *rq_wait; |
| 126 | |
| 127 | struct { |
| 128 | /* |
| 129 | * This counts the number of available slots in the ringbuffer, |
| 130 | * so we avoid overflowing it: it's decremented (if positive) |
| 131 | * when allocating a kiocb and incremented when the resulting |
| 132 | * io_event is pulled off the ringbuffer. |
| 133 | * |
| 134 | * We batch accesses to it with a percpu version. |
| 135 | */ |
| 136 | atomic_t reqs_available; |
| 137 | } ____cacheline_aligned_in_smp; |
| 138 | |
| 139 | struct { |
| 140 | spinlock_t ctx_lock; |
| 141 | struct list_head active_reqs; /* used for cancellation */ |
| 142 | } ____cacheline_aligned_in_smp; |
| 143 | |
| 144 | struct { |
| 145 | struct mutex ring_lock; |
| 146 | wait_queue_head_t wait; |
| 147 | } ____cacheline_aligned_in_smp; |
| 148 | |
| 149 | struct { |
| 150 | unsigned tail; |
| 151 | unsigned completed_events; |
| 152 | spinlock_t completion_lock; |
| 153 | } ____cacheline_aligned_in_smp; |
| 154 | |
| 155 | struct page *internal_pages[AIO_RING_PAGES]; |
| 156 | struct file *aio_ring_file; |
| 157 | |
| 158 | unsigned id; |
| 159 | }; |
| 160 | |
| 161 | /* |
| 162 | * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either |
| 163 | * cancelled or completed (this makes a certain amount of sense because |
| 164 | * successful cancellation - io_cancel() - does deliver the completion to |
| 165 | * userspace). |
| 166 | * |
| 167 | * And since most things don't implement kiocb cancellation and we'd really like |
| 168 | * kiocb completion to be lockless when possible, we use ki_cancel to |
| 169 | * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED |
| 170 | * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel(). |
| 171 | */ |
| 172 | #define KIOCB_CANCELLED ((void *) (~0ULL)) |
| 173 | |
| 174 | struct aio_kiocb { |
| 175 | struct kiocb common; |
| 176 | |
| 177 | struct kioctx *ki_ctx; |
| 178 | kiocb_cancel_fn *ki_cancel; |
| 179 | |
| 180 | struct iocb __user *ki_user_iocb; /* user's aiocb */ |
| 181 | __u64 ki_user_data; /* user's data for completion */ |
| 182 | |
| 183 | struct list_head ki_list; /* the aio core uses this |
| 184 | * for cancellation */ |
| 185 | |
| 186 | /* |
| 187 | * If the aio_resfd field of the userspace iocb is not zero, |
| 188 | * this is the underlying eventfd context to deliver events to. |
| 189 | */ |
| 190 | struct eventfd_ctx *ki_eventfd; |
| 191 | }; |
| 192 | |
| 193 | /*------ sysctl variables----*/ |
| 194 | static DEFINE_SPINLOCK(aio_nr_lock); |
| 195 | unsigned long aio_nr; /* current system wide number of aio requests */ |
| 196 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ |
| 197 | /*----end sysctl variables---*/ |
| 198 | |
| 199 | static struct kmem_cache *kiocb_cachep; |
| 200 | static struct kmem_cache *kioctx_cachep; |
| 201 | |
| 202 | static struct vfsmount *aio_mnt; |
| 203 | |
| 204 | static const struct file_operations aio_ring_fops; |
| 205 | static const struct address_space_operations aio_ctx_aops; |
| 206 | |
| 207 | static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) |
| 208 | { |
| 209 | struct qstr this = QSTR_INIT("[aio]", 5); |
| 210 | struct file *file; |
| 211 | struct path path; |
| 212 | struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); |
| 213 | if (IS_ERR(inode)) |
| 214 | return ERR_CAST(inode); |
| 215 | |
| 216 | inode->i_mapping->a_ops = &aio_ctx_aops; |
| 217 | inode->i_mapping->private_data = ctx; |
| 218 | inode->i_size = PAGE_SIZE * nr_pages; |
| 219 | |
| 220 | path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); |
| 221 | if (!path.dentry) { |
| 222 | iput(inode); |
| 223 | return ERR_PTR(-ENOMEM); |
| 224 | } |
| 225 | path.mnt = mntget(aio_mnt); |
| 226 | |
| 227 | d_instantiate(path.dentry, inode); |
| 228 | file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops); |
| 229 | if (IS_ERR(file)) { |
| 230 | path_put(&path); |
| 231 | return file; |
| 232 | } |
| 233 | |
| 234 | file->f_flags = O_RDWR; |
| 235 | return file; |
| 236 | } |
| 237 | |
| 238 | static struct dentry *aio_mount(struct file_system_type *fs_type, |
| 239 | int flags, const char *dev_name, void *data) |
| 240 | { |
| 241 | static const struct dentry_operations ops = { |
| 242 | .d_dname = simple_dname, |
| 243 | }; |
| 244 | struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops, |
| 245 | AIO_RING_MAGIC); |
| 246 | |
| 247 | if (!IS_ERR(root)) |
| 248 | root->d_sb->s_iflags |= SB_I_NOEXEC; |
| 249 | return root; |
| 250 | } |
| 251 | |
| 252 | /* aio_setup |
| 253 | * Creates the slab caches used by the aio routines, panic on |
| 254 | * failure as this is done early during the boot sequence. |
| 255 | */ |
| 256 | static int __init aio_setup(void) |
| 257 | { |
| 258 | static struct file_system_type aio_fs = { |
| 259 | .name = "aio", |
| 260 | .mount = aio_mount, |
| 261 | .kill_sb = kill_anon_super, |
| 262 | }; |
| 263 | aio_mnt = kern_mount(&aio_fs); |
| 264 | if (IS_ERR(aio_mnt)) |
| 265 | panic("Failed to create aio fs mount."); |
| 266 | |
| 267 | kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
| 268 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
| 269 | |
| 270 | pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); |
| 271 | |
| 272 | return 0; |
| 273 | } |
| 274 | __initcall(aio_setup); |
| 275 | |
| 276 | static void put_aio_ring_file(struct kioctx *ctx) |
| 277 | { |
| 278 | struct file *aio_ring_file = ctx->aio_ring_file; |
| 279 | struct address_space *i_mapping; |
| 280 | |
| 281 | if (aio_ring_file) { |
| 282 | truncate_setsize(file_inode(aio_ring_file), 0); |
| 283 | |
| 284 | /* Prevent further access to the kioctx from migratepages */ |
| 285 | i_mapping = aio_ring_file->f_mapping; |
| 286 | spin_lock(&i_mapping->private_lock); |
| 287 | i_mapping->private_data = NULL; |
| 288 | ctx->aio_ring_file = NULL; |
| 289 | spin_unlock(&i_mapping->private_lock); |
| 290 | |
| 291 | fput(aio_ring_file); |
| 292 | } |
| 293 | } |
| 294 | |
| 295 | static void aio_free_ring(struct kioctx *ctx) |
| 296 | { |
| 297 | int i; |
| 298 | |
| 299 | /* Disconnect the kiotx from the ring file. This prevents future |
| 300 | * accesses to the kioctx from page migration. |
| 301 | */ |
| 302 | put_aio_ring_file(ctx); |
| 303 | |
| 304 | for (i = 0; i < ctx->nr_pages; i++) { |
| 305 | struct page *page; |
| 306 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, |
| 307 | page_count(ctx->ring_pages[i])); |
| 308 | page = ctx->ring_pages[i]; |
| 309 | if (!page) |
| 310 | continue; |
| 311 | ctx->ring_pages[i] = NULL; |
| 312 | put_page(page); |
| 313 | } |
| 314 | |
| 315 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { |
| 316 | kfree(ctx->ring_pages); |
| 317 | ctx->ring_pages = NULL; |
| 318 | } |
| 319 | } |
| 320 | |
| 321 | static int aio_ring_mremap(struct vm_area_struct *vma) |
| 322 | { |
| 323 | struct file *file = vma->vm_file; |
| 324 | struct mm_struct *mm = vma->vm_mm; |
| 325 | struct kioctx_table *table; |
| 326 | int i, res = -EINVAL; |
| 327 | |
| 328 | spin_lock(&mm->ioctx_lock); |
| 329 | rcu_read_lock(); |
| 330 | table = rcu_dereference(mm->ioctx_table); |
| 331 | for (i = 0; i < table->nr; i++) { |
| 332 | struct kioctx *ctx; |
| 333 | |
| 334 | ctx = rcu_dereference(table->table[i]); |
| 335 | if (ctx && ctx->aio_ring_file == file) { |
| 336 | if (!atomic_read(&ctx->dead)) { |
| 337 | ctx->user_id = ctx->mmap_base = vma->vm_start; |
| 338 | res = 0; |
| 339 | } |
| 340 | break; |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | rcu_read_unlock(); |
| 345 | spin_unlock(&mm->ioctx_lock); |
| 346 | return res; |
| 347 | } |
| 348 | |
