io_uring/rsrc.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/mm.h>
   7 #include <linux/slab.h>
   8 #include <linux/nospec.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/compat.h>
  11 #include <linux/io_uring.h>
  12
  13 #include <uapi/linux/io_uring.h>
  14
  15 #include "io_uring.h"
  16 #include "openclose.h"
  17 #include "rsrc.h"
  18
  19 struct io_rsrc_update {
  20         struct file                     *file;
  21         u64                             arg;
  22         u32                             nr_args;
  23         u32                             offset;
  24 };
  25
  26 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
  27 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
  28 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
  29                                   struct io_mapped_ubuf **pimu,
  30                                   struct page **last_hpage);
  31
  32 /* only define max */
  33 #define IORING_MAX_FIXED_FILES  (1U << 20)
  34 #define IORING_MAX_REG_BUFFERS  (1U << 14)
  35
  36 int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
  37 {
  38         unsigned long page_limit, cur_pages, new_pages;
  39
  40         if (!nr_pages)
  41                 return 0;
  42
  43         /* Don't allow more pages than we can safely lock */
  44         page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
  45
  46         cur_pages = atomic_long_read(&user->locked_vm);
  47         do {
  48                 new_pages = cur_pages + nr_pages;
  49                 if (new_pages > page_limit)
  50                         return -ENOMEM;
  51         } while (!atomic_long_try_cmpxchg(&user->locked_vm,
  52                                           &cur_pages, new_pages));
  53         return 0;
  54 }
  55
  56 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
  57 {
  58         if (ctx->user)
  59                 __io_unaccount_mem(ctx->user, nr_pages);
  60
  61         if (ctx->mm_account)
  62                 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
  63 }
  64
  65 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
  66 {
  67         int ret;
  68
  69         if (ctx->user) {
  70                 ret = __io_account_mem(ctx->user, nr_pages);
  71                 if (ret)
  72                         return ret;
  73         }
  74
  75         if (ctx->mm_account)
  76                 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
  77
  78         return 0;
  79 }
  80
  81 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
  82                        void __user *arg, unsigned index)
  83 {
  84         struct iovec __user *src;
  85
  86 #ifdef CONFIG_COMPAT
  87         if (ctx->compat) {
  88                 struct compat_iovec __user *ciovs;
  89                 struct compat_iovec ciov;
  90
  91                 ciovs = (struct compat_iovec __user *) arg;
  92                 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
  93                         return -EFAULT;
  94
  95                 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
  96                 dst->iov_len = ciov.iov_len;
  97                 return 0;
  98         }
  99 #endif
 100         src = (struct iovec __user *) arg;
 101         if (copy_from_user(dst, &src[index], sizeof(*dst)))
 102                 return -EFAULT;
 103         return 0;
 104 }
 105
 106 static int io_buffer_validate(struct iovec *iov)
 107 {
 108         unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
 109
 110         /*
 111          * Don't impose further limits on the size and buffer
 112          * constraints here, we'll -EINVAL later when IO is
 113          * submitted if they are wrong.
 114          */
 115         if (!iov->iov_base)
 116                 return iov->iov_len ? -EFAULT : 0;
 117         if (!iov->iov_len)
 118                 return -EFAULT;
 119
 120         /* arbitrary limit, but we need something */
 121         if (iov->iov_len > SZ_1G)
 122                 return -EFAULT;
 123
 124         if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
 125                 return -EOVERFLOW;
 126
 127         return 0;
 128 }
 129
 130 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
 131 {
 132         struct io_mapped_ubuf *imu = *slot;
 133         unsigned int i;
 134
 135         if (imu != ctx->dummy_ubuf) {
 136                 for (i = 0; i < imu->nr_bvecs; i++)
 137                         unpin_user_page(imu->bvec[i].bv_page);
 138                 if (imu->acct_pages)
 139                         io_unaccount_mem(ctx, imu->acct_pages);
 140                 kvfree(imu);
 141         }
 142         *slot = NULL;
 143 }
 144
 145 static void io_rsrc_put_work(struct io_rsrc_node *node)
 146 {
 147         struct io_rsrc_put *prsrc = &node->item;
 148
 149         if (prsrc->tag)
 150                 io_post_aux_cqe(node->ctx, prsrc->tag, 0, 0);
 151
 152         switch (node->type) {
 153         case IORING_RSRC_FILE:
 154                 io_rsrc_file_put(node->ctx, prsrc);
 155                 break;
 156         case IORING_RSRC_BUFFER:
 157                 io_rsrc_buf_put(node->ctx, prsrc);
 158                 break;
 159         default:
 160                 WARN_ON_ONCE(1);
 161                 break;
 162         }
 163 }
 164
 165 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node)
 166 {
 167         if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache))
 168                 kfree(node);
 169 }
 170
 171 void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
 172         __must_hold(&node->ctx->uring_lock)
 173 {
 174         struct io_ring_ctx *ctx = node->ctx;
 175
 176         while (!list_empty(&ctx->rsrc_ref_list)) {
