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
  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                 sqd->sq_cpu = raw_smp_processor_id();
 218         }
 219         return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
 220 }
 221
 222 static int io_sq_thread(void *data)
 223 {
 224         struct io_sq_data *sqd = data;
 225         struct io_ring_ctx *ctx;
 226         unsigned long timeout = 0;
 227         char buf[TASK_COMM_LEN];
 228         DEFINE_WAIT(wait);
 229
 230         snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
 231         set_task_comm(current, buf);
 232
 233         /* reset to our pid after we've set task_comm, for fdinfo */
 234         sqd->task_pid = current->pid;
 235
 236         if (sqd->sq_cpu != -1) {
 237                 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
 238         } else {
 239                 set_cpus_allowed_ptr(current, cpu_online_mask);
 240                 sqd->sq_cpu = raw_smp_processor_id();
 241         }
 242
 243         mutex_lock(&sqd->lock);
 244         while (1) {
 245                 bool cap_entries, sqt_spin = false;
 246
 247                 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
 248                         if (io_sqd_handle_event(sqd))
 249                                 break;
 250                         timeout = jiffies + sqd->sq_thread_idle;
 251                 }
 252
 253                 cap_entries = !list_is_singular(&sqd->ctx_list);
 254                 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
 255                         int ret = __io_sq_thread(ctx, cap_entries);
 256
 257                         if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
 258                                 sqt_spin = true;
 259                 }
 260                 if (io_run_task_work())
 261                         sqt_spin = true;
 262
 263                 if (sqt_spin || !time_after(jiffies, timeout)) {
 264                         if (sqt_spin)
 265                                 timeout = jiffies + sqd->sq_thread_idle;
 266                         if (unlikely(need_resched())) {
 267                                 mutex_unlock(&sqd->lock);
 268                                 cond_resched();
 269                                 mutex_lock(&sqd->lock);
 270                                 sqd->sq_cpu = raw_smp_processor_id();
 271                         }
 272                         continue;
 273                 }
 274
 275                 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
 276                 if (!io_sqd_events_pending(sqd) && !task_work_pending(current)) {
 277                         bool needs_sched = true;
 278
 279                         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
 280                                 atomic_or(IORING_SQ_NEED_WAKEUP,
 281                                                 &ctx->rings->sq_flags);
 282                                 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
 283                                     !wq_list_empty(&ctx->iopoll_list)) {
 284                                         needs_sched = false;
 285                                         break;
 286                                 }
 287
 288                                 /*
 289                                  * Ensure the store of the wakeup flag is not
 290                                  * reordered with the load of the SQ tail
 291                                  */
 292                                 smp_mb__after_atomic();
 293
 294                                 if (io_sqring_entries(ctx)) {
 295                                         needs_sched = false;
 296                                         break;
 297                                 }
 298                         }
 299
 300                         if (needs_sched) {
 301                                 mutex_unlock(&sqd->lock);
 302                                 schedule();
 303                                 mutex_lock(&sqd->lock);
 304                                 sqd->sq_cpu = raw_smp_processor_id();
 305                         }
 306                         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 307                                 atomic_andnot(IORING_SQ_NEED_WAKEUP,
 308                                                 &ctx->rings->sq_flags);
 309                 }
 310
 311                 finish_wait(&sqd->wait, &wait);
 312                 timeout = jiffies + sqd->sq_thread_idle;
 313         }
 314
 315         io_uring_cancel_generic(true, sqd);
 316         sqd->thread = NULL;
 317         list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
