io_uring/sqpoll.c

   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/cpuset.h>
  14 #include <linux/io_uring.h>
  15
  16 #include <uapi/linux/io_uring.h>
  17
  18 #include "io_uring.h"
  19 #include "napi.h"
  20 #include "sqpoll.h"
  21
  22 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
  23 #define IORING_TW_CAP_ENTRIES_VALUE     8
  24
  25 enum {
  26         IO_SQ_THREAD_SHOULD_STOP = 0,
  27         IO_SQ_THREAD_SHOULD_PARK,
  28 };
  29
  30 void io_sq_thread_unpark(struct io_sq_data *sqd)
  31         __releases(&sqd->lock)
  32 {
  33         WARN_ON_ONCE(sqd->thread == current);
  34
  35         /*
  36          * Do the dance but not conditional clear_bit() because it'd race with
  37          * other threads incrementing park_pending and setting the bit.
  38          */
  39         clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
  40         if (atomic_dec_return(&sqd->park_pending))
  41                 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
  42         mutex_unlock(&sqd->lock);
  43         wake_up(&sqd->wait);
  44 }
  45
  46 void io_sq_thread_park(struct io_sq_data *sqd)
  47         __acquires(&sqd->lock)
  48 {
  49         WARN_ON_ONCE(data_race(sqd->thread) == current);
  50
  51         atomic_inc(&sqd->park_pending);
  52         set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
  53         mutex_lock(&sqd->lock);
  54         if (sqd->thread)
  55                 wake_up_process(sqd->thread);
  56 }
  57
  58 void io_sq_thread_stop(struct io_sq_data *sqd)
  59 {
  60         WARN_ON_ONCE(sqd->thread == current);
  61         WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
  62
  63         set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
  64         mutex_lock(&sqd->lock);
  65         if (sqd->thread)
  66                 wake_up_process(sqd->thread);
  67         mutex_unlock(&sqd->lock);
  68         wait_for_completion(&sqd->exited);
  69 }
  70
  71 void io_put_sq_data(struct io_sq_data *sqd)
  72 {
  73         if (refcount_dec_and_test(&sqd->refs)) {
  74                 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
  75
  76                 io_sq_thread_stop(sqd);
  77                 kfree(sqd);
  78         }
  79 }
  80
  81 static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
  82 {
  83         struct io_ring_ctx *ctx;
  84         unsigned sq_thread_idle = 0;
  85
  86         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
  87                 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
  88         sqd->sq_thread_idle = sq_thread_idle;
  89 }
  90
  91 void io_sq_thread_finish(struct io_ring_ctx *ctx)
  92 {
  93         struct io_sq_data *sqd = ctx->sq_data;
  94
  95         if (sqd) {
  96                 io_sq_thread_park(sqd);
  97                 list_del_init(&ctx->sqd_list);
  98                 io_sqd_update_thread_idle(sqd);
  99                 io_sq_thread_unpark(sqd);
 100
 101                 io_put_sq_data(sqd);
 102                 ctx->sq_data = NULL;
 103         }
 104 }
 105
 106 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
 107 {
 108         struct io_ring_ctx *ctx_attach;
 109         struct io_sq_data *sqd;
 110         CLASS(fd, f)(p->wq_fd);
 111
 112         if (fd_empty(f))
 113                 return ERR_PTR(-ENXIO);
 114         if (!io_is_uring_fops(fd_file(f)))
 115                 return ERR_PTR(-EINVAL);
 116
 117         ctx_attach = fd_file(f)->private_data;
 118         sqd = ctx_attach->sq_data;
 119         if (!sqd)
 120                 return ERR_PTR(-EINVAL);
 121         if (sqd->task_tgid != current->tgid)
 122                 return ERR_PTR(-EPERM);
 123
 124         refcount_inc(&sqd->refs);
 125         return sqd;
 126 }
 127
 128 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
 129                                          bool *attached)
 130 {
 131         struct io_sq_data *sqd;
 132
 133         *attached = false;
 134         if (p->flags & IORING_SETUP_ATTACH_WQ) {
 135                 sqd = io_attach_sq_data(p);
 136                 if (!IS_ERR(sqd)) {
 137                         *attached = true;
 138                         return sqd;
 139                 }
 140                 /* fall through for EPERM case, setup new sqd/task */
 141                 if (PTR_ERR(sqd) != -EPERM)
 142                         return sqd;
 143         }
 144
 145         sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
 146         if (!sqd)
 147                 return ERR_PTR(-ENOMEM);
 148
 149         atomic_set(&sqd->park_pending, 0);
 150         refcount_set(&sqd->refs, 1);
 151         INIT_LIST_HEAD(&sqd->ctx_list);
 152         mutex_init(&sqd->lock);
 153         init_waitqueue_head(&sqd->wait);
 154         init_completion(&sqd->exited);
 155         return sqd;
 156 }
 157
 158 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
 159 {
 160         return READ_ONCE(sqd->state);
 161 }
 162
