lib/iov_iter.c

   1 #include <linux/export.h>
   2 #include <linux/bvec.h>
   3 #include <linux/uio.h>
   4 #include <linux/pagemap.h>
   5 #include <linux/slab.h>
   6 #include <linux/vmalloc.h>
   7 #include <linux/splice.h>
   8 #include <net/checksum.h>
   9
  10 #define PIPE_PARANOIA /* for now */
  11
  12 #define iterate_iovec(i, n, __v, __p, skip, STEP) {     \
  13         size_t left;                                    \
  14         size_t wanted = n;                              \
  15         __p = i->iov;                                   \
  16         __v.iov_len = min(n, __p->iov_len - skip);      \
  17         if (likely(__v.iov_len)) {                      \
  18                 __v.iov_base = __p->iov_base + skip;    \
  19                 left = (STEP);                          \
  20                 __v.iov_len -= left;                    \
  21                 skip += __v.iov_len;                    \
  22                 n -= __v.iov_len;                       \
  23         } else {                                        \
  24                 left = 0;                               \
  25         }                                               \
  26         while (unlikely(!left && n)) {                  \
  27                 __p++;                                  \
  28                 __v.iov_len = min(n, __p->iov_len);     \
  29                 if (unlikely(!__v.iov_len))             \
  30                         continue;                       \
  31                 __v.iov_base = __p->iov_base;           \
  32                 left = (STEP);                          \
  33                 __v.iov_len -= left;                    \
  34                 skip = __v.iov_len;                     \
  35                 n -= __v.iov_len;                       \
  36         }                                               \
  37         n = wanted - n;                                 \
  38 }
  39
  40 #define iterate_kvec(i, n, __v, __p, skip, STEP) {      \
  41         size_t wanted = n;                              \
  42         __p = i->kvec;                                  \
  43         __v.iov_len = min(n, __p->iov_len - skip);      \
  44         if (likely(__v.iov_len)) {                      \
  45                 __v.iov_base = __p->iov_base + skip;    \
  46                 (void)(STEP);                           \
  47                 skip += __v.iov_len;                    \
  48                 n -= __v.iov_len;                       \
  49         }                                               \
  50         while (unlikely(n)) {                           \
  51                 __p++;                                  \
  52                 __v.iov_len = min(n, __p->iov_len);     \
  53                 if (unlikely(!__v.iov_len))             \
  54                         continue;                       \
  55                 __v.iov_base = __p->iov_base;           \
  56                 (void)(STEP);                           \
  57                 skip = __v.iov_len;                     \
  58                 n -= __v.iov_len;                       \
  59         }                                               \
  60         n = wanted;                                     \
  61 }
  62
  63 #define iterate_bvec(i, n, __v, __bi, skip, STEP) {     \
  64         struct bvec_iter __start;                       \
  65         __start.bi_size = n;                            \
  66         __start.bi_bvec_done = skip;                    \
  67         __start.bi_idx = 0;                             \
  68         for_each_bvec(__v, i->bvec, __bi, __start) {    \
  69                 if (!__v.bv_len)                        \
  70                         continue;                       \
  71                 (void)(STEP);                           \
  72         }                                               \
  73 }
  74
  75 #define iterate_all_kinds(i, n, v, I, B, K) {                   \
  76         size_t skip = i->iov_offset;                            \
  77         if (unlikely(i->type & ITER_BVEC)) {                    \
  78                 struct bio_vec v;                               \
  79                 struct bvec_iter __bi;                          \
  80                 iterate_bvec(i, n, v, __bi, skip, (B))          \
  81         } else if (unlikely(i->type & ITER_KVEC)) {             \
  82                 const struct kvec *kvec;                        \
  83                 struct kvec v;                                  \
  84                 iterate_kvec(i, n, v, kvec, skip, (K))          \
  85         } else {                                                \
