Jens Axboe | 17437f3 | 2022-05-25 09:13:39 -0600 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Contains the core associated with submission side polling of the SQ |
| 4 | * ring, offloading submissions from the application to a kernel thread. |
| 5 | */ |
| 6 | #include <linux/kernel.h> |
| 7 | #include <linux/errno.h> |
| 8 | #include <linux/file.h> |
| 9 | #include <linux/mm.h> |
| 10 | #include <linux/slab.h> |
| 11 | #include <linux/audit.h> |
| 12 | #include <linux/security.h> |
| 13 | #include <linux/io_uring.h> |
| 14 | |
| 15 | #include <uapi/linux/io_uring.h> |
| 16 | |
Jens Axboe | 17437f3 | 2022-05-25 09:13:39 -0600 | [diff] [blame] | 17 | #include "io_uring.h" |
| 18 | #include "sqpoll.h" |
| 19 | |
| 20 | #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8 |
| 21 | |
| 22 | enum { |
| 23 | IO_SQ_THREAD_SHOULD_STOP = 0, |
| 24 | IO_SQ_THREAD_SHOULD_PARK, |
| 25 | }; |
| 26 | |
| 27 | void io_sq_thread_unpark(struct io_sq_data *sqd) |
| 28 | __releases(&sqd->lock) |
| 29 | { |
| 30 | WARN_ON_ONCE(sqd->thread == current); |
| 31 | |
| 32 | /* |
| 33 | * Do the dance but not conditional clear_bit() because it'd race with |
| 34 | * other threads incrementing park_pending and setting the bit. |
| 35 | */ |
| 36 | clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
| 37 | if (atomic_dec_return(&sqd->park_pending)) |
| 38 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
| 39 | mutex_unlock(&sqd->lock); |
| 40 | } |
| 41 | |
| 42 | void io_sq_thread_park(struct io_sq_data *sqd) |
| 43 | __acquires(&sqd->lock) |
| 44 | { |
| 45 | WARN_ON_ONCE(sqd->thread == current); |
| 46 | |
| 47 | atomic_inc(&sqd->park_pending); |
| 48 | set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state); |
| 49 | mutex_lock(&sqd->lock); |
| 50 | if (sqd->thread) |
| 51 | wake_up_process(sqd->thread); |
| 52 | } |
| 53 | |
| 54 | void io_sq_thread_stop(struct io_sq_data *sqd) |
| 55 | { |
| 56 | WARN_ON_ONCE(sqd->thread == current); |
| 57 | WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)); |
| 58 | |
| 59 | set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
| 60 | mutex_lock(&sqd->lock); |
| 61 | if (sqd->thread) |
| 62 | wake_up_process(sqd->thread); |
| 63 | mutex_unlock(&sqd->lock); |
| 64 | wait_for_completion(&sqd->exited); |
| 65 | } |
| 66 | |
| 67 | void io_put_sq_data(struct io_sq_data *sqd) |
| 68 | { |
| 69 | if (refcount_dec_and_test(&sqd->refs)) { |
| 70 | WARN_ON_ONCE(atomic_read(&sqd->park_pending)); |
| 71 | |
| 72 | io_sq_thread_stop(sqd); |
| 73 | kfree(sqd); |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd) |
| 78 | { |
| 79 | struct io_ring_ctx *ctx; |
| 80 | unsigned sq_thread_idle = 0; |
| 81 | |
| 82 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
| 83 | sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle); |
| 84 | sqd->sq_thread_idle = sq_thread_idle; |
| 85 | } |
| 86 | |
| 87 | void io_sq_thread_finish(struct io_ring_ctx *ctx) |
| 88 | { |
| 89 | struct io_sq_data *sqd = ctx->sq_data; |
| 90 | |
| 91 | if (sqd) { |
| 92 | io_sq_thread_park(sqd); |
| 93 | list_del_init(&ctx->sqd_list); |
| 94 | io_sqd_update_thread_idle(sqd); |
| 95 | io_sq_thread_unpark(sqd); |
| 96 | |
| 97 | io_put_sq_data(sqd); |
| 98 | ctx->sq_data = NULL; |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
| 103 | { |
