lib/iov_iter.c

   1 #include <linux/export.h>
   2 #include <linux/uio.h>
   3 #include <linux/pagemap.h>
   4 #include <linux/slab.h>
   5 #include <linux/vmalloc.h>
   6 #include <linux/splice.h>
   7 #include <net/checksum.h>
   8
   9 #define PIPE_PARANOIA /* for now */
  10
  11 #define iterate_iovec(i, n, __v, __p, skip, STEP) {     \
  12         size_t left;                                    \
  13         size_t wanted = n;                              \
  14         __p = i->iov;                                   \
  15         __v.iov_len = min(n, __p->iov_len - skip);      \
  16         if (likely(__v.iov_len)) {                      \
  17                 __v.iov_base = __p->iov_base + skip;    \
  18                 left = (STEP);                          \
  19                 __v.iov_len -= left;                    \
  20                 skip += __v.iov_len;                    \
  21                 n -= __v.iov_len;                       \
  22         } else {                                        \
  23                 left = 0;                               \
  24         }                                               \
  25         while (unlikely(!left && n)) {                  \
  26                 __p++;                                  \
  27                 __v.iov_len = min(n, __p->iov_len);     \
  28                 if (unlikely(!__v.iov_len))             \
  29                         continue;                       \
  30                 __v.iov_base = __p->iov_base;           \
  31                 left = (STEP);                          \
  32                 __v.iov_len -= left;                    \
  33                 skip = __v.iov_len;                     \
  34                 n -= __v.iov_len;                       \
  35         }                                               \
  36         n = wanted - n;                                 \
  37 }
  38
  39 #define iterate_kvec(i, n, __v, __p, skip, STEP) {      \
  40         size_t wanted = n;                              \
  41         __p = i->kvec;                                  \
  42         __v.iov_len = min(n, __p->iov_len - skip);      \
  43         if (likely(__v.iov_len)) {                      \
  44                 __v.iov_base = __p->iov_base + skip;    \
  45                 (void)(STEP);                           \
  46                 skip += __v.iov_len;                    \
  47                 n -= __v.iov_len;                       \
  48         }                                               \
  49         while (unlikely(n)) {                           \
  50                 __p++;                                  \
  51                 __v.iov_len = min(n, __p->iov_len);     \
  52                 if (unlikely(!__v.iov_len))             \
  53                         continue;                       \
  54                 __v.iov_base = __p->iov_base;           \
  55                 (void)(STEP);                           \
  56                 skip = __v.iov_len;                     \
  57                 n -= __v.iov_len;                       \
  58         }                                               \
  59         n = wanted;                                     \
  60 }
  61
  62 #define iterate_bvec(i, n, __v, __bi, skip, STEP) {     \
  63         struct bvec_iter __start;                       \
  64         __start.bi_size = n;                            \
  65         __start.bi_bvec_done = skip;                    \
  66         __start.bi_idx = 0;                             \
  67         for_each_bvec(__v, i->bvec, __bi, __start) {    \
  68                 if (!__v.bv_len)                        \
  69                         continue;                       \
  70                 (void)(STEP);                           \
  71         }                                               \
  72 }
  73
  74 #define iterate_all_kinds(i, n, v, I, B, K) {                   \
  75         size_t skip = i->iov_offset;                            \
  76         if (unlikely(i->type & ITER_BVEC)) {                    \
  77                 struct bio_vec v;                               \
  78                 struct bvec_iter __bi;                          \
  79                 iterate_bvec(i, n, v, __bi, skip, (B))          \
  80         } else if (unlikely(i->type & ITER_KVEC)) {             \
  81                 const struct kvec *kvec;                        \
  82                 struct kvec v;                                  \
  83                 iterate_kvec(i, n, v, kvec, skip, (K))          \
  84         } else {                                                \
  85                 const struct iovec *iov;                        \
  86                 struct iovec v;                                 \
  87                 iterate_iovec(i, n, v, iov, skip, (I))          \
  88         }                                                       \
  89 }
  90
  91 #define iterate_and_advance(i, n, v, I, B, K) {                 \
  92         if (unlikely(i->count < n))                             \
  93                 n = i->count;                                   \
  94         if (i->count) {                                         \
  95                 size_t skip = i->iov_offset;                    \
  96                 if (unlikely(i->type & ITER_BVEC)) {            \
  97                         const struct bio_vec *bvec = i->bvec;   \
  98                         struct bio_vec v;                       \
  99                         struct bvec_iter __bi;                  \
 100                         iterate_bvec(i, n, v, __bi, skip, (B))  \
 101                         i->bvec = __bvec_iter_bvec(i->bvec, __bi);      \
 102                         i->nr_segs -= i->bvec - bvec;           \
 103                         skip = __bi.bi_bvec_done;               \
 104                 } else if (unlikely(i->type & ITER_KVEC)) {     \
 105                         const struct kvec *kvec;                \
 106                         struct kvec v;                          \
 107                         iterate_kvec(i, n, v, kvec, skip, (K))  \
 108                         if (skip == kvec->iov_len) {            \
 109                                 kvec++;                         \
 110                                 skip = 0;                       \
 111                         }                                       \
 112                         i->nr_segs -= kvec - i->kvec;           \
 113                         i->kvec = kvec;                         \
 114                 } else {                                        \
 115                         const struct iovec *iov;                \
 116                         struct iovec v;                         \
 117                         iterate_iovec(i, n, v, iov, skip, (I))  \
 118                         if (skip == iov->iov_len) {             \
 119                                 iov++;                          \