 163 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
 164 {
 165         unsigned int to_submit;
 166         int ret = 0;
 167
 168         to_submit = io_sqring_entries(ctx);
 169         /* if we're handling multiple rings, cap submit size for fairness */
 170         if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
 171                 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
 172
 173         if (to_submit || !wq_list_empty(&ctx->iopoll_list)) {
 174                 const struct cred *creds = NULL;
 175
 176                 if (ctx->sq_creds != current_cred())
 177                         creds = override_creds(ctx->sq_creds);
 178
 179                 mutex_lock(&ctx->uring_lock);
 180                 if (!wq_list_empty(&ctx->iopoll_list))
 181                         io_do_iopoll(ctx, true);
 182
 183                 /*
 184                  * Don't submit if refs are dying, good for io_uring_register(),
 185                  * but also it is relied upon by io_ring_exit_work()
 186                  */
 187                 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
 188                     !(ctx->flags & IORING_SETUP_R_DISABLED))
 189                         ret = io_submit_sqes(ctx, to_submit);
 190                 mutex_unlock(&ctx->uring_lock);
 191
 192                 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
 193                         wake_up(&ctx->sqo_sq_wait);
 194                 if (creds)
 195                         revert_creds(creds);
 196         }
 197
 198         return ret;
 199 }
 200
 201 static bool io_sqd_handle_event(struct io_sq_data *sqd)
 202 {
 203         bool did_sig = false;
 204         struct ksignal ksig;
 205
 206         if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
 207             signal_pending(current)) {
 208                 mutex_unlock(&sqd->lock);
 209                 if (signal_pending(current))
 210                         did_sig = get_signal(&ksig);
 211                 wait_event(sqd->wait, !atomic_read(&sqd->park_pending));
 212                 mutex_lock(&sqd->lock);
 213                 sqd->sq_cpu = raw_smp_processor_id();
 214         }
 215         return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
 216 }
 217
 218 /*
 219  * Run task_work, processing the retry_list first. The retry_list holds
 220  * entries that we passed on in the previous run, if we had more task_work
 221  * than we were asked to process. Newly queued task_work isn't run until the
 222  * retry list has been fully processed.
 223  */
 224 static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)
 225 {
 226         struct io_uring_task *tctx = current->io_uring;
 227         unsigned int count = 0;
 228
 229         if (*retry_list) {
 230                 *retry_list = io_handle_tw_list(*retry_list, &count, max_entries);
 231                 if (count >= max_entries)
 232                         goto out;
 233                 max_entries -= count;
 234         }
 235         *retry_list = tctx_task_work_run(tctx, max_entries, &count);
 236 out:
 237         if (task_work_pending(current))
 238                 task_work_run();
 239         return count;
 240 }
 241
 242 static bool io_sq_tw_pending(struct llist_node *retry_list)
 243 {
 244         struct io_uring_task *tctx = current->io_uring;
 245
 246         return retry_list || !llist_empty(&tctx->task_list);
 247 }
 248
 249 static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
 250 {
 251         struct rusage end;
 252
 253         getrusage(current, RUSAGE_SELF, &end);
 254         end.ru_stime.tv_sec -= start->ru_stime.tv_sec;
 255         end.ru_stime.tv_usec -= start->ru_stime.tv_usec;
 256
 257         sqd->work_time += end.ru_stime.tv_usec + end.ru_stime.tv_sec * 1000000;
 258 }
 259
 260 static int io_sq_thread(void *data)
 261 {
 262         struct llist_node *retry_list = NULL;
 263         struct io_sq_data *sqd = data;
 264         struct io_ring_ctx *ctx;
 265         struct rusage start;
 266         unsigned long timeout = 0;
 267         char buf[TASK_COMM_LEN];
 268         DEFINE_WAIT(wait);
 269
 270         /* offload context creation failed, just exit */
 271         if (!current->io_uring)
 272                 goto err_out;
 273
 274         snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
 275         set_task_comm(current, buf);
 276
 277         /* reset to our pid after we've set task_comm, for fdinfo */
 278         sqd->task_pid = current->pid;
 279
 280         if (sqd->sq_cpu != -1) {
 281                 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
 282         } else {
 283                 set_cpus_allowed_ptr(current, cpu_online_mask);
 284                 sqd->sq_cpu = raw_smp_processor_id();
 285         }
 286
 287         /*
 288          * Force audit context to get setup, in case we do prep side async
 289          * operations that would trigger an audit call before any issue side
 290          * audit has been done.