 318                 atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
 319         io_run_task_work();
 320         mutex_unlock(&sqd->lock);
 321
 322         complete(&sqd->exited);
 323         do_exit(0);
 324 }
 325
 326 void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
 327 {
 328         DEFINE_WAIT(wait);
 329
 330         do {
 331                 if (!io_sqring_full(ctx))
 332                         break;
 333                 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
 334
 335                 if (!io_sqring_full(ctx))
 336                         break;
 337                 schedule();
 338         } while (!signal_pending(current));
 339
 340         finish_wait(&ctx->sqo_sq_wait, &wait);
 341 }
 342
 343 __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
 344                                 struct io_uring_params *p)
 345 {
 346         int ret;
 347
 348         /* Retain compatibility with failing for an invalid attach attempt */
 349         if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
 350                                 IORING_SETUP_ATTACH_WQ) {
 351                 struct fd f;
 352
 353                 f = fdget(p->wq_fd);
 354                 if (!f.file)
 355                         return -ENXIO;
 356                 if (!io_is_uring_fops(f.file)) {
 357                         fdput(f);
 358                         return -EINVAL;
 359                 }
 360                 fdput(f);
 361         }
 362         if (ctx->flags & IORING_SETUP_SQPOLL) {
 363                 struct task_struct *tsk;
 364                 struct io_sq_data *sqd;
 365                 bool attached;
 366
 367                 ret = security_uring_sqpoll();
 368                 if (ret)
 369                         return ret;
 370
 371                 sqd = io_get_sq_data(p, &attached);
 372                 if (IS_ERR(sqd)) {
 373                         ret = PTR_ERR(sqd);
 374                         goto err;
 375                 }
 376
 377                 ctx->sq_creds = get_current_cred();
 378                 ctx->sq_data = sqd;
 379                 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
 380                 if (!ctx->sq_thread_idle)
 381                         ctx->sq_thread_idle = HZ;
 382
 383                 io_sq_thread_park(sqd);
 384                 list_add(&ctx->sqd_list, &sqd->ctx_list);
 385                 io_sqd_update_thread_idle(sqd);
 386                 /* don't attach to a dying SQPOLL thread, would be racy */
 387                 ret = (attached && !sqd->thread) ? -ENXIO : 0;
 388                 io_sq_thread_unpark(sqd);
 389
 390                 if (ret < 0)
 391                         goto err;
 392                 if (attached)
 393                         return 0;
 394
 395                 if (p->flags & IORING_SETUP_SQ_AFF) {
 396                         int cpu = p->sq_thread_cpu;
 397
 398                         ret = -EINVAL;
 399                         if (cpu >= nr_cpu_ids || !cpu_online(cpu))
 400                                 goto err_sqpoll;
 401                         sqd->sq_cpu = cpu;
 402                 } else {
 403                         sqd->sq_cpu = -1;
 404                 }
 405
 406                 sqd->task_pid = current->pid;
 407                 sqd->task_tgid = current->tgid;
 408                 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
 409                 if (IS_ERR(tsk)) {
 410                         ret = PTR_ERR(tsk);
 411                         goto err_sqpoll;
 412                 }
 413
 414                 sqd->thread = tsk;
 415                 ret = io_uring_alloc_task_context(tsk, ctx);
 416                 wake_up_new_task(tsk);
 417                 if (ret)
 418                         goto err;
 419         } else if (p->flags & IORING_SETUP_SQ_AFF) {
 420                 /* Can't have SQ_AFF without SQPOLL */
 421                 ret = -EINVAL;
 422                 goto err;
 423         }
 424
 425         return 0;
 426 err_sqpoll:
 427         complete(&ctx->sq_data->exited);
 428 err:
 429         io_sq_thread_finish(ctx);
 430         return ret;
 431 }
 432
 433 __cold int io_sqpoll_wq_cpu_affinity(struct io_ring_ctx *ctx,
 434                                      cpumask_var_t mask)
 435 {
 436         struct io_sq_data *sqd = ctx->sq_data;
 437         int ret = -EINVAL;
 438
 439         if (sqd) {
 440                 io_sq_thread_park(sqd);
 441                 /* Don't set affinity for a dying thread */
 442                 if (sqd->thread)
 443                         ret = io_wq_cpu_affinity(sqd->thread->io_uring, mask);
 444                 io_sq_thread_unpark(sqd);
 445         }
 446
 447         return ret;
 448 }