 177                 node = list_first_entry(&ctx->rsrc_ref_list,
 178                                             struct io_rsrc_node, node);
 179                 /* recycle ref nodes in order */
 180                 if (node->refs)
 181                         break;
 182                 list_del(&node->node);
 183
 184                 if (likely(!node->empty))
 185                         io_rsrc_put_work(node);
 186                 io_rsrc_node_destroy(ctx, node);
 187         }
 188         if (list_empty(&ctx->rsrc_ref_list) && unlikely(ctx->rsrc_quiesce))
 189                 wake_up_all(&ctx->rsrc_quiesce_wq);
 190 }
 191
 192 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
 193 {
 194         struct io_rsrc_node *ref_node;
 195         struct io_cache_entry *entry;
 196
 197         entry = io_alloc_cache_get(&ctx->rsrc_node_cache);
 198         if (entry) {
 199                 ref_node = container_of(entry, struct io_rsrc_node, cache);
 200         } else {
 201                 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
 202                 if (!ref_node)
 203                         return NULL;
 204         }
 205
 206         ref_node->ctx = ctx;
 207         ref_node->empty = 0;
 208         ref_node->refs = 1;
 209         return ref_node;
 210 }
 211
 212 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
 213                                       struct io_ring_ctx *ctx)
 214 {
 215         struct io_rsrc_node *backup;
 216         DEFINE_WAIT(we);
 217         int ret;
 218
 219         /* As We may drop ->uring_lock, other task may have started quiesce */
 220         if (data->quiesce)
 221                 return -ENXIO;
 222
 223         backup = io_rsrc_node_alloc(ctx);
 224         if (!backup)
 225                 return -ENOMEM;
 226         ctx->rsrc_node->empty = true;
 227         ctx->rsrc_node->type = -1;
 228         list_add_tail(&ctx->rsrc_node->node, &ctx->rsrc_ref_list);
 229         io_put_rsrc_node(ctx, ctx->rsrc_node);
 230         ctx->rsrc_node = backup;
 231
 232         if (list_empty(&ctx->rsrc_ref_list))
 233                 return 0;
 234
 235         if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
 236                 atomic_set(&ctx->cq_wait_nr, 1);
 237                 smp_mb();
 238         }
 239
 240         ctx->rsrc_quiesce++;
 241         data->quiesce = true;
 242         do {
 243                 prepare_to_wait(&ctx->rsrc_quiesce_wq, &we, TASK_INTERRUPTIBLE);
 244                 mutex_unlock(&ctx->uring_lock);
 245
 246                 ret = io_run_task_work_sig(ctx);
 247                 if (ret < 0) {
 248                         mutex_lock(&ctx->uring_lock);
 249                         if (list_empty(&ctx->rsrc_ref_list))
 250                                 ret = 0;
 251                         break;
 252                 }
 253
 254                 schedule();
 255                 __set_current_state(TASK_RUNNING);
 256                 mutex_lock(&ctx->uring_lock);
 257                 ret = 0;
 258         } while (!list_empty(&ctx->rsrc_ref_list));
 259
 260         finish_wait(&ctx->rsrc_quiesce_wq, &we);
 261         data->quiesce = false;
 262         ctx->rsrc_quiesce--;
 263
 264         if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
 265                 atomic_set(&ctx->cq_wait_nr, 0);
 266                 smp_mb();
 267         }
 268         return ret;
 269 }
 270
 271 static void io_free_page_table(void **table, size_t size)
 272 {
 273         unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
 274
 275         for (i = 0; i < nr_tables; i++)
 276                 kfree(table[i]);
 277         kfree(table);
 278 }
 279
 280 static void io_rsrc_data_free(struct io_rsrc_data *data)
 281 {
 282         size_t size = data->nr * sizeof(data->tags[0][0]);
 283
 284         if (data->tags)
 285                 io_free_page_table((void **)data->tags, size);
 286         kfree(data);
 287 }
 288
 289 static __cold void **io_alloc_page_table(size_t size)
 290 {
 291         unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
 292         size_t init_size = size;
 293         void **table;
 294
 295         table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
 296         if (!table)
 297                 return NULL;
 298
 299         for (i = 0; i < nr_tables; i++) {
 300                 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
 301
 302                 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
 303                 if (!table[i]) {
 304                         io_free_page_table(table, init_size);
 305                         return NULL;
 306                 }
 307                 size -= this_size;
 308         }
 309         return table;
 310 }
 311
 312 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, int type,
 313                                      u64 __user *utags,
 314                                      unsigned nr, struct io_rsrc_data **pdata)
 315 {
 316         struct io_rsrc_data *data;
 317         int ret = 0;
 318         unsigned i;
 319
 320         data = kzalloc(sizeof(*data), GFP_KERNEL);
 321         if (!data)
 322                 return -ENOMEM;
 323         data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
 324         if (!data->tags) {
 325                 kfree(data);
 326                 return -ENOMEM;
 327         }
 328
 329         data->nr = nr;
 330         data->ctx = ctx;
 331         data->rsrc_type = type;
 332         if (utags) {
 333                 ret = -EFAULT;
 334                 for (i = 0; i < nr; i++) {
 335                         u64 *tag_slot = io_get_tag_slot(data, i);
 336
 337                         if (copy_from_user(tag_slot, &utags[i],
