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_multipages_readable(v.iov_base,
 400                                         v.iov_len);
 401                         if (unlikely(err))
 402                         return err;
 403                 0;}))
 404         }
 405         return 0;
 406 }
 407 EXPORT_SYMBOL(iov_iter_fault_in_readable);
 408
 409 void iov_iter_init(struct iov_iter *i, int direction,
 410                         const struct iovec *iov, unsigned long nr_segs,
 411                         size_t count)
 412 {
 413         /* It will get better.  Eventually... */
 414         if (segment_eq(get_fs(), KERNEL_DS)) {
 415                 direction |= ITER_KVEC;
 416                 i->type = direction;
 417                 i->kvec = (struct kvec *)iov;
 418         } else {
 419                 i->type = direction;
 420                 i->iov = iov;
 421         }
 422         i->nr_segs = nr_segs;
 423         i->iov_offset = 0;
 424         i->count = count;
 425 }
 426 EXPORT_SYMBOL(iov_iter_init);
 427
 428 static void memcpy_from_page(char *to, struct page *page, size_t offset, size_t len)
 429 {
 430         char *from = kmap_atomic(page);
 431         memcpy(to, from + offset, len);
 432         kunmap_atomic(from);
 433 }
 434
 435 static void memcpy_to_page(struct page *page, size_t offset, const char *from, size_t len)
 436 {
 437         char *to = kmap_atomic(page);
 438         memcpy(to + offset, from, len);
 439         kunmap_atomic(to);
 440 }
 441
 442 static void memzero_page(struct page *page, size_t offset, size_t len)
 443 {
 444         char *addr = kmap_atomic(page);
 445         memset(addr + offset, 0, len);
 446         kunmap_atomic(addr);
 447 }
 448
 449 static inline bool allocated(struct pipe_buffer *buf)
 450 {
 451         return buf->ops == &default_pipe_buf_ops;
 452 }
 453
 454 static inline void data_start(const struct iov_iter *i, int *idxp, size_t *offp)
 455 {
 456         size_t off = i->iov_offset;
 457         int idx = i->idx;
 458         if (off && (!allocated(&i->pipe->bufs[idx]) || off == PAGE_SIZE)) {
 459                 idx = next_idx(idx, i->pipe);
 460                 off = 0;
 461         }
 462         *idxp = idx;
 463         *offp = off;
 464 }
 465
 466 static size_t push_pipe(struct iov_iter *i, size_t size,
 467                         int *idxp, size_t *offp)
 468 {
 469         struct pipe_inode_info *pipe = i->pipe;
 470         size_t off;
 471         int idx;
 472         ssize_t left;
 473
 474         if (unlikely(size > i->count))
 475                 size = i->count;
 476         if (unlikely(!size))
 477                 return 0;
 478
 479         left = size;
 480         data_start(i, &idx, &off);
 481         *idxp = idx;
 482         *offp = off;
 483         if (off) {
 484                 left -= PAGE_SIZE - off;
 485                 if (left <= 0) {
 486                         pipe->bufs[idx].len += size;
 487                         return size;
 488                 }
 489                 pipe->bufs[idx].len = PAGE_SIZE;
 490                 idx = next_idx(idx, pipe);
 491         }
 492         while (idx != pipe->curbuf || !pipe->nrbufs) {
 493                 struct page *page = alloc_page(GFP_USER);
 494                 if (!page)
 495                         break;
 496                 pipe->nrbufs++;
 497                 pipe->bufs[idx].ops = &default_pipe_buf_ops;
 498                 pipe->bufs[idx].page = page;
 499                 pipe->bufs[idx].offset = 0;
 500                 if (left <= PAGE_SIZE) {
 501                         pipe->bufs[idx].len = left;
 502                         return size;
 503                 }
 504                 pipe->bufs[idx].len = PAGE_SIZE;
 505                 left -= PAGE_SIZE;
 506                 idx = next_idx(idx, pipe);
 507         }
 508         return size - left;
 509 }
 510
 511 static size_t copy_pipe_to_iter(const void *addr, size_t bytes,
 512                                 struct iov_iter *i)
 513 {
 514         struct pipe_inode_info *pipe = i->pipe;
 515         size_t n, off;
 516         int idx;
 517
