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 #ifdef CONFIG_COMPAT
  37 static int io_iov_compat_buffer_select_prep(struct io_rw *rw)
  38 {
  39         struct compat_iovec __user *uiov;
  40         compat_ssize_t clen;
  41
  42         uiov = u64_to_user_ptr(rw->addr);
  43         if (!access_ok(uiov, sizeof(*uiov)))
  44                 return -EFAULT;
  45         if (__get_user(clen, &uiov->iov_len))
  46                 return -EFAULT;
  47         if (clen < 0)
  48                 return -EINVAL;
  49
  50         rw->len = clen;
  51         return 0;
  52 }
  53 #endif
  54
  55 static int io_iov_buffer_select_prep(struct io_kiocb *req)
  56 {
  57         struct iovec __user *uiov;
  58         struct iovec iov;
  59         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
  60
  61         if (rw->len != 1)
  62                 return -EINVAL;
  63
  64 #ifdef CONFIG_COMPAT
  65         if (req->ctx->compat)
  66                 return io_iov_compat_buffer_select_prep(rw);
  67 #endif
  68
  69         uiov = u64_to_user_ptr(rw->addr);
  70         if (copy_from_user(&iov, uiov, sizeof(*uiov)))
  71                 return -EFAULT;
  72         rw->len = iov.iov_len;
  73         return 0;
  74 }
  75
  76 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
  77 {
  78         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
  79         unsigned ioprio;
  80         int ret;
  81
  82         rw->kiocb.ki_pos = READ_ONCE(sqe->off);
  83         /* used for fixed read/write too - just read unconditionally */
  84         req->buf_index = READ_ONCE(sqe->buf_index);
  85
  86         ioprio = READ_ONCE(sqe->ioprio);
  87         if (ioprio) {
  88                 ret = ioprio_check_cap(ioprio);
  89                 if (ret)
  90                         return ret;
  91
  92                 rw->kiocb.ki_ioprio = ioprio;
  93         } else {
  94                 rw->kiocb.ki_ioprio = get_current_ioprio();
  95         }
  96         rw->kiocb.dio_complete = NULL;
  97
  98         rw->addr = READ_ONCE(sqe->addr);
  99         rw->len = READ_ONCE(sqe->len);
 100         rw->flags = READ_ONCE(sqe->rw_flags);
 101         return 0;
 102 }
 103
 104 int io_prep_rwv(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 105 {
 106         int ret;
 107
 108         ret = io_prep_rw(req, sqe);
 109         if (unlikely(ret))
 110                 return ret;
 111
 112         /*
 113          * Have to do this validation here, as this is in io_read() rw->len
 114          * might have chanaged due to buffer selection
 115          */
 116         if (req->flags & REQ_F_BUFFER_SELECT)
 117                 return io_iov_buffer_select_prep(req);
 118
 119         return 0;
 120 }
 121
 122 int io_prep_rw_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 123 {
 124         struct io_ring_ctx *ctx = req->ctx;
 125         u16 index;
 126         int ret;
 127
 128         ret = io_prep_rw(req, sqe);
 129         if (unlikely(ret))
 130                 return ret;
 131
 132         if (unlikely(req->buf_index >= ctx->nr_user_bufs))
 133                 return -EFAULT;
 134         index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
 135         req->imu = ctx->user_bufs[index];
 136         io_req_set_rsrc_node(req, ctx, 0);
 137         return 0;
 138 }
 139
 140 /*
 141  * Multishot read is prepared just like a normal read/write request, only
 142  * difference is that we set the MULTISHOT flag.
 143  */
 144 int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 145 {
 146         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 147         int ret;
 148
 149         /* must be used with provided buffers */
 150         if (!(req->flags & REQ_F_BUFFER_SELECT))
 151                 return -EINVAL;
 152
 153         ret = io_prep_rw(req, sqe);
 154         if (unlikely(ret))
 155                 return ret;
 156
 157         if (rw->addr || rw->len)
 158                 return -EINVAL;
 159
 160         req->flags |= REQ_F_APOLL_MULTISHOT;
 161         return 0;
 162 }
 163
 164 void io_readv_writev_cleanup(struct io_kiocb *req)
 165 {
 166         struct io_async_rw *io = req->async_data;
 167
 168         kfree(io->free_iovec);
 169 }
 170
 171 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
 172 {
 173         switch (ret) {
 174         case -EIOCBQUEUED:
 175                 break;
 176         case -ERESTARTSYS:
 177         case -ERESTARTNOINTR:
 178         case -ERESTARTNOHAND:
 179         case -ERESTART_RESTARTBLOCK:
 180                 /*
 181                  * We can't just restart the syscall, since previously
 182                  * submitted sqes may already be in progress. Just fail this
 183                  * IO with EINTR.
