1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/kernel.h> 3 #include <linux/errno.h> 4 #include <linux/file.h> 5 #include <linux/mm.h> 6 #include <linux/slab.h> 7 #include <linux/nospec.h> 8 #include <linux/io_uring.h> 9 10 #include <uapi/linux/io_uring.h> 11 12 #include "io_uring.h" 13 #include "tctx.h" 14 #include "bpf_filter.h" 15 16 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx, 17 struct task_struct *task) 18 { 19 struct io_wq_hash *hash; 20 struct io_wq_data data; 21 unsigned int concurrency; 22 23 mutex_lock(&ctx->uring_lock); 24 hash = ctx->hash_map; 25 if (!hash) { 26 hash = kzalloc_obj(*hash); 27 if (!hash) { 28 mutex_unlock(&ctx->uring_lock); 29 return ERR_PTR(-ENOMEM); 30 } 31 refcount_set(&hash->refs, 1); 32 init_waitqueue_head(&hash->wait); 33 ctx->hash_map = hash; 34 } 35 mutex_unlock(&ctx->uring_lock); 36 37 data.hash = hash; 38 data.task = task; 39 40 /* Do QD, or 4 * CPUS, whatever is smallest */ 41 concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); 42 43 return io_wq_create(concurrency, &data); 44 } 45 46 void __io_uring_free(struct task_struct *tsk) 47 { 48 struct io_uring_task *tctx = tsk->io_uring; 49 struct io_tctx_node *node; 50 unsigned long index; 51 52 /* 53 * Fault injection forcing allocation errors in the xa_store() path 54 * can lead to xa_empty() returning false, even though no actual 55 * node is stored in the xarray. Until that gets sorted out, attempt 56 * an iteration here and warn if any entries are found. 57 */ 58 if (tctx) { 59 xa_for_each(&tctx->xa, index, node) { 60 WARN_ON_ONCE(1); 61 break; 62 } 63 WARN_ON_ONCE(tctx->io_wq); 64 WARN_ON_ONCE(tctx->cached_refs); 65 66 percpu_counter_destroy(&tctx->inflight); 67 kfree(tctx); 68 tsk->io_uring = NULL; 69 } 70 if (tsk->io_uring_restrict) { 71 io_put_bpf_filters(tsk->io_uring_restrict); 72 kfree(tsk->io_uring_restrict); 73 tsk->io_uring_restrict = NULL; 74 } 75 } 76 77 __cold struct io_uring_task *io_uring_alloc_task_context(struct task_struct *task, 78 struct io_ring_ctx *ctx) 79 { 80 struct io_uring_task *tctx; 81 int ret; 82 83 tctx = kzalloc_obj(*tctx); 84 if (unlikely(!tctx)) 85 return ERR_PTR(-ENOMEM); 86 87 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); 88 if (unlikely(ret)) { 89 kfree(tctx); 90 return ERR_PTR(ret); 91 } 92 93 tctx->io_wq = io_init_wq_offload(ctx, task); 94 if (IS_ERR(tctx->io_wq)) { 95 ret = PTR_ERR(tctx->io_wq); 96 percpu_counter_destroy(&tctx->inflight); 97 kfree(tctx); 98 return ERR_PTR(ret); 99 } 100 101 tctx->task = task; 102 xa_init(&tctx->xa); 103 init_waitqueue_head(&tctx->wait); 104 atomic_set(&tctx->in_cancel, 0); 105 atomic_set(&tctx->inflight_tracked, 0); 106 mpscq_init(&tctx->task_list, &tctx->task_head); 107 INIT_WORK(&tctx->fallback_work, io_tctx_fallback_work); 108 init_task_work(&tctx->task_work, tctx_task_work); 109 return tctx; 110 } 111 112 static int io_tctx_install_node(struct io_ring_ctx *ctx, 113 struct io_uring_task *tctx) 114 { 115 struct io_tctx_node *node; 116 int ret; 117 118 if (xa_load(&tctx->xa, (unsigned long)ctx)) 119 return 0; 120 121 node = kmalloc_obj(*node); 122 if (!node) 123 return -ENOMEM; 124 node->ctx = ctx; 125 node->task = current; 126 127 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx, 128 node, GFP_KERNEL)); 129 if (ret) { 130 kfree(node); 131 return ret; 132 } 133 134 mutex_lock(&ctx->tctx_lock); 135 list_add(&node->ctx_node, &ctx->tctx_list); 136 mutex_unlock(&ctx->tctx_lock); 137 