io_uring/kbuf.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/namei.h>
   9 #include <linux/poll.h>
  10 #include <linux/io_uring.h>
  11
  12 #include <uapi/linux/io_uring.h>
  13
  14 #include "io_uring.h"
  15 #include "opdef.h"
  16 #include "kbuf.h"
  17
  18 #define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
  19
  20 #define BGID_ARRAY      64
  21
  22 struct io_provide_buf {
  23         struct file                     *file;
  24         __u64                           addr;
  25         __u32                           len;
  26         __u32                           bgid;
  27         __u16                           nbufs;
  28         __u16                           bid;
  29 };
  30
  31 static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
  32                                                         unsigned int bgid)
  33 {
  34         if (ctx->io_bl && bgid < BGID_ARRAY)
  35                 return &ctx->io_bl[bgid];
  36
  37         return xa_load(&ctx->io_bl_xa, bgid);
  38 }
  39
  40 static int io_buffer_add_list(struct io_ring_ctx *ctx,
  41                               struct io_buffer_list *bl, unsigned int bgid)
  42 {
  43         bl->bgid = bgid;
  44         if (bgid < BGID_ARRAY)
  45                 return 0;
  46
  47         return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
  48 }
  49
  50 void io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags)
  51 {
  52         struct io_ring_ctx *ctx = req->ctx;
  53         struct io_buffer_list *bl;
  54         struct io_buffer *buf;
  55
  56         /*
  57          * For legacy provided buffer mode, don't recycle if we already did
  58          * IO to this buffer. For ring-mapped provided buffer mode, we should
  59          * increment ring->head to explicitly monopolize the buffer to avoid
  60          * multiple use.
  61          */
  62         if (req->flags & REQ_F_PARTIAL_IO)
  63                 return;
  64
  65         io_ring_submit_lock(ctx, issue_flags);
  66
  67         buf = req->kbuf;
  68         bl = io_buffer_get_list(ctx, buf->bgid);
  69         list_add(&buf->list, &bl->buf_list);
  70         req->flags &= ~REQ_F_BUFFER_SELECTED;
  71         req->buf_index = buf->bgid;
  72
  73         io_ring_submit_unlock(ctx, issue_flags);
  74         return;
  75 }
  76
  77 unsigned int __io_put_kbuf(struct io_kiocb *req, unsigned issue_flags)
  78 {
  79         unsigned int cflags;
  80
  81         /*
  82          * We can add this buffer back to two lists:
  83          *
  84          * 1) The io_buffers_cache list. This one is protected by the
  85          *    ctx->uring_lock. If we already hold this lock, add back to this
  86          *    list as we can grab it from issue as well.
  87          * 2) The io_buffers_comp list. This one is protected by the
  88          *    ctx->completion_lock.
  89          *
  90          * We migrate buffers from the comp_list to the issue cache list
  91          * when we need one.
  92          */
  93         if (req->flags & REQ_F_BUFFER_RING) {
  94                 /* no buffers to recycle for this case */
  95                 cflags = __io_put_kbuf_list(req, NULL);
  96         } else if (issue_flags & IO_URING_F_UNLOCKED) {
  97                 struct io_ring_ctx *ctx = req->ctx;
  98
  99                 spin_lock(&ctx->completion_lock);
 100                 cflags = __io_put_kbuf_list(req, &ctx->io_buffers_comp);
 101                 spin_unlock(&ctx->completion_lock);
 102         } else {
 103                 lockdep_assert_held(&req->ctx->uring_lock);
 104
 105                 cflags = __io_put_kbuf_list(req, &req->ctx->io_buffers_cache);
 106         }
 107         return cflags;
 108 }
 109
 110 static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
 111                                               struct io_buffer_list *bl)
 112 {
 113         if (!list_empty(&bl->buf_list)) {
 114                 struct io_buffer *kbuf;
 115
 116                 kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
 117                 list_del(&kbuf->list);
 118                 if (*len == 0 || *len > kbuf->len)
 119                         *len = kbuf->len;
 120                 req->flags |= REQ_F_BUFFER_SELECTED;
 121                 req->kbuf = kbuf;
 122                 req->buf_index = kbuf->bid;
 123                 return u64_to_user_ptr(kbuf->addr);
 124         }
 125         return NULL;
 126 }
 127
 128 static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