| 104 | struct io_ring_ctx *ctx_attach; |
| 105 | struct io_sq_data *sqd; |
| 106 | struct fd f; |
| 107 | |
| 108 | f = fdget(p->wq_fd); |
| 109 | if (!f.file) |
| 110 | return ERR_PTR(-ENXIO); |
| 111 | if (!io_is_uring_fops(f.file)) { |
| 112 | fdput(f); |
| 113 | return ERR_PTR(-EINVAL); |
| 114 | } |
| 115 | |
| 116 | ctx_attach = f.file->private_data; |
| 117 | sqd = ctx_attach->sq_data; |
| 118 | if (!sqd) { |
| 119 | fdput(f); |
| 120 | return ERR_PTR(-EINVAL); |
| 121 | } |
| 122 | if (sqd->task_tgid != current->tgid) { |
| 123 | fdput(f); |
| 124 | return ERR_PTR(-EPERM); |
| 125 | } |
| 126 | |
| 127 | refcount_inc(&sqd->refs); |
| 128 | fdput(f); |
| 129 | return sqd; |
| 130 | } |
| 131 | |
| 132 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p, |
| 133 | bool *attached) |
| 134 | { |
| 135 | struct io_sq_data *sqd; |
| 136 | |
| 137 | *attached = false; |
| 138 | if (p->flags & IORING_SETUP_ATTACH_WQ) { |
| 139 | sqd = io_attach_sq_data(p); |
| 140 | if (!IS_ERR(sqd)) { |
| 141 | *attached = true; |
| 142 | return sqd; |
| 143 | } |
| 144 | /* fall through for EPERM case, setup new sqd/task */ |
| 145 | if (PTR_ERR(sqd) != -EPERM) |
| 146 | return sqd; |
| 147 | } |
| 148 | |
| 149 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
| 150 | if (!sqd) |
| 151 | return ERR_PTR(-ENOMEM); |
| 152 | |
| 153 | atomic_set(&sqd->park_pending, 0); |
| 154 | refcount_set(&sqd->refs, 1); |
| 155 | INIT_LIST_HEAD(&sqd->ctx_list); |
| 156 | mutex_init(&sqd->lock); |
| 157 | init_waitqueue_head(&sqd->wait); |
| 158 | init_completion(&sqd->exited); |
| 159 | return sqd; |
| 160 | } |
| 161 | |
| 162 | static inline bool io_sqd_events_pending(struct io_sq_data *sqd) |
| 163 | { |
| 164 | return READ_ONCE(sqd->state); |
| 165 | } |
| 166 | |
| 167 | static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) |
| 168 | { |
| 169 | unsigned int to_submit; |
| 170 | int ret = 0; |
| 171 | |
| 172 | to_submit = io_sqring_entries(ctx); |
| 173 | /* if we're handling multiple rings, cap submit size for fairness */ |
| 174 | if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE) |
| 175 | to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE; |
| 176 | |
| 177 | if (!wq_list_empty(&ctx->iopoll_list) || to_submit) { |
| 178 | const struct cred *creds = NULL; |
| 179 | |
| 180 | if (ctx->sq_creds != current_cred()) |
| 181 | creds = override_creds(ctx->sq_creds); |
| 182 | |
| 183 | mutex_lock(&ctx->uring_lock); |
| 184 | if (!wq_list_empty(&ctx->iopoll_list)) |
| 185 | io_do_iopoll(ctx, true); |
| 186 | |
| 187 | /* |
| 188 | * Don't submit if refs are dying, good for io_uring_register(), |
| 189 | * but also it is relied upon by io_ring_exit_work() |
| 190 | */ |
| 191 | if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) && |
| 192 | !(ctx->flags & IORING_SETUP_R_DISABLED)) |
| 193 | ret = io_submit_sqes(ctx, to_submit); |
| 194 | mutex_unlock(&ctx->uring_lock); |
| 195 | |
| 196 | if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait)) |
| 197 | wake_up(&ctx->sqo_sq_wait); |
| 198 | if (creds) |
| 199 | revert_creds(creds); |
| 200 | } |
| 201 | |
| 202 | return ret; |
| 203 | } |
| 204 | |
| 205 | static bool io_sqd_handle_event(struct io_sq_data *sqd) |
| 206 | { |
| 207 | bool did_sig = false; |
| 208 | struct ksignal ksig; |
| 209 | |