 518         if (!sanity(i))
 519                 return 0;
 520
 521         bytes = n = push_pipe(i, bytes, &idx, &off);
 522         if (unlikely(!n))
 523                 return 0;
 524         for ( ; n; idx = next_idx(idx, pipe), off = 0) {
 525                 size_t chunk = min_t(size_t, n, PAGE_SIZE - off);
 526                 memcpy_to_page(pipe->bufs[idx].page, off, addr, chunk);
 527                 i->idx = idx;
 528                 i->iov_offset = off + chunk;
 529                 n -= chunk;
 530                 addr += chunk;
 531         }
 532         i->count -= bytes;
 533         return bytes;
 534 }
 535
 536 size_t copy_to_iter(const void *addr, size_t bytes, struct iov_iter *i)
 537 {
 538         const char *from = addr;
 539         if (unlikely(i->type & ITER_PIPE))
 540                 return copy_pipe_to_iter(addr, bytes, i);
 541         iterate_and_advance(i, bytes, v,
 542                 __copy_to_user(v.iov_base, (from += v.iov_len) - v.iov_len,
 543                                v.iov_len),
 544                 memcpy_to_page(v.bv_page, v.bv_offset,
 545                                (from += v.bv_len) - v.bv_len, v.bv_len),
 546                 memcpy(v.iov_base, (from += v.iov_len) - v.iov_len, v.iov_len)
 547         )
 548
 549         return bytes;
 550 }
 551 EXPORT_SYMBOL(copy_to_iter);
 552
 553 size_t copy_from_iter(void *addr, size_t bytes, struct iov_iter *i)
 554 {
 555         char *to = addr;
 556         if (unlikely(i->type & ITER_PIPE)) {
 557                 WARN_ON(1);
 558                 return 0;
 559         }
 560         iterate_and_advance(i, bytes, v,
 561                 __copy_from_user((to += v.iov_len) - v.iov_len, v.iov_base,
 562                                  v.iov_len),
 563                 memcpy_from_page((to += v.bv_len) - v.bv_len, v.bv_page,
 564                                  v.bv_offset, v.bv_len),
 565                 memcpy((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
 566         )
 567
 568         return bytes;
 569 }
 570 EXPORT_SYMBOL(copy_from_iter);
 571
 572 size_t copy_from_iter_nocache(void *addr, size_t bytes, struct iov_iter *i)
 573 {
 574         char *to = addr;
 575         if (unlikely(i->type & ITER_PIPE)) {
 576                 WARN_ON(1);
 577                 return 0;
 578         }
 579         iterate_and_advance(i, bytes, v,
 580                 __copy_from_user_nocache((to += v.iov_len) - v.iov_len,
 581                                          v.iov_base, v.iov_len),
 582                 memcpy_from_page((to += v.bv_len) - v.bv_len, v.bv_page,
 583                                  v.bv_offset, v.bv_len),
 584                 memcpy((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
 585         )
 586
 587         return bytes;
 588 }
 589 EXPORT_SYMBOL(copy_from_iter_nocache);
 590
 591 size_t copy_page_to_iter(struct page *page, size_t offset, size_t bytes,
 592                          struct iov_iter *i)
 593 {
 594         if (i->type & (ITER_BVEC|ITER_KVEC)) {
 595                 void *kaddr = kmap_atomic(page);
 596                 size_t wanted = copy_to_iter(kaddr + offset, bytes, i);
 597                 kunmap_atomic(kaddr);
 598                 return wanted;
 599         } else if (likely(!(i->type & ITER_PIPE)))
 600                 return copy_page_to_iter_iovec(page, offset, bytes, i);
 601         else
 602                 return copy_page_to_iter_pipe(page, offset, bytes, i);
 603 }
 604 EXPORT_SYMBOL(copy_page_to_iter);
 605
 606 size_t copy_page_from_iter(struct page *page, size_t offset, size_t bytes,
 607                          struct iov_iter *i)
 608 {
 609         if (unlikely(i->type & ITER_PIPE)) {
 610                 WARN_ON(1);
 611                 return 0;
 612         }
 613         if (i->type & (ITER_BVEC|ITER_KVEC)) {
 614                 void *kaddr = kmap_atomic(page);
 615                 size_t wanted = copy_from_iter(kaddr + offset, bytes, i);
 616                 kunmap_atomic(kaddr);
 617                 return wanted;
 618         } else
 619                 return copy_page_from_iter_iovec(page, offset, bytes, i);