 120                                 skip = 0;                       \
 121                         }                                       \
 122                         i->nr_segs -= iov - i->iov;             \
 123                         i->iov = iov;                           \
 124                 }                                               \
 125                 i->count -= n;                                  \
 126                 i->iov_offset = skip;                           \
 127         }                                                       \
 128 }
 129
 130 static size_t copy_page_to_iter_iovec(struct page *page, size_t offset, size_t bytes,
 131                          struct iov_iter *i)
 132 {
 133         size_t skip, copy, left, wanted;
 134         const struct iovec *iov;
 135         char __user *buf;
 136         void *kaddr, *from;
 137
 138         if (unlikely(bytes > i->count))
 139                 bytes = i->count;
 140
 141         if (unlikely(!bytes))
 142                 return 0;
 143
 144         wanted = bytes;
 145         iov = i->iov;
 146         skip = i->iov_offset;
 147         buf = iov->iov_base + skip;
 148         copy = min(bytes, iov->iov_len - skip);
 149
 150         if (IS_ENABLED(CONFIG_HIGHMEM) && !fault_in_pages_writeable(buf, copy)) {
 151                 kaddr = kmap_atomic(page);
 152                 from = kaddr + offset;
 153
 154                 /* first chunk, usually the only one */
 155                 left = __copy_to_user_inatomic(buf, from, copy);
 156                 copy -= left;
 157                 skip += copy;
 158                 from += copy;
 159                 bytes -= copy;
 160
 161                 while (unlikely(!left && bytes)) {
 162                         iov++;
 163                         buf = iov->iov_base;
 164                         copy = min(bytes, iov->iov_len);
 165                         left = __copy_to_user_inatomic(buf, from, copy);
 166                         copy -= left;
 167                         skip = copy;
 168                         from += copy;
 169                         bytes -= copy;
 170                 }
 171                 if (likely(!bytes)) {
 172                         kunmap_atomic(kaddr);
 173                         goto done;
 174                 }
 175                 offset = from - kaddr;
 176                 buf += copy;
 177                 kunmap_atomic(kaddr);
 178                 copy = min(bytes, iov->iov_len - skip);
 179         }
 180         /* Too bad - revert to non-atomic kmap */
 181
 182         kaddr = kmap(page);
 183         from = kaddr + offset;
 184         left = __copy_to_user(buf, from, copy);
 185         copy -= left;
 186         skip += copy;
 187         from += copy;
 188         bytes -= copy;
 189         while (unlikely(!left && bytes)) {
 190                 iov++;
 191                 buf = iov->iov_base;
 192                 copy = min(bytes, iov->iov_len);
 193                 left = __copy_to_user(buf, from, copy);
 194                 copy -= left;
 195                 skip = copy;
 196                 from += copy;
 197                 bytes -= copy;
 198         }
 199         kunmap(page);
 200
 201 done:
 202         if (skip == iov->iov_len) {
 203                 iov++;
 204                 skip = 0;
 205         }
 206         i->count -= wanted - bytes;
 207         i->nr_segs -= iov - i->iov;
 208         i->iov = iov;
 209         i->iov_offset = skip;
 210         return wanted - bytes;
 211 }
 212
 213 static size_t copy_page_from_iter_iovec(struct page *page, size_t offset, size_t bytes,
 214                          struct iov_iter *i)
 215 {
 216         size_t skip, copy, left, wanted;
 217         const struct iovec *iov;
 218         char __user *buf;
 219         void *kaddr, *to;
 220
 221         if (unlikely(bytes > i->count))
 222                 bytes = i->count;
 223
 224         if (unlikely(!bytes))
 225                 return 0;
 226
 227         wanted = bytes;
 228         iov = i->iov;
 229         skip = i->iov_offset;
 230         buf = iov->iov_base + skip;
 231         copy = min(bytes, iov->iov_len - skip);
 232
 233         if (IS_ENABLED(CONFIG_HIGHMEM) && !fault_in_pages_readable(buf, copy)) {
 234                 kaddr = kmap_atomic(page);
 235                 to = kaddr + offset;
 236
 237                 /* first chunk, usually the only one */
 238                 left = __copy_from_user_inatomic(to, buf, copy);
 239                 copy -= left;
 240                 skip += copy;
 241                 to += copy;
 242                 bytes -= copy;
 243
 244                 while (unlikely(!left && bytes)) {
 245                         iov++;
 246                         buf = iov->iov_base;
 247                         copy = min(bytes, iov->iov_len);
 248                         left = __copy_from_user_inatomic(to, buf, copy);
 249                         copy -= left;
 250                         skip = copy;
 251                         to += copy;
 252                         bytes -= copy;
 253                 }
 254                 if (likely(!bytes)) {
 255                         kunmap_atomic(kaddr);
 256                         goto done;
 257                 }
 258                 offset = to - kaddr;
 259                 buf += copy;
 260                 kunmap_atomic(kaddr);
 261                 copy = min(bytes, iov->iov_len - skip);
 262         }
 263         /* Too bad - revert to non-atomic kmap */
 264
 265         kaddr = kmap(page);
 266         to = kaddr + offset;
 267         left = __copy_from_user(to, buf, copy);
 268         copy -= left;
 269         skip += copy;
 270         to += copy;
 271         bytes -= copy;
 272         while (unlikely(!left && bytes)) {
 273                 iov++;
 274                 buf = iov->iov_base;
 275                 copy = min(bytes, iov->iov_len);
 276                 left = __copy_from_user(to, buf, copy);
 277                 copy -= left;
 278                 skip = copy;
 279                 to += copy;
 280                 bytes -= copy;
 281         }
 282         kunmap(page);
 283
 284 done:
 285         if (skip == iov->iov_len) {
 286                 iov++;
 287                 skip = 0;
 288         }
 289         i->count -= wanted - bytes;
 290         i->nr_segs -= iov - i->iov;
 291         i->iov = iov;
 292         i->iov_offset = skip;
 293         return wanted - bytes;
 294 }
 295
 296 #ifdef PIPE_PARANOIA
 297 static bool sanity(const struct iov_iter *i)
 298 {
 299         struct pipe_inode_info *pipe = i->pipe;
 300         int idx = i->idx;
 301         int next = pipe->curbuf + pipe->nrbufs;
 302         if (i->iov_offset) {
 303                 struct pipe_buffer *p;
 304                 if (unlikely(!pipe->nrbufs))
 305                         goto Bad;       // pipe must be non-empty
 306                 if (unlikely(idx != ((next - 1) & (pipe->buffers - 1))))
 307                         goto Bad;       // must be at the last buffer...