 338                                            sizeof(*tag_slot)))
 339                                 goto fail;
 340                 }
 341         }
 342         *pdata = data;
 343         return 0;
 344 fail:
 345         io_rsrc_data_free(data);
 346         return ret;
 347 }
 348
 349 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
 350                                  struct io_uring_rsrc_update2 *up,
 351                                  unsigned nr_args)
 352 {
 353         u64 __user *tags = u64_to_user_ptr(up->tags);
 354         __s32 __user *fds = u64_to_user_ptr(up->data);
 355         struct io_rsrc_data *data = ctx->file_data;
 356         struct io_fixed_file *file_slot;
 357         int fd, i, err = 0;
 358         unsigned int done;
 359
 360         if (!ctx->file_data)
 361                 return -ENXIO;
 362         if (up->offset + nr_args > ctx->nr_user_files)
 363                 return -EINVAL;
 364
 365         for (done = 0; done < nr_args; done++) {
 366                 u64 tag = 0;
 367
 368                 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
 369                     copy_from_user(&fd, &fds[done], sizeof(fd))) {
 370                         err = -EFAULT;
 371                         break;
 372                 }
 373                 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
 374                         err = -EINVAL;
 375                         break;
 376                 }
 377                 if (fd == IORING_REGISTER_FILES_SKIP)
 378                         continue;
 379
 380                 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
 381                 file_slot = io_fixed_file_slot(&ctx->file_table, i);
 382
 383                 if (file_slot->file_ptr) {
 384                         err = io_queue_rsrc_removal(data, i,
 385                                                     io_slot_file(file_slot));
 386                         if (err)
 387                                 break;
 388                         file_slot->file_ptr = 0;
 389                         io_file_bitmap_clear(&ctx->file_table, i);
 390                 }
 391                 if (fd != -1) {
 392                         struct file *file = fget(fd);
 393
 394                         if (!file) {
 395                                 err = -EBADF;
 396                                 break;
 397                         }
 398                         /*
 399                          * Don't allow io_uring instances to be registered. If
 400                          * UNIX isn't enabled, then this causes a reference
 401                          * cycle and this instance can never get freed. If UNIX
 402                          * is enabled we'll handle it just fine, but there's
 403                          * still no point in allowing a ring fd as it doesn't
 404                          * support regular read/write anyway.
 405                          */
 406                         if (io_is_uring_fops(file)) {
 407                                 fput(file);
 408                                 err = -EBADF;
 409                                 break;
 410                         }
 411                         err = io_scm_file_account(ctx, file);
 412                         if (err) {
 413                                 fput(file);
 414                                 break;
 415                         }
 416                         *io_get_tag_slot(data, i) = tag;
 417                         io_fixed_file_set(file_slot, file);
 418                         io_file_bitmap_set(&ctx->file_table, i);
 419                 }
 420         }
 421         return done ? done : err;
 422 }
 423
 424 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
 425                                    struct io_uring_rsrc_update2 *up,
 426                                    unsigned int nr_args)
 427 {
 428         u64 __user *tags = u64_to_user_ptr(up->tags);
 429         struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
 430         struct page *last_hpage = NULL;
 431         __u32 done;
 432         int i, err;
 433
 434         if (!ctx->buf_data)
 435                 return -ENXIO;
 436         if (up->offset + nr_args > ctx->nr_user_bufs)
 437                 return -EINVAL;
 438
 439         for (done = 0; done < nr_args; done++) {
 440                 struct io_mapped_ubuf *imu;
 441                 u64 tag = 0;
 442
 443                 err = io_copy_iov(ctx, &iov, iovs, done);
 444                 if (err)
 445                         break;
 446                 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
 447                         err = -EFAULT;
 448                         break;
 449                 }
 450                 err = io_buffer_validate(&iov);
 451                 if (err)
 452                         break;
 453                 if (!iov.iov_base && tag) {
 454                         err = -EINVAL;
 455                         break;
 456                 }
 457                 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
 458                 if (err)
 459                         break;
 460
 461                 i = array_index_nospec(up->offset + done, ctx->nr_user_bufs);
 462                 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
 463                         err = io_queue_rsrc_removal(ctx->buf_data, i,
 464                                                     ctx->user_bufs[i]);
 465                         if (unlikely(err)) {
 466                                 io_buffer_unmap(ctx, &imu);
 467                                 break;
 468                         }
 469                         ctx->user_bufs[i] = ctx->dummy_ubuf;
 470                 }