 620 }
 621 EXPORT_SYMBOL(copy_page_from_iter);
 622
 623 static size_t pipe_zero(size_t bytes, struct iov_iter *i)
 624 {
 625         struct pipe_inode_info *pipe = i->pipe;
 626         size_t n, off;
 627         int idx;
 628
 629         if (!sanity(i))
 630                 return 0;
 631
 632         bytes = n = push_pipe(i, bytes, &idx, &off);
 633         if (unlikely(!n))
 634                 return 0;
 635
 636         for ( ; n; idx = next_idx(idx, pipe), off = 0) {
 637                 size_t chunk = min_t(size_t, n, PAGE_SIZE - off);
 638                 memzero_page(pipe->bufs[idx].page, off, chunk);
 639                 i->idx = idx;
 640                 i->iov_offset = off + chunk;
 641                 n -= chunk;
 642         }
 643         i->count -= bytes;
 644         return bytes;
 645 }
 646
 647 size_t iov_iter_zero(size_t bytes, struct iov_iter *i)
 648 {
 649         if (unlikely(i->type & ITER_PIPE))
 650                 return pipe_zero(bytes, i);
 651         iterate_and_advance(i, bytes, v,
 652                 __clear_user(v.iov_base, v.iov_len),
 653                 memzero_page(v.bv_page, v.bv_offset, v.bv_len),
 654                 memset(v.iov_base, 0, v.iov_len)
 655         )
 656
 657         return bytes;
 658 }
 659 EXPORT_SYMBOL(iov_iter_zero);
 660
 661 size_t iov_iter_copy_from_user_atomic(struct page *page,
 662                 struct iov_iter *i, unsigned long offset, size_t bytes)
 663 {
 664         char *kaddr = kmap_atomic(page), *p = kaddr + offset;
 665         if (unlikely(i->type & ITER_PIPE)) {
 666                 kunmap_atomic(kaddr);
 667                 WARN_ON(1);
 668                 return 0;
 669         }
 670         iterate_all_kinds(i, bytes, v,
 671                 __copy_from_user_inatomic((p += v.iov_len) - v.iov_len,
 672                                           v.iov_base, v.iov_len),
 673                 memcpy_from_page((p += v.bv_len) - v.bv_len, v.bv_page,
 674                                  v.bv_offset, v.bv_len),
 675                 memcpy((p += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
 676         )
 677         kunmap_atomic(kaddr);
 678         return bytes;
 679 }
 680 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
 681
 682 static void pipe_advance(struct iov_iter *i, size_t size)
 683 {
 684         struct pipe_inode_info *pipe = i->pipe;
 685         struct pipe_buffer *buf;
 686         int idx = i->idx;
 687         size_t off = i->iov_offset;
 688
 689         if (unlikely(i->count < size))
 690                 size = i->count;
 691
 692         if (size) {
 693                 if (off) /* make it relative to the beginning of buffer */
 694                         size += off - pipe->bufs[idx].offset;
 695                 while (1) {
 696                         buf = &pipe->bufs[idx];
 697                         if (size <= buf->len)
 698                                 break;
 699                         size -= buf->len;
 700                         idx = next_idx(idx, pipe);
 701                 }
 702                 buf->len = size;
 703                 i->idx = idx;
 704                 off = i->iov_offset = buf->offset + size;
 705         }
 706         if (off)
 707                 idx = next_idx(idx, pipe);
 708         if (pipe->nrbufs) {
 709                 int unused = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 710                 /* [curbuf,unused) is in use.  Free [idx,unused) */
 711                 while (idx != unused) {
 712                         pipe_buf_release(pipe, &pipe->bufs[idx]);
 713                         idx = next_idx(idx, pipe);
 714                         pipe->nrbufs--;
 715                 }
 716         }
 717 }
 718
 719 void iov_iter_advance(struct iov_iter *i, size_t size)
 720 {
 721         if (unlikely(i->type & ITER_PIPE)) {
 722                 pipe_advance(i, size);
 723                 return;
 724         }
 725         iterate_and_advance(i, size, v, 0, 0, 0)
 726 }
 727 EXPORT_SYMBOL(iov_iter_advance);
 728
 729 /*
 730  * Return the count of just the current iov_iter segment.