 308
 309                 p = &pipe->bufs[idx];
 310                 if (unlikely(p->offset + p->len != i->iov_offset))
 311                         goto Bad;       // ... at the end of segment
 312         } else {
 313                 if (idx != (next & (pipe->buffers - 1)))
 314                         goto Bad;       // must be right after the last buffer
 315         }
 316         return true;
 317 Bad:
 318         printk(KERN_ERR "idx = %d, offset = %zd\n", i->idx, i->iov_offset);
 319         printk(KERN_ERR "curbuf = %d, nrbufs = %d, buffers = %d\n",
 320                         pipe->curbuf, pipe->nrbufs, pipe->buffers);
 321         for (idx = 0; idx < pipe->buffers; idx++)
 322                 printk(KERN_ERR "[%p %p %d %d]\n",
 323                         pipe->bufs[idx].ops,
 324                         pipe->bufs[idx].page,
 325                         pipe->bufs[idx].offset,
 326                         pipe->bufs[idx].len);
 327         WARN_ON(1);
 328         return false;
 329 }
 330 #else
 331 #define sanity(i) true
 332 #endif
 333
 334 static inline int next_idx(int idx, struct pipe_inode_info *pipe)
 335 {
 336         return (idx + 1) & (pipe->buffers - 1);
 337 }
 338
 339 static size_t copy_page_to_iter_pipe(struct page *page, size_t offset, size_t bytes,
 340                          struct iov_iter *i)
 341 {
 342         struct pipe_inode_info *pipe = i->pipe;
 343         struct pipe_buffer *buf;
 344         size_t off;
 345         int idx;
 346
 347         if (unlikely(bytes > i->count))
 348                 bytes = i->count;
 349
 350         if (unlikely(!bytes))
 351                 return 0;
 352
 353         if (!sanity(i))
 354                 return 0;
 355
 356         off = i->iov_offset;
 357         idx = i->idx;
 358         buf = &pipe->bufs[idx];
 359         if (off) {
 360                 if (offset == off && buf->page == page) {
 361                         /* merge with the last one */
 362                         buf->len += bytes;
 363                         i->iov_offset += bytes;
 364                         goto out;
 365                 }
 366                 idx = next_idx(idx, pipe);
 367                 buf = &pipe->bufs[idx];
 368         }
 369         if (idx == pipe->curbuf && pipe->nrbufs)
 370                 return 0;
 371         pipe->nrbufs++;
 372         buf->ops = &page_cache_pipe_buf_ops;
 373         get_page(buf->page = page);
 374         buf->offset = offset;
 375         buf->len = bytes;
 376         i->iov_offset = offset + bytes;
 377         i->idx = idx;
 378 out:
 379         i->count -= bytes;
 380         return bytes;
 381 }
 382
 383 /*
 384  * Fault in one or more iovecs of the given iov_iter, to a maximum length of
 385  * bytes.  For each iovec, fault in each page that constitutes the iovec.
 386  *
 387  * Return 0 on success, or non-zero if the memory could not be accessed (i.e.
 388  * because it is an invalid address).
 389  */
 390 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
 391 {
 392         size_t skip = i->iov_offset;
 393         const struct iovec *iov;
 394         int err;
 395         struct iovec v;
 396
 397         if (!(i->type & (ITER_BVEC|ITER_KVEC))) {
 398                 iterate_iovec(i, bytes, v, iov, skip, ({
 399                         err = fault_in_pages_readable(v.iov_base, v.iov_len);
 400                         if (unlikely(err))
 401                         return err;
 402                 0;}))
 403         }
 404         return 0;
 405 }
 406 EXPORT_SYMBOL(iov_iter_fault_in_readable);
 407
 408 void iov_iter_init(struct iov_iter *i, int direction,
 409                         const struct iovec *iov, unsigned long nr_segs,
 410                         size_t count)
 411 {
 412         /* It will get better.  Eventually... */
 413         if (segment_eq(get_fs(), KERNEL_DS)) {
 414                 direction |= ITER_KVEC;
 415                 i->type = direction;
 416                 i->kvec = (struct kvec *)iov;
 417         } else {
 418                 i->type = direction;
 419                 i->iov = iov;
 420         }
 421         i->nr_segs = nr_segs;
 422         i->iov_offset = 0;
 423         i->count = count;
 424 }
 425 EXPORT_SYMBOL(iov_iter_init);
 426
 427 static void memcpy_from_page(char *to, struct page *page, size_t offset, size_t len)
 428 {
 429         char *from = kmap_atomic(page);
 430         memcpy(to, from + offset, len);
 431         kunmap_atomic(from);
 432 }
 433
 434 static void memcpy_to_page(struct page *page, size_t offset, const char *from, size_t len)
 435 {
 436         char *to = kmap_atomic(page);
 437         memcpy(to + offset, from, len);
 438         kunmap_atomic(to);
 439 }
 440
 441 static void memzero_page(struct page *page, size_t offset, size_t len)
 442 {
 443         char *addr = kmap_atomic(page);
 444         memset(addr + offset, 0, len);
 445         kunmap_atomic(addr);
 446 }
 447
 448 static inline bool allocated(struct pipe_buffer *buf)
 449 {
 450         return buf->ops == &default_pipe_buf_ops;
 451 }
 452
 453 static inline void data_start(const struct iov_iter *i, int *idxp, size_t *offp)
 454 {
 455         size_t off = i->iov_offset;
 456         int idx = i->idx;
 457         if (off && (!allocated(&i->pipe->bufs[idx]) || off == PAGE_SIZE)) {
 458                 idx = next_idx(idx, i->pipe);
 459                 off = 0;
 460         }
 461         *idxp = idx;
 462         *offp = off;
 463 }
 464
 465 static size_t push_pipe(struct iov_iter *i, size_t size,
 466                         int *idxp, size_t *offp)
 467 {
 468         struct pipe_inode_info *pipe = i->pipe;
 469         size_t off;
 470         int idx;
 471         ssize_t left;
 472
 473         if (unlikely(size > i->count))
 474                 size = i->count;
 475         if (unlikely(!size))
 476                 return 0;
 477
 478         left = size;
 479         data_start(i, &idx, &off);
 480         *idxp = idx;
 481         *offp = off;
 482         if (off) {
 483                 left -= PAGE_SIZE - off;
 484                 if (left <= 0) {
 485                         pipe->bufs[idx].len += size;