 129                                           struct io_buffer_list *bl,
 130                                           unsigned int issue_flags)
 131 {
 132         struct io_uring_buf_ring *br = bl->buf_ring;
 133         struct io_uring_buf *buf;
 134         __u16 head = bl->head;
 135
 136         if (unlikely(smp_load_acquire(&br->tail) == head))
 137                 return NULL;
 138
 139         head &= bl->mask;
 140         /* mmaped buffers are always contig */
 141         if (bl->is_mmap || head < IO_BUFFER_LIST_BUF_PER_PAGE) {
 142                 buf = &br->bufs[head];
 143         } else {
 144                 int off = head & (IO_BUFFER_LIST_BUF_PER_PAGE - 1);
 145                 int index = head / IO_BUFFER_LIST_BUF_PER_PAGE;
 146                 buf = page_address(bl->buf_pages[index]);
 147                 buf += off;
 148         }
 149         if (*len == 0 || *len > buf->len)
 150                 *len = buf->len;
 151         req->flags |= REQ_F_BUFFER_RING;
 152         req->buf_list = bl;
 153         req->buf_index = buf->bid;
 154
 155         if (issue_flags & IO_URING_F_UNLOCKED || !file_can_poll(req->file)) {
 156                 /*
 157                  * If we came in unlocked, we have no choice but to consume the
 158                  * buffer here, otherwise nothing ensures that the buffer won't
 159                  * get used by others. This does mean it'll be pinned until the
 160                  * IO completes, coming in unlocked means we're being called from
 161                  * io-wq context and there may be further retries in async hybrid
 162                  * mode. For the locked case, the caller must call commit when
 163                  * the transfer completes (or if we get -EAGAIN and must poll of
 164                  * retry).
 165                  */
 166                 req->buf_list = NULL;
 167                 bl->head++;
 168         }
 169         return u64_to_user_ptr(buf->addr);
 170 }
 171
 172 void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 173                               unsigned int issue_flags)
 174 {
 175         struct io_ring_ctx *ctx = req->ctx;
 176         struct io_buffer_list *bl;
 177         void __user *ret = NULL;
 178
 179         io_ring_submit_lock(req->ctx, issue_flags);
 180
 181         bl = io_buffer_get_list(ctx, req->buf_index);
 182         if (likely(bl)) {
 183                 if (bl->is_mapped)
 184                         ret = io_ring_buffer_select(req, len, bl, issue_flags);
 185                 else
 186                         ret = io_provided_buffer_select(req, len, bl);
 187         }
 188         io_ring_submit_unlock(req->ctx, issue_flags);
 189         return ret;
 190 }
 191
 192 static __cold int io_init_bl_list(struct io_ring_ctx *ctx)
 193 {
 194         int i;
 195
 196         ctx->io_bl = kcalloc(BGID_ARRAY, sizeof(struct io_buffer_list),
 197                                 GFP_KERNEL);
 198         if (!ctx->io_bl)
 199                 return -ENOMEM;
 200
 201         for (i = 0; i < BGID_ARRAY; i++) {
 202                 INIT_LIST_HEAD(&ctx->io_bl[i].buf_list);
 203                 ctx->io_bl[i].bgid = i;
 204         }
 205
 206         return 0;
 207 }
 208
 209 static int __io_remove_buffers(struct io_ring_ctx *ctx,
 210                                struct io_buffer_list *bl, unsigned nbufs)
 211 {
 212         unsigned i = 0;
 213
 214         /* shouldn't happen */
 215         if (!nbufs)
 216                 return 0;
 217
 218         if (bl->is_mapped) {
 219                 i = bl->buf_ring->tail - bl->head;
 220                 if (bl->is_mmap) {
 221                         folio_put(virt_to_folio(bl->buf_ring));
 222                         bl->buf_ring = NULL;
 223                         bl->is_mmap = 0;
 224                 } else if (bl->buf_nr_pages) {
 225                         int j;
 226
 227                         for (j = 0; j < bl->buf_nr_pages; j++)
 228                                 unpin_user_page(bl->buf_pages[j]);
 229                         kvfree(bl->buf_pages);
 230                         bl->buf_pages = NULL;
 231                         bl->buf_nr_pages = 0;
 232                 }
 233                 /* make sure it's seen as empty */
 234                 INIT_LIST_HEAD(&bl->buf_list);
 235                 bl->is_mapped = 0;
 236                 return i;
 237         }
 238
 239         /* protects io_buffers_cache */
