io_uring/rw.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/kernel.h>
   3 #include <linux/errno.h>
   4 #include <linux/fs.h>
   5 #include <linux/file.h>
   6 #include <linux/blk-mq.h>
   7 #include <linux/mm.h>
   8 #include <linux/slab.h>
   9 #include <linux/fsnotify.h>
  10 #include <linux/poll.h>
  11 #include <linux/nospec.h>
  12 #include <linux/compat.h>
  13 #include <linux/io_uring.h>
  14
  15 #include <uapi/linux/io_uring.h>
  16
  17 #include "io_uring.h"
  18 #include "opdef.h"
  19 #include "kbuf.h"
  20 #include "rsrc.h"
  21 #include "rw.h"
  22
  23 struct io_rw {
  24         /* NOTE: kiocb has the file as the first member, so don't do it here */
  25         struct kiocb                    kiocb;
  26         u64                             addr;
  27         u32                             len;
  28         rwf_t                           flags;
  29 };
  30
  31 static inline bool io_file_supports_nowait(struct io_kiocb *req)
  32 {
  33         return req->flags & REQ_F_SUPPORT_NOWAIT;
  34 }
  35
  36 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
  37 {
  38         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
  39         unsigned ioprio;
  40         int ret;
  41
  42         rw->kiocb.ki_pos = READ_ONCE(sqe->off);
  43         /* used for fixed read/write too - just read unconditionally */
  44         req->buf_index = READ_ONCE(sqe->buf_index);
  45
  46         if (req->opcode == IORING_OP_READ_FIXED ||
  47             req->opcode == IORING_OP_WRITE_FIXED) {
  48                 struct io_ring_ctx *ctx = req->ctx;
  49                 u16 index;
  50
  51                 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
  52                         return -EFAULT;
  53                 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
  54                 req->imu = ctx->user_bufs[index];
  55                 io_req_set_rsrc_node(req, ctx, 0);
  56         }
  57
  58         ioprio = READ_ONCE(sqe->ioprio);
  59         if (ioprio) {
  60                 ret = ioprio_check_cap(ioprio);
  61                 if (ret)
  62                         return ret;
  63
  64                 rw->kiocb.ki_ioprio = ioprio;
  65         } else {
  66                 rw->kiocb.ki_ioprio = get_current_ioprio();
  67         }
  68
  69         rw->addr = READ_ONCE(sqe->addr);
  70         rw->len = READ_ONCE(sqe->len);
  71         rw->flags = READ_ONCE(sqe->rw_flags);
  72         return 0;
  73 }
  74
  75 void io_readv_writev_cleanup(struct io_kiocb *req)
  76 {
  77         struct io_async_rw *io = req->async_data;
  78
  79         kfree(io->free_iovec);
  80 }
  81
  82 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
  83 {
  84         switch (ret) {
  85         case -EIOCBQUEUED:
  86                 break;
  87         case -ERESTARTSYS:
  88         case -ERESTARTNOINTR:
  89         case -ERESTARTNOHAND:
  90         case -ERESTART_RESTARTBLOCK:
  91                 /*
  92                  * We can't just restart the syscall, since previously
  93                  * submitted sqes may already be in progress. Just fail this
  94                  * IO with EINTR.
  95                  */
  96                 ret = -EINTR;
  97                 fallthrough;
  98         default:
  99                 kiocb->ki_complete(kiocb, ret);
 100         }
 101 }
 102
 103 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
 104 {
 105         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 106
 107         if (rw->kiocb.ki_pos != -1)
 108                 return &rw->kiocb.ki_pos;
 109
 110         if (!(req->file->f_mode & FMODE_STREAM)) {
 111                 req->flags |= REQ_F_CUR_POS;
 112                 rw->kiocb.ki_pos = req->file->f_pos;
 113                 return &rw->kiocb.ki_pos;
 114         }
 115
 116         rw->kiocb.ki_pos = 0;
 117         return NULL;
 118 }
 119
 120 static void io_req_task_queue_reissue(struct io_kiocb *req)
 121 {
 122         req->io_task_work.func = io_queue_iowq;
 123         io_req_task_work_add(req);
 124 }
 125
 126 #ifdef CONFIG_BLOCK
 127 static bool io_resubmit_prep(struct io_kiocb *req)
 128 {
 129         struct io_async_rw *io = req->async_data;
 130
 131         if (!req_has_async_data(req))
 132                 return !io_req_prep_async(req);
 133         iov_iter_restore(&io->s.iter, &io->s.iter_state);
 134         return true;
 135 }
 136
 137 static bool io_rw_should_reissue(struct io_kiocb *req)
 138 {
 139         umode_t mode = file_inode(req->file)->i_mode;
 140         struct io_ring_ctx *ctx = req->ctx;
 141
 142         if (!S_ISBLK(mode) && !S_ISREG(mode))
 143                 return false;
 144         if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
 145             !(ctx->flags & IORING_SETUP_IOPOLL)))
 146                 return false;
 147         /*
 148          * If ref is dying, we might be running poll reap from the exit work.