 731  */
 732 size_t iov_iter_single_seg_count(const struct iov_iter *i)
 733 {
 734         if (unlikely(i->type & ITER_PIPE))
 735                 return i->count;        // it is a silly place, anyway
 736         if (i->nr_segs == 1)
 737                 return i->count;
 738         else if (i->type & ITER_BVEC)
 739                 return min(i->count, i->bvec->bv_len - i->iov_offset);
 740         else
 741                 return min(i->count, i->iov->iov_len - i->iov_offset);
 742 }
 743 EXPORT_SYMBOL(iov_iter_single_seg_count);
 744
 745 void iov_iter_kvec(struct iov_iter *i, int direction,
 746                         const struct kvec *kvec, unsigned long nr_segs,
 747                         size_t count)
 748 {
 749         BUG_ON(!(direction & ITER_KVEC));
 750         i->type = direction;
 751         i->kvec = kvec;
 752         i->nr_segs = nr_segs;
 753         i->iov_offset = 0;
 754         i->count = count;
 755 }
 756 EXPORT_SYMBOL(iov_iter_kvec);
 757
 758 void iov_iter_bvec(struct iov_iter *i, int direction,
 759                         const struct bio_vec *bvec, unsigned long nr_segs,
 760                         size_t count)
 761 {
 762         BUG_ON(!(direction & ITER_BVEC));
 763         i->type = direction;
 764         i->bvec = bvec;
 765         i->nr_segs = nr_segs;
 766         i->iov_offset = 0;
 767         i->count = count;
 768 }
 769 EXPORT_SYMBOL(iov_iter_bvec);
 770
 771 void iov_iter_pipe(struct iov_iter *i, int direction,
 772                         struct pipe_inode_info *pipe,
 773                         size_t count)
 774 {
 775         BUG_ON(direction != ITER_PIPE);
 776         i->type = direction;
 777         i->pipe = pipe;
 778         i->idx = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 779         i->iov_offset = 0;
 780         i->count = count;
 781 }
 782 EXPORT_SYMBOL(iov_iter_pipe);
 783
 784 unsigned long iov_iter_alignment(const struct iov_iter *i)
 785 {
 786         unsigned long res = 0;
 787         size_t size = i->count;
 788
 789         if (!size)
 790                 return 0;
 791
 792         if (unlikely(i->type & ITER_PIPE)) {
 793                 if (i->iov_offset && allocated(&i->pipe->bufs[i->idx]))
 794                         return size | i->iov_offset;
 795                 return size;
 796         }
 797         iterate_all_kinds(i, size, v,
 798                 (res |= (unsigned long)v.iov_base | v.iov_len, 0),
 799                 res |= v.bv_offset | v.bv_len,
 800                 res |= (unsigned long)v.iov_base | v.iov_len
 801         )
 802         return res;
 803 }
 804 EXPORT_SYMBOL(iov_iter_alignment);
 805
 806 unsigned long iov_iter_gap_alignment(const struct iov_iter *i)
 807 {
 808         unsigned long res = 0;
 809         size_t size = i->count;
 810         if (!size)
 811                 return 0;
 812
 813         if (unlikely(i->type & ITER_PIPE)) {
 814                 WARN_ON(1);
 815                 return ~0U;
 816         }
 817
 818         iterate_all_kinds(i, size, v,
 819                 (res |= (!res ? 0 : (unsigned long)v.iov_base) |
 820                         (size != v.iov_len ? size : 0), 0),
 821                 (res |= (!res ? 0 : (unsigned long)v.bv_offset) |
 822                         (size != v.bv_len ? size : 0)),
 823                 (res |= (!res ? 0 : (unsigned long)v.iov_base) |
 824                         (size != v.iov_len ? size : 0))
 825                 );
 826                 return res;
 827 }
 828 EXPORT_SYMBOL(iov_iter_gap_alignment);
 829
 830 static inline size_t __pipe_get_pages(struct iov_iter *i,
 831                                 size_t maxsize,
 832                                 struct page **pages,
 833                                 int idx,
 834                                 size_t *start)
 835 {
 836         struct pipe_inode_info *pipe = i->pipe;
 837         size_t n = push_pipe(i, maxsize, &idx, start);
 838         if (!n)
 839                 return -EFAULT;
 840
 841         maxsize = n;
 842         n += *start;
 843         while (n >= PAGE_SIZE) {
 844                 get_page(*pages++ = pipe->bufs[idx].page);
 845                 idx = next_idx(idx, pipe);
 846                 n -= PAGE_SIZE;
 847         }
 848
 849         return maxsize;
 850 }
 851
 852 static ssize_t pipe_get_pages(struct iov_iter *i,
 853                    struct page **pages, size_t maxsize, unsigned maxpages,