| 349 | static const struct vm_operations_struct aio_ring_vm_ops = { |
| 350 | .mremap = aio_ring_mremap, |
| 351 | #if IS_ENABLED(CONFIG_MMU) |
| 352 | .fault = filemap_fault, |
| 353 | .map_pages = filemap_map_pages, |
| 354 | .page_mkwrite = filemap_page_mkwrite, |
| 355 | #endif |
| 356 | }; |
| 357 | |
| 358 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) |
| 359 | { |
| 360 | vma->vm_flags |= VM_DONTEXPAND; |
| 361 | vma->vm_ops = &aio_ring_vm_ops; |
| 362 | return 0; |
| 363 | } |
| 364 | |
| 365 | static const struct file_operations aio_ring_fops = { |
| 366 | .mmap = aio_ring_mmap, |
| 367 | }; |
| 368 | |
| 369 | #if IS_ENABLED(CONFIG_MIGRATION) |
| 370 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
| 371 | struct page *old, enum migrate_mode mode) |
| 372 | { |
| 373 | struct kioctx *ctx; |
| 374 | unsigned long flags; |
| 375 | pgoff_t idx; |
| 376 | int rc; |
| 377 | |
| 378 | /* |
| 379 | * We cannot support the _NO_COPY case here, because copy needs to |
| 380 | * happen under the ctx->completion_lock. That does not work with the |
| 381 | * migration workflow of MIGRATE_SYNC_NO_COPY. |
| 382 | */ |
| 383 | if (mode == MIGRATE_SYNC_NO_COPY) |
| 384 | return -EINVAL; |
| 385 | |
| 386 | rc = 0; |
| 387 | |
| 388 | /* mapping->private_lock here protects against the kioctx teardown. */ |
| 389 | spin_lock(&mapping->private_lock); |
| 390 | ctx = mapping->private_data; |
| 391 | if (!ctx) { |
| 392 | rc = -EINVAL; |
| 393 | goto out; |
| 394 | } |
| 395 | |
| 396 | /* The ring_lock mutex. The prevents aio_read_events() from writing |
| 397 | * to the ring's head, and prevents page migration from mucking in |
| 398 | * a partially initialized kiotx. |
| 399 | */ |
| 400 | if (!mutex_trylock(&ctx->ring_lock)) { |
| 401 | rc = -EAGAIN; |
| 402 | goto out; |
| 403 | } |
| 404 | |
| 405 | idx = old->index; |
| 406 | if (idx < (pgoff_t)ctx->nr_pages) { |
| 407 | /* Make sure the old page hasn't already been changed */ |
| 408 | if (ctx->ring_pages[idx] != old) |
| 409 | rc = -EAGAIN; |
| 410 | } else |
| 411 | rc = -EINVAL; |
| 412 | |
| 413 | if (rc != 0) |
| 414 | goto out_unlock; |
| 415 | |
| 416 | /* Writeback must be complete */ |
| 417 | BUG_ON(PageWriteback(old)); |
| 418 | get_page(new); |
| 419 | |
| 420 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); |
| 421 | if (rc != MIGRATEPAGE_SUCCESS) { |
| 422 | put_page(new); |
| 423 | goto out_unlock; |
| 424 | } |
| 425 | |
| 426 | /* Take completion_lock to prevent other writes to the ring buffer |
| 427 | * while the old page is copied to the new. This prevents new |
| 428 | * events from being lost. |
| 429 | */ |
| 430 | spin_lock_irqsave(&ctx->completion_lock, flags); |
| 431 | migrate_page_copy(new, old); |
| 432 | BUG_ON(ctx->ring_pages[idx] != old); |
| 433 | ctx->ring_pages[idx] = new; |
| 434 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
| 435 | |
| 436 | /* The old page is no longer accessible. */ |
| 437 | put_page(old); |
| 438 | |
| 439 | out_unlock: |
| 440 | mutex_unlock(&ctx->ring_lock); |
| 441 | out: |
| 442 | spin_unlock(&mapping->private_lock); |
| 443 | return rc; |
| 444 | } |
| 445 | #endif |
| 446 | |
| 447 | static const struct address_space_operations aio_ctx_aops = { |
| 448 | .set_page_dirty = __set_page_dirty_no_writeback, |
| 449 | #if IS_ENABLED(CONFIG_MIGRATION) |
| 450 | .migratepage = aio_migratepage, |
| 451 | #endif |
| 452 | }; |
| 453 | |
| 454 | static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events) |
| 455 | { |
| 456 | struct aio_ring *ring; |
| 457 | struct mm_struct *mm = current->mm; |
| 458 | unsigned long size, unused; |
| 459 | int nr_pages; |
| 460 | int i; |
| 461 | struct file *file; |
| 462 | |
| 463 | /* Compensate for the ring buffer's head/tail overlap entry */ |
| 464 | nr_events += 2; /* 1 is required, 2 for good luck */ |
| 465 | |
| 466 | size = sizeof(struct aio_ring); |
| 467 | size += sizeof(struct io_event) * nr_events; |
| 468 | |
| 469 | nr_pages = PFN_UP(size); |
| 470 | if (nr_pages < 0) |
| 471 | return -EINVAL; |
| 472 | |
| 473 | file = aio_private_file(ctx, nr_pages); |
| 474 | if (IS_ERR(file)) { |
| 475 | ctx->aio_ring_file = NULL; |
| 476 | return -ENOMEM; |
| 477 | } |
| 478 | |
| 479 | ctx->aio_ring_file = file; |
| 480 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) |
| 481 | / sizeof(struct io_event); |
| 482 | |
| 483 | ctx->ring_pages = ctx->internal_pages; |
| 484 | if (nr_pages > AIO_RING_PAGES) { |
| 485 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), |
| 486 | GFP_KERNEL); |
| 487 | if (!ctx->ring_pages) { |
| 488 | put_aio_ring_file(ctx); |
| 489 | return -ENOMEM; |
| 490 | } |
| 491 | } |
| 492 | |
| 493 | for (i = 0; i < nr_pages; i++) { |
| 494 | struct page *page; |
| 495 | page = find_or_create_page(file->f_mapping, |
| 496 | i, GFP_HIGHUSER | __GFP_ZERO); |
| 497 | if (!page) |
| 498 | break; |
| 499 | pr_debug("pid(%d) page[%d]->count=%d\n", |
| 500 | current->pid, i, page_count(page)); |
| 501 | SetPageUptodate(page); |
| 502 | unlock_page(page); |
| 503 | |
| 504 | ctx->ring_pages[i] = page; |
| 505 | } |
| 506 | ctx->nr_pages = i; |
| 507 | |
| 508 | if (unlikely(i != nr_pages)) { |
| 509 | aio_free_ring(ctx); |
| 510 | return -ENOMEM; |
| 511 | } |
| 512 | |
| 513 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
| 514 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); |
| 515 | |
| 516 | if (down_write_killable(&mm->mmap_sem)) { |
| 517 | ctx->mmap_size = 0; |
| 518 | aio_free_ring(ctx); |
| 519 | return -EINTR; |
| 520 | } |
| 521 | |
| 522 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
| 523 | PROT_READ | PROT_WRITE, |
| 524 | MAP_SHARED, 0, &unused, NULL); |
| 525 | up_write(&mm->mmap_sem); |
| 526 | if (IS_ERR((void *)ctx->mmap_base)) { |
| 527 | ctx->mmap_size = 0; |
| 528 | aio_free_ring(ctx); |
| 529 | return -ENOMEM; |
| 530 | } |
| 531 | |
| 532 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
| 533 | |
| 534 | ctx->user_id = ctx->mmap_base; |
| 535 | ctx->nr_events = nr_events; /* trusted copy */ |
| 536 | |
| 537 | ring = kmap_atomic(ctx->ring_pages[0]); |
| 538 | ring->nr = nr_events; /* user copy */ |
| 539 | ring->id = ~0U; |
| 540 | ring->head = ring->tail = 0; |
| 541 | ring->magic = AIO_RING_MAGIC; |
| 542 | ring->compat_features = AIO_RING_COMPAT_FEATURES; |
| 543 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; |
| 544 | ring->header_length = sizeof(struct aio_ring); |
| 545 | kunmap_atomic(ring); |
| 546 | flush_dcache_page(ctx->ring_pages[0]); |
| 547 | |
| 548 | return 0; |
| 549 | } |
| 550 | |
| 551 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
| 552 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) |
| 553 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) |
| 554 | |
| 555 | void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel) |
| 556 | { |
| 557 | struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common); |
| 558 | struct kioctx *ctx = req->ki_ctx; |
| 559 | unsigned long flags; |
| 560 | |
| 561 | spin_lock_irqsave(&ctx->ctx_lock, flags); |
| 562 | |
| 563 | if (!req->ki_list.next) |
| 564 | list_add(&req->ki_list, &ctx->active_reqs); |
| 565 | |
| 566 | req->ki_cancel = cancel; |
| 567 | |
| 568 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); |
| 569 | } |
| 570 | EXPORT_SYMBOL(kiocb_set_cancel_fn); |
| 571 | |
| 572 | static int kiocb_cancel(struct aio_kiocb *kiocb) |
| 573 | { |
| 574 | kiocb_cancel_fn *old, *cancel; |
| 575 | |
| 576 | /* |
| 577 | * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it |