 149          * Don't attempt to reissue from that path, just let it fail with
 150          * -EAGAIN.
 151          */
 152         if (percpu_ref_is_dying(&ctx->refs))
 153                 return false;
 154         /*
 155          * Play it safe and assume not safe to re-import and reissue if we're
 156          * not in the original thread group (or in task context).
 157          */
 158         if (!same_thread_group(req->task, current) || !in_task())
 159                 return false;
 160         return true;
 161 }
 162 #else
 163 static bool io_resubmit_prep(struct io_kiocb *req)
 164 {
 165         return false;
 166 }
 167 static bool io_rw_should_reissue(struct io_kiocb *req)
 168 {
 169         return false;
 170 }
 171 #endif
 172
 173 static void kiocb_end_write(struct io_kiocb *req)
 174 {
 175         /*
 176          * Tell lockdep we inherited freeze protection from submission
 177          * thread.
 178          */
 179         if (req->flags & REQ_F_ISREG) {
 180                 struct super_block *sb = file_inode(req->file)->i_sb;
 181
 182                 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
 183                 sb_end_write(sb);
 184         }
 185 }
 186
 187 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
 188 {
 189         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 190
 191         if (rw->kiocb.ki_flags & IOCB_WRITE) {
 192                 kiocb_end_write(req);
 193                 fsnotify_modify(req->file);
 194         } else {
 195                 fsnotify_access(req->file);
 196         }
 197         if (unlikely(res != req->cqe.res)) {
 198                 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
 199                     io_rw_should_reissue(req)) {
 200                         req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
 201                         return true;
 202                 }
 203                 req_set_fail(req);
 204                 req->cqe.res = res;
 205         }
 206         return false;
 207 }
 208
 209 static inline unsigned io_fixup_rw_res(struct io_kiocb *req, unsigned res)
 210 {
 211         struct io_async_rw *io = req->async_data;
 212
 213         /* add previously done IO, if any */
 214         if (req_has_async_data(req) && io->bytes_done > 0) {
 215                 if (res < 0)
 216                         res = io->bytes_done;
 217                 else
 218                         res += io->bytes_done;
 219         }
 220         return res;
 221 }
 222
 223 static void io_complete_rw(struct kiocb *kiocb, long res)
 224 {
 225         struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
 226         struct io_kiocb *req = cmd_to_io_kiocb(rw);
 227
 228         if (__io_complete_rw_common(req, res))
 229                 return;
 230         io_req_set_res(req, io_fixup_rw_res(req, res), 0);
 231         req->io_task_work.func = io_req_task_complete;
 232         io_req_task_work_add(req);
 233 }
 234
 235 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
 236 {
 237         struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
 238         struct io_kiocb *req = cmd_to_io_kiocb(rw);
 239
 240         if (kiocb->ki_flags & IOCB_WRITE)
 241                 kiocb_end_write(req);
 242         if (unlikely(res != req->cqe.res)) {
 243                 if (res == -EAGAIN && io_rw_should_reissue(req)) {
 244                         req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
 245                         return;
 246                 }
 247                 req->cqe.res = res;
 248         }
 249
 250         /* order with io_iopoll_complete() checking ->iopoll_completed */
 251         smp_store_release(&req->iopoll_completed, 1);
 252 }
 253
 254 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
 255                        unsigned int issue_flags)
 256 {
 257         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 258         unsigned final_ret = io_fixup_rw_res(req, ret);
 259
 260         if (req->flags & REQ_F_CUR_POS)
 261                 req->file->f_pos = rw->kiocb.ki_pos;
 262         if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
 263                 if (!__io_complete_rw_common(req, ret)) {
 264                         io_req_set_res(req, final_ret,
 265                                        io_put_kbuf(req, issue_flags));
 266                         return IOU_OK;
 267                 }
 268         } else {
 269                 io_rw_done(&rw->kiocb, ret);
 270         }
 271
 272         if (req->flags & REQ_F_REISSUE) {
 273                 req->flags &= ~REQ_F_REISSUE;
 274                 if (io_resubmit_prep(req))
 275                         io_req_task_queue_reissue(req);
 276                 else
 277                         io_req_task_queue_fail(req, final_ret);
 278         }
 279         return IOU_ISSUE_SKIP_COMPLETE;
 280 }
 281
 282 #ifdef CONFIG_COMPAT
 283 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
 284                                 unsigned int issue_flags)
 285 {
 286         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 287         struct compat_iovec __user *uiov;