 291          */
 292         audit_uring_entry(IORING_OP_NOP);
 293         audit_uring_exit(true, 0);
 294
 295         mutex_lock(&sqd->lock);
 296         while (1) {
 297                 bool cap_entries, sqt_spin = false;
 298
 299                 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
 300                         if (io_sqd_handle_event(sqd))
 301                                 break;
 302                         timeout = jiffies + sqd->sq_thread_idle;
 303                 }
 304
 305                 cap_entries = !list_is_singular(&sqd->ctx_list);
 306                 getrusage(current, RUSAGE_SELF, &start);
 307                 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
 308                         int ret = __io_sq_thread(ctx, cap_entries);
 309
 310                         if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
 311                                 sqt_spin = true;
 312                 }
 313                 if (io_sq_tw(&retry_list, IORING_TW_CAP_ENTRIES_VALUE))
 314                         sqt_spin = true;
 315
 316                 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 317                         if (io_napi(ctx))
 318                                 io_napi_sqpoll_busy_poll(ctx);
 319
 320                 if (sqt_spin || !time_after(jiffies, timeout)) {
 321                         if (sqt_spin) {
 322                                 io_sq_update_worktime(sqd, &start);
 323                                 timeout = jiffies + sqd->sq_thread_idle;
 324                         }
 325                         if (unlikely(need_resched())) {
 326                                 mutex_unlock(&sqd->lock);
 327                                 cond_resched();
 328                                 mutex_lock(&sqd->lock);
 329                                 sqd->sq_cpu = raw_smp_processor_id();
 330                         }
 331                         continue;
 332                 }
 333
 334                 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
 335                 if (!io_sqd_events_pending(sqd) && !io_sq_tw_pending(retry_list)) {
 336                         bool needs_sched = true;
 337
 338                         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
 339                                 atomic_or(IORING_SQ_NEED_WAKEUP,
 340                                                 &ctx->rings->sq_flags);
 341                                 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
 342                                     !wq_list_empty(&ctx->iopoll_list)) {
 343                                         needs_sched = false;
 344                                         break;
 345                                 }
 346
 347                                 /*
 348                                  * Ensure the store of the wakeup flag is not
 349                                  * reordered with the load of the SQ tail
 350                                  */
 351                                 smp_mb__after_atomic();
 352
 353                                 if (io_sqring_entries(ctx)) {
 354                                         needs_sched = false;
 355                                         break;
 356                                 }
 357                         }
 358
 359                         if (needs_sched) {
 360                                 mutex_unlock(&sqd->lock);
 361                                 schedule();
 362                                 mutex_lock(&sqd->lock);
 363                                 sqd->sq_cpu = raw_smp_processor_id();
 364                         }
 365                         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 366                                 atomic_andnot(IORING_SQ_NEED_WAKEUP,
 367                                                 &ctx->rings->sq_flags);
 368                 }
 369
 370                 finish_wait(&sqd->wait, &wait);
 371                 timeout = jiffies + sqd->sq_thread_idle;
 372         }
 373
 374         if (retry_list)
 375                 io_sq_tw(&retry_list, UINT_MAX);
 376
 377         io_uring_cancel_generic(true, sqd);
 378         sqd->thread = NULL;
 379         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 380                 atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
 381         io_run_task_work();
 382         mutex_unlock(&sqd->lock);
 383 err_out:
 384         complete(&sqd->exited);
 385         do_exit(0);
 386 }
 387
 388 void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
 389 {
 390         DEFINE_WAIT(wait);
 391
 392         do {
 393                 if (!io_sqring_full(ctx))
 394                         break;
 395                 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