| 578 | * actually has a cancel function, hence the cmpxchg() |
| 579 | */ |
| 580 | |
| 581 | cancel = ACCESS_ONCE(kiocb->ki_cancel); |
| 582 | do { |
| 583 | if (!cancel || cancel == KIOCB_CANCELLED) |
| 584 | return -EINVAL; |
| 585 | |
| 586 | old = cancel; |
| 587 | cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); |
| 588 | } while (cancel != old); |
| 589 | |
| 590 | return cancel(&kiocb->common); |
| 591 | } |
| 592 | |
| 593 | /* |
| 594 | * free_ioctx() should be RCU delayed to synchronize against the RCU |
| 595 | * protected lookup_ioctx() and also needs process context to call |
| 596 | * aio_free_ring(), so the double bouncing through kioctx->free_rcu and |
| 597 | * ->free_work. |
| 598 | */ |
| 599 | static void free_ioctx(struct work_struct *work) |
| 600 | { |
| 601 | struct kioctx *ctx = container_of(work, struct kioctx, free_work); |
| 602 | |
| 603 | pr_debug("freeing %p\n", ctx); |
| 604 | |
| 605 | aio_free_ring(ctx); |
| 606 | free_percpu(ctx->cpu); |
| 607 | percpu_ref_exit(&ctx->reqs); |
| 608 | percpu_ref_exit(&ctx->users); |
| 609 | kmem_cache_free(kioctx_cachep, ctx); |
| 610 | } |
| 611 | |
| 612 | static void free_ioctx_rcufn(struct rcu_head *head) |
| 613 | { |
| 614 | struct kioctx *ctx = container_of(head, struct kioctx, free_rcu); |
| 615 | |
| 616 | INIT_WORK(&ctx->free_work, free_ioctx); |
| 617 | schedule_work(&ctx->free_work); |
| 618 | } |
| 619 | |
| 620 | static void free_ioctx_reqs(struct percpu_ref *ref) |
| 621 | { |
| 622 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); |
| 623 | |
| 624 | /* At this point we know that there are no any in-flight requests */ |
| 625 | if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count)) |
| 626 | complete(&ctx->rq_wait->comp); |
| 627 | |
| 628 | /* Synchronize against RCU protected table->table[] dereferences */ |
| 629 | call_rcu(&ctx->free_rcu, free_ioctx_rcufn); |
| 630 | } |
| 631 | |
| 632 | /* |
| 633 | * When this function runs, the kioctx has been removed from the "hash table" |
| 634 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - |
| 635 | * now it's safe to cancel any that need to be. |
| 636 | */ |
| 637 | static void free_ioctx_users(struct percpu_ref *ref) |
| 638 | { |
| 639 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
| 640 | struct aio_kiocb *req; |
| 641 | |
| 642 | spin_lock_irq(&ctx->ctx_lock); |
| 643 | |
| 644 | while (!list_empty(&ctx->active_reqs)) { |
| 645 | req = list_first_entry(&ctx->active_reqs, |
| 646 | struct aio_kiocb, ki_list); |
| 647 | kiocb_cancel(req); |
| 648 | list_del_init(&req->ki_list); |
| 649 | } |
| 650 | |
| 651 | spin_unlock_irq(&ctx->ctx_lock); |
| 652 | |
| 653 | percpu_ref_kill(&ctx->reqs); |
| 654 | percpu_ref_put(&ctx->reqs); |
| 655 | } |
| 656 | |
| 657 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
| 658 | { |
| 659 | unsigned i, new_nr; |
| 660 | struct kioctx_table *table, *old; |
| 661 | struct aio_ring *ring; |
| 662 | |
| 663 | spin_lock(&mm->ioctx_lock); |
| 664 | table = rcu_dereference_raw(mm->ioctx_table); |
| 665 | |
| 666 | while (1) { |
| 667 | if (table) |
| 668 | for (i = 0; i < table->nr; i++) |
| 669 | if (!rcu_access_pointer(table->table[i])) { |
| 670 | ctx->id = i; |
| 671 | rcu_assign_pointer(table->table[i], ctx); |
| 672 | spin_unlock(&mm->ioctx_lock); |
| 673 | |
| 674 | /* While kioctx setup is in progress, |
| 675 | * we are protected from page migration |
| 676 | * changes ring_pages by ->ring_lock. |
| 677 | */ |
| 678 | ring = kmap_atomic(ctx->ring_pages[0]); |
| 679 | ring->id = ctx->id; |
| 680 | kunmap_atomic(ring); |
| 681 | return 0; |
| 682 | } |
| 683 | |
| 684 | new_nr = (table ? table->nr : 1) * 4; |
| 685 | spin_unlock(&mm->ioctx_lock); |
| 686 | |
| 687 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * |
| 688 | new_nr, GFP_KERNEL); |
| 689 | if (!table) |
| 690 | return -ENOMEM; |
| 691 | |
| 692 | table->nr = new_nr; |
| 693 | |
| 694 | spin_lock(&mm->ioctx_lock); |
| 695 | old = rcu_dereference_raw(mm->ioctx_table); |
| 696 | |
| 697 | if (!old) { |
| 698 | rcu_assign_pointer(mm->ioctx_table, table); |
| 699 | } else if (table->nr > old->nr) { |
| 700 | memcpy(table->table, old->table, |
| 701 | old->nr * sizeof(struct kioctx *)); |
| 702 | |
| 703 | rcu_assign_pointer(mm->ioctx_table, table); |
| 704 | kfree_rcu(old, rcu); |
| 705 | } else { |
| 706 | kfree(table); |
| 707 | table = old; |
| 708 | } |
| 709 | } |
| 710 | } |
| 711 | |
| 712 | static void aio_nr_sub(unsigned nr) |
| 713 | { |
| 714 | spin_lock(&aio_nr_lock); |
| 715 | if (WARN_ON(aio_nr - nr > aio_nr)) |
| 716 | aio_nr = 0; |
| 717 | else |
| 718 | aio_nr -= nr; |
| 719 | spin_unlock(&aio_nr_lock); |
| 720 | } |
| 721 | |
| 722 | /* ioctx_alloc |
| 723 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. |
| 724 | */ |
| 725 | static struct kioctx *ioctx_alloc(unsigned nr_events) |
| 726 | { |
| 727 | struct mm_struct *mm = current->mm; |
| 728 | struct kioctx *ctx; |
| 729 | int err = -ENOMEM; |
| 730 | |
| 731 | /* |
| 732 | * Store the original nr_events -- what userspace passed to io_setup(), |
| 733 | * for counting against the global limit -- before it changes. |
| 734 | */ |
| 735 | unsigned int max_reqs = nr_events; |
| 736 | |
| 737 | /* |
| 738 | * We keep track of the number of available ringbuffer slots, to prevent |
| 739 | * overflow (reqs_available), and we also use percpu counters for this. |
| 740 | * |
| 741 | * So since up to half the slots might be on other cpu's percpu counters |
| 742 | * and unavailable, double nr_events so userspace sees what they |
| 743 | * expected: additionally, we move req_batch slots to/from percpu |
| 744 | * counters at a time, so make sure that isn't 0: |
| 745 | */ |
| 746 | nr_events = max(nr_events, num_possible_cpus() * 4); |
| 747 | nr_events *= 2; |
| 748 | |
| 749 | /* Prevent overflows */ |
| 750 | if (nr_events > (0x10000000U / sizeof(struct io_event))) { |
| 751 | pr_debug("ENOMEM: nr_events too high\n"); |
| 752 | return ERR_PTR(-EINVAL); |
| 753 | } |
| 754 | |
| 755 | if (!nr_events || (unsigned long)max_reqs > aio_max_nr) |
| 756 | return ERR_PTR(-EAGAIN); |
| 757 | |
| 758 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
| 759 | if (!ctx) |
| 760 | return ERR_PTR(-ENOMEM); |
| 761 | |
| 762 | ctx->max_reqs = max_reqs; |
| 763 | |
| 764 | spin_lock_init(&ctx->ctx_lock); |
| 765 | spin_lock_init(&ctx->completion_lock); |
| 766 | mutex_init(&ctx->ring_lock); |
| 767 | /* Protect against page migration throughout kiotx setup by keeping |
| 768 | * the ring_lock mutex held until setup is complete. */ |
| 769 | mutex_lock(&ctx->ring_lock); |
| 770 | init_waitqueue_head(&ctx->wait); |
| 771 | |
| 772 | INIT_LIST_HEAD(&ctx->active_reqs); |
| 773 | |
| 774 | if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) |
| 775 | goto err; |
| 776 | |
| 777 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) |
| 778 | goto err; |
| 779 | |
| 780 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
| 781 | if (!ctx->cpu) |
| 782 | goto err; |
| 783 | |
| 784 | err = aio_setup_ring(ctx, nr_events); |
| 785 | if (err < 0) |
| 786 | goto err; |
| 787 | |
| 788 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
| 789 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
| 790 | if (ctx->req_batch < 1) |
| 791 | ctx->req_batch = 1; |
| 792 | |
| 793 | /* limit the number of system wide aios */ |
| 794 | spin_lock(&aio_nr_lock); |
| 795 | if (aio_nr + ctx->max_reqs > aio_max_nr || |
| 796 | aio_nr + ctx->max_reqs < aio_nr) { |
| 797 | spin_unlock(&aio_nr_lock); |
| 798 | err = -EAGAIN; |
| 799 | goto err_ctx; |
| 800 | } |