 288         compat_ssize_t clen;
 289         void __user *buf;
 290         size_t len;
 291
 292         uiov = u64_to_user_ptr(rw->addr);
 293         if (!access_ok(uiov, sizeof(*uiov)))
 294                 return -EFAULT;
 295         if (__get_user(clen, &uiov->iov_len))
 296                 return -EFAULT;
 297         if (clen < 0)
 298                 return -EINVAL;
 299
 300         len = clen;
 301         buf = io_buffer_select(req, &len, issue_flags);
 302         if (!buf)
 303                 return -ENOBUFS;
 304         rw->addr = (unsigned long) buf;
 305         iov[0].iov_base = buf;
 306         rw->len = iov[0].iov_len = (compat_size_t) len;
 307         return 0;
 308 }
 309 #endif
 310
 311 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
 312                                       unsigned int issue_flags)
 313 {
 314         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 315         struct iovec __user *uiov = u64_to_user_ptr(rw->addr);
 316         void __user *buf;
 317         ssize_t len;
 318
 319         if (copy_from_user(iov, uiov, sizeof(*uiov)))
 320                 return -EFAULT;
 321
 322         len = iov[0].iov_len;
 323         if (len < 0)
 324                 return -EINVAL;
 325         buf = io_buffer_select(req, &len, issue_flags);
 326         if (!buf)
 327                 return -ENOBUFS;
 328         rw->addr = (unsigned long) buf;
 329         iov[0].iov_base = buf;
 330         rw->len = iov[0].iov_len = len;
 331         return 0;
 332 }
 333
 334 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
 335                                     unsigned int issue_flags)
 336 {
 337         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 338
 339         if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
 340                 iov[0].iov_base = u64_to_user_ptr(rw->addr);
 341                 iov[0].iov_len = rw->len;
 342                 return 0;
 343         }
 344         if (rw->len != 1)
 345                 return -EINVAL;
 346
 347 #ifdef CONFIG_COMPAT
 348         if (req->ctx->compat)
 349                 return io_compat_import(req, iov, issue_flags);
 350 #endif
 351
 352         return __io_iov_buffer_select(req, iov, issue_flags);
 353 }
 354
 355 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
 356                                        struct io_rw_state *s,
 357                                        unsigned int issue_flags)
 358 {
 359         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 360         struct iov_iter *iter = &s->iter;
 361         u8 opcode = req->opcode;
 362         struct iovec *iovec;
 363         void __user *buf;
 364         size_t sqe_len;
 365         ssize_t ret;
 366
 367         if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
 368                 ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
 369                 if (ret)
 370                         return ERR_PTR(ret);
 371                 return NULL;
 372         }
 373
 374         buf = u64_to_user_ptr(rw->addr);
 375         sqe_len = rw->len;
 376
 377         if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
 378                 if (io_do_buffer_select(req)) {
 379                         buf = io_buffer_select(req, &sqe_len, issue_flags);
 380                         if (!buf)
 381                                 return ERR_PTR(-ENOBUFS);
 382                         rw->addr = (unsigned long) buf;
 383                         rw->len = sqe_len;
 384                 }
 385
 386                 ret = import_single_range(ddir, buf, sqe_len, s->fast_iov, iter);
 387                 if (ret)
 388                         return ERR_PTR(ret);
 389                 return NULL;
 390         }
 391
 392         iovec = s->fast_iov;
 393         if (req->flags & REQ_F_BUFFER_SELECT) {
 394                 ret = io_iov_buffer_select(req, iovec, issue_flags);
 395                 if (ret)
 396                         return ERR_PTR(ret);
 397                 iov_iter_init(iter, ddir, iovec, 1, iovec->iov_len);
 398                 return NULL;
 399         }
 400
 401         ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
 402                               req->ctx->compat);
 403         if (unlikely(ret < 0))
 404                 return ERR_PTR(ret);
 405         return iovec;
 406 }
 407
 408 static inline int io_import_iovec(int rw, struct io_kiocb *req,
 409                                   struct iovec **iovec, struct io_rw_state *s,
 410                                   unsigned int issue_flags)
 411 {
 412         *iovec = __io_import_iovec(rw, req, s, issue_flags);
 413         if (unlikely(IS_ERR(*iovec)))
 414                 return PTR_ERR(*iovec);
 415
 416         iov_iter_save_state(&s->iter, &s->iter_state);
 417         return 0;
 418 }
 419
 420 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
 421 {
 422         return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
 423 }
 424
 425 /*
 426  * For files that don't have ->read_iter() and ->write_iter(), handle them
 427  * by looping over ->read() or ->write() manually.