 240         lockdep_assert_held(&ctx->uring_lock);
 241
 242         while (!list_empty(&bl->buf_list)) {
 243                 struct io_buffer *nxt;
 244
 245                 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
 246                 list_move(&nxt->list, &ctx->io_buffers_cache);
 247                 if (++i == nbufs)
 248                         return i;
 249                 cond_resched();
 250         }
 251
 252         return i;
 253 }
 254
 255 void io_destroy_buffers(struct io_ring_ctx *ctx)
 256 {
 257         struct io_buffer_list *bl;
 258         unsigned long index;
 259         int i;
 260
 261         for (i = 0; i < BGID_ARRAY; i++) {
 262                 if (!ctx->io_bl)
 263                         break;
 264                 __io_remove_buffers(ctx, &ctx->io_bl[i], -1U);
 265         }
 266
 267         xa_for_each(&ctx->io_bl_xa, index, bl) {
 268                 xa_erase(&ctx->io_bl_xa, bl->bgid);
 269                 __io_remove_buffers(ctx, bl, -1U);
 270                 kfree(bl);
 271         }
 272
 273         while (!list_empty(&ctx->io_buffers_pages)) {
 274                 struct page *page;
 275
 276                 page = list_first_entry(&ctx->io_buffers_pages, struct page, lru);
 277                 list_del_init(&page->lru);
 278                 __free_page(page);
 279         }
 280 }
 281
 282 int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 283 {
 284         struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
 285         u64 tmp;
 286
 287         if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
 288             sqe->splice_fd_in)
 289                 return -EINVAL;
 290
 291         tmp = READ_ONCE(sqe->fd);
 292         if (!tmp || tmp > USHRT_MAX)
 293                 return -EINVAL;
 294
 295         memset(p, 0, sizeof(*p));
 296         p->nbufs = tmp;
 297         p->bgid = READ_ONCE(sqe->buf_group);
 298         return 0;
 299 }
 300
 301 int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
 302 {
 303         struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
 304         struct io_ring_ctx *ctx = req->ctx;
 305         struct io_buffer_list *bl;
 306         int ret = 0;
 307
 308         io_ring_submit_lock(ctx, issue_flags);
 309
 310         ret = -ENOENT;
 311         bl = io_buffer_get_list(ctx, p->bgid);
 312         if (bl) {
 313                 ret = -EINVAL;
 314                 /* can't use provide/remove buffers command on mapped buffers */
 315                 if (!bl->is_mapped)
 316                         ret = __io_remove_buffers(ctx, bl, p->nbufs);
 317         }
 318         io_ring_submit_unlock(ctx, issue_flags);
 319         if (ret < 0)
 320                 req_set_fail(req);
 321         io_req_set_res(req, ret, 0);
 322         return IOU_OK;
 323 }
 324
 325 int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 326 {
 327         unsigned long size, tmp_check;
 328         struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
 329         u64 tmp;
 330
 331         if (sqe->rw_flags || sqe->splice_fd_in)
 332                 return -EINVAL;
 333
 334         tmp = READ_ONCE(sqe->fd);
 335         if (!tmp || tmp > USHRT_MAX)
 336                 return -E2BIG;
 337         p->nbufs = tmp;
 338         p->addr = READ_ONCE(sqe->addr);
 339         p->len = READ_ONCE(sqe->len);
 340
 341         if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
 342                                 &size))
 343                 return -EOVERFLOW;
 344         if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
 345                 return -EOVERFLOW;
 346
 347         size = (unsigned long)p->len * p->nbufs;
 348         if (!access_ok(u64_to_user_ptr(p->addr), size))
 349                 return -EFAULT;
 350
 351         p->bgid = READ_ONCE(sqe->buf_group);
 352         tmp = READ_ONCE(sqe->off);
 353         if (tmp > USHRT_MAX)
 354                 return -E2BIG;
 355         if (tmp + p->nbufs >= USHRT_MAX)
 356                 return -EINVAL;
 357         p->bid = tmp;
 358         return 0;
 359 }
 360
 361 static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
 362 {
 363         struct io_buffer *buf;
 364         struct page *page;
 365         int bufs_in_page;
 366
 367         /*
 368          * Completions that don't happen inline (eg not under uring_lock) will
 369          * add to ->io_buffers_comp. If we don't have any free buffers, check
 370          * the completion list and splice those entries first.