| 210 | if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) || |
| 211 | signal_pending(current)) { |
| 212 | mutex_unlock(&sqd->lock); |
| 213 | if (signal_pending(current)) |
| 214 | did_sig = get_signal(&ksig); |
| 215 | cond_resched(); |
| 216 | mutex_lock(&sqd->lock); |
| 217 | } |
| 218 | return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state); |
| 219 | } |
| 220 | |
| 221 | static int io_sq_thread(void *data) |
| 222 | { |
| 223 | struct io_sq_data *sqd = data; |
| 224 | struct io_ring_ctx *ctx; |
| 225 | unsigned long timeout = 0; |
| 226 | char buf[TASK_COMM_LEN]; |
| 227 | DEFINE_WAIT(wait); |
| 228 | |
| 229 | snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid); |
| 230 | set_task_comm(current, buf); |
| 231 | |
| 232 | if (sqd->sq_cpu != -1) |
| 233 | set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu)); |
| 234 | else |
| 235 | set_cpus_allowed_ptr(current, cpu_online_mask); |
| 236 | current->flags |= PF_NO_SETAFFINITY; |
| 237 | |
| 238 | audit_alloc_kernel(current); |
| 239 | |
| 240 | mutex_lock(&sqd->lock); |
| 241 | while (1) { |
| 242 | bool cap_entries, sqt_spin = false; |
| 243 | |
| 244 | if (io_sqd_events_pending(sqd) || signal_pending(current)) { |
| 245 | if (io_sqd_handle_event(sqd)) |
| 246 | break; |
| 247 | timeout = jiffies + sqd->sq_thread_idle; |
| 248 | } |
| 249 | |
| 250 | cap_entries = !list_is_singular(&sqd->ctx_list); |
| 251 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
| 252 | int ret = __io_sq_thread(ctx, cap_entries); |
| 253 | |
| 254 | if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list))) |
| 255 | sqt_spin = true; |
| 256 | } |
| 257 | if (io_run_task_work()) |
| 258 | sqt_spin = true; |
| 259 | |
| 260 | if (sqt_spin || !time_after(jiffies, timeout)) { |
| 261 | cond_resched(); |
| 262 | if (sqt_spin) |
| 263 | timeout = jiffies + sqd->sq_thread_idle; |
| 264 | continue; |
| 265 | } |
| 266 | |
| 267 | prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE); |
| 268 | if (!io_sqd_events_pending(sqd) && !task_work_pending(current)) { |
| 269 | bool needs_sched = true; |
| 270 | |
| 271 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
| 272 | atomic_or(IORING_SQ_NEED_WAKEUP, |
| 273 | &ctx->rings->sq_flags); |
| 274 | if ((ctx->flags & IORING_SETUP_IOPOLL) && |
| 275 | !wq_list_empty(&ctx->iopoll_list)) { |
| 276 | needs_sched = false; |
| 277 | break; |
| 278 | } |
| 279 | |
| 280 | /* |
| 281 | * Ensure the store of the wakeup flag is not |
| 282 | * reordered with the load of the SQ tail |
| 283 | */ |
| 284 | smp_mb__after_atomic(); |
| 285 | |
| 286 | if (io_sqring_entries(ctx)) { |
| 287 | needs_sched = false; |
| 288 | break; |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | if (needs_sched) { |
| 293 | mutex_unlock(&sqd->lock); |
| 294 | schedule(); |
| 295 | mutex_lock(&sqd->lock); |
| 296 | } |
| 297 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
| 298 | atomic_andnot(IORING_SQ_NEED_WAKEUP, |
| 299 | &ctx->rings->sq_flags); |
| 300 | } |
| 301 | |
| 302 | finish_wait(&sqd->wait, &wait); |
| 303 | timeout = jiffies + sqd->sq_thread_idle; |
| 304 | } |
| 305 | |
| 306 | io_uring_cancel_generic(true, sqd); |
| 307 | sqd->thread = NULL; |
| 308 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
| 309 | atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags); |