 428  */
 429 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
 430 {
 431         struct kiocb *kiocb = &rw->kiocb;
 432         struct file *file = kiocb->ki_filp;
 433         ssize_t ret = 0;
 434         loff_t *ppos;
 435
 436         /*
 437          * Don't support polled IO through this interface, and we can't
 438          * support non-blocking either. For the latter, this just causes
 439          * the kiocb to be handled from an async context.
 440          */
 441         if (kiocb->ki_flags & IOCB_HIPRI)
 442                 return -EOPNOTSUPP;
 443         if ((kiocb->ki_flags & IOCB_NOWAIT) &&
 444             !(kiocb->ki_filp->f_flags & O_NONBLOCK))
 445                 return -EAGAIN;
 446
 447         ppos = io_kiocb_ppos(kiocb);
 448
 449         while (iov_iter_count(iter)) {
 450                 struct iovec iovec;
 451                 ssize_t nr;
 452
 453                 if (!iov_iter_is_bvec(iter)) {
 454                         iovec = iov_iter_iovec(iter);
 455                 } else {
 456                         iovec.iov_base = u64_to_user_ptr(rw->addr);
 457                         iovec.iov_len = rw->len;
 458                 }
 459
 460                 if (ddir == READ) {
 461                         nr = file->f_op->read(file, iovec.iov_base,
 462                                               iovec.iov_len, ppos);
 463                 } else {
 464                         nr = file->f_op->write(file, iovec.iov_base,
 465                                                iovec.iov_len, ppos);
 466                 }
 467
 468                 if (nr < 0) {
 469                         if (!ret)
 470                                 ret = nr;
 471                         break;
 472                 }
 473                 ret += nr;
 474                 if (!iov_iter_is_bvec(iter)) {
 475                         iov_iter_advance(iter, nr);
 476                 } else {
 477                         rw->addr += nr;
 478                         rw->len -= nr;
 479                         if (!rw->len)
 480                                 break;
 481                 }
 482                 if (nr != iovec.iov_len)
 483                         break;
 484         }
 485
 486         return ret;
 487 }
 488
 489 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
 490                           const struct iovec *fast_iov, struct iov_iter *iter)
 491 {
 492         struct io_async_rw *io = req->async_data;
 493
 494         memcpy(&io->s.iter, iter, sizeof(*iter));
 495         io->free_iovec = iovec;
 496         io->bytes_done = 0;
 497         /* can only be fixed buffers, no need to do anything */
 498         if (iov_iter_is_bvec(iter))
 499                 return;
 500         if (!iovec) {
 501                 unsigned iov_off = 0;
 502
 503                 io->s.iter.iov = io->s.fast_iov;
 504                 if (iter->iov != fast_iov) {
 505                         iov_off = iter->iov - fast_iov;
 506                         io->s.iter.iov += iov_off;
 507                 }
 508                 if (io->s.fast_iov != fast_iov)
 509                         memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
 510                                sizeof(struct iovec) * iter->nr_segs);
 511         } else {
 512                 req->flags |= REQ_F_NEED_CLEANUP;
 513         }
 514 }
 515
 516 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
 517                              struct io_rw_state *s, bool force)
 518 {
 519         if (!force && !io_op_defs[req->opcode].prep_async)
 520                 return 0;
 521         if (!req_has_async_data(req)) {
 522                 struct io_async_rw *iorw;
 523
 524                 if (io_alloc_async_data(req)) {
 525                         kfree(iovec);
 526                         return -ENOMEM;
 527                 }
 528
 529                 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
 530                 iorw = req->async_data;
 531                 /* we've copied and mapped the iter, ensure state is saved */
 532                 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
 533         }
 534         return 0;
 535 }
 536
 537 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
 538 {
 539         struct io_async_rw *iorw = req->async_data;
 540         struct iovec *iov;
 541         int ret;
 542
 543         /* submission path, ->uring_lock should already be taken */
 544         ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
 545         if (unlikely(ret < 0))
 546                 return ret;
 547
 548         iorw->bytes_done = 0;
 549         iorw->free_iovec = iov;
 550         if (iov)
 551                 req->flags |= REQ_F_NEED_CLEANUP;
 552         return 0;
 553 }
 554
 555 int io_readv_prep_async(struct io_kiocb *req)
 556 {
 557         return io_rw_prep_async(req, READ);
 558 }
 559
 560 int io_writev_prep_async(struct io_kiocb *req)
 561 {
 562         return io_rw_prep_async(req, WRITE);
 563 }
 564
 565 /*
 566  * This is our waitqueue callback handler, registered through __folio_lock_async()
 567  * when we initially tried to do the IO with the iocb armed our waitqueue.
