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