 396
 397                 if (!io_sqring_full(ctx))
 398                         break;
 399                 schedule();
 400         } while (!signal_pending(current));
 401
 402         finish_wait(&ctx->sqo_sq_wait, &wait);
 403 }
 404
 405 __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
 406                                 struct io_uring_params *p)
 407 {
 408         int ret;
 409
 410         /* Retain compatibility with failing for an invalid attach attempt */
 411         if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
 412                                 IORING_SETUP_ATTACH_WQ) {
 413                 CLASS(fd, f)(p->wq_fd);
 414                 if (fd_empty(f))
 415                         return -ENXIO;
 416                 if (!io_is_uring_fops(fd_file(f)))
 417                         return -EINVAL;
 418         }
 419         if (ctx->flags & IORING_SETUP_SQPOLL) {
 420                 struct task_struct *tsk;
 421                 struct io_sq_data *sqd;
 422                 bool attached;
 423
 424                 ret = security_uring_sqpoll();
 425                 if (ret)
 426                         return ret;
 427
 428                 sqd = io_get_sq_data(p, &attached);
 429                 if (IS_ERR(sqd)) {
 430                         ret = PTR_ERR(sqd);
 431                         goto err;
 432                 }
 433
 434                 ctx->sq_creds = get_current_cred();
 435                 ctx->sq_data = sqd;
 436                 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
 437                 if (!ctx->sq_thread_idle)
 438                         ctx->sq_thread_idle = HZ;
 439
 440                 io_sq_thread_park(sqd);
 441                 list_add(&ctx->sqd_list, &sqd->ctx_list);
 442                 io_sqd_update_thread_idle(sqd);
 443                 /* don't attach to a dying SQPOLL thread, would be racy */
 444                 ret = (attached && !sqd->thread) ? -ENXIO : 0;
 445                 io_sq_thread_unpark(sqd);
 446
 447                 if (ret < 0)
 448                         goto err;
 449                 if (attached)
 450                         return 0;
 451
 452                 if (p->flags & IORING_SETUP_SQ_AFF) {
 453                         cpumask_var_t allowed_mask;
 454                         int cpu = p->sq_thread_cpu;
 455
 456                         ret = -EINVAL;
 457                         if (cpu >= nr_cpu_ids || !cpu_online(cpu))
 458                                 goto err_sqpoll;
 459                         ret = -ENOMEM;
 460                         if (!alloc_cpumask_var(&allowed_mask, GFP_KERNEL))
 461                                 goto err_sqpoll;
 462                         ret = -EINVAL;
 463                         cpuset_cpus_allowed(current, allowed_mask);
 464                         if (!cpumask_test_cpu(cpu, allowed_mask)) {
 465                                 free_cpumask_var(allowed_mask);
 466                                 goto err_sqpoll;
 467                         }
 468                         free_cpumask_var(allowed_mask);
 469                         sqd->sq_cpu = cpu;
 470                 } else {
 471                         sqd->sq_cpu = -1;
 472                 }
 473
 474                 sqd->task_pid = current->pid;
 475                 sqd->task_tgid = current->tgid;
 476                 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
 477                 if (IS_ERR(tsk)) {
 478                         ret = PTR_ERR(tsk);
 479                         goto err_sqpoll;
 480                 }
 481
 482                 sqd->thread = tsk;
 483                 ret = io_uring_alloc_task_context(tsk, ctx);
 484                 wake_up_new_task(tsk);
 485                 if (ret)
 486                         goto err;
 487         } else if (p->flags & IORING_SETUP_SQ_AFF) {
 488                 /* Can't have SQ_AFF without SQPOLL */
 489                 ret = -EINVAL;
 490                 goto err;
 491         }
 492
 493         return 0;
 494 err_sqpoll:
 495         complete(&ctx->sq_data->exited);
 496 err:
 497         io_sq_thread_finish(ctx);
 498         return ret;
 499 }
 500
 501 __cold int io_sqpoll_wq_cpu_affinity(struct io_ring_ctx *ctx,
 502                                      cpumask_var_t mask)
 503 {
 504         struct io_sq_data *sqd = ctx->sq_data;
 505         int ret = -EINVAL;
 506
 507         if (sqd) {
 508                 io_sq_thread_park(sqd);
 509                 /* Don't set affinity for a dying thread */
 510                 if (sqd->thread)
 511                         ret = io_wq_cpu_affinity(sqd->thread->io_uring, mask);
 512                 io_sq_thread_unpark(sqd);
 513         }
 514
 515         return ret;
 516 }