 471
 472                 ctx->user_bufs[i] = imu;
 473                 *io_get_tag_slot(ctx->buf_data, i) = tag;
 474         }
 475         return done ? done : err;
 476 }
 477
 478 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
 479                                      struct io_uring_rsrc_update2 *up,
 480                                      unsigned nr_args)
 481 {
 482         __u32 tmp;
 483
 484         lockdep_assert_held(&ctx->uring_lock);
 485
 486         if (check_add_overflow(up->offset, nr_args, &tmp))
 487                 return -EOVERFLOW;
 488
 489         switch (type) {
 490         case IORING_RSRC_FILE:
 491                 return __io_sqe_files_update(ctx, up, nr_args);
 492         case IORING_RSRC_BUFFER:
 493                 return __io_sqe_buffers_update(ctx, up, nr_args);
 494         }
 495         return -EINVAL;
 496 }
 497
 498 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
 499                              unsigned nr_args)
 500 {
 501         struct io_uring_rsrc_update2 up;
 502
 503         if (!nr_args)
 504                 return -EINVAL;
 505         memset(&up, 0, sizeof(up));
 506         if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
 507                 return -EFAULT;
 508         if (up.resv || up.resv2)
 509                 return -EINVAL;
 510         return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
 511 }
 512
 513 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
 514                             unsigned size, unsigned type)
 515 {
 516         struct io_uring_rsrc_update2 up;
 517
 518         if (size != sizeof(up))
 519                 return -EINVAL;
 520         if (copy_from_user(&up, arg, sizeof(up)))
 521                 return -EFAULT;
 522         if (!up.nr || up.resv || up.resv2)
 523                 return -EINVAL;
 524         return __io_register_rsrc_update(ctx, type, &up, up.nr);
 525 }
 526
 527 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
 528                             unsigned int size, unsigned int type)
 529 {
 530         struct io_uring_rsrc_register rr;
 531
 532         /* keep it extendible */
 533         if (size != sizeof(rr))
 534                 return -EINVAL;
 535
 536         memset(&rr, 0, sizeof(rr));
 537         if (copy_from_user(&rr, arg, size))
 538                 return -EFAULT;
 539         if (!rr.nr || rr.resv2)
 540                 return -EINVAL;
 541         if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
 542                 return -EINVAL;
 543
 544         switch (type) {
 545         case IORING_RSRC_FILE:
 546                 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
 547                         break;
 548                 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
 549                                              rr.nr, u64_to_user_ptr(rr.tags));
 550         case IORING_RSRC_BUFFER:
 551                 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
 552                         break;
 553                 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
 554                                                rr.nr, u64_to_user_ptr(rr.tags));
 555         }
 556         return -EINVAL;
 557 }
 558
 559 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 560 {
 561         struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
 562
 563         if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
 564                 return -EINVAL;
 565         if (sqe->rw_flags || sqe->splice_fd_in)
 566                 return -EINVAL;
 567
 568         up->offset = READ_ONCE(sqe->off);
 569         up->nr_args = READ_ONCE(sqe->len);
 570         if (!up->nr_args)
 571                 return -EINVAL;
 572         up->arg = READ_ONCE(sqe->addr);
 573         return 0;
 574 }
 575
 576 static int io_files_update_with_index_alloc(struct io_kiocb *req,
 577                                             unsigned int issue_flags)
 578 {
 579         struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
 580         __s32 __user *fds = u64_to_user_ptr(up->arg);
 581         unsigned int done;
 582         struct file *file;
 583         int ret, fd;
 584
 585         if (!req->ctx->file_data)
 586                 return -ENXIO;
 587
 588         for (done = 0; done < up->nr_args; done++) {
 589                 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
 590                         ret = -EFAULT;
 591                         break;
 592                 }
 593
 594                 file = fget(fd);
 595                 if (!file) {
 596                         ret = -EBADF;
 597                         break;
 598                 }
 599                 ret = io_fixed_fd_install(req, issue_flags, file,
 600                                           IORING_FILE_INDEX_ALLOC);
 601                 if (ret < 0)
 602                         break;
 603                 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
 604                         __io_close_fixed(req->ctx, issue_flags, ret);
 605                         ret = -EFAULT;
 606                         break;
 607                 }
 608         }
 609
 610         if (done)
 611                 return done;
 612         return ret;
 613 }
 614
 615 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
 616 {
 617         struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
 618         struct io_ring_ctx *ctx = req->ctx;
 619         struct io_uring_rsrc_update2 up2;
 620         int ret;
 621
 622         up2.offset = up->offset;
 623         up2.data = up->arg;
 624         up2.nr = 0;
 625         up2.tags = 0;
 626         up2.resv = 0;
 627         up2.resv2 = 0;
 628
 629         if (up->offset == IORING_FILE_INDEX_ALLOC) {
