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