  86                 const struct iovec *iov;                        \
  87                 struct iovec v;                                 \
  88                 iterate_iovec(i, n, v, iov, skip, (I))          \
  89         }                                                       \
  90 }
  91
  92 #define iterate_and_advance(i, n, v, I, B, K) {                 \
  93         if (unlikely(i->count < n))                             \
  94                 n = i->count;                                   \
  95         if (i->count) {                                         \
  96                 size_t skip = i->iov_offset;                    \
  97                 if (unlikely(i->type & ITER_BVEC)) {            \
  98                         const struct bio_vec *bvec = i->bvec;   \
  99                         struct bio_vec v;                       \
 100                         struct bvec_iter __bi;                  \
 101                         iterate_bvec(i, n, v, __bi, skip, (B))  \
 102                         i->bvec = __bvec_iter_bvec(i->bvec, __bi);      \
 103                         i->nr_segs -= i->bvec - bvec;           \
 104                         skip = __bi.bi_bvec_done;               \
 105                 } else if (unlikely(i->type & ITER_KVEC)) {     \
 106                         const struct kvec *kvec;                \
 107                         struct kvec v;                          \
 108                         iterate_kvec(i, n, v, kvec, skip, (K))  \
 109                         if (skip == kvec->iov_len) {            \
 110                                 kvec++;                         \
 111                                 skip = 0;                       \
 112                         }                                       \
 113                         i->nr_segs -= kvec - i->kvec;           \
 114                         i->kvec = kvec;                         \
 115                 } else {                                        \
 116                         const struct iovec *iov;                \
 117                         struct iovec v;                         \
 118                         iterate_iovec(i, n, v, iov, skip, (I))  \
 119                         if (skip == iov->iov_len) {             \
 120                                 iov++;                          \
 121                                 skip = 0;                       \
 122                         }                                       \
 123                         i->nr_segs -= iov - i->iov;             \
 124                         i->iov = iov;                           \
 125                 }                                               \
 126                 i->count -= n;                                  \
 127                 i->iov_offset = skip;                           \
 128         }                                                       \
 129 }
 130
 131 static size_t copy_page_to_iter_iovec(struct page *page, size_t offset, size_t bytes,
 132                          struct iov_iter *i)
 133 {
 134         size_t skip, copy, left, wanted;
 135         const struct iovec *iov;
 136         char __user *buf;
 137         void *kaddr, *from;
 138
 139         if (unlikely(bytes > i->count))
 140                 bytes = i->count;
 141
 142         if (unlikely(!bytes))
 143                 return 0;
 144
 145         wanted = bytes;
 146         iov = i->iov;
 147         skip = i->iov_offset;
 148         buf = iov->iov_base + skip;
 149         copy = min(bytes, iov->iov_len - skip);
 150
 151         if (IS_ENABLED(CONFIG_HIGHMEM) && !fault_in_pages_writeable(buf, copy)) {
 152                 kaddr = kmap_atomic(page);
 153                 from = kaddr + offset;
 154
 155                 /* first chunk, usually the only one */
 156                 left = __copy_to_user_inatomic(buf, from, copy);
 157                 copy -= left;
 158                 skip += copy;
 159                 from += copy;
 160                 bytes -= copy;
 161
 162                 while (unlikely(!left && bytes)) {
 163                         iov++;
 164                         buf = iov->iov_base;
 165                         copy = min(bytes, iov->iov_len);
 166                         left = __copy_to_user_inatomic(buf, from, copy);
 167                         copy -= left;
 168                         skip = copy;
 169                         from += copy;
 170                         bytes -= copy;
 171                 }
 172                 if (likely(!bytes)) {
 173                         kunmap_atomic(kaddr);
 174                         goto done;
 175                 }
 176                 offset = from - kaddr;