 630                 ret = io_files_update_with_index_alloc(req, issue_flags);
 631         } else {
 632                 io_ring_submit_lock(ctx, issue_flags);
 633                 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
 634                                                 &up2, up->nr_args);
 635                 io_ring_submit_unlock(ctx, issue_flags);
 636         }
 637
 638         if (ret < 0)
 639                 req_set_fail(req);
 640         io_req_set_res(req, ret, 0);
 641         return IOU_OK;
 642 }
 643
 644 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, void *rsrc)
 645 {
 646         struct io_ring_ctx *ctx = data->ctx;
 647         struct io_rsrc_node *node = ctx->rsrc_node;
 648         u64 *tag_slot = io_get_tag_slot(data, idx);
 649
 650         ctx->rsrc_node = io_rsrc_node_alloc(ctx);
 651         if (unlikely(!ctx->rsrc_node)) {
 652                 ctx->rsrc_node = node;
 653                 return -ENOMEM;
 654         }
 655
 656         node->item.rsrc = rsrc;
 657         node->type = data->rsrc_type;
 658         node->item.tag = *tag_slot;
 659         *tag_slot = 0;
 660         list_add_tail(&node->node, &ctx->rsrc_ref_list);
 661         io_put_rsrc_node(ctx, node);
 662         return 0;
 663 }
 664
 665 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
 666 {
 667         int i;
 668
 669         for (i = 0; i < ctx->nr_user_files; i++) {
 670                 struct file *file = io_file_from_index(&ctx->file_table, i);
 671
 672                 /* skip scm accounted files, they'll be freed by ->ring_sock */
 673                 if (!file || io_file_need_scm(file))
 674                         continue;
 675                 io_file_bitmap_clear(&ctx->file_table, i);
 676                 fput(file);
 677         }
 678
 679 #if defined(CONFIG_UNIX)
 680         if (ctx->ring_sock) {
 681                 struct sock *sock = ctx->ring_sock->sk;
 682                 struct sk_buff *skb;
 683
 684                 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
 685                         kfree_skb(skb);
 686         }
 687 #endif
 688         io_free_file_tables(&ctx->file_table);
 689         io_file_table_set_alloc_range(ctx, 0, 0);
 690         io_rsrc_data_free(ctx->file_data);
 691         ctx->file_data = NULL;
 692         ctx->nr_user_files = 0;
 693 }
 694
 695 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
 696 {
 697         unsigned nr = ctx->nr_user_files;
 698         int ret;
 699
 700         if (!ctx->file_data)
 701                 return -ENXIO;
 702
 703         /*
 704          * Quiesce may unlock ->uring_lock, and while it's not held
 705          * prevent new requests using the table.
 706          */
 707         ctx->nr_user_files = 0;
 708         ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
 709         ctx->nr_user_files = nr;
 710         if (!ret)
 711                 __io_sqe_files_unregister(ctx);
 712         return ret;
 713 }
 714
 715 /*
 716  * Ensure the UNIX gc is aware of our file set, so we are certain that
 717  * the io_uring can be safely unregistered on process exit, even if we have
 718  * loops in the file referencing. We account only files that can hold other
 719  * files because otherwise they can't form a loop and so are not interesting
 720  * for GC.
 721  */
 722 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
 723 {
 724 #if defined(CONFIG_UNIX)
 725         struct sock *sk = ctx->ring_sock->sk;
 726         struct sk_buff_head *head = &sk->sk_receive_queue;
 727         struct scm_fp_list *fpl;
 728         struct sk_buff *skb;
 729
 730         if (likely(!io_file_need_scm(file)))
 731                 return 0;
 732
 733         /*
 734          * See if we can merge this file into an existing skb SCM_RIGHTS
 735          * file set. If there's no room, fall back to allocating a new skb
 736          * and filling it in.
 737          */
 738         spin_lock_irq(&head->lock);
 739         skb = skb_peek(head);
 740         if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
 741                 __skb_unlink(skb, head);
 742         else
 743                 skb = NULL;
 744         spin_unlock_irq(&head->lock);
 745
 746         if (!skb) {
 747                 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
 748                 if (!fpl)
 749                         return -ENOMEM;
 750
 751                 skb = alloc_skb(0, GFP_KERNEL);
 752                 if (!skb) {
 753                         kfree(fpl);
 754                         return -ENOMEM;
 755                 }
 756
 757                 fpl->user = get_uid(current_user());
 758                 fpl->max = SCM_MAX_FD;
 759                 fpl->count = 0;
 760
 761                 UNIXCB(skb).fp = fpl;
 762                 skb->sk = sk;
 763                 skb->destructor = io_uring_destruct_scm;
 764                 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
 765         }
 766
 767         fpl = UNIXCB(skb).fp;
 768         fpl->fp[fpl->count++] = get_file(file);
 769         unix_inflight(fpl->user, file);
 770         skb_queue_head(head, skb);
 771         fput(file);
 772 #endif
 773         return 0;
 774 }
 775
 776 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file)
 777 {
 778 #if defined(CONFIG_UNIX)
 779         struct sock *sock = ctx->ring_sock->sk;
 780         struct sk_buff_head list, *head = &sock->sk_receive_queue;
 781         struct sk_buff *skb;
 782         int i;
 783
 784         __skb_queue_head_init(&list);
 785
 786         /*
 787          * Find the skb that holds this file in its SCM_RIGHTS. When found,
 788          * remove this entry and rearrange the file array.