 568  * This gets called when the page is unlocked, and we generally expect that to
 569  * happen when the page IO is completed and the page is now uptodate. This will
 570  * queue a task_work based retry of the operation, attempting to copy the data
 571  * again. If the latter fails because the page was NOT uptodate, then we will
 572  * do a thread based blocking retry of the operation. That's the unexpected
 573  * slow path.
 574  */
 575 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
 576                              int sync, void *arg)
 577 {
 578         struct wait_page_queue *wpq;
 579         struct io_kiocb *req = wait->private;
 580         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 581         struct wait_page_key *key = arg;
 582
 583         wpq = container_of(wait, struct wait_page_queue, wait);
 584
 585         if (!wake_page_match(wpq, key))
 586                 return 0;
 587
 588         rw->kiocb.ki_flags &= ~IOCB_WAITQ;
 589         list_del_init(&wait->entry);
 590         io_req_task_queue(req);
 591         return 1;
 592 }
 593
 594 /*
 595  * This controls whether a given IO request should be armed for async page
 596  * based retry. If we return false here, the request is handed to the async
 597  * worker threads for retry. If we're doing buffered reads on a regular file,
 598  * we prepare a private wait_page_queue entry and retry the operation. This
 599  * will either succeed because the page is now uptodate and unlocked, or it
 600  * will register a callback when the page is unlocked at IO completion. Through
 601  * that callback, io_uring uses task_work to setup a retry of the operation.
 602  * That retry will attempt the buffered read again. The retry will generally
 603  * succeed, or in rare cases where it fails, we then fall back to using the
 604  * async worker threads for a blocking retry.
 605  */
 606 static bool io_rw_should_retry(struct io_kiocb *req)
 607 {
 608         struct io_async_rw *io = req->async_data;
 609         struct wait_page_queue *wait = &io->wpq;
 610         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 611         struct kiocb *kiocb = &rw->kiocb;
 612
 613         /* never retry for NOWAIT, we just complete with -EAGAIN */
 614         if (req->flags & REQ_F_NOWAIT)
 615                 return false;
 616
 617         /* Only for buffered IO */
 618         if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
 619                 return false;
 620
 621         /*
 622          * just use poll if we can, and don't attempt if the fs doesn't
 623          * support callback based unlocks
 624          */
 625         if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
 626                 return false;
 627
 628         wait->wait.func = io_async_buf_func;
 629         wait->wait.private = req;
 630         wait->wait.flags = 0;
 631         INIT_LIST_HEAD(&wait->wait.entry);
 632         kiocb->ki_flags |= IOCB_WAITQ;
 633         kiocb->ki_flags &= ~IOCB_NOWAIT;
 634         kiocb->ki_waitq = wait;
 635         return true;
 636 }
 637
 638 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
 639 {
 640         struct file *file = rw->kiocb.ki_filp;
 641
 642         if (likely(file->f_op->read_iter))
 643                 return call_read_iter(file, &rw->kiocb, iter);
 644         else if (file->f_op->read)
 645                 return loop_rw_iter(READ, rw, iter);
 646         else
 647                 return -EINVAL;
 648 }
 649
 650 static bool need_complete_io(struct io_kiocb *req)
 651 {
 652         return req->flags & REQ_F_ISREG ||
 653                 S_ISBLK(file_inode(req->file)->i_mode);
 654 }
 655
 656 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
 657 {
 658         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 659         struct kiocb *kiocb = &rw->kiocb;
 660         struct io_ring_ctx *ctx = req->ctx;
 661         struct file *file = req->file;
 662         int ret;
 663
 664         if (unlikely(!file || !(file->f_mode & mode)))
 665                 return -EBADF;
 666
 667         if (!io_req_ffs_set(req))
 668                 req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
 669
 670         kiocb->ki_flags = file->f_iocb_flags;
 671         ret = kiocb_set_rw_flags(kiocb, rw->flags);
 672         if (unlikely(ret))
 673                 return ret;
 674
 675         /*
 676          * If the file is marked O_NONBLOCK, still allow retry for it if it
 677          * supports async. Otherwise it's impossible to use O_NONBLOCK files
 678          * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
 679          */
 680         if ((kiocb->ki_flags & IOCB_NOWAIT) ||
 681             ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
 682                 req->flags |= REQ_F_NOWAIT;
 683
 684         if (ctx->flags & IORING_SETUP_IOPOLL) {
 685                 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