| 310 | io_run_task_work(); |
| 311 | mutex_unlock(&sqd->lock); |
| 312 | |
| 313 | audit_free(current); |
| 314 | |
| 315 | complete(&sqd->exited); |
| 316 | do_exit(0); |
| 317 | } |
| 318 | |
| 319 | int io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
| 320 | { |
| 321 | DEFINE_WAIT(wait); |
| 322 | |
| 323 | do { |
| 324 | if (!io_sqring_full(ctx)) |
| 325 | break; |
| 326 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); |
| 327 | |
| 328 | if (!io_sqring_full(ctx)) |
| 329 | break; |
| 330 | schedule(); |
| 331 | } while (!signal_pending(current)); |
| 332 | |
| 333 | finish_wait(&ctx->sqo_sq_wait, &wait); |
| 334 | return 0; |
| 335 | } |
| 336 | |
| 337 | __cold int io_sq_offload_create(struct io_ring_ctx *ctx, |
| 338 | struct io_uring_params *p) |
| 339 | { |
| 340 | int ret; |
| 341 | |
| 342 | /* Retain compatibility with failing for an invalid attach attempt */ |
| 343 | if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) == |
| 344 | IORING_SETUP_ATTACH_WQ) { |
| 345 | struct fd f; |
| 346 | |
| 347 | f = fdget(p->wq_fd); |
| 348 | if (!f.file) |
| 349 | return -ENXIO; |
| 350 | if (!io_is_uring_fops(f.file)) { |
| 351 | fdput(f); |
| 352 | return -EINVAL; |
| 353 | } |
| 354 | fdput(f); |
| 355 | } |
| 356 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
| 357 | struct task_struct *tsk; |
| 358 | struct io_sq_data *sqd; |
| 359 | bool attached; |
| 360 | |
| 361 | ret = security_uring_sqpoll(); |
| 362 | if (ret) |
| 363 | return ret; |
| 364 | |
| 365 | sqd = io_get_sq_data(p, &attached); |
| 366 | if (IS_ERR(sqd)) { |
| 367 | ret = PTR_ERR(sqd); |
| 368 | goto err; |
| 369 | } |
| 370 | |
| 371 | ctx->sq_creds = get_current_cred(); |
| 372 | ctx->sq_data = sqd; |
| 373 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
| 374 | if (!ctx->sq_thread_idle) |
| 375 | ctx->sq_thread_idle = HZ; |
| 376 | |
| 377 | io_sq_thread_park(sqd); |
| 378 | list_add(&ctx->sqd_list, &sqd->ctx_list); |
| 379 | io_sqd_update_thread_idle(sqd); |
| 380 | /* don't attach to a dying SQPOLL thread, would be racy */ |
| 381 | ret = (attached && !sqd->thread) ? -ENXIO : 0; |
| 382 | io_sq_thread_unpark(sqd); |
| 383 | |
| 384 | if (ret < 0) |
| 385 | goto err; |
| 386 | if (attached) |
| 387 | return 0; |
| 388 | |
| 389 | if (p->flags & IORING_SETUP_SQ_AFF) { |
| 390 | int cpu = p->sq_thread_cpu; |
| 391 | |
| 392 | ret = -EINVAL; |
| 393 | if (cpu >= nr_cpu_ids || !cpu_online(cpu)) |
| 394 | goto err_sqpoll; |
| 395 | sqd->sq_cpu = cpu; |
| 396 | } else { |
| 397 | sqd->sq_cpu = -1; |
| 398 | } |
| 399 | |
| 400 | sqd->task_pid = current->pid; |
| 401 | sqd->task_tgid = current->tgid; |
| 402 | tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE); |
| 403 | if (IS_ERR(tsk)) { |
| 404 | ret = PTR_ERR(tsk); |
| 405 | goto err_sqpoll; |
| 406 | } |
| 407 | |
| 408 | sqd->thread = tsk; |
| 409 | ret = io_uring_alloc_task_context(tsk, ctx); |
| 410 | wake_up_new_task(tsk); |
| 411 | if (ret) |
| 412 | goto err; |
| 413 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
| 414 | /* Can't have SQ_AFF without SQPOLL */ |
| 415 | ret = -EINVAL; |
| 416 | goto err; |
| 417 | } |
| 418 | |
| 419 | return 0; |
| 420 | err_sqpoll: |
| 421 | complete(&ctx->sq_data->exited); |
| 422 | err: |
| 423 | io_sq_thread_finish(ctx); |
| 424 | return ret; |
| 425 | } |