 184                  */
 185                 ret = -EINTR;
 186                 fallthrough;
 187         default:
 188                 kiocb->ki_complete(kiocb, ret);
 189         }
 190 }
 191
 192 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
 193 {
 194         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 195
 196         if (rw->kiocb.ki_pos != -1)
 197                 return &rw->kiocb.ki_pos;
 198
 199         if (!(req->file->f_mode & FMODE_STREAM)) {
 200                 req->flags |= REQ_F_CUR_POS;
 201                 rw->kiocb.ki_pos = req->file->f_pos;
 202                 return &rw->kiocb.ki_pos;
 203         }
 204
 205         rw->kiocb.ki_pos = 0;
 206         return NULL;
 207 }
 208
 209 static void io_req_task_queue_reissue(struct io_kiocb *req)
 210 {
 211         req->io_task_work.func = io_queue_iowq;
 212         io_req_task_work_add(req);
 213 }
 214
 215 #ifdef CONFIG_BLOCK
 216 static bool io_resubmit_prep(struct io_kiocb *req)
 217 {
 218         struct io_async_rw *io = req->async_data;
 219
 220         if (!req_has_async_data(req))
 221                 return !io_req_prep_async(req);
 222         iov_iter_restore(&io->s.iter, &io->s.iter_state);
 223         return true;
 224 }
 225
 226 static bool io_rw_should_reissue(struct io_kiocb *req)
 227 {
 228         umode_t mode = file_inode(req->file)->i_mode;
 229         struct io_ring_ctx *ctx = req->ctx;
 230
 231         if (!S_ISBLK(mode) && !S_ISREG(mode))
 232                 return false;
 233         if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
 234             !(ctx->flags & IORING_SETUP_IOPOLL)))
 235                 return false;
 236         /*
 237          * If ref is dying, we might be running poll reap from the exit work.
 238          * Don't attempt to reissue from that path, just let it fail with
 239          * -EAGAIN.
 240          */
 241         if (percpu_ref_is_dying(&ctx->refs))
 242                 return false;
 243         /*
 244          * Play it safe and assume not safe to re-import and reissue if we're
 245          * not in the original thread group (or in task context).
 246          */
 247         if (!same_thread_group(req->task, current) || !in_task())
 248                 return false;
 249         return true;
 250 }
 251 #else
 252 static bool io_resubmit_prep(struct io_kiocb *req)
 253 {
 254         return false;
 255 }
 256 static bool io_rw_should_reissue(struct io_kiocb *req)
 257 {
 258         return false;
 259 }
 260 #endif
 261
 262 static void io_req_end_write(struct io_kiocb *req)
 263 {
 264         if (req->flags & REQ_F_ISREG) {
 265                 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 266
 267                 kiocb_end_write(&rw->kiocb);
 268         }
 269 }
 270
 271 /*
 272  * Trigger the notifications after having done some IO, and finish the write
 273  * accounting, if any.
 274  */
 275 static void io_req_io_end(struct io_kiocb *req)
 276 {
 277         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 278
 279         if (rw->kiocb.ki_flags & IOCB_WRITE) {
 280                 io_req_end_write(req);
 281                 fsnotify_modify(req->file);
 282         } else {
 283                 fsnotify_access(req->file);
 284         }
 285 }
 286
 287 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
 288 {
 289         if (unlikely(res != req->cqe.res)) {
 290                 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
 291                     io_rw_should_reissue(req)) {
 292                         /*
 293                          * Reissue will start accounting again, finish the
 294                          * current cycle.
 295                          */
 296                         io_req_io_end(req);
 297                         req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
 298                         return true;
 299                 }
 300                 req_set_fail(req);
 301                 req->cqe.res = res;
 302         }
 303         return false;
 304 }
 305
 306 static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
 307 {
 308         struct io_async_rw *io = req->async_data;
 309
 310         /* add previously done IO, if any */
 311         if (req_has_async_data(req) && io->bytes_done > 0) {
 312                 if (res < 0)
 313                         res = io->bytes_done;
 314                 else
 315                         res += io->bytes_done;
 316         }
 317         return res;
 318 }
 319
 320 void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts)
 321 {
 322         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 323         struct kiocb *kiocb = &rw->kiocb;
 324
 325         if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
 326                 long res = kiocb->dio_complete(rw->kiocb.private);
 327
 328                 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
 329         }
 330
 331         io_req_io_end(req);
 332
 333         if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
 334                 unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
 335
 336                 req->cqe.flags |= io_put_kbuf(req, issue_flags);
 337         }
 338         io_req_task_complete(req, ts);
 339 }
 340
 341 static void io_complete_rw(struct kiocb *kiocb, long res)
 342 {
 343         struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
 344         struct io_kiocb *req = cmd_to_io_kiocb(rw);
 345
 346         if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
 347                 if (__io_complete_rw_common(req, res))
 348                         return;
 349                 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
 350         }
 351         req->io_task_work.func = io_req_rw_complete;
 352         __io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
 353 }
 354
 355 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
 356 {
 357         struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
 358         struct io_kiocb *req = cmd_to_io_kiocb(rw);
 359
 360         if (kiocb->ki_flags & IOCB_WRITE)
 361                 io_req_end_write(req);
 362         if (unlikely(res != req->cqe.res)) {
 363                 if (res == -EAGAIN && io_rw_should_reissue(req)) {
 364                         req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
 365                         return;
 366                 }
 367                 req->cqe.res = res;
 368         }
 369
 370         /* order with io_iopoll_complete() checking ->iopoll_completed */
 371         smp_store_release(&req->iopoll_completed, 1);
 372 }
 373
 374 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
 375                        unsigned int issue_flags)
 376 {
 377         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 378         unsigned final_ret = io_fixup_rw_res(req, ret);
 379
 380         if (ret >= 0 && req->flags & REQ_F_CUR_POS)
 381                 req->file->f_pos = rw->kiocb.ki_pos;
 382         if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
 383                 if (!__io_complete_rw_common(req, ret)) {
 384                         /*
 385                          * Safe to call io_end from here as we're inline
 386                          * from the submission path.