| 801 | aio_nr += ctx->max_reqs; |
| 802 | spin_unlock(&aio_nr_lock); |
| 803 | |
| 804 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
| 805 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ |
| 806 | |
| 807 | err = ioctx_add_table(ctx, mm); |
| 808 | if (err) |
| 809 | goto err_cleanup; |
| 810 | |
| 811 | /* Release the ring_lock mutex now that all setup is complete. */ |
| 812 | mutex_unlock(&ctx->ring_lock); |
| 813 | |
| 814 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
| 815 | ctx, ctx->user_id, mm, ctx->nr_events); |
| 816 | return ctx; |
| 817 | |
| 818 | err_cleanup: |
| 819 | aio_nr_sub(ctx->max_reqs); |
| 820 | err_ctx: |
| 821 | atomic_set(&ctx->dead, 1); |
| 822 | if (ctx->mmap_size) |
| 823 | vm_munmap(ctx->mmap_base, ctx->mmap_size); |
| 824 | aio_free_ring(ctx); |
| 825 | err: |
| 826 | mutex_unlock(&ctx->ring_lock); |
| 827 | free_percpu(ctx->cpu); |
| 828 | percpu_ref_exit(&ctx->reqs); |
| 829 | percpu_ref_exit(&ctx->users); |
| 830 | kmem_cache_free(kioctx_cachep, ctx); |
| 831 | pr_debug("error allocating ioctx %d\n", err); |
| 832 | return ERR_PTR(err); |
| 833 | } |
| 834 | |
| 835 | /* kill_ioctx |
| 836 | * Cancels all outstanding aio requests on an aio context. Used |
| 837 | * when the processes owning a context have all exited to encourage |
| 838 | * the rapid destruction of the kioctx. |
| 839 | */ |
| 840 | static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, |
| 841 | struct ctx_rq_wait *wait) |
| 842 | { |
| 843 | struct kioctx_table *table; |
| 844 | |
| 845 | spin_lock(&mm->ioctx_lock); |
| 846 | if (atomic_xchg(&ctx->dead, 1)) { |
| 847 | spin_unlock(&mm->ioctx_lock); |
| 848 | return -EINVAL; |
| 849 | } |
| 850 | |
| 851 | table = rcu_dereference_raw(mm->ioctx_table); |
| 852 | WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id])); |
| 853 | RCU_INIT_POINTER(table->table[ctx->id], NULL); |
| 854 | spin_unlock(&mm->ioctx_lock); |
| 855 | |
| 856 | /* free_ioctx_reqs() will do the necessary RCU synchronization */ |
| 857 | wake_up_all(&ctx->wait); |
| 858 | |
| 859 | /* |
| 860 | * It'd be more correct to do this in free_ioctx(), after all |
| 861 | * the outstanding kiocbs have finished - but by then io_destroy |
| 862 | * has already returned, so io_setup() could potentially return |
| 863 | * -EAGAIN with no ioctxs actually in use (as far as userspace |
| 864 | * could tell). |
| 865 | */ |
| 866 | aio_nr_sub(ctx->max_reqs); |
| 867 | |
| 868 | if (ctx->mmap_size) |
| 869 | vm_munmap(ctx->mmap_base, ctx->mmap_size); |
| 870 | |
| 871 | ctx->rq_wait = wait; |
| 872 | percpu_ref_kill(&ctx->users); |
| 873 | return 0; |
| 874 | } |
| 875 | |
| 876 | /* |
| 877 | * exit_aio: called when the last user of mm goes away. At this point, there is |
| 878 | * no way for any new requests to be submited or any of the io_* syscalls to be |
| 879 | * called on the context. |
| 880 | * |
| 881 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on |
| 882 | * them. |
| 883 | */ |
| 884 | void exit_aio(struct mm_struct *mm) |
| 885 | { |
| 886 | struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); |
| 887 | struct ctx_rq_wait wait; |
| 888 | int i, skipped; |
| 889 | |
| 890 | if (!table) |
| 891 | return; |
| 892 | |
| 893 | atomic_set(&wait.count, table->nr); |
| 894 | init_completion(&wait.comp); |
| 895 | |
| 896 | skipped = 0; |
| 897 | for (i = 0; i < table->nr; ++i) { |
| 898 | struct kioctx *ctx = |
| 899 | rcu_dereference_protected(table->table[i], true); |
| 900 | |
| 901 | if (!ctx) { |
| 902 | skipped++; |
| 903 | continue; |
| 904 | } |
| 905 | |
| 906 | /* |
| 907 | * We don't need to bother with munmap() here - exit_mmap(mm) |
| 908 | * is coming and it'll unmap everything. And we simply can't, |
| 909 | * this is not necessarily our ->mm. |
| 910 | * Since kill_ioctx() uses non-zero ->mmap_size as indicator |
| 911 | * that it needs to unmap the area, just set it to 0. |
| 912 | */ |
| 913 | ctx->mmap_size = 0; |
| 914 | kill_ioctx(mm, ctx, &wait); |
| 915 | } |
| 916 | |
| 917 | if (!atomic_sub_and_test(skipped, &wait.count)) { |
| 918 | /* Wait until all IO for the context are done. */ |
| 919 | wait_for_completion(&wait.comp); |
| 920 | } |
| 921 | |
| 922 | RCU_INIT_POINTER(mm->ioctx_table, NULL); |
| 923 | kfree(table); |
| 924 | } |
| 925 | |
| 926 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
| 927 | { |
| 928 | struct kioctx_cpu *kcpu; |
| 929 | unsigned long flags; |
| 930 | |
| 931 | local_irq_save(flags); |
| 932 | kcpu = this_cpu_ptr(ctx->cpu); |
| 933 | kcpu->reqs_available += nr; |
| 934 | |
| 935 | while (kcpu->reqs_available >= ctx->req_batch * 2) { |
| 936 | kcpu->reqs_available -= ctx->req_batch; |
| 937 | atomic_add(ctx->req_batch, &ctx->reqs_available); |
| 938 | } |
| 939 | |
| 940 | local_irq_restore(flags); |
| 941 | } |
| 942 | |
| 943 | static bool get_reqs_available(struct kioctx *ctx) |
| 944 | { |
| 945 | struct kioctx_cpu *kcpu; |
| 946 | bool ret = false; |
| 947 | unsigned long flags; |
| 948 | |
| 949 | local_irq_save(flags); |
| 950 | kcpu = this_cpu_ptr(ctx->cpu); |
| 951 | if (!kcpu->reqs_available) { |
| 952 | int old, avail = atomic_read(&ctx->reqs_available); |
| 953 | |
| 954 | do { |
| 955 | if (avail < ctx->req_batch) |
| 956 | goto out; |
| 957 | |
| 958 | old = avail; |
| 959 | avail = atomic_cmpxchg(&ctx->reqs_available, |
| 960 | avail, avail - ctx->req_batch); |
| 961 | } while (avail != old); |
| 962 | |
| 963 | kcpu->reqs_available += ctx->req_batch; |
| 964 | } |
| 965 | |
| 966 | ret = true; |
| 967 | kcpu->reqs_available--; |
| 968 | out: |
| 969 | local_irq_restore(flags); |
| 970 | return ret; |
| 971 | } |
| 972 | |
| 973 | /* refill_reqs_available |
| 974 | * Updates the reqs_available reference counts used for tracking the |
| 975 | * number of free slots in the completion ring. This can be called |
| 976 | * from aio_complete() (to optimistically update reqs_available) or |
| 977 | * from aio_get_req() (the we're out of events case). It must be |
| 978 | * called holding ctx->completion_lock. |
| 979 | */ |
| 980 | static void refill_reqs_available(struct kioctx *ctx, unsigned head, |
| 981 | unsigned tail) |
| 982 | { |
| 983 | unsigned events_in_ring, completed; |
| 984 | |
| 985 | /* Clamp head since userland can write to it. */ |
| 986 | head %= ctx->nr_events; |
| 987 | if (head <= tail) |
| 988 | events_in_ring = tail - head; |
| 989 | else |
| 990 | events_in_ring = ctx->nr_events - (head - tail); |
| 991 | |
| 992 | completed = ctx->completed_events; |
| 993 | if (events_in_ring < completed) |
| 994 | completed -= events_in_ring; |
| 995 | else |
| 996 | completed = 0; |
| 997 | |
| 998 | if (!completed) |
| 999 | return; |
| 1000 | |
| 1001 | ctx->completed_events -= completed; |
| 1002 | put_reqs_available(ctx, completed); |
| 1003 | } |
| 1004 | |
| 1005 | /* user_refill_reqs_available |
| 1006 | * Called to refill reqs_available when aio_get_req() encounters an |
| 1007 | * out of space in the completion ring. |
| 1008 | */ |
| 1009 | static void user_refill_reqs_available(struct kioctx *ctx) |
| 1010 | { |
| 1011 | spin_lock_irq(&ctx->completion_lock); |
| 1012 | if (ctx->completed_events) { |
| 1013 | struct aio_ring *ring; |
| 1014 | unsigned head; |
| 1015 | |
| 1016 | /* Access of ring->head may race with aio_read_events_ring() |
| 1017 | * here, but that's okay since whether we read the old version |
| 1018 | * or the new version, and either will be valid. The important |
| 1019 | * part is that head cannot pass tail since we prevent |
| 1020 | * aio_complete() from updating tail by holding |