 177                 buf += copy;
 178                 kunmap_atomic(kaddr);
 179                 copy = min(bytes, iov->iov_len - skip);
 180         }
 181         /* Too bad - revert to non-atomic kmap */
 182
 183         kaddr = kmap(page);
 184         from = kaddr + offset;
 185         left = __copy_to_user(buf, from, copy);
 186         copy -= left;
 187         skip += copy;
 188         from += copy;
 189         bytes -= copy;
 190         while (unlikely(!left && bytes)) {
 191                 iov++;
 192                 buf = iov->iov_base;
 193                 copy = min(bytes, iov->iov_len);
 194                 left = __copy_to_user(buf, from, copy);
 195                 copy -= left;
 196                 skip = copy;
 197                 from += copy;
 198                 bytes -= copy;
 199         }
 200         kunmap(page);
 201
 202 done:
 203         if (skip == iov->iov_len) {
 204                 iov++;
 205                 skip = 0;
 206         }
 207         i->count -= wanted - bytes;
 208         i->nr_segs -= iov - i->iov;
 209         i->iov = iov;
 210         i->iov_offset = skip;
 211         return wanted - bytes;
 212 }
 213
 214 static size_t copy_page_from_iter_iovec(struct page *page, size_t offset, size_t bytes,
 215                          struct iov_iter *i)
 216 {
 217         size_t skip, copy, left, wanted;
 218         const struct iovec *iov;
 219         char __user *buf;
 220         void *kaddr, *to;
 221
 222         if (unlikely(bytes > i->count))
 223                 bytes = i->count;
 224
 225         if (unlikely(!bytes))
 226                 return 0;
 227
 228         wanted = bytes;
 229         iov = i->iov;
 230         skip = i->iov_offset;
 231         buf = iov->iov_base + skip;
 232         copy = min(bytes, iov->iov_len - skip);
 233
 234         if (IS_ENABLED(CONFIG_HIGHMEM) && !fault_in_pages_readable(buf, copy)) {
 235                 kaddr = kmap_atomic(page);
 236                 to = kaddr + offset;
 237
 238                 /* first chunk, usually the only one */
 239                 left = __copy_from_user_inatomic(to, buf, copy);
 240                 copy -= left;
 241                 skip += copy;
 242                 to += copy;
 243                 bytes -= copy;
 244
 245                 while (unlikely(!left && bytes)) {
 246                         iov++;
 247                         buf = iov->iov_base;
 248                         copy = min(bytes, iov->iov_len);
 249                         left = __copy_from_user_inatomic(to, buf, copy);
 250                         copy -= left;
 251                         skip = copy;
 252                         to += copy;
 253                         bytes -= copy;
 254                 }
 255                 if (likely(!bytes)) {
 256                         kunmap_atomic(kaddr);
 257                         goto done;
 258                 }
 259                 offset = to - kaddr;
 260                 buf += copy;
 261                 kunmap_atomic(kaddr);
 262                 copy = min(bytes, iov->iov_len - skip);
 263         }
 264         /* Too bad - revert to non-atomic kmap */
 265
 266         kaddr = kmap(page);
 267         to = kaddr + offset;
 268         left = __copy_from_user(to, buf, copy);
 269         copy -= left;
 270         skip += copy;
 271         to += copy;
 272         bytes -= copy;
 273         while (unlikely(!left && bytes)) {
 274                 iov++;
 275                 buf = iov->iov_base;
 276                 copy = min(bytes, iov->iov_len);
 277                 left = __copy_from_user(to, buf, copy);
 278                 copy -= left;
 279                 skip = copy;
 280                 to += copy;
 281                 bytes -= copy;
 282         }
 283         kunmap(page);
 284
 285 done:
 286         if (skip == iov->iov_len) {
 287                 iov++;
 288                 skip = 0;
 289         }
 290         i->count -= wanted - bytes;
 291         i->nr_segs -= iov - i->iov;
 292         i->iov = iov;
 293         i->iov_offset = skip;
 294         return wanted - bytes;
 295 }
 296
 297 #ifdef PIPE_PARANOIA
 298 static bool sanity(const struct iov_iter *i)
 299 {
 300         struct pipe_inode_info *pipe = i->pipe;
 301         int idx = i->idx;
 302         int next = pipe->curbuf + pipe->nrbufs;
 303         if (i->iov_offset) {
 304                 struct pipe_buffer *p;
 305                 if (unlikely(!pipe->nrbufs))
 306                         goto Bad;       // pipe must be non-empty
 307                 if (unlikely(idx != ((next - 1) & (pipe->buffers - 1))))
 308                         goto Bad;       // must be at the last buffer...