return 0; 138 } 139 140 int __io_uring_add_tctx_node(struct io_ring_ctx *ctx) 141 { 142 struct io_uring_task *tctx = current->io_uring; 143 bool new_tctx = false; 144 int ret; 145 146 if (unlikely(!tctx)) { 147 tctx = io_uring_alloc_task_context(current, ctx); 148 if (IS_ERR(tctx)) 149 return PTR_ERR(tctx); 150 new_tctx = true; 151 152 if (data_race(ctx->int_flags) & IO_RING_F_IOWQ_LIMITS_SET) { 153 unsigned int limits[2]; 154 155 mutex_lock(&ctx->uring_lock); 156 limits[0] = ctx->iowq_limits[0]; 157 limits[1] = ctx->iowq_limits[1]; 158 mutex_unlock(&ctx->uring_lock); 159 160 ret = io_wq_max_workers(tctx->io_wq, limits); 161 if (ret) 162 goto err_free; 163 } 164 } 165 166 /* 167 * Re-activate io-wq keepalive on any new io_uring usage. The wq may have 168 * been marked for idle-exit when the task temporarily had no active 169 * io_uring instances. 170 */ 171 if (tctx->io_wq) 172 io_wq_set_exit_on_idle(tctx->io_wq, false); 173 174 if (new_tctx) 175 current->io_uring = tctx; 176 177 ret = io_tctx_install_node(ctx, tctx); 178 if (!ret) 179 return 0; 180 err_free: 181 if (new_tctx) { 182 current->io_uring = NULL; 183 if (tctx->io_wq) { 184 io_wq_exit_start(tctx->io_wq); 185 io_wq_put_and_exit(tctx->io_wq); 186 } 187 percpu_counter_destroy(&tctx->inflight); 188 kfree(tctx); 189 } 190 return ret; 191 } 192 193 int __io_uring_add_tctx_node_from_submit(struct io_ring_ctx *ctx) 194 { 195 int ret; 196 197 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER 198 && ctx->submitter_task != current) 199 return -EEXIST; 200 201 ret = __io_uring_add_tctx_node(ctx); 202 if (ret) 203 return ret; 204 205 current->io_uring->last = ctx; 206 return 0; 207 } 208 209 /* 210 * Remove this io_uring_file -> task mapping. 211 */ 212 __cold void io_uring_del_tctx_node(unsigned long index) 213 { 214 struct io_uring_task *tctx = current->io_uring; 215 struct io_tctx_node *node; 216 217 if (!tctx) 218 return; 219 node = xa_erase(&tctx->xa, index); 220 if (!node) 221 return; 222 223 WARN_ON_ONCE(current != node->task); 224 WARN_ON_ONCE(list_empty(&node->ctx_node)); 225 226 mutex_lock(&node->ctx->tctx_lock); 227 list_del(&node->ctx_node); 228 mutex_unlock(&node->ctx->tctx_lock); 229 230 if (tctx->last == node->ctx) 231 tctx->last = NULL; 232 kfree(node); 233 234 if (xa_empty(&tctx->xa) && tctx->io_wq) 235 io_wq_set_exit_on_idle(tctx->io_wq, true); 236 } 237 238 __cold void io_uring_clean_tctx(struct io_uring_task *tctx) 239 { 240 struct io_wq *wq = tctx->io_wq; 241 struct io_tctx_node *node; 242 unsigned long index; 243 244 xa_for_each(&tctx->xa, index, node) { 245 io_uring_del_tctx_node(index); 246 cond_resched(); 247 } 248 if (wq) { 249 /* 250 * Must be after io_uring_del_tctx_node() (removes nodes under 251 * uring_lock) to avoid race with io_uring_try_cancel_iowq(). 252 */ 253 io_wq_put_and_exit(wq); 254 tctx->io_wq = NULL; 255 } 256 } 257 258 void io_uring_unreg_ringfd(void) 259 { 260 struct io_uring_task *tctx = current->io_uring; 261 int i; 262 263 for (i = 0; i < IO_RINGFD_REG_MAX; i++) { 264 if (tctx->registered_rings[i]) { 265 fput(tctx->registered_rings[i]); 266 tctx->registered_rings[i] = NULL; 267 } 268 } 269 } 270 271 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file, 272 int start, int end) 273 { 274 int offset, idx; 275 for (offset = start; offset < end; offset++) { 276 idx = array_index_nospec(offset, IO_RINGFD_REG_MAX); 277 if (tctx->registered_rings[idx]) 278 continue; 279 280 