 387                          */
 388                         io_req_io_end(req);
 389                         io_req_set_res(req, final_ret,
 390                                        io_put_kbuf(req, issue_flags));
 391                         return IOU_OK;
 392                 }
 393         } else {
 394                 io_rw_done(&rw->kiocb, ret);
 395         }
 396
 397         if (req->flags & REQ_F_REISSUE) {
 398                 req->flags &= ~REQ_F_REISSUE;
 399                 if (io_resubmit_prep(req))
 400                         io_req_task_queue_reissue(req);
 401                 else
 402                         io_req_task_queue_fail(req, final_ret);
 403         }
 404         return IOU_ISSUE_SKIP_COMPLETE;
 405 }
 406
 407 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
 408                                        struct io_rw_state *s,
 409                                        unsigned int issue_flags)
 410 {
 411         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 412         struct iov_iter *iter = &s->iter;
 413         u8 opcode = req->opcode;
 414         struct iovec *iovec;
 415         void __user *buf;
 416         size_t sqe_len;
 417         ssize_t ret;
 418
 419         if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
 420                 ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
 421                 if (ret)
 422                         return ERR_PTR(ret);
 423                 return NULL;
 424         }
 425
 426         buf = u64_to_user_ptr(rw->addr);
 427         sqe_len = rw->len;
 428
 429         if (!io_issue_defs[opcode].vectored || req->flags & REQ_F_BUFFER_SELECT) {
 430                 if (io_do_buffer_select(req)) {
 431                         buf = io_buffer_select(req, &sqe_len, issue_flags);
 432                         if (!buf)
 433                                 return ERR_PTR(-ENOBUFS);
 434                         rw->addr = (unsigned long) buf;
 435                         rw->len = sqe_len;
 436                 }
 437
 438                 ret = import_ubuf(ddir, buf, sqe_len, iter);
 439                 if (ret)
 440                         return ERR_PTR(ret);
 441                 return NULL;
 442         }
 443
 444         iovec = s->fast_iov;
 445         ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
 446                               req->ctx->compat);
 447         if (unlikely(ret < 0))
 448                 return ERR_PTR(ret);
 449         return iovec;
 450 }
 451
 452 static inline int io_import_iovec(int rw, struct io_kiocb *req,
 453                                   struct iovec **iovec, struct io_rw_state *s,
 454                                   unsigned int issue_flags)
 455 {
 456         *iovec = __io_import_iovec(rw, req, s, issue_flags);
 457         if (IS_ERR(*iovec))
 458                 return PTR_ERR(*iovec);
 459
 460         iov_iter_save_state(&s->iter, &s->iter_state);
 461         return 0;
 462 }
 463
 464 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
 465 {
 466         return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
 467 }
 468
 469 /*
 470  * For files that don't have ->read_iter() and ->write_iter(), handle them
 471  * by looping over ->read() or ->write() manually.
 472  */
 473 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
 474 {
 475         struct kiocb *kiocb = &rw->kiocb;
 476         struct file *file = kiocb->ki_filp;
 477         ssize_t ret = 0;
 478         loff_t *ppos;
 479
 480         /*
 481          * Don't support polled IO through this interface, and we can't
 482          * support non-blocking either. For the latter, this just causes
 483          * the kiocb to be handled from an async context.