 854                    size_t *start)
 855 {
 856         unsigned npages;
 857         size_t capacity;
 858         int idx;
 859
 860         if (!sanity(i))
 861                 return -EFAULT;
 862
 863         data_start(i, &idx, start);
 864         /* some of this one + all after this one */
 865         npages = ((i->pipe->curbuf - idx - 1) & (i->pipe->buffers - 1)) + 1;
 866         capacity = min(npages,maxpages) * PAGE_SIZE - *start;
 867
 868         return __pipe_get_pages(i, min(maxsize, capacity), pages, idx, start);
 869 }
 870
 871 ssize_t iov_iter_get_pages(struct iov_iter *i,
 872                    struct page **pages, size_t maxsize, unsigned maxpages,
 873                    size_t *start)
 874 {
 875         if (maxsize > i->count)
 876                 maxsize = i->count;
 877
 878         if (!maxsize)
 879                 return 0;
 880
 881         if (unlikely(i->type & ITER_PIPE))
 882                 return pipe_get_pages(i, pages, maxsize, maxpages, start);
 883         iterate_all_kinds(i, maxsize, v, ({
 884                 unsigned long addr = (unsigned long)v.iov_base;
 885                 size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1));
 886                 int n;
 887                 int res;
 888
 889                 if (len > maxpages * PAGE_SIZE)
 890                         len = maxpages * PAGE_SIZE;
 891                 addr &= ~(PAGE_SIZE - 1);
 892                 n = DIV_ROUND_UP(len, PAGE_SIZE);
 893                 res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, pages);
 894                 if (unlikely(res < 0))
 895                         return res;
 896                 return (res == n ? len : res * PAGE_SIZE) - *start;
 897         0;}),({
 898                 /* can't be more than PAGE_SIZE */
 899                 *start = v.bv_offset;
 900                 get_page(*pages = v.bv_page);
 901                 return v.bv_len;
 902         }),({
 903                 return -EFAULT;
 904         })
 905         )
 906         return 0;
 907 }
 908 EXPORT_SYMBOL(iov_iter_get_pages);
 909
 910 static struct page **get_pages_array(size_t n)
 911 {
 912         struct page **p = kmalloc(n * sizeof(struct page *), GFP_KERNEL);
 913         if (!p)
 914                 p = vmalloc(n * sizeof(struct page *));
 915         return p;
 916 }
 917
 918 static ssize_t pipe_get_pages_alloc(struct iov_iter *i,
 919                    struct page ***pages, size_t maxsize,
 920                    size_t *start)
 921 {
 922         struct page **p;
 923         size_t n;
 924         int idx;
 925         int npages;
 926
 927         if (!sanity(i))
 928                 return -EFAULT;
 929
 930         data_start(i, &idx, start);
 931         /* some of this one + all after this one */
 932         npages = ((i->pipe->curbuf - idx - 1) & (i->pipe->buffers - 1)) + 1;
 933         n = npages * PAGE_SIZE - *start;
 934         if (maxsize > n)
 935                 maxsize = n;
 936         else
 937                 npages = DIV_ROUND_UP(maxsize + *start, PAGE_SIZE);
 938         p = get_pages_array(npages);
 939         if (!p)
 940                 return -ENOMEM;
 941         n = __pipe_get_pages(i, maxsize, p, idx, start);
 942         if (n > 0)
 943                 *pages = p;
 944         else
 945                 kvfree(p);
 946         return n;
 947 }
 948
 949 ssize_t iov_iter_get_pages_alloc(struct iov_iter *i,
 950                    struct page ***pages, size_t maxsize,
 951                    size_t *start)
 952 {
 953         struct page **p;
 954
 955         if (maxsize > i->count)
 956                 maxsize = i->count;
 957
 958         if (!maxsize)
 959                 return 0;
 960
 961         if (unlikely(i->type & ITER_PIPE))
 962                 return pipe_get_pages_alloc(i, pages, maxsize, start);
 963         iterate_all_kinds(i, maxsize, v, ({
 964                 unsigned long addr = (unsigned long)v.iov_base;
 965                 size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1));
 966                 int n;
 967                 int res;
 968
 969                 addr &= ~(PAGE_SIZE - 1);
 970                 n = DIV_ROUND_UP(len, PAGE_SIZE);
 971                 p = get_pages_array(n);
 972                 if (!p)
 973                         return -ENOMEM;
 974                 res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, p);