 686                         return -EOPNOTSUPP;
 687
 688                 kiocb->private = NULL;
 689                 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
 690                 kiocb->ki_complete = io_complete_rw_iopoll;
 691                 req->iopoll_completed = 0;
 692         } else {
 693                 if (kiocb->ki_flags & IOCB_HIPRI)
 694                         return -EINVAL;
 695                 kiocb->ki_complete = io_complete_rw;
 696         }
 697
 698         return 0;
 699 }
 700
 701 int io_read(struct io_kiocb *req, unsigned int issue_flags)
 702 {
 703         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 704         struct io_rw_state __s, *s = &__s;
 705         struct iovec *iovec;
 706         struct kiocb *kiocb = &rw->kiocb;
 707         bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
 708         struct io_async_rw *io;
 709         ssize_t ret, ret2;
 710         loff_t *ppos;
 711
 712         if (!req_has_async_data(req)) {
 713                 ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
 714                 if (unlikely(ret < 0))
 715                         return ret;
 716         } else {
 717                 io = req->async_data;
 718                 s = &io->s;
 719
 720                 /*
 721                  * Safe and required to re-import if we're using provided
 722                  * buffers, as we dropped the selected one before retry.
 723                  */
 724                 if (io_do_buffer_select(req)) {
 725                         ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
 726                         if (unlikely(ret < 0))
 727                                 return ret;
 728                 }
 729
 730                 /*
 731                  * We come here from an earlier attempt, restore our state to
 732                  * match in case it doesn't. It's cheap enough that we don't
 733                  * need to make this conditional.
 734                  */
 735                 iov_iter_restore(&s->iter, &s->iter_state);
 736                 iovec = NULL;
 737         }
 738         ret = io_rw_init_file(req, FMODE_READ);
 739         if (unlikely(ret)) {
 740                 kfree(iovec);
 741                 return ret;
 742         }
 743         req->cqe.res = iov_iter_count(&s->iter);
 744
 745         if (force_nonblock) {
 746                 /* If the file doesn't support async, just async punt */
 747                 if (unlikely(!io_file_supports_nowait(req))) {
 748                         ret = io_setup_async_rw(req, iovec, s, true);
 749                         return ret ?: -EAGAIN;
 750                 }
 751                 kiocb->ki_flags |= IOCB_NOWAIT;
 752         } else {
 753                 /* Ensure we clear previously set non-block flag */
 754                 kiocb->ki_flags &= ~IOCB_NOWAIT;
 755         }
 756
 757         ppos = io_kiocb_update_pos(req);
 758
 759         ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
 760         if (unlikely(ret)) {
 761                 kfree(iovec);
 762                 return ret;
 763         }
 764
 765         ret = io_iter_do_read(rw, &s->iter);
 766
 767         if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
 768                 req->flags &= ~REQ_F_REISSUE;
 769                 /* if we can poll, just do that */
 770                 if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
 771                         return -EAGAIN;
 772                 /* IOPOLL retry should happen for io-wq threads */
 773                 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
 774                         goto done;
 775                 /* no retry on NONBLOCK nor RWF_NOWAIT */
 776                 if (req->flags & REQ_F_NOWAIT)
 777                         goto done;
 778                 ret = 0;
 779         } else if (ret == -EIOCBQUEUED) {
 780                 if (iovec)
 781                         kfree(iovec);
 782                 return IOU_ISSUE_SKIP_COMPLETE;
 783         } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
 784                    (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
 785                 /* read all, failed, already did sync or don't want to retry */
 786                 goto done;
 787         }
 788
 789         /*
 790          * Don't depend on the iter state matching what was consumed, or being
 791          * untouched in case of error. Restore it and we'll advance it
 792          * manually if we need to.
 793          */
 794         iov_iter_restore(&s->iter, &s->iter_state);
 795
 796         ret2 = io_setup_async_rw(req, iovec, s, true);
 797         if (ret2)
 798                 return ret2;
 799
 800         iovec = NULL;
 801         io = req->async_data;
 802         s = &io->s;
 803         /*
 804          * Now use our persistent iterator and state, if we aren't already.