 484          */
 485         if (kiocb->ki_flags & IOCB_HIPRI)
 486                 return -EOPNOTSUPP;
 487         if ((kiocb->ki_flags & IOCB_NOWAIT) &&
 488             !(kiocb->ki_filp->f_flags & O_NONBLOCK))
 489                 return -EAGAIN;
 490
 491         ppos = io_kiocb_ppos(kiocb);
 492
 493         while (iov_iter_count(iter)) {
 494                 void __user *addr;
 495                 size_t len;
 496                 ssize_t nr;
 497
 498                 if (iter_is_ubuf(iter)) {
 499                         addr = iter->ubuf + iter->iov_offset;
 500                         len = iov_iter_count(iter);
 501                 } else if (!iov_iter_is_bvec(iter)) {
 502                         addr = iter_iov_addr(iter);
 503                         len = iter_iov_len(iter);
 504                 } else {
 505                         addr = u64_to_user_ptr(rw->addr);
 506                         len = rw->len;
 507                 }
 508
 509                 if (ddir == READ)
 510                         nr = file->f_op->read(file, addr, len, ppos);
 511                 else
 512                         nr = file->f_op->write(file, addr, len, ppos);
 513
 514                 if (nr < 0) {
 515                         if (!ret)
 516                                 ret = nr;
 517                         break;
 518                 }
 519                 ret += nr;
 520                 if (!iov_iter_is_bvec(iter)) {
 521                         iov_iter_advance(iter, nr);
 522                 } else {
 523                         rw->addr += nr;
 524                         rw->len -= nr;
 525                         if (!rw->len)
 526                                 break;
 527                 }
 528                 if (nr != len)
 529                         break;
 530         }
 531
 532         return ret;
 533 }
 534
 535 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
 536                           const struct iovec *fast_iov, struct iov_iter *iter)
 537 {
 538         struct io_async_rw *io = req->async_data;
 539
 540         memcpy(&io->s.iter, iter, sizeof(*iter));
 541         io->free_iovec = iovec;
 542         io->bytes_done = 0;
 543         /* can only be fixed buffers, no need to do anything */
 544         if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter))
 545                 return;
 546         if (!iovec) {
 547                 unsigned iov_off = 0;
 548
 549                 io->s.iter.__iov = io->s.fast_iov;
 550                 if (iter->__iov != fast_iov) {
 551                         iov_off = iter_iov(iter) - fast_iov;
 552                         io->s.iter.__iov += iov_off;
 553                 }
 554                 if (io->s.fast_iov != fast_iov)
 555                         memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
 556                                sizeof(struct iovec) * iter->nr_segs);
 557         } else {
 558                 req->flags |= REQ_F_NEED_CLEANUP;
 559         }
 560 }
 561
 562 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
 563                              struct io_rw_state *s, bool force)
 564 {
 565         if (!force && !io_cold_defs[req->opcode].prep_async)
 566                 return 0;
 567         /* opcode type doesn't need async data */
 568         if (!io_cold_defs[req->opcode].async_size)
 569                 return 0;
 570         if (!req_has_async_data(req)) {
 571                 struct io_async_rw *iorw;
 572
 573                 if (io_alloc_async_data(req)) {
 574                         kfree(iovec);
 575                         return -ENOMEM;
 576                 }
 577
 578                 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
 579                 iorw = req->async_data;
 580                 /* we've copied and mapped the iter, ensure state is saved */
 581                 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
 582         }
 583         return 0;
 584 }
 585
 586 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
 587 {
 588         struct io_async_rw *iorw = req->async_data;
 589         struct iovec *iov;
 590         int ret;
 591
 592         /* submission path, ->uring_lock should already be taken */
 593         ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
 594         if (unlikely(ret < 0))
 595                 return ret;
 596
 597         iorw->bytes_done = 0;
 598         iorw->free_iovec = iov;
 599         if (iov)
 600                 req->flags |= REQ_F_NEED_CLEANUP;
 601         return 0;
 602 }
 603
 604 int io_readv_prep_async(struct io_kiocb *req)
 605 {
 606         return io_rw_prep_async(req, ITER_DEST);
 607 }
 608
 609 int io_writev_prep_async(struct io_kiocb *req)
 610 {
 611         return io_rw_prep_async(req, ITER_SOURCE);
 612 }
 613
 614 /*
 615  * This is our waitqueue callback handler, registered through __folio_lock_async()
 616  * when we initially tried to do the IO with the iocb armed our waitqueue.
 617  * This gets called when the page is unlocked, and we generally expect that to
 618  * happen when the page IO is completed and the page is now uptodate. This will
 619  * queue a task_work based retry of the operation, attempting to copy the data
 620  * again. If the latter fails because the page was NOT uptodate, then we will
 621  * do a thread based blocking retry of the operation. That's the unexpected
 622  * slow path.