tctx->registered_rings[idx] = file; 281 return idx; 282 } 283 return -EBUSY; 284 } 285 286 static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd, 287 int start, int end) 288 { 289 struct file *file; 290 int offset; 291 292 file = fget(fd); 293 if (!file) { 294 return -EBADF; 295 } else if (!io_is_uring_fops(file)) { 296 fput(file); 297 return -EOPNOTSUPP; 298 } 299 offset = io_ring_add_registered_file(tctx, file, start, end); 300 if (offset < 0) 301 fput(file); 302 return offset; 303 } 304 305 /* 306 * Register a ring fd to avoid fdget/fdput for each io_uring_enter() 307 * invocation. User passes in an array of struct io_uring_rsrc_update 308 * with ->data set to the ring_fd, and ->offset given for the desired 309 * index. If no index is desired, application may set ->offset == -1U 310 * and we'll find an available index. Returns number of entries 311 * successfully processed, or < 0 on error if none were processed. 312 */ 313 int io_ringfd_register(struct io_ring_ctx *ctx, void __user *__arg, 314 unsigned nr_args) 315 { 316 struct io_uring_rsrc_update __user *arg = __arg; 317 struct io_uring_rsrc_update reg; 318 struct io_uring_task *tctx; 319 int ret, i; 320 321 if (!nr_args || nr_args > IO_RINGFD_REG_MAX) 322 return -EINVAL; 323 324 mutex_unlock(&ctx->uring_lock); 325 ret = __io_uring_add_tctx_node(ctx); 326 mutex_lock(&ctx->uring_lock); 327 if (ret) 328 return ret; 329 330 tctx = current->io_uring; 331 for (i = 0; i < nr_args; i++) { 332 int start, end; 333 334 if (copy_from_user(®, &arg[i], sizeof(reg))) { 335 ret = -EFAULT; 336 break; 337 } 338 339 if (reg.resv) { 340 ret = -EINVAL; 341 break; 342 } 343 344 if (reg.offset == -1U) { 345 start = 0; 346 end = IO_RINGFD_REG_MAX; 347 } else { 348 if (reg.offset >= IO_RINGFD_REG_MAX) { 349 ret = -EINVAL; 350 break; 351 } 352 start = reg.offset; 353 end = start + 1; 354 } 355 356 ret = io_ring_add_registered_fd(tctx, reg.data, start, end); 357 if (ret < 0) 358 break; 359 360 reg.offset = ret; 361 if (copy_to_user(&arg[i], ®, sizeof(reg))) { 362 fput(tctx->registered_rings[reg.offset]); 363 tctx->registered_rings[reg.offset] = NULL; 364 ret = -EFAULT; 365 break; 366 } 367 } 368 369 return i ? i : ret; 370 } 371 372 int io_ringfd_unregister(struct io_ring_ctx *ctx, void __user *__arg, 373 unsigned nr_args) 374 { 375 struct io_uring_rsrc_update __user *arg = __arg; 376 struct io_uring_task *tctx = current->io_uring; 377 struct io_uring_rsrc_update reg; 378 int ret = 0, i; 379 380 if (!nr_args || nr_args > IO_RINGFD_REG_MAX) 381 return -EINVAL; 382 if (!tctx) 383 return 0; 384 385 for (i = 0; i < nr_args; i++) { 386 if (copy_from_user(®, &arg[i], sizeof(reg))) { 387 ret = -EFAULT; 388 break; 389 } 390 if (reg.resv || reg.data || reg.offset >= IO_RINGFD_REG_MAX) { 391 ret = -EINVAL; 392 break; 393 } 394 395 reg.offset = array_index_nospec(reg.offset, IO_RINGFD_REG_MAX); 396 if (tctx->registered_rings[reg.offset]) { 397 fput(tctx->registered_rings[reg.offset]); 398 tctx->registered_rings[reg.offset] = NULL; 399 } 400 } 401 402 return i ? i : ret; 403 } 404 405 int __io_uring_fork(struct task_struct *tsk) 406 { 407 struct io_restriction *res, *src = tsk->io_uring_restrict; 408 409 /* Don't leave it dangling on error */ 410 tsk->io_uring_restrict = NULL; 411 412 res = kzalloc_obj(*res, GFP_KERNEL_ACCOUNT); 413 if (!res) 414 return -ENOMEM; 415 416 tsk->io_uring_restrict = res; 417 io_restriction_clone(res, src); 418 return 0; 419 } 420