| 1021 | * ctx->completion_lock. Even if head is invalid, the check |
| 1022 | * against ctx->completed_events below will make sure we do the |
| 1023 | * safe/right thing. |
| 1024 | */ |
| 1025 | ring = kmap_atomic(ctx->ring_pages[0]); |
| 1026 | head = ring->head; |
| 1027 | kunmap_atomic(ring); |
| 1028 | |
| 1029 | refill_reqs_available(ctx, head, ctx->tail); |
| 1030 | } |
| 1031 | |
| 1032 | spin_unlock_irq(&ctx->completion_lock); |
| 1033 | } |
| 1034 | |
| 1035 | /* aio_get_req |
| 1036 | * Allocate a slot for an aio request. |
| 1037 | * Returns NULL if no requests are free. |
| 1038 | */ |
| 1039 | static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx) |
| 1040 | { |
| 1041 | struct aio_kiocb *req; |
| 1042 | |
| 1043 | if (!get_reqs_available(ctx)) { |
| 1044 | user_refill_reqs_available(ctx); |
| 1045 | if (!get_reqs_available(ctx)) |
| 1046 | return NULL; |
| 1047 | } |
| 1048 | |
| 1049 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
| 1050 | if (unlikely(!req)) |
| 1051 | goto out_put; |
| 1052 | |
| 1053 | percpu_ref_get(&ctx->reqs); |
| 1054 | |
| 1055 | req->ki_ctx = ctx; |
| 1056 | return req; |
| 1057 | out_put: |
| 1058 | put_reqs_available(ctx, 1); |
| 1059 | return NULL; |
| 1060 | } |
| 1061 | |
| 1062 | static void kiocb_free(struct aio_kiocb *req) |
| 1063 | { |
| 1064 | if (req->common.ki_filp) |
| 1065 | fput(req->common.ki_filp); |
| 1066 | if (req->ki_eventfd != NULL) |
| 1067 | eventfd_ctx_put(req->ki_eventfd); |
| 1068 | kmem_cache_free(kiocb_cachep, req); |
| 1069 | } |
| 1070 | |
| 1071 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
| 1072 | { |
| 1073 | struct aio_ring __user *ring = (void __user *)ctx_id; |
| 1074 | struct mm_struct *mm = current->mm; |
| 1075 | struct kioctx *ctx, *ret = NULL; |
| 1076 | struct kioctx_table *table; |
| 1077 | unsigned id; |
| 1078 | |
| 1079 | if (get_user(id, &ring->id)) |
| 1080 | return NULL; |
| 1081 | |
| 1082 | rcu_read_lock(); |
| 1083 | table = rcu_dereference(mm->ioctx_table); |
| 1084 | |
| 1085 | if (!table || id >= table->nr) |
| 1086 | goto out; |
| 1087 | |
| 1088 | id = array_index_nospec(id, table->nr); |
| 1089 | ctx = rcu_dereference(table->table[id]); |
| 1090 | if (ctx && ctx->user_id == ctx_id) { |
| 1091 | if (percpu_ref_tryget_live(&ctx->users)) |
| 1092 | ret = ctx; |
| 1093 | } |
| 1094 | out: |
| 1095 | rcu_read_unlock(); |
| 1096 | return ret; |
| 1097 | } |
| 1098 | |
| 1099 | /* aio_complete |
| 1100 | * Called when the io request on the given iocb is complete. |
| 1101 | */ |
| 1102 | static void aio_complete(struct kiocb *kiocb, long res, long res2) |
| 1103 | { |
| 1104 | struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common); |
| 1105 | struct kioctx *ctx = iocb->ki_ctx; |
| 1106 | struct aio_ring *ring; |
| 1107 | struct io_event *ev_page, *event; |
| 1108 | unsigned tail, pos, head; |
| 1109 | unsigned long flags; |
| 1110 | |
| 1111 | if (kiocb->ki_flags & IOCB_WRITE) { |
| 1112 | struct file *file = kiocb->ki_filp; |
| 1113 | |
| 1114 | /* |
| 1115 | * Tell lockdep we inherited freeze protection from submission |
| 1116 | * thread. |
| 1117 | */ |
| 1118 | if (S_ISREG(file_inode(file)->i_mode)) |
| 1119 | __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE); |
| 1120 | file_end_write(file); |
| 1121 | } |
| 1122 | |
| 1123 | /* |
| 1124 | * Special case handling for sync iocbs: |
| 1125 | * - events go directly into the iocb for fast handling |
| 1126 | * - the sync task with the iocb in its stack holds the single iocb |
| 1127 | * ref, no other paths have a way to get another ref |
| 1128 | * - the sync task helpfully left a reference to itself in the iocb |
| 1129 | */ |
| 1130 | BUG_ON(is_sync_kiocb(kiocb)); |
| 1131 | |
| 1132 | if (iocb->ki_list.next) { |
| 1133 | unsigned long flags; |
| 1134 | |
| 1135 | spin_lock_irqsave(&ctx->ctx_lock, flags); |
| 1136 | list_del(&iocb->ki_list); |
| 1137 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); |
| 1138 | } |
| 1139 | |
| 1140 | /* |
| 1141 | * Add a completion event to the ring buffer. Must be done holding |
| 1142 | * ctx->completion_lock to prevent other code from messing with the tail |
| 1143 | * pointer since we might be called from irq context. |
| 1144 | */ |
| 1145 | spin_lock_irqsave(&ctx->completion_lock, flags); |
| 1146 | |
| 1147 | tail = ctx->tail; |
| 1148 | pos = tail + AIO_EVENTS_OFFSET; |
| 1149 | |
| 1150 | if (++tail >= ctx->nr_events) |
| 1151 | tail = 0; |
| 1152 | |
| 1153 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
| 1154 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
| 1155 | |
| 1156 | event->obj = (u64)(unsigned long)iocb->ki_user_iocb; |
| 1157 | event->data = iocb->ki_user_data; |
| 1158 | event->res = res; |
| 1159 | event->res2 = res2; |
| 1160 | |
| 1161 | kunmap_atomic(ev_page); |
| 1162 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
| 1163 | |
| 1164 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", |
| 1165 | ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data, |
| 1166 | res, res2); |
| 1167 | |
| 1168 | /* after flagging the request as done, we |
| 1169 | * must never even look at it again |
| 1170 | */ |
| 1171 | smp_wmb(); /* make event visible before updating tail */ |
| 1172 | |
| 1173 | ctx->tail = tail; |
| 1174 | |
| 1175 | ring = kmap_atomic(ctx->ring_pages[0]); |
| 1176 | head = ring->head; |
| 1177 | ring->tail = tail; |
| 1178 | kunmap_atomic(ring); |
| 1179 | flush_dcache_page(ctx->ring_pages[0]); |
| 1180 | |
| 1181 | ctx->completed_events++; |
| 1182 | if (ctx->completed_events > 1) |
| 1183 | refill_reqs_available(ctx, head, tail); |
| 1184 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
| 1185 | |
| 1186 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
| 1187 | |
| 1188 | /* |
| 1189 | * Check if the user asked us to deliver the result through an |
| 1190 | * eventfd. The eventfd_signal() function is safe to be called |
| 1191 | * from IRQ context. |
| 1192 | */ |
| 1193 | if (iocb->ki_eventfd != NULL) |
| 1194 | eventfd_signal(iocb->ki_eventfd, 1); |
| 1195 | |
| 1196 | /* everything turned out well, dispose of the aiocb. */ |
| 1197 | kiocb_free(iocb); |
| 1198 | |
| 1199 | /* |
| 1200 | * We have to order our ring_info tail store above and test |
| 1201 | * of the wait list below outside the wait lock. This is |
| 1202 | * like in wake_up_bit() where clearing a bit has to be |
| 1203 | * ordered with the unlocked test. |
| 1204 | */ |
| 1205 | smp_mb(); |
| 1206 | |
| 1207 | if (waitqueue_active(&ctx->wait)) |
| 1208 | wake_up(&ctx->wait); |
| 1209 | |
| 1210 | percpu_ref_put(&ctx->reqs); |
| 1211 | } |
| 1212 | |
| 1213 | /* aio_read_events_ring |
| 1214 | * Pull an event off of the ioctx's event ring. Returns the number of |
| 1215 | * events fetched |
| 1216 | */ |
| 1217 | static long aio_read_events_ring(struct kioctx *ctx, |
| 1218 | struct io_event __user *event, long nr) |
| 1219 | { |
| 1220 | struct aio_ring *ring; |
| 1221 | unsigned head, tail, pos; |
| 1222 | long ret = 0; |
| 1223 | int copy_ret; |
| 1224 | |
| 1225 | /* |
| 1226 | * The mutex can block and wake us up and that will cause |
| 1227 | * wait_event_interruptible_hrtimeout() to schedule without sleeping |
| 1228 | * and repeat. This should be rare enough that it doesn't cause |
| 1229 | * peformance issues. See the comment in read_events() for more detail. |
| 1230 | */ |
| 1231 | sched_annotate_sleep(); |
| 1232 | mutex_lock(&ctx->ring_lock); |
| 1233 | |
| 1234 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
| 1235 | ring = kmap_atomic(ctx->ring_pages[0]); |