 486                         return size;
 487                 }
 488                 pipe->bufs[idx].len = PAGE_SIZE;
 489                 idx = next_idx(idx, pipe);
 490         }
 491         while (idx != pipe->curbuf || !pipe->nrbufs) {
 492                 struct page *page = alloc_page(GFP_USER);
 493                 if (!page)
 494                         break;
 495                 pipe->nrbufs++;
 496                 pipe->bufs[idx].ops = &default_pipe_buf_ops;
 497                 pipe->bufs[idx].page = page;
 498                 pipe->bufs[idx].offset = 0;
 499                 if (left <= PAGE_SIZE) {
 500                         pipe->bufs[idx].len = left;
 501                         return size;
 502                 }
 503                 pipe->bufs[idx].len = PAGE_SIZE;
 504                 left -= PAGE_SIZE;
 505                 idx = next_idx(idx, pipe);
 506         }
 507         return size - left;
 508 }
 509
 510 static size_t copy_pipe_to_iter(const void *addr, size_t bytes,
 511                                 struct iov_iter *i)
 512 {
 513         struct pipe_inode_info *pipe = i->pipe;
 514         size_t n, off;
 515         int idx;
 516
 517         if (!sanity(i))
 518                 return 0;
 519
 520         bytes = n = push_pipe(i, bytes, &idx, &off);
 521         if (unlikely(!n))
 522                 return 0;
 523         for ( ; n; idx = next_idx(idx, pipe), off = 0) {
 524                 size_t chunk = min_t(size_t, n, PAGE_SIZE - off);
 525                 memcpy_to_page(pipe->bufs[idx].page, off, addr, chunk);
 526                 i->idx = idx;
 527                 i->iov_offset = off + chunk;
 528                 n -= chunk;
 529                 addr += chunk;
 530         }
 531         i->count -= bytes;
 532         return bytes;
 533 }
 534
 535 size_t copy_to_iter(const void *addr, size_t bytes, struct iov_iter *i)
 536 {
 537         const char *from = addr;
 538         if (unlikely(i->type & ITER_PIPE))
 539                 return copy_pipe_to_iter(addr, bytes, i);
 540         iterate_and_advance(i, bytes, v,
 541                 __copy_to_user(v.iov_base, (from += v.iov_len) - v.iov_len,
 542                                v.iov_len),
 543                 memcpy_to_page(v.bv_page, v.bv_offset,
 544                                (from += v.bv_len) - v.bv_len, v.bv_len),
 545                 memcpy(v.iov_base, (from += v.iov_len) - v.iov_len, v.iov_len)
 546         )
 547
 548         return bytes;
 549 }
 550 EXPORT_SYMBOL(copy_to_iter);
 551
 552 size_t copy_from_iter(void *addr, size_t bytes, struct iov_iter *i)
 553 {
 554         char *to = addr;
 555         if (unlikely(i->type & ITER_PIPE)) {
 556                 WARN_ON(1);
 557                 return 0;
 558         }
 559         iterate_and_advance(i, bytes, v,
 560                 __copy_from_user((to += v.iov_len) - v.iov_len, v.iov_base,
 561                                  v.iov_len),
 562                 memcpy_from_page((to += v.bv_len) - v.bv_len, v.bv_page,
 563                                  v.bv_offset, v.bv_len),
 564                 memcpy((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
 565         )
 566
 567         return bytes;
 568 }
 569 EXPORT_SYMBOL(copy_from_iter);
 570
 571 size_t copy_from_iter_nocache(void *addr, size_t bytes, struct iov_iter *i)
 572 {
 573         char *to = addr;
 574         if (unlikely(i->type & ITER_PIPE)) {
 575                 WARN_ON(1);
 576                 return 0;
 577         }
 578         iterate_and_advance(i, bytes, v,
 579                 __copy_from_user_nocache((to += v.iov_len) - v.iov_len,
 580                                          v.iov_base, v.iov_len),
 581                 memcpy_from_page((to += v.bv_len) - v.bv_len, v.bv_page,
 582                                  v.bv_offset, v.bv_len),
 583                 memcpy((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
 584         )
 585
 586         return bytes;
 587 }
 588 EXPORT_SYMBOL(copy_from_iter_nocache);
 589
 590 size_t copy_page_to_iter(struct page *page, size_t offset, size_t bytes,
 591                          struct iov_iter *i)
 592 {
 593         if (i->type & (ITER_BVEC|ITER_KVEC)) {
 594                 void *kaddr = kmap_atomic(page);
 595                 size_t wanted = copy_to_iter(kaddr + offset, bytes, i);
 596                 kunmap_atomic(kaddr);
 597                 return wanted;
 598         } else if (likely(!(i->type & ITER_PIPE)))
 599                 return copy_page_to_iter_iovec(page, offset, bytes, i);
 600         else
 601                 return copy_page_to_iter_pipe(page, offset, bytes, i);
 602 }
 603 EXPORT_SYMBOL(copy_page_to_iter);
 604
 605 size_t copy_page_from_iter(struct page *page, size_t offset, size_t bytes,
 606                          struct iov_iter *i)
 607 {
 608         if (unlikely(i->type & ITER_PIPE)) {
 609                 WARN_ON(1);
 610                 return 0;
 611         }
 612         if (i->type & (ITER_BVEC|ITER_KVEC)) {
 613                 void *kaddr = kmap_atomic(page);
 614                 size_t wanted = copy_from_iter(kaddr + offset, bytes, i);
 615                 kunmap_atomic(kaddr);
 616                 return wanted;
 617         } else
 618                 return copy_page_from_iter_iovec(page, offset, bytes, i);
 619 }
 620 EXPORT_SYMBOL(copy_page_from_iter);
 621
 622 static size_t pipe_zero(size_t bytes, struct iov_iter *i)
 623 {
 624         struct pipe_inode_info *pipe = i->pipe;
 625         size_t n, off;
 626         int idx;
 627
 628         if (!sanity(i))
 629                 return 0;
 630
 631         bytes = n = push_pipe(i, bytes, &idx, &off);
 632         if (unlikely(!n))
 633                 return 0;
 634
 635         for ( ; n; idx = next_idx(idx, pipe), off = 0) {
 636                 size_t chunk = min_t(size_t, n, PAGE_SIZE - off);
 637                 memzero_page(pipe->bufs[idx].page, off, chunk);
 638                 i->idx = idx;
 639                 i->iov_offset = off + chunk;
 640                 n -= chunk;
 641         }
 642         i->count -= bytes;
 643         return bytes;
 644 }
 645
 646 size_t iov_iter_zero(size_t bytes, struct iov_iter *i)