 789          */
 790         skb = skb_dequeue(head);
 791         while (skb) {
 792                 struct scm_fp_list *fp;
 793
 794                 fp = UNIXCB(skb).fp;
 795                 for (i = 0; i < fp->count; i++) {
 796                         int left;
 797
 798                         if (fp->fp[i] != file)
 799                                 continue;
 800
 801                         unix_notinflight(fp->user, fp->fp[i]);
 802                         left = fp->count - 1 - i;
 803                         if (left) {
 804                                 memmove(&fp->fp[i], &fp->fp[i + 1],
 805                                                 left * sizeof(struct file *));
 806                         }
 807                         fp->count--;
 808                         if (!fp->count) {
 809                                 kfree_skb(skb);
 810                                 skb = NULL;
 811                         } else {
 812                                 __skb_queue_tail(&list, skb);
 813                         }
 814                         fput(file);
 815                         file = NULL;
 816                         break;
 817                 }
 818
 819                 if (!file)
 820                         break;
 821
 822                 __skb_queue_tail(&list, skb);
 823
 824                 skb = skb_dequeue(head);
 825         }
 826
 827         if (skb_peek(&list)) {
 828                 spin_lock_irq(&head->lock);
 829                 while ((skb = __skb_dequeue(&list)) != NULL)
 830                         __skb_queue_tail(head, skb);
 831                 spin_unlock_irq(&head->lock);
 832         }
 833 #endif
 834 }
 835
 836 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
 837 {
 838         struct file *file = prsrc->file;
 839
 840         if (likely(!io_file_need_scm(file)))
 841                 fput(file);
 842         else
 843                 io_rsrc_file_scm_put(ctx, file);
 844 }
 845
 846 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
 847                           unsigned nr_args, u64 __user *tags)
 848 {
 849         __s32 __user *fds = (__s32 __user *) arg;
 850         struct file *file;
 851         int fd, ret;
 852         unsigned i;
 853
 854         if (ctx->file_data)
 855                 return -EBUSY;
 856         if (!nr_args)
 857                 return -EINVAL;
 858         if (nr_args > IORING_MAX_FIXED_FILES)
 859                 return -EMFILE;
 860         if (nr_args > rlimit(RLIMIT_NOFILE))
 861                 return -EMFILE;
 862         ret = io_rsrc_data_alloc(ctx, IORING_RSRC_FILE, tags, nr_args,
 863                                  &ctx->file_data);
 864         if (ret)
 865                 return ret;
 866
 867         if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
 868                 io_rsrc_data_free(ctx->file_data);
 869                 ctx->file_data = NULL;
 870                 return -ENOMEM;
 871         }
 872
 873         for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
 874                 struct io_fixed_file *file_slot;
 875
 876                 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
 877                         ret = -EFAULT;
 878                         goto fail;
 879                 }
 880                 /* allow sparse sets */
 881                 if (!fds || fd == -1) {
 882                         ret = -EINVAL;
 883                         if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
 884                                 goto fail;
 885                         continue;
 886                 }
 887
 888                 file = fget(fd);
 889                 ret = -EBADF;
 890                 if (unlikely(!file))
 891                         goto fail;
 892
 893                 /*
 894                  * Don't allow io_uring instances to be registered. If UNIX
 895                  * isn't enabled, then this causes a reference cycle and this
 896                  * instance can never get freed. If UNIX is enabled we'll
 897                  * handle it just fine, but there's still no point in allowing
 898                  * a ring fd as it doesn't support regular read/write anyway.
 899                  */
 900                 if (io_is_uring_fops(file)) {
 901                         fput(file);
 902                         goto fail;
 903                 }
 904                 ret = io_scm_file_account(ctx, file);
 905                 if (ret) {
 906                         fput(file);
 907                         goto fail;
 908                 }
 909                 file_slot = io_fixed_file_slot(&ctx->file_table, i);
 910                 io_fixed_file_set(file_slot, file);
 911                 io_file_bitmap_set(&ctx->file_table, i);
 912         }
 913
 914         /* default it to the whole table */
 915         io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
 916         return 0;
 917 fail:
 918         __io_sqe_files_unregister(ctx);
 919         return ret;
 920 }
 921
 922 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
 923 {
 924         io_buffer_unmap(ctx, &prsrc->buf);
 925         prsrc->buf = NULL;
 926 }
 927
 928 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
 929 {
 930         unsigned int i;
 931
 932         for (i = 0; i < ctx->nr_user_bufs; i++)
 933                 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
 934         kfree(ctx->user_bufs);
 935         io_rsrc_data_free(ctx->buf_data);
 936         ctx->user_bufs = NULL;
 937         ctx->buf_data = NULL;
 938         ctx->nr_user_bufs = 0;
 939 }
 940
 941 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
 942 {
 943         unsigned nr = ctx->nr_user_bufs;
 944         int ret;
 945
 946         if (!ctx->buf_data)
 947                 return -ENXIO;
 948
 949         /*
 950          * Quiesce may unlock ->uring_lock, and while it's not held
 951          * prevent new requests using the table.