| 1236 | head = ring->head; |
| 1237 | tail = ring->tail; |
| 1238 | kunmap_atomic(ring); |
| 1239 | |
| 1240 | /* |
| 1241 | * Ensure that once we've read the current tail pointer, that |
| 1242 | * we also see the events that were stored up to the tail. |
| 1243 | */ |
| 1244 | smp_rmb(); |
| 1245 | |
| 1246 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
| 1247 | |
| 1248 | if (head == tail) |
| 1249 | goto out; |
| 1250 | |
| 1251 | head %= ctx->nr_events; |
| 1252 | tail %= ctx->nr_events; |
| 1253 | |
| 1254 | while (ret < nr) { |
| 1255 | long avail; |
| 1256 | struct io_event *ev; |
| 1257 | struct page *page; |
| 1258 | |
| 1259 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
| 1260 | if (head == tail) |
| 1261 | break; |
| 1262 | |
| 1263 | avail = min(avail, nr - ret); |
| 1264 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - |
| 1265 | ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); |
| 1266 | |
| 1267 | pos = head + AIO_EVENTS_OFFSET; |
| 1268 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
| 1269 | pos %= AIO_EVENTS_PER_PAGE; |
| 1270 | |
| 1271 | ev = kmap(page); |
| 1272 | copy_ret = copy_to_user(event + ret, ev + pos, |
| 1273 | sizeof(*ev) * avail); |
| 1274 | kunmap(page); |
| 1275 | |
| 1276 | if (unlikely(copy_ret)) { |
| 1277 | ret = -EFAULT; |
| 1278 | goto out; |
| 1279 | } |
| 1280 | |
| 1281 | ret += avail; |
| 1282 | head += avail; |
| 1283 | head %= ctx->nr_events; |
| 1284 | } |
| 1285 | |
| 1286 | ring = kmap_atomic(ctx->ring_pages[0]); |
| 1287 | ring->head = head; |
| 1288 | kunmap_atomic(ring); |
| 1289 | flush_dcache_page(ctx->ring_pages[0]); |
| 1290 | |
| 1291 | pr_debug("%li h%u t%u\n", ret, head, tail); |
| 1292 | out: |
| 1293 | mutex_unlock(&ctx->ring_lock); |
| 1294 | |
| 1295 | return ret; |
| 1296 | } |
| 1297 | |
| 1298 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
| 1299 | struct io_event __user *event, long *i) |
| 1300 | { |
| 1301 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
| 1302 | |
| 1303 | if (ret > 0) |
| 1304 | *i += ret; |
| 1305 | |
| 1306 | if (unlikely(atomic_read(&ctx->dead))) |
| 1307 | ret = -EINVAL; |
| 1308 | |
| 1309 | if (!*i) |
| 1310 | *i = ret; |
| 1311 | |
| 1312 | return ret < 0 || *i >= min_nr; |
| 1313 | } |
| 1314 | |
| 1315 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
| 1316 | struct io_event __user *event, |
| 1317 | struct timespec __user *timeout) |
| 1318 | { |
| 1319 | ktime_t until = KTIME_MAX; |
| 1320 | long ret = 0; |
| 1321 | |
| 1322 | if (timeout) { |
| 1323 | struct timespec ts; |
| 1324 | |
| 1325 | if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) |
| 1326 | return -EFAULT; |
| 1327 | |
| 1328 | until = timespec_to_ktime(ts); |
| 1329 | } |
| 1330 | |
| 1331 | /* |
| 1332 | * Note that aio_read_events() is being called as the conditional - i.e. |
| 1333 | * we're calling it after prepare_to_wait() has set task state to |
| 1334 | * TASK_INTERRUPTIBLE. |
| 1335 | * |
| 1336 | * But aio_read_events() can block, and if it blocks it's going to flip |
| 1337 | * the task state back to TASK_RUNNING. |
| 1338 | * |
| 1339 | * This should be ok, provided it doesn't flip the state back to |
| 1340 | * TASK_RUNNING and return 0 too much - that causes us to spin. That |
| 1341 | * will only happen if the mutex_lock() call blocks, and we then find |
| 1342 | * the ringbuffer empty. So in practice we should be ok, but it's |
| 1343 | * something to be aware of when touching this code. |
| 1344 | */ |
| 1345 | if (until == 0) |
| 1346 | aio_read_events(ctx, min_nr, nr, event, &ret); |
| 1347 | else |
| 1348 | wait_event_interruptible_hrtimeout(ctx->wait, |
| 1349 | aio_read_events(ctx, min_nr, nr, event, &ret), |
| 1350 | until); |
| 1351 | |
| 1352 | if (!ret && signal_pending(current)) |
| 1353 | ret = -EINTR; |
| 1354 | |
| 1355 | return ret; |
| 1356 | } |
| 1357 | |
| 1358 | /* sys_io_setup: |
| 1359 | * Create an aio_context capable of receiving at least nr_events. |
| 1360 | * ctxp must not point to an aio_context that already exists, and |
| 1361 | * must be initialized to 0 prior to the call. On successful |
| 1362 | * creation of the aio_context, *ctxp is filled in with the resulting |
| 1363 | * handle. May fail with -EINVAL if *ctxp is not initialized, |
| 1364 | * if the specified nr_events exceeds internal limits. May fail |
| 1365 | * with -EAGAIN if the specified nr_events exceeds the user's limit |
| 1366 | * of available events. May fail with -ENOMEM if insufficient kernel |
| 1367 | * resources are available. May fail with -EFAULT if an invalid |
| 1368 | * pointer is passed for ctxp. Will fail with -ENOSYS if not |
| 1369 | * implemented. |
| 1370 | */ |
| 1371 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
| 1372 | { |
| 1373 | struct kioctx *ioctx = NULL; |
| 1374 | unsigned long ctx; |
| 1375 | long ret; |
| 1376 | |
| 1377 | ret = get_user(ctx, ctxp); |
| 1378 | if (unlikely(ret)) |
| 1379 | goto out; |
| 1380 | |
| 1381 | ret = -EINVAL; |
| 1382 | if (unlikely(ctx || nr_events == 0)) { |
| 1383 | pr_debug("EINVAL: ctx %lu nr_events %u\n", |
| 1384 | ctx, nr_events); |
| 1385 | goto out; |
| 1386 | } |
| 1387 | |
| 1388 | ioctx = ioctx_alloc(nr_events); |
| 1389 | ret = PTR_ERR(ioctx); |
| 1390 | if (!IS_ERR(ioctx)) { |
| 1391 | ret = put_user(ioctx->user_id, ctxp); |
| 1392 | if (ret) |
| 1393 | kill_ioctx(current->mm, ioctx, NULL); |
| 1394 | percpu_ref_put(&ioctx->users); |
| 1395 | } |
| 1396 | |
| 1397 | out: |
| 1398 | return ret; |
| 1399 | } |
| 1400 | |
| 1401 | #ifdef CONFIG_COMPAT |
| 1402 | COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p) |
| 1403 | { |
| 1404 | struct kioctx *ioctx = NULL; |
| 1405 | unsigned long ctx; |
| 1406 | long ret; |
| 1407 | |
| 1408 | ret = get_user(ctx, ctx32p); |
| 1409 | if (unlikely(ret)) |
| 1410 | goto out; |
| 1411 | |
| 1412 | ret = -EINVAL; |
| 1413 | if (unlikely(ctx || nr_events == 0)) { |
| 1414 | pr_debug("EINVAL: ctx %lu nr_events %u\n", |
| 1415 | ctx, nr_events); |
| 1416 | goto out; |
| 1417 | } |
| 1418 | |
| 1419 | ioctx = ioctx_alloc(nr_events); |
| 1420 | ret = PTR_ERR(ioctx); |
| 1421 | if (!IS_ERR(ioctx)) { |
| 1422 | /* truncating is ok because it's a user address */ |
| 1423 | ret = put_user((u32)ioctx->user_id, ctx32p); |
| 1424 | if (ret) |
| 1425 | kill_ioctx(current->mm, ioctx, NULL); |
| 1426 | percpu_ref_put(&ioctx->users); |
| 1427 | } |
| 1428 | |
| 1429 | out: |
| 1430 | return ret; |
| 1431 | } |
| 1432 | #endif |
| 1433 | |
| 1434 | /* sys_io_destroy: |
| 1435 | * Destroy the aio_context specified. May cancel any outstanding |
| 1436 | * AIOs and block on completion. Will fail with -ENOSYS if not |
| 1437 | * implemented. May fail with -EINVAL if the context pointed to |
| 1438 | * is invalid. |
| 1439 | */ |
| 1440 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
| 1441 | { |
| 1442 | struct kioctx *ioctx = lookup_ioctx(ctx); |
| 1443 | if (likely(NULL != ioctx)) { |
| 1444 | struct ctx_rq_wait wait; |
| 1445 | int ret; |
| 1446 | |
| 1447 | init_completion(&wait.comp); |
| 1448 | atomic_set(&wait.count, 1); |
| 1449 | |
| 1450 | /* Pass requests_done to kill_ioctx() where it can be set |
| 1451 | * in a thread-safe way. If we try to set it here then we have |
| 1452 | * a race condition if two io_destroy() called simultaneously. |
| 1453 | */ |
| 1454 | ret = kill_ioctx(current->mm, ioctx, &wait); |
| 1455 | percpu_ref_put(&ioctx->users); |
| 1456 | |
| 1457 | /* Wait until all IO for the context are done. Otherwise kernel |
| 1458 | * keep using user-space buffers even if user thinks the context |