 647 {
 648         if (unlikely(i->type & ITER_PIPE))
 649                 return pipe_zero(bytes, i);
 650         iterate_and_advance(i, bytes, v,
 651                 __clear_user(v.iov_base, v.iov_len),
 652                 memzero_page(v.bv_page, v.bv_offset, v.bv_len),
 653                 memset(v.iov_base, 0, v.iov_len)
 654         )
 655
 656         return bytes;
 657 }
 658 EXPORT_SYMBOL(iov_iter_zero);
 659
 660 size_t iov_iter_copy_from_user_atomic(struct page *page,
 661                 struct iov_iter *i, unsigned long offset, size_t bytes)
 662 {
 663         char *kaddr = kmap_atomic(page), *p = kaddr + offset;
 664         if (unlikely(i->type & ITER_PIPE)) {
 665                 kunmap_atomic(kaddr);
 666                 WARN_ON(1);
 667                 return 0;
 668         }
 669         iterate_all_kinds(i, bytes, v,
 670                 __copy_from_user_inatomic((p += v.iov_len) - v.iov_len,
 671                                           v.iov_base, v.iov_len),
 672                 memcpy_from_page((p += v.bv_len) - v.bv_len, v.bv_page,
 673                                  v.bv_offset, v.bv_len),
 674                 memcpy((p += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
 675         )
 676         kunmap_atomic(kaddr);
 677         return bytes;
 678 }
 679 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
 680
 681 static inline void pipe_truncate(struct iov_iter *i)
 682 {
 683         struct pipe_inode_info *pipe = i->pipe;
 684         if (pipe->nrbufs) {
 685                 size_t off = i->iov_offset;
 686                 int idx = i->idx;
 687                 int nrbufs = (idx - pipe->curbuf) & (pipe->buffers - 1);
 688                 if (off) {
 689                         pipe->bufs[idx].len = off - pipe->bufs[idx].offset;
 690                         idx = next_idx(idx, pipe);
 691                         nrbufs++;
 692                 }
 693                 while (pipe->nrbufs > nrbufs) {
 694                         pipe_buf_release(pipe, &pipe->bufs[idx]);
 695                         idx = next_idx(idx, pipe);
 696                         pipe->nrbufs--;
 697                 }
 698         }
 699 }
 700
 701 static void pipe_advance(struct iov_iter *i, size_t size)
 702 {
 703         struct pipe_inode_info *pipe = i->pipe;
 704         if (unlikely(i->count < size))
 705                 size = i->count;
 706         if (size) {
 707                 struct pipe_buffer *buf;
 708                 size_t off = i->iov_offset, left = size;
 709                 int idx = i->idx;
 710                 if (off) /* make it relative to the beginning of buffer */
 711                         left += off - pipe->bufs[idx].offset;
 712                 while (1) {
 713                         buf = &pipe->bufs[idx];
 714                         if (left <= buf->len)
 715                                 break;
 716                         left -= buf->len;
 717                         idx = next_idx(idx, pipe);
 718                 }
 719                 i->idx = idx;
 720                 i->iov_offset = buf->offset + left;
 721         }
 722         i->count -= size;
 723         /* ... and discard everything past that point */
 724         pipe_truncate(i);
 725 }
 726
 727 void iov_iter_advance(struct iov_iter *i, size_t size)
 728 {
 729         if (unlikely(i->type & ITER_PIPE)) {
 730                 pipe_advance(i, size);
 731                 return;
 732         }
 733         iterate_and_advance(i, size, v, 0, 0, 0)
 734 }
 735 EXPORT_SYMBOL(iov_iter_advance);
 736
 737 void iov_iter_revert(struct iov_iter *i, size_t unroll)
 738 {
 739         if (!unroll)
 740                 return;
 741         i->count += unroll;
 742         if (unlikely(i->type & ITER_PIPE)) {
 743                 struct pipe_inode_info *pipe = i->pipe;
 744                 int idx = i->idx;
 745                 size_t off = i->iov_offset;
 746                 while (1) {
 747                         size_t n = off - pipe->bufs[idx].offset;
 748                         if (unroll < n) {
 749                                 off -= (n - unroll);
 750                                 break;
 751                         }
 752                         unroll -= n;
 753                         if (!unroll && idx == i->start_idx) {
 754                                 off = 0;
 755                                 break;
 756                         }
 757                         if (!idx--)
 758                                 idx = pipe->buffers - 1;
 759                         off = pipe->bufs[idx].offset + pipe->bufs[idx].len;
 760                 }
 761                 i->iov_offset = off;
 762                 i->idx = idx;
 763                 pipe_truncate(i);
 764                 return;
 765         }
 766         if (unroll <= i->iov_offset) {
 767                 i->iov_offset -= unroll;
 768                 return;
 769         }
 770         unroll -= i->iov_offset;
 771         if (i->type & ITER_BVEC) {
 772                 const struct bio_vec *bvec = i->bvec;
 773                 while (1) {
 774                         size_t n = (--bvec)->bv_len;
 775                         i->nr_segs++;
 776                         if (unroll <= n) {
 777                                 i->bvec = bvec;
 778                                 i->iov_offset = n - unroll;
 779                                 return;
 780                         }
 781                         unroll -= n;
 782                 }
 783         } else { /* same logics for iovec and kvec */
 784                 const struct iovec *iov = i->iov;
 785                 while (1) {
 786                         size_t n = (--iov)->iov_len;
 787                         i->nr_segs++;
 788                         if (unroll <= n) {
 789                                 i->iov = iov;
 790                                 i->iov_offset = n - unroll;
 791                                 return;
 792                         }
 793                         unroll -= n;
 794                 }
 795         }
 796 }
 797 EXPORT_SYMBOL(iov_iter_revert);
 798
 799 /*
 800  * Return the count of just the current iov_iter segment.