 975                 if (unlikely(res < 0)) {
 976                         kvfree(p);
 977                         return res;
 978                 }
 979                 *pages = p;
 980                 return (res == n ? len : res * PAGE_SIZE) - *start;
 981         0;}),({
 982                 /* can't be more than PAGE_SIZE */
 983                 *start = v.bv_offset;
 984                 *pages = p = get_pages_array(1);
 985                 if (!p)
 986                         return -ENOMEM;
 987                 get_page(*p = v.bv_page);
 988                 return v.bv_len;
 989         }),({
 990                 return -EFAULT;
 991         })
 992         )
 993         return 0;
 994 }
 995 EXPORT_SYMBOL(iov_iter_get_pages_alloc);
 996
 997 size_t csum_and_copy_from_iter(void *addr, size_t bytes, __wsum *csum,
 998                                struct iov_iter *i)
 999 {
1000         char *to = addr;
1001         __wsum sum, next;
1002         size_t off = 0;
1003         sum = *csum;
1004         if (unlikely(i->type & ITER_PIPE)) {
1005                 WARN_ON(1);
1006                 return 0;
1007         }
1008         iterate_and_advance(i, bytes, v, ({
1009                 int err = 0;
1010                 next = csum_and_copy_from_user(v.iov_base,
1011                                                (to += v.iov_len) - v.iov_len,
1012                                                v.iov_len, 0, &err);
1013                 if (!err) {
1014                         sum = csum_block_add(sum, next, off);
1015                         off += v.iov_len;
1016                 }
1017                 err ? v.iov_len : 0;
1018         }), ({
1019                 char *p = kmap_atomic(v.bv_page);
1020                 next = csum_partial_copy_nocheck(p + v.bv_offset,
1021                                                  (to += v.bv_len) - v.bv_len,
1022                                                  v.bv_len, 0);
1023                 kunmap_atomic(p);
1024                 sum = csum_block_add(sum, next, off);
1025                 off += v.bv_len;
1026         }),({
1027                 next = csum_partial_copy_nocheck(v.iov_base,
1028                                                  (to += v.iov_len) - v.iov_len,
1029                                                  v.iov_len, 0);
1030                 sum = csum_block_add(sum, next, off);
1031                 off += v.iov_len;
1032         })
1033         )
1034         *csum = sum;
1035         return bytes;
1036 }
1037 EXPORT_SYMBOL(csum_and_copy_from_iter);
1038
1039 size_t csum_and_copy_to_iter(const void *addr, size_t bytes, __wsum *csum,
1040                              struct iov_iter *i)
1041 {
1042         const char *from = addr;
1043         __wsum sum, next;
1044         size_t off = 0;
1045         sum = *csum;
1046         if (unlikely(i->type & ITER_PIPE)) {
1047                 WARN_ON(1);     /* for now */
1048                 return 0;
1049         }
1050         iterate_and_advance(i, bytes, v, ({
1051                 int err = 0;
1052                 next = csum_and_copy_to_user((from += v.iov_len) - v.iov_len,
1053                                              v.iov_base,
1054                                              v.iov_len, 0, &err);
1055                 if (!err) {
1056                         sum = csum_block_add(sum, next, off);
1057                         off += v.iov_len;
1058                 }
1059                 err ? v.iov_len : 0;
1060         }), ({
1061                 char *p = kmap_atomic(v.bv_page);
1062                 next = csum_partial_copy_nocheck((from += v.bv_len) - v.bv_len,
1063                                                  p + v.bv_offset,
1064                                                  v.bv_len, 0);
1065                 kunmap_atomic(p);
1066                 sum = csum_block_add(sum, next, off);
1067                 off += v.bv_len;
1068         }),({
1069                 next = csum_partial_copy_nocheck((from += v.iov_len) - v.iov_len,
1070                                                  v.iov_base,
1071                                                  v.iov_len, 0);
1072                 sum = csum_block_add(sum, next, off);
1073                 off += v.iov_len;
1074         })
1075         )
1076         *csum = sum;
1077         return bytes;
1078 }
1079 EXPORT_SYMBOL(csum_and_copy_to_iter);