| 1459 | * is destroyed. |
| 1460 | */ |
| 1461 | if (!ret) |
| 1462 | wait_for_completion(&wait.comp); |
| 1463 | |
| 1464 | return ret; |
| 1465 | } |
| 1466 | pr_debug("EINVAL: invalid context id\n"); |
| 1467 | return -EINVAL; |
| 1468 | } |
| 1469 | |
| 1470 | static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec, |
| 1471 | bool vectored, bool compat, struct iov_iter *iter) |
| 1472 | { |
| 1473 | void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf; |
| 1474 | size_t len = iocb->aio_nbytes; |
| 1475 | |
| 1476 | if (!vectored) { |
| 1477 | ssize_t ret = import_single_range(rw, buf, len, *iovec, iter); |
| 1478 | *iovec = NULL; |
| 1479 | return ret; |
| 1480 | } |
| 1481 | #ifdef CONFIG_COMPAT |
| 1482 | if (compat) |
| 1483 | return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec, |
| 1484 | iter); |
| 1485 | #endif |
| 1486 | return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter); |
| 1487 | } |
| 1488 | |
| 1489 | static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret) |
| 1490 | { |
| 1491 | switch (ret) { |
| 1492 | case -EIOCBQUEUED: |
| 1493 | return ret; |
| 1494 | case -ERESTARTSYS: |
| 1495 | case -ERESTARTNOINTR: |
| 1496 | case -ERESTARTNOHAND: |
| 1497 | case -ERESTART_RESTARTBLOCK: |
| 1498 | /* |
| 1499 | * There's no easy way to restart the syscall since other AIO's |
| 1500 | * may be already running. Just fail this IO with EINTR. |
| 1501 | */ |
| 1502 | ret = -EINTR; |
| 1503 | /*FALLTHRU*/ |
| 1504 | default: |
| 1505 | aio_complete(req, ret, 0); |
| 1506 | return 0; |
| 1507 | } |
| 1508 | } |
| 1509 | |
| 1510 | static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored, |
| 1511 | bool compat) |
| 1512 | { |
| 1513 | struct file *file = req->ki_filp; |
| 1514 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
| 1515 | struct iov_iter iter; |
| 1516 | ssize_t ret; |
| 1517 | |
| 1518 | if (unlikely(!(file->f_mode & FMODE_READ))) |
| 1519 | return -EBADF; |
| 1520 | if (unlikely(!file->f_op->read_iter)) |
| 1521 | return -EINVAL; |
| 1522 | |
| 1523 | ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter); |
| 1524 | if (ret) |
| 1525 | return ret; |
| 1526 | ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter)); |
| 1527 | if (!ret) |
| 1528 | ret = aio_ret(req, call_read_iter(file, req, &iter)); |
| 1529 | kfree(iovec); |
| 1530 | return ret; |
| 1531 | } |
| 1532 | |
| 1533 | static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored, |
| 1534 | bool compat) |
| 1535 | { |
| 1536 | struct file *file = req->ki_filp; |
| 1537 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
| 1538 | struct iov_iter iter; |
| 1539 | ssize_t ret; |
| 1540 | |
| 1541 | if (unlikely(!(file->f_mode & FMODE_WRITE))) |
| 1542 | return -EBADF; |
| 1543 | if (unlikely(!file->f_op->write_iter)) |
| 1544 | return -EINVAL; |
| 1545 | |
| 1546 | ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter); |
| 1547 | if (ret) |
| 1548 | return ret; |
| 1549 | ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter)); |
| 1550 | if (!ret) { |
| 1551 | req->ki_flags |= IOCB_WRITE; |
| 1552 | file_start_write(file); |
| 1553 | ret = aio_ret(req, call_write_iter(file, req, &iter)); |
| 1554 | /* |
| 1555 | * We release freeze protection in aio_complete(). Fool lockdep |
| 1556 | * by telling it the lock got released so that it doesn't |
| 1557 | * complain about held lock when we return to userspace. |
| 1558 | */ |
| 1559 | if (S_ISREG(file_inode(file)->i_mode)) |
| 1560 | __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE); |
| 1561 | } |
| 1562 | kfree(iovec); |
| 1563 | return ret; |
| 1564 | } |
| 1565 | |
| 1566 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
| 1567 | struct iocb *iocb, bool compat) |
| 1568 | { |
| 1569 | struct aio_kiocb *req; |
| 1570 | struct file *file; |
| 1571 | ssize_t ret; |
| 1572 | |
| 1573 | /* enforce forwards compatibility on users */ |
| 1574 | if (unlikely(iocb->aio_reserved2)) { |
| 1575 | pr_debug("EINVAL: reserve field set\n"); |
| 1576 | return -EINVAL; |
| 1577 | } |
| 1578 | |
| 1579 | /* prevent overflows */ |
| 1580 | if (unlikely( |
| 1581 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || |
| 1582 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || |
| 1583 | ((ssize_t)iocb->aio_nbytes < 0) |
| 1584 | )) { |
| 1585 | pr_debug("EINVAL: overflow check\n"); |
| 1586 | return -EINVAL; |
| 1587 | } |
| 1588 | |
| 1589 | req = aio_get_req(ctx); |
| 1590 | if (unlikely(!req)) |
| 1591 | return -EAGAIN; |
| 1592 | |
| 1593 | req->common.ki_filp = file = fget(iocb->aio_fildes); |
| 1594 | if (unlikely(!req->common.ki_filp)) { |
| 1595 | ret = -EBADF; |
| 1596 | goto out_put_req; |
| 1597 | } |
| 1598 | req->common.ki_pos = iocb->aio_offset; |
| 1599 | req->common.ki_complete = aio_complete; |
| 1600 | req->common.ki_flags = iocb_flags(req->common.ki_filp); |
| 1601 | req->common.ki_hint = file_write_hint(file); |
| 1602 | |
| 1603 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
| 1604 | /* |
| 1605 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an |
| 1606 | * instance of the file* now. The file descriptor must be |
| 1607 | * an eventfd() fd, and will be signaled for each completed |
| 1608 | * event using the eventfd_signal() function. |
| 1609 | */ |
| 1610 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
| 1611 | if (IS_ERR(req->ki_eventfd)) { |
| 1612 | ret = PTR_ERR(req->ki_eventfd); |
| 1613 | req->ki_eventfd = NULL; |
| 1614 | goto out_put_req; |
| 1615 | } |
| 1616 | |
| 1617 | req->common.ki_flags |= IOCB_EVENTFD; |
| 1618 | } |
| 1619 | |
| 1620 | ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags); |
| 1621 | if (unlikely(ret)) { |
| 1622 | pr_debug("EINVAL: aio_rw_flags\n"); |
| 1623 | goto out_put_req; |
| 1624 | } |
| 1625 | |
| 1626 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
| 1627 | if (unlikely(ret)) { |
| 1628 | pr_debug("EFAULT: aio_key\n"); |
| 1629 | goto out_put_req; |
| 1630 | } |
| 1631 | |
| 1632 | req->ki_user_iocb = user_iocb; |
| 1633 | req->ki_user_data = iocb->aio_data; |
| 1634 | |
| 1635 | get_file(file); |
| 1636 | switch (iocb->aio_lio_opcode) { |
| 1637 | case IOCB_CMD_PREAD: |
| 1638 | ret = aio_read(&req->common, iocb, false, compat); |
| 1639 | break; |
| 1640 | case IOCB_CMD_PWRITE: |
| 1641 | ret = aio_write(&req->common, iocb, false, compat); |
| 1642 | break; |
| 1643 | case IOCB_CMD_PREADV: |
| 1644 | ret = aio_read(&req->common, iocb, true, compat); |
| 1645 | break; |
| 1646 | case IOCB_CMD_PWRITEV: |
| 1647 | ret = aio_write(&req->common, iocb, true, compat); |
| 1648 | break; |
| 1649 | default: |
| 1650 | pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode); |
| 1651 | ret = -EINVAL; |
| 1652 | break; |
| 1653 | } |
| 1654 | fput(file); |
| 1655 | |
| 1656 | if (ret && ret != -EIOCBQUEUED) |
| 1657 | goto out_put_req; |
| 1658 | return 0; |
| 1659 | out_put_req: |
| 1660 | put_reqs_available(ctx, 1); |
| 1661 | percpu_ref_put(&ctx->reqs); |
| 1662 | kiocb_free(req); |
| 1663 | return ret; |
| 1664 | } |
| 1665 | |
| 1666 | static long do_io_submit(aio_context_t ctx_id, long nr, |
| 1667 | struct iocb __user *__user *iocbpp, bool compat) |
| 1668 | { |
| 1669 | struct kioctx *ctx; |
| 1670 | long ret = 0; |
| 1671 | int i = 0; |
| 1672 | struct blk_plug plug; |
| 1673 | |
| 1674 | if (unlikely(nr < 0)) |
| 1675 | return -EINVAL; |
| 1676 | |
| 1677 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
| 1678 | nr = LONG_MAX/sizeof(*iocbpp); |
| 1679 | |
| 1680 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
| 1681 | return -EFAULT; |