 801  */
 802 size_t iov_iter_single_seg_count(const struct iov_iter *i)
 803 {
 804         if (unlikely(i->type & ITER_PIPE))
 805                 return i->count;        // it is a silly place, anyway
 806         if (i->nr_segs == 1)
 807                 return i->count;
 808         else if (i->type & ITER_BVEC)
 809                 return min(i->count, i->bvec->bv_len - i->iov_offset);
 810         else
 811                 return min(i->count, i->iov->iov_len - i->iov_offset);
 812 }
 813 EXPORT_SYMBOL(iov_iter_single_seg_count);
 814
 815 void iov_iter_kvec(struct iov_iter *i, int direction,
 816                         const struct kvec *kvec, unsigned long nr_segs,
 817                         size_t count)
 818 {
 819         BUG_ON(!(direction & ITER_KVEC));
 820         i->type = direction;
 821         i->kvec = kvec;
 822         i->nr_segs = nr_segs;
 823         i->iov_offset = 0;
 824         i->count = count;
 825 }
 826 EXPORT_SYMBOL(iov_iter_kvec);
 827
 828 void iov_iter_bvec(struct iov_iter *i, int direction,
 829                         const struct bio_vec *bvec, unsigned long nr_segs,
 830                         size_t count)
 831 {
 832         BUG_ON(!(direction & ITER_BVEC));
 833         i->type = direction;
 834         i->bvec = bvec;
 835         i->nr_segs = nr_segs;
 836         i->iov_offset = 0;
 837         i->count = count;
 838 }
 839 EXPORT_SYMBOL(iov_iter_bvec);
 840
 841 void iov_iter_pipe(struct iov_iter *i, int direction,
 842                         struct pipe_inode_info *pipe,
 843                         size_t count)
 844 {
 845         BUG_ON(direction != ITER_PIPE);
 846         WARN_ON(pipe->nrbufs == pipe->buffers);
 847         i->type = direction;
 848         i->pipe = pipe;
 849         i->idx = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 850         i->iov_offset = 0;
 851         i->count = count;
 852         i->start_idx = i->idx;
 853 }
 854 EXPORT_SYMBOL(iov_iter_pipe);
 855
 856 unsigned long iov_iter_alignment(const struct iov_iter *i)
 857 {
 858         unsigned long res = 0;
 859         size_t size = i->count;
 860
 861         if (!size)
 862                 return 0;
 863
 864         if (unlikely(i->type & ITER_PIPE)) {
 865                 if (i->iov_offset && allocated(&i->pipe->bufs[i->idx]))
 866                         return size | i->iov_offset;
 867                 return size;
 868         }
 869         iterate_all_kinds(i, size, v,
 870                 (res |= (unsigned long)v.iov_base | v.iov_len, 0),
 871                 res |= v.bv_offset | v.bv_len,
 872                 res |= (unsigned long)v.iov_base | v.iov_len
 873         )
 874         return res;
 875 }
 876 EXPORT_SYMBOL(iov_iter_alignment);
 877
 878 unsigned long iov_iter_gap_alignment(const struct iov_iter *i)
 879 {
 880         unsigned long res = 0;
 881         size_t size = i->count;
 882         if (!size)
 883                 return 0;
 884
 885         if (unlikely(i->type & ITER_PIPE)) {
 886                 WARN_ON(1);
 887                 return ~0U;
 888         }
 889
 890         iterate_all_kinds(i, size, v,
 891                 (res |= (!res ? 0 : (unsigned long)v.iov_base) |
 892                         (size != v.iov_len ? size : 0), 0),
 893                 (res |= (!res ? 0 : (unsigned long)v.bv_offset) |
 894                         (size != v.bv_len ? size : 0)),
 895                 (res |= (!res ? 0 : (unsigned long)v.iov_base) |
 896                         (size != v.iov_len ? size : 0))
 897                 );
 898                 return res;
 899 }
 900 EXPORT_SYMBOL(iov_iter_gap_alignment);
 901
 902 static inline size_t __pipe_get_pages(struct iov_iter *i,
 903                                 size_t maxsize,
 904                                 struct page **pages,
 905                                 int idx,
 906                                 size_t *start)
 907 {
 908         struct pipe_inode_info *pipe = i->pipe;
 909         ssize_t n = push_pipe(i, maxsize, &idx, start);
 910         if (!n)
 911                 return -EFAULT;
 912
 913         maxsize = n;
 914         n += *start;
 915         while (n > 0) {
 916                 get_page(*pages++ = pipe->bufs[idx].page);
 917                 idx = next_idx(idx, pipe);
 918                 n -= PAGE_SIZE;
 919         }
 920
 921         return maxsize;
 922 }
 923
 924 static ssize_t pipe_get_pages(struct iov_iter *i,
 925                    struct page **pages, size_t maxsize, unsigned maxpages,
 926                    size_t *start)
 927 {
 928         unsigned npages;
 929         size_t capacity;
 930         int idx;
 931
 932         if (!sanity(i))
 933                 return -EFAULT;
 934
 935         data_start(i, &idx, start);
 936         /* some of this one + all after this one */
 937         npages = ((i->pipe->curbuf - idx - 1) & (i->pipe->buffers - 1)) + 1;
 938         capacity = min(npages,maxpages) * PAGE_SIZE - *start;
 939
 940         return __pipe_get_pages(i, min(maxsize, capacity), pages, idx, start);
 941 }
 942
 943 ssize_t iov_iter_get_pages(struct iov_iter *i,
 944                    struct page **pages, size_t maxsize, unsigned maxpages,
 945                    size_t *start)
 946 {
 947         if (maxsize > i->count)
 948                 maxsize = i->count;
 949
 950         if (!maxsize)
 951                 return 0;
 952
 953         if (unlikely(i->type & ITER_PIPE))
 954                 return pipe_get_pages(i, pages, maxsize, maxpages, start);
 955         iterate_all_kinds(i, maxsize, v, ({
 956                 unsigned long addr = (unsigned long)v.iov_base;
 957                 size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1));
 958                 int n;
 959                 int res;
 960
 961                 if (len > maxpages * PAGE_SIZE)
 962                         len = maxpages * PAGE_SIZE;
 963                 addr &= ~(PAGE_SIZE - 1);
 964                 n = DIV_ROUND_UP(len, PAGE_SIZE);
 965                 res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, pages);
 966                 if (unlikely(res < 0))
 967                         return res;
 968                 return (res == n ? len : res * PAGE_SIZE) - *start;
 969         0;}),({