1080
1081 int iov_iter_npages(const struct iov_iter *i, int maxpages)
1082 {
1083         size_t size = i->count;
1084         int npages = 0;
1085
1086         if (!size)
1087                 return 0;
1088
1089         if (unlikely(i->type & ITER_PIPE)) {
1090                 struct pipe_inode_info *pipe = i->pipe;
1091                 size_t off;
1092                 int idx;
1093
1094                 if (!sanity(i))
1095                         return 0;
1096
1097                 data_start(i, &idx, &off);
1098                 /* some of this one + all after this one */
1099                 npages = ((pipe->curbuf - idx - 1) & (pipe->buffers - 1)) + 1;
1100                 if (npages >= maxpages)
1101                         return maxpages;
1102         } else iterate_all_kinds(i, size, v, ({
1103                 unsigned long p = (unsigned long)v.iov_base;
1104                 npages += DIV_ROUND_UP(p + v.iov_len, PAGE_SIZE)
1105                         - p / PAGE_SIZE;
1106                 if (npages >= maxpages)
1107                         return maxpages;
1108         0;}),({
1109                 npages++;
1110                 if (npages >= maxpages)
1111                         return maxpages;
1112         }),({
1113                 unsigned long p = (unsigned long)v.iov_base;
1114                 npages += DIV_ROUND_UP(p + v.iov_len, PAGE_SIZE)
1115                         - p / PAGE_SIZE;
1116                 if (npages >= maxpages)
1117                         return maxpages;
1118         })
1119         )
1120         return npages;
1121 }
1122 EXPORT_SYMBOL(iov_iter_npages);
1123
1124 const void *dup_iter(struct iov_iter *new, struct iov_iter *old, gfp_t flags)
1125 {
1126         *new = *old;
1127         if (unlikely(new->type & ITER_PIPE)) {
1128                 WARN_ON(1);
1129                 return NULL;
1130         }
1131         if (new->type & ITER_BVEC)
1132                 return new->bvec = kmemdup(new->bvec,
1133                                     new->nr_segs * sizeof(struct bio_vec),
1134                                     flags);
1135         else
1136                 /* iovec and kvec have identical layout */
1137                 return new->iov = kmemdup(new->iov,
1138                                    new->nr_segs * sizeof(struct iovec),
1139                                    flags);
1140 }
1141 EXPORT_SYMBOL(dup_iter);
1142
1143 int import_iovec(int type, const struct iovec __user * uvector,
1144                  unsigned nr_segs, unsigned fast_segs,
1145                  struct iovec **iov, struct iov_iter *i)
1146 {
1147         ssize_t n;
1148         struct iovec *p;
1149         n = rw_copy_check_uvector(type, uvector, nr_segs, fast_segs,
1150                                   *iov, &p);
1151         if (n < 0) {
1152                 if (p != *iov)
1153                         kfree(p);
1154                 *iov = NULL;
1155                 return n;
1156         }
1157         iov_iter_init(i, type, p, nr_segs, n);
1158         *iov = p == *iov ? NULL : p;
1159         return 0;
1160 }
1161 EXPORT_SYMBOL(import_iovec);
1162
1163 #ifdef CONFIG_COMPAT
1164 #include <linux/compat.h>
1165
1166 int compat_import_iovec(int type, const struct compat_iovec __user * uvector,
1167                  unsigned nr_segs, unsigned fast_segs,
1168                  struct iovec **iov, struct iov_iter *i)
1169 {
1170         ssize_t n;
1171         struct iovec *p;
1172         n = compat_rw_copy_check_uvector(type, uvector, nr_segs, fast_segs,
1173                                   *iov, &p);
1174         if (n < 0) {
1175                 if (p != *iov)
1176                         kfree(p);
1177                 *iov = NULL;
1178                 return n;
1179         }
1180         iov_iter_init(i, type, p, nr_segs, n);
1181         *iov = p == *iov ? NULL : p;
1182         return 0;
1183 }
1184 #endif
1185
1186 int import_single_range(int rw, void __user *buf, size_t len,
1187                  struct iovec *iov, struct iov_iter *i)
1188 {
1189         if (len > MAX_RW_COUNT)
1190                 len = MAX_RW_COUNT;
1191         if (unlikely(!access_ok(!rw, buf, len)))
1192                 return -EFAULT;
1193
1194         iov->iov_base = buf;
1195         iov->iov_len = len;
1196         iov_iter_init(i, rw, iov, 1, len);
1197         return 0;
1198 }
1199 EXPORT_SYMBOL(import_single_range);