 952          */
 953         ctx->nr_user_bufs = 0;
 954         ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
 955         ctx->nr_user_bufs = nr;
 956         if (!ret)
 957                 __io_sqe_buffers_unregister(ctx);
 958         return ret;
 959 }
 960
 961 /*
 962  * Not super efficient, but this is just a registration time. And we do cache
 963  * the last compound head, so generally we'll only do a full search if we don't
 964  * match that one.
 965  *
 966  * We check if the given compound head page has already been accounted, to
 967  * avoid double accounting it. This allows us to account the full size of the
 968  * page, not just the constituent pages of a huge page.
 969  */
 970 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
 971                                   int nr_pages, struct page *hpage)
 972 {
 973         int i, j;
 974
 975         /* check current page array */
 976         for (i = 0; i < nr_pages; i++) {
 977                 if (!PageCompound(pages[i]))
 978                         continue;
 979                 if (compound_head(pages[i]) == hpage)
 980                         return true;
 981         }
 982
 983         /* check previously registered pages */
 984         for (i = 0; i < ctx->nr_user_bufs; i++) {
 985                 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
 986
 987                 for (j = 0; j < imu->nr_bvecs; j++) {
 988                         if (!PageCompound(imu->bvec[j].bv_page))
 989                                 continue;
 990                         if (compound_head(imu->bvec[j].bv_page) == hpage)
 991                                 return true;
 992                 }
 993         }
 994
 995         return false;
 996 }
 997
 998 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
 999                                  int nr_pages, struct io_mapped_ubuf *imu,
1000                                  struct page **last_hpage)
1001 {
1002         int i, ret;
1003
1004         imu->acct_pages = 0;
1005         for (i = 0; i < nr_pages; i++) {
1006                 if (!PageCompound(pages[i])) {
1007                         imu->acct_pages++;
1008                 } else {
1009                         struct page *hpage;
1010
1011                         hpage = compound_head(pages[i]);
1012                         if (hpage == *last_hpage)
1013                                 continue;
1014                         *last_hpage = hpage;
1015                         if (headpage_already_acct(ctx, pages, i, hpage))
1016                                 continue;
1017                         imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1018                 }
1019         }
1020
1021         if (!imu->acct_pages)
1022                 return 0;
1023
1024         ret = io_account_mem(ctx, imu->acct_pages);
1025         if (ret)
1026                 imu->acct_pages = 0;
1027         return ret;
1028 }
1029
1030 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1031 {
1032         unsigned long start, end, nr_pages;
1033         struct page **pages = NULL;
1034         int pret, ret = -ENOMEM;
1035
1036         end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1037         start = ubuf >> PAGE_SHIFT;
1038         nr_pages = end - start;
1039
1040         pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1041         if (!pages)
1042                 goto done;
1043
1044         ret = 0;
1045         mmap_read_lock(current->mm);
1046         pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1047                               pages);
1048         if (pret == nr_pages)
1049                 *npages = nr_pages;
1050         else
1051                 ret = pret < 0 ? pret : -EFAULT;
1052
1053         mmap_read_unlock(current->mm);
1054         if (ret) {
1055                 /* if we did partial map, release any pages we did get */
1056                 if (pret > 0)
1057                         unpin_user_pages(pages, pret);
1058                 goto done;
1059         }
1060         ret = 0;
1061 done:
1062         if (ret < 0) {
1063                 kvfree(pages);
1064                 pages = ERR_PTR(ret);
1065         }
1066         return pages;
1067 }
1068
1069 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1070                                   struct io_mapped_ubuf **pimu,
1071                                   struct page **last_hpage)
1072 {
1073         struct io_mapped_ubuf *imu = NULL;
1074         struct page **pages = NULL;
1075         unsigned long off;
1076         size_t size;
1077         int ret, nr_pages, i;
1078         struct folio *folio = NULL;
1079
1080         *pimu = ctx->dummy_ubuf;
1081         if (!iov->iov_base)
1082                 return 0;
1083
1084         ret = -ENOMEM;
1085         pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1086                                 &nr_pages);
1087         if (IS_ERR(pages)) {
1088                 ret = PTR_ERR(pages);
1089                 pages = NULL;
1090                 goto done;
1091         }
1092
1093         /* If it's a huge page, try to coalesce them into a single bvec entry */
1094         if (nr_pages > 1) {
1095                 folio = page_folio(pages[0]);
1096                 for (i = 1; i < nr_pages; i++) {
1097                         /*
1098                          * Pages must be consecutive and on the same folio for
1099                          * this to work
1100                          */
1101                         if (page_folio(pages[i]) != folio ||
1102                             pages[i] != pages[i - 1] + 1) {
1103                                 folio = NULL;
1104                                 break;
1105                         }
1106                 }
1107                 if (folio) {
1108                         /*
1109                          * The pages are bound to the folio, it doesn't
1110                          * actually unpin them but drops all but one reference,
1111                          * which is usually put down by io_buffer_unmap().
1112                          * Note, needs a better helper.