 623  */
 624 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
 625                              int sync, void *arg)
 626 {
 627         struct wait_page_queue *wpq;
 628         struct io_kiocb *req = wait->private;
 629         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 630         struct wait_page_key *key = arg;
 631
 632         wpq = container_of(wait, struct wait_page_queue, wait);
 633
 634         if (!wake_page_match(wpq, key))
 635                 return 0;
 636
 637         rw->kiocb.ki_flags &= ~IOCB_WAITQ;
 638         list_del_init(&wait->entry);
 639         io_req_task_queue(req);
 640         return 1;
 641 }
 642
 643 /*
 644  * This controls whether a given IO request should be armed for async page
 645  * based retry. If we return false here, the request is handed to the async
 646  * worker threads for retry. If we're doing buffered reads on a regular file,
 647  * we prepare a private wait_page_queue entry and retry the operation. This
 648  * will either succeed because the page is now uptodate and unlocked, or it
 649  * will register a callback when the page is unlocked at IO completion. Through
 650  * that callback, io_uring uses task_work to setup a retry of the operation.
 651  * That retry will attempt the buffered read again. The retry will generally
 652  * succeed, or in rare cases where it fails, we then fall back to using the
 653  * async worker threads for a blocking retry.
 654  */
 655 static bool io_rw_should_retry(struct io_kiocb *req)
 656 {
 657         struct io_async_rw *io = req->async_data;
 658         struct wait_page_queue *wait = &io->wpq;
 659         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 660         struct kiocb *kiocb = &rw->kiocb;
 661
 662         /* never retry for NOWAIT, we just complete with -EAGAIN */
 663         if (req->flags & REQ_F_NOWAIT)
 664                 return false;
 665
 666         /* Only for buffered IO */
 667         if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
 668                 return false;
 669
 670         /*
 671          * just use poll if we can, and don't attempt if the fs doesn't
 672          * support callback based unlocks
 673          */
 674         if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
 675                 return false;
 676
 677         wait->wait.func = io_async_buf_func;
 678         wait->wait.private = req;
 679         wait->wait.flags = 0;
 680         INIT_LIST_HEAD(&wait->wait.entry);
 681         kiocb->ki_flags |= IOCB_WAITQ;
 682         kiocb->ki_flags &= ~IOCB_NOWAIT;
 683         kiocb->ki_waitq = wait;
 684         return true;
 685 }
 686
 687 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
 688 {
 689         struct file *file = rw->kiocb.ki_filp;
 690
 691         if (likely(file->f_op->read_iter))
 692                 return call_read_iter(file, &rw->kiocb, iter);
 693         else if (file->f_op->read)
 694                 return loop_rw_iter(READ, rw, iter);
 695         else
 696                 return -EINVAL;
 697 }
 698
 699 static bool need_complete_io(struct io_kiocb *req)
 700 {
 701         return req->flags & REQ_F_ISREG ||
 702                 S_ISBLK(file_inode(req->file)->i_mode);
 703 }
 704
 705 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
 706 {
 707         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 708         struct kiocb *kiocb = &rw->kiocb;
 709         struct io_ring_ctx *ctx = req->ctx;
 710         struct file *file = req->file;
 711         int ret;
 712
 713         if (unlikely(!file || !(file->f_mode & mode)))
 714                 return -EBADF;
 715
 716         if (!(req->flags & REQ_F_FIXED_FILE))
 717                 req->flags |= io_file_get_flags(file);
 718
 719         kiocb->ki_flags = file->f_iocb_flags;
 720         ret = kiocb_set_rw_flags(kiocb, rw->flags);
 721         if (unlikely(ret))
 722                 return ret;
 723         kiocb->ki_flags |= IOCB_ALLOC_CACHE;
 724
 725         /*
 726          * If the file is marked O_NONBLOCK, still allow retry for it if it
 727          * supports async. Otherwise it's impossible to use O_NONBLOCK files
 728          * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
 729          */
 730         if ((kiocb->ki_flags & IOCB_NOWAIT) ||
 731             ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
 732                 req->flags |= REQ_F_NOWAIT;
 733
 734         if (ctx->flags & IORING_SETUP_IOPOLL) {
 735                 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
 736                         return -EOPNOTSUPP;
 737
 738                 kiocb->private = NULL;
 739                 kiocb->ki_flags |= IOCB_HIPRI;
 740                 kiocb->ki_complete = io_complete_rw_iopoll;
 741                 req->iopoll_completed = 0;
 742         } else {
 743                 if (kiocb->ki_flags & IOCB_HIPRI)
 744                         return -EINVAL;
 745                 kiocb->ki_complete = io_complete_rw;
 746         }
 747
 748         return 0;
 749 }
 750
 751 static int __io_read(struct io_kiocb *req, unsigned int issue_flags)
 752 {
 753         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 754         struct io_rw_state __s, *s = &__s;
 755         struct iovec *iovec;
 756         struct kiocb *kiocb = &rw->kiocb;
 757         bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
 758         struct io_async_rw *io;
 759         ssize_t ret, ret2;
 760         loff_t *ppos;
 761
 762         if (!req_has_async_data(req)) {
 763                 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
 764                 if (unlikely(ret < 0))
 765                         return ret;
 766         } else {
 767                 io = req->async_data;
 768                 s = &io->s;
 769
 770                 /*
 771                  * Safe and required to re-import if we're using provided
 772                  * buffers, as we dropped the selected one before retry.