| 1682 | |
| 1683 | ctx = lookup_ioctx(ctx_id); |
| 1684 | if (unlikely(!ctx)) { |
| 1685 | pr_debug("EINVAL: invalid context id\n"); |
| 1686 | return -EINVAL; |
| 1687 | } |
| 1688 | |
| 1689 | blk_start_plug(&plug); |
| 1690 | |
| 1691 | /* |
| 1692 | * AKPM: should this return a partial result if some of the IOs were |
| 1693 | * successfully submitted? |
| 1694 | */ |
| 1695 | for (i=0; i<nr; i++) { |
| 1696 | struct iocb __user *user_iocb; |
| 1697 | struct iocb tmp; |
| 1698 | |
| 1699 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { |
| 1700 | ret = -EFAULT; |
| 1701 | break; |
| 1702 | } |
| 1703 | |
| 1704 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { |
| 1705 | ret = -EFAULT; |
| 1706 | break; |
| 1707 | } |
| 1708 | |
| 1709 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
| 1710 | if (ret) |
| 1711 | break; |
| 1712 | } |
| 1713 | blk_finish_plug(&plug); |
| 1714 | |
| 1715 | percpu_ref_put(&ctx->users); |
| 1716 | return i ? i : ret; |
| 1717 | } |
| 1718 | |
| 1719 | /* sys_io_submit: |
| 1720 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns |
| 1721 | * the number of iocbs queued. May return -EINVAL if the aio_context |
| 1722 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at |
| 1723 | * *iocbpp[0] is not properly initialized, if the operation specified |
| 1724 | * is invalid for the file descriptor in the iocb. May fail with |
| 1725 | * -EFAULT if any of the data structures point to invalid data. May |
| 1726 | * fail with -EBADF if the file descriptor specified in the first |
| 1727 | * iocb is invalid. May fail with -EAGAIN if insufficient resources |
| 1728 | * are available to queue any iocbs. Will return 0 if nr is 0. Will |
| 1729 | * fail with -ENOSYS if not implemented. |
| 1730 | */ |
| 1731 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, |
| 1732 | struct iocb __user * __user *, iocbpp) |
| 1733 | { |
| 1734 | return do_io_submit(ctx_id, nr, iocbpp, 0); |
| 1735 | } |
| 1736 | |
| 1737 | #ifdef CONFIG_COMPAT |
| 1738 | static inline long |
| 1739 | copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64) |
| 1740 | { |
| 1741 | compat_uptr_t uptr; |
| 1742 | int i; |
| 1743 | |
| 1744 | for (i = 0; i < nr; ++i) { |
| 1745 | if (get_user(uptr, ptr32 + i)) |
| 1746 | return -EFAULT; |
| 1747 | if (put_user(compat_ptr(uptr), ptr64 + i)) |
| 1748 | return -EFAULT; |
| 1749 | } |
| 1750 | return 0; |
| 1751 | } |
| 1752 | |
| 1753 | #define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *)) |
| 1754 | |
| 1755 | COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id, |
| 1756 | int, nr, u32 __user *, iocb) |
| 1757 | { |
| 1758 | struct iocb __user * __user *iocb64; |
| 1759 | long ret; |
| 1760 | |
| 1761 | if (unlikely(nr < 0)) |
| 1762 | return -EINVAL; |
| 1763 | |
| 1764 | if (nr > MAX_AIO_SUBMITS) |
| 1765 | nr = MAX_AIO_SUBMITS; |
| 1766 | |
| 1767 | iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64)); |
| 1768 | ret = copy_iocb(nr, iocb, iocb64); |
| 1769 | if (!ret) |
| 1770 | ret = do_io_submit(ctx_id, nr, iocb64, 1); |
| 1771 | return ret; |
| 1772 | } |
| 1773 | #endif |
| 1774 | |
| 1775 | /* lookup_kiocb |
| 1776 | * Finds a given iocb for cancellation. |
| 1777 | */ |
| 1778 | static struct aio_kiocb * |
| 1779 | lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key) |
| 1780 | { |
| 1781 | struct aio_kiocb *kiocb; |
| 1782 | |
| 1783 | assert_spin_locked(&ctx->ctx_lock); |
| 1784 | |
| 1785 | if (key != KIOCB_KEY) |
| 1786 | return NULL; |
| 1787 | |
| 1788 | /* TODO: use a hash or array, this sucks. */ |
| 1789 | list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) { |
| 1790 | if (kiocb->ki_user_iocb == iocb) |
| 1791 | return kiocb; |
| 1792 | } |
| 1793 | return NULL; |
| 1794 | } |
| 1795 | |
| 1796 | /* sys_io_cancel: |
| 1797 | * Attempts to cancel an iocb previously passed to io_submit. If |
| 1798 | * the operation is successfully cancelled, the resulting event is |
| 1799 | * copied into the memory pointed to by result without being placed |
| 1800 | * into the completion queue and 0 is returned. May fail with |
| 1801 | * -EFAULT if any of the data structures pointed to are invalid. |
| 1802 | * May fail with -EINVAL if aio_context specified by ctx_id is |
| 1803 | * invalid. May fail with -EAGAIN if the iocb specified was not |
| 1804 | * cancelled. Will fail with -ENOSYS if not implemented. |
| 1805 | */ |
| 1806 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
| 1807 | struct io_event __user *, result) |
| 1808 | { |
| 1809 | struct kioctx *ctx; |
| 1810 | struct aio_kiocb *kiocb; |
| 1811 | u32 key; |
| 1812 | int ret; |
| 1813 | |
| 1814 | ret = get_user(key, &iocb->aio_key); |
| 1815 | if (unlikely(ret)) |
| 1816 | return -EFAULT; |
| 1817 | |
| 1818 | ctx = lookup_ioctx(ctx_id); |
| 1819 | if (unlikely(!ctx)) |
| 1820 | return -EINVAL; |
| 1821 | |
| 1822 | spin_lock_irq(&ctx->ctx_lock); |
| 1823 | |
| 1824 | kiocb = lookup_kiocb(ctx, iocb, key); |
| 1825 | if (kiocb) |
| 1826 | ret = kiocb_cancel(kiocb); |
| 1827 | else |
| 1828 | ret = -EINVAL; |
| 1829 | |
| 1830 | spin_unlock_irq(&ctx->ctx_lock); |
| 1831 | |
| 1832 | if (!ret) { |
| 1833 | /* |
| 1834 | * The result argument is no longer used - the io_event is |
| 1835 | * always delivered via the ring buffer. -EINPROGRESS indicates |
| 1836 | * cancellation is progress: |
| 1837 | */ |
| 1838 | ret = -EINPROGRESS; |
| 1839 | } |
| 1840 | |
| 1841 | percpu_ref_put(&ctx->users); |
| 1842 | |
| 1843 | return ret; |
| 1844 | } |
| 1845 | |
| 1846 | /* io_getevents: |
| 1847 | * Attempts to read at least min_nr events and up to nr events from |
| 1848 | * the completion queue for the aio_context specified by ctx_id. If |
| 1849 | * it succeeds, the number of read events is returned. May fail with |
| 1850 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is |
| 1851 | * out of range, if timeout is out of range. May fail with -EFAULT |
| 1852 | * if any of the memory specified is invalid. May return 0 or |
| 1853 | * < min_nr if the timeout specified by timeout has elapsed |
| 1854 | * before sufficient events are available, where timeout == NULL |
| 1855 | * specifies an infinite timeout. Note that the timeout pointed to by |
| 1856 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
| 1857 | */ |
| 1858 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
| 1859 | long, min_nr, |
| 1860 | long, nr, |
| 1861 | struct io_event __user *, events, |
| 1862 | struct timespec __user *, timeout) |
| 1863 | { |
| 1864 | struct kioctx *ioctx = lookup_ioctx(ctx_id); |
| 1865 | long ret = -EINVAL; |
| 1866 | |
| 1867 | if (likely(ioctx)) { |
| 1868 | if (likely(min_nr <= nr && min_nr >= 0)) |
| 1869 | ret = read_events(ioctx, min_nr, nr, events, timeout); |
| 1870 | percpu_ref_put(&ioctx->users); |
| 1871 | } |
| 1872 | return ret; |
| 1873 | } |
| 1874 | |
| 1875 | #ifdef CONFIG_COMPAT |
| 1876 | COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id, |
| 1877 | compat_long_t, min_nr, |
| 1878 | compat_long_t, nr, |
| 1879 | struct io_event __user *, events, |
| 1880 | struct compat_timespec __user *, timeout) |
| 1881 | { |
| 1882 | struct timespec t; |
| 1883 | struct timespec __user *ut = NULL; |
| 1884 | |
| 1885 | if (timeout) { |
| 1886 | if (compat_get_timespec(&t, timeout)) |
| 1887 | return -EFAULT; |
| 1888 | |
| 1889 | ut = compat_alloc_user_space(sizeof(*ut)); |
| 1890 | if (copy_to_user(ut, &t, sizeof(t))) |
| 1891 | return -EFAULT; |
| 1892 | } |
| 1893 | return sys_io_getevents(ctx_id, min_nr, nr, events, ut); |
| 1894 | } |
| 1895 | #endif |