1113                          */
1114                         unpin_user_pages(&pages[1], nr_pages - 1);
1115                         nr_pages = 1;
1116                 }
1117         }
1118
1119         imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1120         if (!imu)
1121                 goto done;
1122
1123         ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1124         if (ret) {
1125                 unpin_user_pages(pages, nr_pages);
1126                 goto done;
1127         }
1128
1129         off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1130         size = iov->iov_len;
1131         /* store original address for later verification */
1132         imu->ubuf = (unsigned long) iov->iov_base;
1133         imu->ubuf_end = imu->ubuf + iov->iov_len;
1134         imu->nr_bvecs = nr_pages;
1135         *pimu = imu;
1136         ret = 0;
1137
1138         if (folio) {
1139                 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1140                 goto done;
1141         }
1142         for (i = 0; i < nr_pages; i++) {
1143                 size_t vec_len;
1144
1145                 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1146                 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
1147                 off = 0;
1148                 size -= vec_len;
1149         }
1150 done:
1151         if (ret)
1152                 kvfree(imu);
1153         kvfree(pages);
1154         return ret;
1155 }
1156
1157 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1158 {
1159         ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1160         return ctx->user_bufs ? 0 : -ENOMEM;
1161 }
1162
1163 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1164                             unsigned int nr_args, u64 __user *tags)
1165 {
1166         struct page *last_hpage = NULL;
1167         struct io_rsrc_data *data;
1168         int i, ret;
1169         struct iovec iov;
1170
1171         BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1172
1173         if (ctx->user_bufs)
1174                 return -EBUSY;
1175         if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1176                 return -EINVAL;
1177         ret = io_rsrc_data_alloc(ctx, IORING_RSRC_BUFFER, tags, nr_args, &data);
1178         if (ret)
1179                 return ret;
1180         ret = io_buffers_map_alloc(ctx, nr_args);
1181         if (ret) {
1182                 io_rsrc_data_free(data);
1183                 return ret;
1184         }
1185
1186         for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1187                 if (arg) {
1188                         ret = io_copy_iov(ctx, &iov, arg, i);
1189                         if (ret)
1190                                 break;
1191                         ret = io_buffer_validate(&iov);
1192                         if (ret)
1193                                 break;
1194                 } else {
1195                         memset(&iov, 0, sizeof(iov));
1196                 }
1197
1198                 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1199                         ret = -EINVAL;
1200                         break;
1201                 }
1202
1203                 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1204                                              &last_hpage);
1205                 if (ret)
1206                         break;
1207         }
1208
1209         WARN_ON_ONCE(ctx->buf_data);
1210
1211         ctx->buf_data = data;
1212         if (ret)
1213                 __io_sqe_buffers_unregister(ctx);
1214         return ret;
1215 }
1216
1217 int io_import_fixed(int ddir, struct iov_iter *iter,
1218                            struct io_mapped_ubuf *imu,
1219                            u64 buf_addr, size_t len)
1220 {
1221         u64 buf_end;
1222         size_t offset;
1223
1224         if (WARN_ON_ONCE(!imu))
1225                 return -EFAULT;
1226         if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1227                 return -EFAULT;
1228         /* not inside the mapped region */
1229         if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1230                 return -EFAULT;
1231
1232         /*
1233          * Might not be a start of buffer, set size appropriately
1234          * and advance us to the beginning.
1235          */
1236         offset = buf_addr - imu->ubuf;
1237         iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1238
1239         if (offset) {
1240                 /*
1241                  * Don't use iov_iter_advance() here, as it's really slow for
1242                  * using the latter parts of a big fixed buffer - it iterates
1243                  * over each segment manually. We can cheat a bit here, because
1244                  * we know that:
1245                  *
1246                  * 1) it's a BVEC iter, we set it up
1247                  * 2) all bvecs are PAGE_SIZE in size, except potentially the
1248                  *    first and last bvec
1249                  *
1250                  * So just find our index, and adjust the iterator afterwards.
1251                  * If the offset is within the first bvec (or the whole first
1252                  * bvec, just use iov_iter_advance(). This makes it easier
1253                  * since we can just skip the first segment, which may not
1254                  * be PAGE_SIZE aligned.
1255                  */
1256                 const struct bio_vec *bvec = imu->bvec;
1257
1258                 if (offset <= bvec->bv_len) {
1259                         /*
1260                          * Note, huge pages buffers consists of one large
1261                          * bvec entry and should always go this way. The other
1262                          * branch doesn't expect non PAGE_SIZE'd chunks.
1263                          */
1264                         iter->bvec = bvec;
1265                         iter->nr_segs = bvec->bv_len;
1266                         iter->count -= offset;
1267                         iter->iov_offset = offset;
1268                 } else {
1269                         unsigned long seg_skip;
1270
1271                         /* skip first vec */
1272                         offset -= bvec->bv_len;
1273                         seg_skip = 1 + (offset >> PAGE_SHIFT);
1274
1275                         iter->bvec = bvec + seg_skip;
1276                         iter->nr_segs -= seg_skip;
1277                         iter->count -= bvec->bv_len + offset;
1278                         iter->iov_offset = offset & ~PAGE_MASK;
1279                 }
1280         }
1281
1282         return 0;
1283 }