 970                 /* can't be more than PAGE_SIZE */
 971                 *start = v.bv_offset;
 972                 get_page(*pages = v.bv_page);
 973                 return v.bv_len;
 974         }),({
 975                 return -EFAULT;
 976         })
 977         )
 978         return 0;
 979 }
 980 EXPORT_SYMBOL(iov_iter_get_pages);
 981
 982 static struct page **get_pages_array(size_t n)
 983 {
 984         struct page **p = kmalloc(n * sizeof(struct page *), GFP_KERNEL);
 985         if (!p)
 986                 p = vmalloc(n * sizeof(struct page *));
 987         return p;
 988 }
 989
 990 static ssize_t pipe_get_pages_alloc(struct iov_iter *i,
 991                    struct page ***pages, size_t maxsize,
 992                    size_t *start)
 993 {
 994         struct page **p;
 995         size_t n;
 996         int idx;
 997         int npages;
 998
 999         if (!sanity(i))
1000                 return -EFAULT;
1001
1002         data_start(i, &idx, start);
1003         /* some of this one + all after this one */
1004         npages = ((i->pipe->curbuf - idx - 1) & (i->pipe->buffers - 1)) + 1;
1005         n = npages * PAGE_SIZE - *start;
1006         if (maxsize > n)
1007                 maxsize = n;
1008         else
1009                 npages = DIV_ROUND_UP(maxsize + *start, PAGE_SIZE);
1010         p = get_pages_array(npages);
1011         if (!p)
1012                 return -ENOMEM;
1013         n = __pipe_get_pages(i, maxsize, p, idx, start);
1014         if (n > 0)
1015                 *pages = p;
1016         else
1017                 kvfree(p);
1018         return n;
1019 }
1020
1021 ssize_t iov_iter_get_pages_alloc(struct iov_iter *i,
1022                    struct page ***pages, size_t maxsize,
1023                    size_t *start)
1024 {
1025         struct page **p;
1026
1027         if (maxsize > i->count)
1028                 maxsize = i->count;
1029
1030         if (!maxsize)
1031                 return 0;
1032
1033         if (unlikely(i->type & ITER_PIPE))
1034                 return pipe_get_pages_alloc(i, pages, maxsize, start);
1035         iterate_all_kinds(i, maxsize, v, ({
1036                 unsigned long addr = (unsigned long)v.iov_base;
1037                 size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1));
1038                 int n;
1039                 int res;
1040
1041                 addr &= ~(PAGE_SIZE - 1);
1042                 n = DIV_ROUND_UP(len, PAGE_SIZE);
1043                 p = get_pages_array(n);
1044                 if (!p)
1045                         return -ENOMEM;
1046                 res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, p);
1047                 if (unlikely(res < 0)) {
1048                         kvfree(p);
1049                         return res;
1050                 }
1051                 *pages = p;
1052                 return (res == n ? len : res * PAGE_SIZE) - *start;
1053         0;}),({
1054                 /* can't be more than PAGE_SIZE */
1055                 *start = v.bv_offset;
1056                 *pages = p = get_pages_array(1);
1057                 if (!p)
1058                         return -ENOMEM;
1059                 get_page(*p = v.bv_page);
1060                 return v.bv_len;
1061         }),({
1062                 return -EFAULT;
1063         })
1064         )
1065         return 0;
1066 }
1067 EXPORT_SYMBOL(iov_iter_get_pages_alloc);
1068
1069 size_t csum_and_copy_from_iter(void *addr, size_t bytes, __wsum *csum,
1070                                struct iov_iter *i)
1071 {
1072         char *to = addr;
1073         __wsum sum, next;
1074         size_t off = 0;
1075         sum = *csum;
1076         if (unlikely(i->type & ITER_PIPE)) {
1077                 WARN_ON(1);
1078                 return 0;
1079         }
1080         iterate_and_advance(i, bytes, v, ({
1081                 int err = 0;
1082                 next = csum_and_copy_from_user(v.iov_base,
1083                                                (to += v.iov_len) - v.iov_len,
1084                                                v.iov_len, 0, &err);
1085                 if (!err) {
1086                         sum = csum_block_add(sum, next, off);
1087                         off += v.iov_len;
1088                 }
1089                 err ? v.iov_len : 0;
1090         }), ({
1091                 char *p = kmap_atomic(v.bv_page);
1092                 next = csum_partial_copy_nocheck(p + v.bv_offset,
1093                                                  (to += v.bv_len) - v.bv_len,
1094                                                  v.bv_len, 0);
1095                 kunmap_atomic(p);
1096                 sum = csum_block_add(sum, next, off);
1097                 off += v.bv_len;
1098         }),({
1099                 next = csum_partial_copy_nocheck(v.iov_base,
1100                                                  (to += v.iov_len) - v.iov_len,
1101                                                  v.iov_len, 0);
1102                 sum = csum_block_add(sum, next, off);
1103                 off += v.iov_len;
1104         })
1105         )
1106         *csum = sum;
1107         return bytes;
1108 }
1109 EXPORT_SYMBOL(csum_and_copy_from_iter);
1110
1111 size_t csum_and_copy_to_iter(const void *addr, size_t bytes, __wsum *csum,
1112                              struct iov_iter *i)
1113 {
1114         const char *from = addr;
1115         __wsum sum, next;
1116         size_t off = 0;
1117         sum = *csum;
1118         if (unlikely(i->type & ITER_PIPE)) {
1119                 WARN_ON(1);     /* for now */
1120                 return 0;
1121         }
1122         iterate_and_advance(i, bytes, v, ({
1123                 int err = 0;
1124                 next = csum_and_copy_to_user((from += v.iov_len) - v.iov_len,
1125                                              v.iov_base,
1126                                              v.iov_len, 0, &err);
1127                 if (!err) {
1128                         sum = csum_block_add(sum, next, off);
1129                         off += v.iov_len;
1130                 }
1131                 err ? v.iov_len : 0;
1132         }), ({
1133                 char *p = kmap_atomic(v.bv_page);
1134                 next = csum_partial_copy_nocheck((from += v.bv_len) - v.bv_len,