 805          * We've restored and mapped the iter to match.
 806          */
 807
 808         do {
 809                 /*
 810                  * We end up here because of a partial read, either from
 811                  * above or inside this loop. Advance the iter by the bytes
 812                  * that were consumed.
 813                  */
 814                 iov_iter_advance(&s->iter, ret);
 815                 if (!iov_iter_count(&s->iter))
 816                         break;
 817                 io->bytes_done += ret;
 818                 iov_iter_save_state(&s->iter, &s->iter_state);
 819
 820                 /* if we can retry, do so with the callbacks armed */
 821                 if (!io_rw_should_retry(req)) {
 822                         kiocb->ki_flags &= ~IOCB_WAITQ;
 823                         return -EAGAIN;
 824                 }
 825
 826                 /*
 827                  * Now retry read with the IOCB_WAITQ parts set in the iocb. If
 828                  * we get -EIOCBQUEUED, then we'll get a notification when the
 829                  * desired page gets unlocked. We can also get a partial read
 830                  * here, and if we do, then just retry at the new offset.
 831                  */
 832                 ret = io_iter_do_read(rw, &s->iter);
 833                 if (ret == -EIOCBQUEUED)
 834                         return IOU_ISSUE_SKIP_COMPLETE;
 835                 /* we got some bytes, but not all. retry. */
 836                 kiocb->ki_flags &= ~IOCB_WAITQ;
 837                 iov_iter_restore(&s->iter, &s->iter_state);
 838         } while (ret > 0);
 839 done:
 840         /* it's faster to check here then delegate to kfree */
 841         if (iovec)
 842                 kfree(iovec);
 843         return kiocb_done(req, ret, issue_flags);
 844 }
 845
 846 int io_write(struct io_kiocb *req, unsigned int issue_flags)
 847 {
 848         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 849         struct io_rw_state __s, *s = &__s;
 850         struct iovec *iovec;
 851         struct kiocb *kiocb = &rw->kiocb;
 852         bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
 853         ssize_t ret, ret2;
 854         loff_t *ppos;
 855
 856         if (!req_has_async_data(req)) {
 857                 ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
 858                 if (unlikely(ret < 0))
 859                         return ret;
 860         } else {
 861                 struct io_async_rw *io = req->async_data;
 862
 863                 s = &io->s;
 864                 iov_iter_restore(&s->iter, &s->iter_state);
 865                 iovec = NULL;
 866         }
 867         ret = io_rw_init_file(req, FMODE_WRITE);
 868         if (unlikely(ret)) {
 869                 kfree(iovec);
 870                 return ret;
 871         }
 872         req->cqe.res = iov_iter_count(&s->iter);
 873
 874         if (force_nonblock) {
 875                 /* If the file doesn't support async, just async punt */
 876                 if (unlikely(!io_file_supports_nowait(req)))
 877                         goto copy_iov;
 878
 879                 /* File path supports NOWAIT for non-direct_IO only for block devices. */
 880                 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
 881                         !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
 882                         (req->flags & REQ_F_ISREG))
 883                         goto copy_iov;
 884
 885                 kiocb->ki_flags |= IOCB_NOWAIT;
 886         } else {
 887                 /* Ensure we clear previously set non-block flag */
 888                 kiocb->ki_flags &= ~IOCB_NOWAIT;
 889         }
 890
 891         ppos = io_kiocb_update_pos(req);
 892
 893         ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
 894         if (unlikely(ret)) {
 895                 kfree(iovec);
 896                 return ret;
 897         }
 898
 899         /*
 900          * Open-code file_start_write here to grab freeze protection,
 901          * which will be released by another thread in
 902          * io_complete_rw().  Fool lockdep by telling it the lock got
 903          * released so that it doesn't complain about the held lock when
 904          * we return to userspace.
 905          */
 906         if (req->flags & REQ_F_ISREG) {
 907                 sb_start_write(file_inode(req->file)->i_sb);
 908                 __sb_writers_release(file_inode(req->file)->i_sb,
 909                                         SB_FREEZE_WRITE);
 910         }
 911         kiocb->ki_flags |= IOCB_WRITE;
 912
 913         if (likely(req->file->f_op->write_iter))
 914                 ret2 = call_write_iter(req->file, kiocb, &s->iter);
 915         else if (req->file->f_op->write)
 916                 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
 917         else
 918                 ret2 = -EINVAL;
 919
 920         if (req->flags & REQ_F_REISSUE) {
 921                 req->flags &= ~REQ_F_REISSUE;
 922                 ret2 = -EAGAIN;
 923         }
 924
 925         /*
 926          * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
 927          * retry them without IOCB_NOWAIT.