 371          */
 372         if (!list_empty_careful(&ctx->io_buffers_comp)) {
 373                 spin_lock(&ctx->completion_lock);
 374                 if (!list_empty(&ctx->io_buffers_comp)) {
 375                         list_splice_init(&ctx->io_buffers_comp,
 376                                                 &ctx->io_buffers_cache);
 377                         spin_unlock(&ctx->completion_lock);
 378                         return 0;
 379                 }
 380                 spin_unlock(&ctx->completion_lock);
 381         }
 382
 383         /*
 384          * No free buffers and no completion entries either. Allocate a new
 385          * page worth of buffer entries and add those to our freelist.
 386          */
 387         page = alloc_page(GFP_KERNEL_ACCOUNT);
 388         if (!page)
 389                 return -ENOMEM;
 390
 391         list_add(&page->lru, &ctx->io_buffers_pages);
 392
 393         buf = page_address(page);
 394         bufs_in_page = PAGE_SIZE / sizeof(*buf);
 395         while (bufs_in_page) {
 396                 list_add_tail(&buf->list, &ctx->io_buffers_cache);
 397                 buf++;
 398                 bufs_in_page--;
 399         }
 400
 401         return 0;
 402 }
 403
 404 static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
 405                           struct io_buffer_list *bl)
 406 {
 407         struct io_buffer *buf;
 408         u64 addr = pbuf->addr;
 409         int i, bid = pbuf->bid;
 410
 411         for (i = 0; i < pbuf->nbufs; i++) {
 412                 if (list_empty(&ctx->io_buffers_cache) &&
 413                     io_refill_buffer_cache(ctx))
 414                         break;
 415                 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
 416                                         list);
 417                 list_move_tail(&buf->list, &bl->buf_list);
 418                 buf->addr = addr;
 419                 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
 420                 buf->bid = bid;
 421                 buf->bgid = pbuf->bgid;
 422                 addr += pbuf->len;
 423                 bid++;
 424                 cond_resched();
 425         }
 426
 427         return i ? 0 : -ENOMEM;
 428 }
 429
 430 int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
 431 {
 432         struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
 433         struct io_ring_ctx *ctx = req->ctx;
 434         struct io_buffer_list *bl;
 435         int ret = 0;
 436
 437         io_ring_submit_lock(ctx, issue_flags);
 438
 439         if (unlikely(p->bgid < BGID_ARRAY && !ctx->io_bl)) {
 440                 ret = io_init_bl_list(ctx);
 441                 if (ret)
 442                         goto err;
 443         }
 444
 445         bl = io_buffer_get_list(ctx, p->bgid);
 446         if (unlikely(!bl)) {
 447                 bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
 448                 if (!bl) {
 449                         ret = -ENOMEM;
 450                         goto err;
 451                 }
 452                 INIT_LIST_HEAD(&bl->buf_list);
 453                 ret = io_buffer_add_list(ctx, bl, p->bgid);
 454                 if (ret) {
 455                         kfree(bl);
 456                         goto err;
 457                 }
 458         }
 459         /* can't add buffers via this command for a mapped buffer ring */
 460         if (bl->is_mapped) {
 461                 ret = -EINVAL;
 462                 goto err;
 463         }
 464
 465         ret = io_add_buffers(ctx, p, bl);
 466 err:
 467         io_ring_submit_unlock(ctx, issue_flags);
 468
 469         if (ret < 0)
 470                 req_set_fail(req);
 471         io_req_set_res(req, ret, 0);
 472         return IOU_OK;
 473 }
 474
 475 static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
 476                             struct io_buffer_list *bl)
 477 {
 478         struct io_uring_buf_ring *br;
 479         struct page **pages;
 480         int nr_pages;
 481
 482         pages = io_pin_pages(reg->ring_addr,
 483                              flex_array_size(br, bufs, reg->ring_entries),
 484                              &nr_pages);
 485         if (IS_ERR(pages))
 486                 return PTR_ERR(pages);
 487
 488         br = page_address(pages[0]);
 489 #ifdef SHM_COLOUR
 490         /*
 491          * On platforms that have specific aliasing requirements, SHM_COLOUR
 492          * is set and we must guarantee that the kernel and user side align
 493          * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
 494          * the application mmap's the provided ring buffer. Fail the request
 495          * if we, by chance, don't end up with aligned addresses. The app
 496          * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
 497          * this transparently.