1135                                                  p + v.bv_offset,
1136                                                  v.bv_len, 0);
1137                 kunmap_atomic(p);
1138                 sum = csum_block_add(sum, next, off);
1139                 off += v.bv_len;
1140         }),({
1141                 next = csum_partial_copy_nocheck((from += v.iov_len) - v.iov_len,
1142                                                  v.iov_base,
1143                                                  v.iov_len, 0);
1144                 sum = csum_block_add(sum, next, off);
1145                 off += v.iov_len;
1146         })
1147         )
1148         *csum = sum;
1149         return bytes;
1150 }
1151 EXPORT_SYMBOL(csum_and_copy_to_iter);
1152
1153 int iov_iter_npages(const struct iov_iter *i, int maxpages)
1154 {
1155         size_t size = i->count;
1156         int npages = 0;
1157
1158         if (!size)
1159                 return 0;
1160
1161         if (unlikely(i->type & ITER_PIPE)) {
1162                 struct pipe_inode_info *pipe = i->pipe;
1163                 size_t off;
1164                 int idx;
1165
1166                 if (!sanity(i))
1167                         return 0;
1168
1169                 data_start(i, &idx, &off);
1170                 /* some of this one + all after this one */
1171                 npages = ((pipe->curbuf - idx - 1) & (pipe->buffers - 1)) + 1;
1172                 if (npages >= maxpages)
1173                         return maxpages;
1174         } else iterate_all_kinds(i, size, v, ({
1175                 unsigned long p = (unsigned long)v.iov_base;
1176                 npages += DIV_ROUND_UP(p + v.iov_len, PAGE_SIZE)
1177                         - p / PAGE_SIZE;
1178                 if (npages >= maxpages)
1179                         return maxpages;
1180         0;}),({
1181                 npages++;
1182                 if (npages >= maxpages)
1183                         return maxpages;
1184         }),({
1185                 unsigned long p = (unsigned long)v.iov_base;
1186                 npages += DIV_ROUND_UP(p + v.iov_len, PAGE_SIZE)
1187                         - p / PAGE_SIZE;
1188                 if (npages >= maxpages)
1189                         return maxpages;
1190         })
1191         )
1192         return npages;
1193 }
1194 EXPORT_SYMBOL(iov_iter_npages);
1195
1196 const void *dup_iter(struct iov_iter *new, struct iov_iter *old, gfp_t flags)
1197 {
1198         *new = *old;
1199         if (unlikely(new->type & ITER_PIPE)) {
1200                 WARN_ON(1);
1201                 return NULL;
1202         }
1203         if (new->type & ITER_BVEC)
1204                 return new->bvec = kmemdup(new->bvec,
1205                                     new->nr_segs * sizeof(struct bio_vec),
1206                                     flags);
1207         else
1208                 /* iovec and kvec have identical layout */
1209                 return new->iov = kmemdup(new->iov,
1210                                    new->nr_segs * sizeof(struct iovec),
1211                                    flags);
1212 }
1213 EXPORT_SYMBOL(dup_iter);
1214
1215 /**
1216  * import_iovec() - Copy an array of &struct iovec from userspace
1217  *     into the kernel, check that it is valid, and initialize a new
1218  *     &struct iov_iter iterator to access it.
1219  *
1220  * @type: One of %READ or %WRITE.
1221  * @uvector: Pointer to the userspace array.
1222  * @nr_segs: Number of elements in userspace array.
1223  * @fast_segs: Number of elements in @iov.
1224  * @iov: (input and output parameter) Pointer to pointer to (usually small
1225  *     on-stack) kernel array.
1226  * @i: Pointer to iterator that will be initialized on success.
1227  *
1228  * If the array pointed to by *@iov is large enough to hold all @nr_segs,
1229  * then this function places %NULL in *@iov on return. Otherwise, a new
1230  * array will be allocated and the result placed in *@iov. This means that
1231  * the caller may call kfree() on *@iov regardless of whether the small
1232  * on-stack array was used or not (and regardless of whether this function
1233  * returns an error or not).
1234  *
1235  * Return: 0 on success or negative error code on error.
1236  */
1237 int import_iovec(int type, const struct iovec __user * uvector,
1238                  unsigned nr_segs, unsigned fast_segs,
1239                  struct iovec **iov, struct iov_iter *i)
1240 {
1241         ssize_t n;
1242         struct iovec *p;
1243         n = rw_copy_check_uvector(type, uvector, nr_segs, fast_segs,
1244                                   *iov, &p);
1245         if (n < 0) {
1246                 if (p != *iov)
1247                         kfree(p);
1248                 *iov = NULL;
1249                 return n;
1250         }
1251         iov_iter_init(i, type, p, nr_segs, n);
1252         *iov = p == *iov ? NULL : p;
1253         return 0;
1254 }
1255 EXPORT_SYMBOL(import_iovec);
1256
1257 #ifdef CONFIG_COMPAT
1258 #include <linux/compat.h>
1259
1260 int compat_import_iovec(int type, const struct compat_iovec __user * uvector,
1261                  unsigned nr_segs, unsigned fast_segs,
1262                  struct iovec **iov, struct iov_iter *i)
1263 {
1264         ssize_t n;
1265         struct iovec *p;
1266         n = compat_rw_copy_check_uvector(type, uvector, nr_segs, fast_segs,
1267                                   *iov, &p);
1268         if (n < 0) {
1269                 if (p != *iov)
1270                         kfree(p);
1271                 *iov = NULL;
1272                 return n;
1273         }
1274         iov_iter_init(i, type, p, nr_segs, n);
1275         *iov = p == *iov ? NULL : p;
1276         return 0;
1277 }
1278 #endif
1279
1280 int import_single_range(int rw, void __user *buf, size_t len,
1281                  struct iovec *iov, struct iov_iter *i)
1282 {
1283         if (len > MAX_RW_COUNT)
1284                 len = MAX_RW_COUNT;
1285         if (unlikely(!access_ok(!rw, buf, len)))
1286                 return -EFAULT;
1287
1288         iov->iov_base = buf;
1289         iov->iov_len = len;
1290         iov_iter_init(i, rw, iov, 1, len);
1291         return 0;
1292 }
1293 EXPORT_SYMBOL(import_single_range);