 773                  */
 774                 if (io_do_buffer_select(req)) {
 775                         ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
 776                         if (unlikely(ret < 0))
 777                                 return ret;
 778                 }
 779
 780                 /*
 781                  * We come here from an earlier attempt, restore our state to
 782                  * match in case it doesn't. It's cheap enough that we don't
 783                  * need to make this conditional.
 784                  */
 785                 iov_iter_restore(&s->iter, &s->iter_state);
 786                 iovec = NULL;
 787         }
 788         ret = io_rw_init_file(req, FMODE_READ);
 789         if (unlikely(ret)) {
 790                 kfree(iovec);
 791                 return ret;
 792         }
 793         req->cqe.res = iov_iter_count(&s->iter);
 794
 795         if (force_nonblock) {
 796                 /* If the file doesn't support async, just async punt */
 797                 if (unlikely(!io_file_supports_nowait(req))) {
 798                         ret = io_setup_async_rw(req, iovec, s, true);
 799                         return ret ?: -EAGAIN;
 800                 }
 801                 kiocb->ki_flags |= IOCB_NOWAIT;
 802         } else {
 803                 /* Ensure we clear previously set non-block flag */
 804                 kiocb->ki_flags &= ~IOCB_NOWAIT;
 805         }
 806
 807         ppos = io_kiocb_update_pos(req);
 808
 809         ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
 810         if (unlikely(ret)) {
 811                 kfree(iovec);
 812                 return ret;
 813         }
 814
 815         ret = io_iter_do_read(rw, &s->iter);
 816
 817         if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
 818                 req->flags &= ~REQ_F_REISSUE;
 819                 /*
 820                  * If we can poll, just do that. For a vectored read, we'll
 821                  * need to copy state first.
 822                  */
 823                 if (file_can_poll(req->file) && !io_issue_defs[req->opcode].vectored)
 824                         return -EAGAIN;
 825                 /* IOPOLL retry should happen for io-wq threads */
 826                 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
 827                         goto done;
 828                 /* no retry on NONBLOCK nor RWF_NOWAIT */
 829                 if (req->flags & REQ_F_NOWAIT)
 830                         goto done;
 831                 ret = 0;
 832         } else if (ret == -EIOCBQUEUED) {
 833                 if (iovec)
 834                         kfree(iovec);
 835                 return IOU_ISSUE_SKIP_COMPLETE;
 836         } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
 837                    (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
 838                 /* read all, failed, already did sync or don't want to retry */
 839                 goto done;
 840         }
 841
 842         /*
 843          * Don't depend on the iter state matching what was consumed, or being
 844          * untouched in case of error. Restore it and we'll advance it
 845          * manually if we need to.
 846          */
 847         iov_iter_restore(&s->iter, &s->iter_state);
 848
 849         ret2 = io_setup_async_rw(req, iovec, s, true);
 850         iovec = NULL;
 851         if (ret2) {
 852                 ret = ret > 0 ? ret : ret2;
 853                 goto done;
 854         }
 855
 856         io = req->async_data;
 857         s = &io->s;
 858         /*
 859          * Now use our persistent iterator and state, if we aren't already.
 860          * We've restored and mapped the iter to match.
 861          */
 862
 863         do {
 864                 /*
 865                  * We end up here because of a partial read, either from
 866                  * above or inside this loop. Advance the iter by the bytes
 867                  * that were consumed.
 868                  */
 869                 iov_iter_advance(&s->iter, ret);
 870                 if (!iov_iter_count(&s->iter))
 871                         break;
 872                 io->bytes_done += ret;
 873                 iov_iter_save_state(&s->iter, &s->iter_state);
 874
 875                 /* if we can retry, do so with the callbacks armed */
 876                 if (!io_rw_should_retry(req)) {
 877                         kiocb->ki_flags &= ~IOCB_WAITQ;
 878                         return -EAGAIN;
 879                 }
 880
 881                 req->cqe.res = iov_iter_count(&s->iter);
 882                 /*
 883                  * Now retry read with the IOCB_WAITQ parts set in the iocb. If
 884                  * we get -EIOCBQUEUED, then we'll get a notification when the
 885                  * desired page gets unlocked. We can also get a partial read
 886                  * here, and if we do, then just retry at the new offset.