 498          */
 499         if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1)) {
 500                 int i;
 501
 502                 for (i = 0; i < nr_pages; i++)
 503                         unpin_user_page(pages[i]);
 504                 return -EINVAL;
 505         }
 506 #endif
 507         bl->buf_pages = pages;
 508         bl->buf_nr_pages = nr_pages;
 509         bl->buf_ring = br;
 510         bl->is_mapped = 1;
 511         bl->is_mmap = 0;
 512         return 0;
 513 }
 514
 515 static int io_alloc_pbuf_ring(struct io_uring_buf_reg *reg,
 516                               struct io_buffer_list *bl)
 517 {
 518         gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
 519         size_t ring_size;
 520         void *ptr;
 521
 522         ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
 523         ptr = (void *) __get_free_pages(gfp, get_order(ring_size));
 524         if (!ptr)
 525                 return -ENOMEM;
 526
 527         bl->buf_ring = ptr;
 528         bl->is_mapped = 1;
 529         bl->is_mmap = 1;
 530         return 0;
 531 }
 532
 533 int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
 534 {
 535         struct io_uring_buf_reg reg;
 536         struct io_buffer_list *bl, *free_bl = NULL;
 537         int ret;
 538
 539         if (copy_from_user(&reg, arg, sizeof(reg)))
 540                 return -EFAULT;
 541
 542         if (reg.resv[0] || reg.resv[1] || reg.resv[2])
 543                 return -EINVAL;
 544         if (reg.flags & ~IOU_PBUF_RING_MMAP)
 545                 return -EINVAL;
 546         if (!(reg.flags & IOU_PBUF_RING_MMAP)) {
 547                 if (!reg.ring_addr)
 548                         return -EFAULT;
 549                 if (reg.ring_addr & ~PAGE_MASK)
 550                         return -EINVAL;
 551         } else {
 552                 if (reg.ring_addr)
 553                         return -EINVAL;
 554         }
 555
 556         if (!is_power_of_2(reg.ring_entries))
 557                 return -EINVAL;
 558
 559         /* cannot disambiguate full vs empty due to head/tail size */
 560         if (reg.ring_entries >= 65536)
 561                 return -EINVAL;
 562
 563         if (unlikely(reg.bgid < BGID_ARRAY && !ctx->io_bl)) {
 564                 int ret = io_init_bl_list(ctx);
 565                 if (ret)
 566                         return ret;
 567         }
 568
 569         bl = io_buffer_get_list(ctx, reg.bgid);
 570         if (bl) {
 571                 /* if mapped buffer ring OR classic exists, don't allow */
 572                 if (bl->is_mapped || !list_empty(&bl->buf_list))
 573                         return -EEXIST;
 574         } else {
 575                 free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
 576                 if (!bl)
 577                         return -ENOMEM;
 578         }
 579
 580         if (!(reg.flags & IOU_PBUF_RING_MMAP))
 581                 ret = io_pin_pbuf_ring(&reg, bl);
 582         else
 583                 ret = io_alloc_pbuf_ring(&reg, bl);
 584
 585         if (!ret) {
 586                 bl->nr_entries = reg.ring_entries;
 587                 bl->mask = reg.ring_entries - 1;
 588
 589                 io_buffer_add_list(ctx, bl, reg.bgid);
 590                 return 0;
 591         }
 592
 593         kfree(free_bl);
 594         return ret;
 595 }
 596
 597 int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
 598 {
 599         struct io_uring_buf_reg reg;
 600         struct io_buffer_list *bl;
 601
 602         if (copy_from_user(&reg, arg, sizeof(reg)))
 603                 return -EFAULT;
 604         if (reg.resv[0] || reg.resv[1] || reg.resv[2])
 605                 return -EINVAL;
 606         if (reg.flags)
 607                 return -EINVAL;
 608
 609         bl = io_buffer_get_list(ctx, reg.bgid);
 610         if (!bl)
 611                 return -ENOENT;
 612         if (!bl->is_mapped)
 613                 return -EINVAL;
 614
 615         __io_remove_buffers(ctx, bl, -1U);
 616         if (bl->bgid >= BGID_ARRAY) {
 617                 xa_erase(&ctx->io_bl_xa, bl->bgid);
 618                 kfree(bl);
 619         }
 620         return 0;
 621 }
 622
 623 void *io_pbuf_get_address(struct io_ring_ctx *ctx, unsigned long bgid)
 624 {
 625         struct io_buffer_list *bl;
 626
 627         bl = io_buffer_get_list(ctx, bgid);
 628         if (!bl || !bl->is_mmap)
 629                 return NULL;
 630
 631         return bl->buf_ring;
 632 }