 928          */
 929         if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
 930                 ret2 = -EAGAIN;
 931         /* no retry on NONBLOCK nor RWF_NOWAIT */
 932         if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
 933                 goto done;
 934         if (!force_nonblock || ret2 != -EAGAIN) {
 935                 /* IOPOLL retry should happen for io-wq threads */
 936                 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
 937                         goto copy_iov;
 938
 939                 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
 940                         struct io_async_rw *rw;
 941
 942                         trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
 943                                                 req->cqe.res, ret2);
 944
 945                         /* This is a partial write. The file pos has already been
 946                          * updated, setup the async struct to complete the request
 947                          * in the worker. Also update bytes_done to account for
 948                          * the bytes already written.
 949                          */
 950                         iov_iter_save_state(&s->iter, &s->iter_state);
 951                         ret = io_setup_async_rw(req, iovec, s, true);
 952
 953                         rw = req->async_data;
 954                         if (rw)
 955                                 rw->bytes_done += ret2;
 956
 957                         if (kiocb->ki_flags & IOCB_WRITE)
 958                                 kiocb_end_write(req);
 959                         return ret ? ret : -EAGAIN;
 960                 }
 961 done:
 962                 ret = kiocb_done(req, ret2, issue_flags);
 963         } else {
 964 copy_iov:
 965                 iov_iter_restore(&s->iter, &s->iter_state);
 966                 ret = io_setup_async_rw(req, iovec, s, false);
 967                 if (!ret) {
 968                         if (kiocb->ki_flags & IOCB_WRITE)
 969                                 kiocb_end_write(req);
 970                         return -EAGAIN;
 971                 }
 972                 return ret;
 973         }
 974         /* it's reportedly faster than delegating the null check to kfree() */
 975         if (iovec)
 976                 kfree(iovec);
 977         return ret;
 978 }
 979
 980 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
 981 {
 982         io_commit_cqring_flush(ctx);
 983         if (ctx->flags & IORING_SETUP_SQPOLL)
 984                 io_cqring_wake(ctx);
 985 }
 986
 987 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
 988 {
 989         struct io_wq_work_node *pos, *start, *prev;
 990         unsigned int poll_flags = BLK_POLL_NOSLEEP;
 991         DEFINE_IO_COMP_BATCH(iob);
 992         int nr_events = 0;
 993
 994         /*
 995          * Only spin for completions if we don't have multiple devices hanging
 996          * off our complete list.
 997          */
 998         if (ctx->poll_multi_queue || force_nonspin)
 999                 poll_flags |= BLK_POLL_ONESHOT;
1000
1001         wq_list_for_each(pos, start, &ctx->iopoll_list) {
1002                 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1003                 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1004                 int ret;
1005
1006                 /*
1007                  * Move completed and retryable entries to our local lists.
1008                  * If we find a request that requires polling, break out
1009                  * and complete those lists first, if we have entries there.
1010                  */
1011                 if (READ_ONCE(req->iopoll_completed))
1012                         break;
1013
1014                 ret = rw->kiocb.ki_filp->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1015                 if (unlikely(ret < 0))
1016                         return ret;
1017                 else if (ret)
1018                         poll_flags |= BLK_POLL_ONESHOT;
1019
1020                 /* iopoll may have completed current req */
1021                 if (!rq_list_empty(iob.req_list) ||
1022                     READ_ONCE(req->iopoll_completed))
1023                         break;
1024         }
1025
1026         if (!rq_list_empty(iob.req_list))
1027                 iob.complete(&iob);
1028         else if (!pos)
1029                 return 0;
1030
1031         prev = start;
1032         wq_list_for_each_resume(pos, prev) {
1033                 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1034
1035                 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1036                 if (!smp_load_acquire(&req->iopoll_completed))
1037                         break;
1038                 nr_events++;
1039                 if (unlikely(req->flags & REQ_F_CQE_SKIP))
1040                         continue;
1041
1042                 req->cqe.flags = io_put_kbuf(req, 0);
1043                 __io_fill_cqe_req(req->ctx, req);
1044         }
1045
1046         if (unlikely(!nr_events))
1047                 return 0;
1048
1049         io_commit_cqring(ctx);
1050         io_cqring_ev_posted_iopoll(ctx);
1051         pos = start ? start->next : ctx->iopoll_list.first;
1052         wq_list_cut(&ctx->iopoll_list, prev, start);
1053         io_free_batch_list(ctx, pos);
1054         return nr_events;
1055 }