 887                  */
 888                 ret = io_iter_do_read(rw, &s->iter);
 889                 if (ret == -EIOCBQUEUED)
 890                         return IOU_ISSUE_SKIP_COMPLETE;
 891                 /* we got some bytes, but not all. retry. */
 892                 kiocb->ki_flags &= ~IOCB_WAITQ;
 893                 iov_iter_restore(&s->iter, &s->iter_state);
 894         } while (ret > 0);
 895 done:
 896         /* it's faster to check here then delegate to kfree */
 897         if (iovec)
 898                 kfree(iovec);
 899         return ret;
 900 }
 901
 902 int io_read(struct io_kiocb *req, unsigned int issue_flags)
 903 {
 904         int ret;
 905
 906         ret = __io_read(req, issue_flags);
 907         if (ret >= 0)
 908                 return kiocb_done(req, ret, issue_flags);
 909
 910         return ret;
 911 }
 912
 913 int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
 914 {
 915         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 916         unsigned int cflags = 0;
 917         int ret;
 918
 919         /*
 920          * Multishot MUST be used on a pollable file
 921          */
 922         if (!file_can_poll(req->file))
 923                 return -EBADFD;
 924
 925         ret = __io_read(req, issue_flags);
 926
 927         /*
 928          * If we get -EAGAIN, recycle our buffer and just let normal poll
 929          * handling arm it.
 930          */
 931         if (ret == -EAGAIN) {
 932                 /*
 933                  * Reset rw->len to 0 again to avoid clamping future mshot
 934                  * reads, in case the buffer size varies.
 935                  */
 936                 if (io_kbuf_recycle(req, issue_flags))
 937                         rw->len = 0;
 938                 return -EAGAIN;
 939         }
 940
 941         /*
 942          * Any successful return value will keep the multishot read armed.
 943          */
 944         if (ret > 0) {
 945                 /*
 946                  * Put our buffer and post a CQE. If we fail to post a CQE, then
 947                  * jump to the termination path. This request is then done.
 948                  */
 949                 cflags = io_put_kbuf(req, issue_flags);
 950                 rw->len = 0; /* similarly to above, reset len to 0 */
 951
 952                 if (io_fill_cqe_req_aux(req,
 953                                         issue_flags & IO_URING_F_COMPLETE_DEFER,
 954                                         ret, cflags | IORING_CQE_F_MORE)) {
 955                         if (issue_flags & IO_URING_F_MULTISHOT)
 956                                 return IOU_ISSUE_SKIP_COMPLETE;
 957                         return -EAGAIN;
 958                 }
 959         }
 960
 961         /*
 962          * Either an error, or we've hit overflow posting the CQE. For any
 963          * multishot request, hitting overflow will terminate it.
 964          */
 965         io_req_set_res(req, ret, cflags);
 966         if (issue_flags & IO_URING_F_MULTISHOT)
 967                 return IOU_STOP_MULTISHOT;
 968         return IOU_OK;
 969 }
 970
 971 int io_write(struct io_kiocb *req, unsigned int issue_flags)
 972 {
 973         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
 974         struct io_rw_state __s, *s = &__s;
 975         struct iovec *iovec;
 976         struct kiocb *kiocb = &rw->kiocb;
 977         bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
 978         ssize_t ret, ret2;
 979         loff_t *ppos;
 980
 981         if (!req_has_async_data(req)) {
 982                 ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
 983                 if (unlikely(ret < 0))
 984                         return ret;
 985         } else {
 986                 struct io_async_rw *io = req->async_data;
 987
 988                 s = &io->s;
 989                 iov_iter_restore(&s->iter, &s->iter_state);
 990                 iovec = NULL;
 991         }
 992         ret = io_rw_init_file(req, FMODE_WRITE);
 993         if (unlikely(ret)) {
 994                 kfree(iovec);
 995                 return ret;
 996         }
 997         req->cqe.res = iov_iter_count(&s->iter);
 998
 999         if (force_nonblock) {
1000                 /* If the file doesn't support async, just async punt */
1001                 if (unlikely(!io_file_supports_nowait(req)))
1002                         goto copy_iov;
1003
1004                 /* File path supports NOWAIT for non-direct_IO only for block devices. */
1005                 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
1006                         !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
1007                         (req->flags & REQ_F_ISREG))
1008                         goto copy_iov;
1009
1010                 kiocb->ki_flags |= IOCB_NOWAIT;
1011         } else {
1012                 /* Ensure we clear previously set non-block flag */
1013                 kiocb->ki_flags &= ~IOCB_NOWAIT;
1014         }
1015
1016         ppos = io_kiocb_update_pos(req);
1017
1018         ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
1019         if (unlikely(ret)) {
1020                 kfree(iovec);
1021                 return ret;
1022         }
1023
1024         if (req->flags & REQ_F_ISREG)
1025                 kiocb_start_write(kiocb);
1026         kiocb->ki_flags |= IOCB_WRITE;
1027
1028         if (likely(req->file->f_op->write_iter))
1029                 ret2 = call_write_iter(req->file, kiocb, &s->iter);
1030         else if (req->file->f_op->write)
1031                 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
1032         else
1033                 ret2 = -EINVAL;
1034
1035         if (req->flags & REQ_F_REISSUE) {
1036                 req->flags &= ~REQ_F_REISSUE;
1037                 ret2 = -EAGAIN;
1038         }
1039
1040         /*
1041          * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
1042          * retry them without IOCB_NOWAIT.
1043          */
1044         if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
1045                 ret2 = -EAGAIN;
1046         /* no retry on NONBLOCK nor RWF_NOWAIT */
1047         if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
1048                 goto done;
1049         if (!force_nonblock || ret2 != -EAGAIN) {
1050                 /* IOPOLL retry should happen for io-wq threads */
1051                 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1052                         goto copy_iov;
1053
1054                 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
1055                         struct io_async_rw *io;
1056
1057                         trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
1058                                                 req->cqe.res, ret2);
1059
1060                         /* This is a partial write. The file pos has already been
1061                          * updated, setup the async struct to complete the request
1062                          * in the worker. Also update bytes_done to account for
1063                          * the bytes already written.
1064                          */
1065                         iov_iter_save_state(&s->iter, &s->iter_state);
1066                         ret = io_setup_async_rw(req, iovec, s, true);
1067
1068                         io = req->async_data;
1069                         if (io)
1070                                 io->bytes_done += ret2;
1071
1072                         if (kiocb->ki_flags & IOCB_WRITE)
1073                                 io_req_end_write(req);
1074                         return ret ? ret : -EAGAIN;
1075                 }
1076 done:
1077                 ret = kiocb_done(req, ret2, issue_flags);
1078         } else {
1079 copy_iov:
1080                 iov_iter_restore(&s->iter, &s->iter_state);
1081                 ret = io_setup_async_rw(req, iovec, s, false);
1082                 if (!ret) {
1083                         if (kiocb->ki_flags & IOCB_WRITE)
1084                                 io_req_end_write(req);
1085                         return -EAGAIN;
1086                 }
1087                 return ret;
1088         }
1089         /* it's reportedly faster than delegating the null check to kfree() */
1090         if (iovec)
1091                 kfree(iovec);
1092         return ret;
1093 }
1094
1095 void io_rw_fail(struct io_kiocb *req)
1096 {
1097         int res;
1098
1099         res = io_fixup_rw_res(req, req->cqe.res);
1100         io_req_set_res(req, res, req->cqe.flags);
1101 }
1102
1103 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1104 {
1105         struct io_wq_work_node *pos, *start, *prev;
1106         unsigned int poll_flags = 0;
1107         DEFINE_IO_COMP_BATCH(iob);
1108         int nr_events = 0;
1109
1110         /*
1111          * Only spin for completions if we don't have multiple devices hanging
1112          * off our complete list.
1113          */
1114         if (ctx->poll_multi_queue || force_nonspin)
1115                 poll_flags |= BLK_POLL_ONESHOT;
1116
1117         wq_list_for_each(pos, start, &ctx->iopoll_list) {
1118                 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1119                 struct file *file = req->file;
1120                 int ret;
1121
1122                 /*
1123                  * Move completed and retryable entries to our local lists.
1124                  * If we find a request that requires polling, break out
1125                  * and complete those lists first, if we have entries there.
1126                  */
1127                 if (READ_ONCE(req->iopoll_completed))
1128                         break;
1129
1130                 if (req->opcode == IORING_OP_URING_CMD) {
1131                         struct io_uring_cmd *ioucmd;
1132
1133                         ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1134                         ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1135                                                                 poll_flags);
1136                 } else {
1137                         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1138
1139                         ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1140                 }
1141                 if (unlikely(ret < 0))
1142                         return ret;
1143                 else if (ret)
1144                         poll_flags |= BLK_POLL_ONESHOT;
1145
1146                 /* iopoll may have completed current req */
1147                 if (!rq_list_empty(iob.req_list) ||
1148                     READ_ONCE(req->iopoll_completed))
1149                         break;
1150         }
1151
1152         if (!rq_list_empty(iob.req_list))
1153                 iob.complete(&iob);
1154         else if (!pos)
1155                 return 0;
1156
1157         prev = start;
1158         wq_list_for_each_resume(pos, prev) {
1159                 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1160
1161                 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1162                 if (!smp_load_acquire(&req->iopoll_completed))
1163                         break;
1164                 nr_events++;
1165                 req->cqe.flags = io_put_kbuf(req, 0);
1166         }
1167         if (unlikely(!nr_events))
1168                 return 0;
1169
1170         pos = start ? start->next : ctx->iopoll_list.first;
1171         wq_list_cut(&ctx->iopoll_list, prev, start);
1172
1173         if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1174                 return 0;
1175         ctx->submit_state.compl_reqs.first = pos;
1176         __io_submit_flush_completions(ctx);
1177         return nr_events;
1178 }