 309
 310                 p = &pipe->bufs[idx];
 311                 if (unlikely(p->offset + p->len != i->iov_offset))
 312                         goto Bad;       // ... at the end of segment
 313         } else {
 314                 if (idx != (next & (pipe->buffers - 1)))
 315                         goto Bad;       // must be right after the last buffer
 316         }
 317         return true;
 318 Bad:
 319         printk(KERN_ERR "idx = %d, offset = %zd\n", i->idx, i->iov_offset);
 320         printk(KERN_ERR "curbuf = %d, nrbufs = %d, buffers = %d\n",
 321                         pipe->curbuf, pipe->nrbufs, pipe->buffers);
 322         for (idx = 0; idx < pipe->buffers; idx++)
 323                 printk(KERN_ERR "[%p %p %d %d]\n",
 324                         pipe->bufs[idx].ops,
 325                         pipe->bufs[idx].page,
 326                         pipe->bufs[idx].offset,
 327                         pipe->bufs[idx].len);
 328         WARN_ON(1);
 329         return false;
 330 }
 331 #else
 332 #define sanity(i) true
 333 #endif
 334
 335 static inline int next_idx(int idx, struct pipe_inode_info *pipe)
 336 {
 337         return (idx + 1) & (pipe->buffers - 1);
 338 }
 339
 340 static size_t copy_page_to_iter_pipe(struct page *page, size_t offset, size_t bytes,
 341                          struct iov_iter *i)
 342 {
 343         struct pipe_inode_info *pipe = i->pipe;
 344         struct pipe_buffer *buf;
 345         size_t off;
 346         int idx;
 347
 348         if (unlikely(bytes > i->count))
 349                 bytes = i->count;
 350
 351         if (unlikely(!bytes))
 352                 return 0;
 353
 354         if (!sanity(i))
 355                 return 0;
 356
 357         off = i->iov_offset;
 358         idx = i->idx;
 359         buf = &pipe->bufs[idx];
 360         if (off) {
 361                 if (offset == off && buf->page == page) {
 362                         /* merge with the last one */
 363                         buf->len += bytes;
 364                         i->iov_offset += bytes;
 365                         goto out;
 366                 }
 367                 idx = next_idx(idx, pipe);
 368                 buf = &pipe->bufs[idx];
 369         }
 370         if (idx == pipe->curbuf && pipe->nrbufs)
 371                 return 0;
 372         pipe->nrbufs++;
 373         buf->ops = &page_cache_pipe_buf_ops;
 374         get_page(buf->page = page);
 375         buf->offset = offset;
 376         buf->len = bytes;
 377         i->iov_offset = offset + bytes;
 378         i->idx = idx;
 379 out:
 380         i->count -= bytes;
 381         return bytes;
 382 }
 383
 384 /*
 385  * Fault in one or more iovecs of the given iov_iter, to a maximum length of
 386  * bytes.  For each iovec, fault in each page that constitutes the iovec.
 387  *
 388  * Return 0 on success, or non-zero if the memory could not be accessed (i.e.
 389  * because it is an invalid address).
 390  */
 391 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
 392 {
 393         size_t skip = i->iov_offset;
 394         const struct iovec *iov;
 395         int err;
 396         struct iovec v;
 397
 398         if (!(i->type & (ITER_BVEC|ITER_KVEC))) {
 399                 iterate_iovec(i, bytes, v, iov, skip, ({
 400                         err = fault_in_pages_readable(v.iov_base, 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, orig_sz;
 688
 689         if (unlikely(i->count < size))
 690                 size = i->count;
 691         orig_sz = size;
 692
 693         if (size) {
 694                 if (off) /* make it relative to the beginning of buffer */
 695                         size += off - pipe->bufs[idx].offset;
 696                 while (1) {
 697                         buf = &pipe->bufs[idx];
 698                         if (size <= buf->len)
 699                                 break;
 700                         size -= buf->len;
 701                         idx = next_idx(idx, pipe);
 702                 }
 703                 buf->len = size;
 704                 i->idx = idx;
 705                 off = i->iov_offset = buf->offset + size;
 706         }
 707         if (off)
 708                 idx = next_idx(idx, pipe);
 709         if (pipe->nrbufs) {
 710                 int unused = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 711                 /* [curbuf,unused) is in use.  Free [idx,unused) */
 712                 while (idx != unused) {
 713                         pipe_buf_release(pipe, &pipe->bufs[idx]);
 714                         idx = next_idx(idx, pipe);
 715                         pipe->nrbufs--;
 716                 }
 717         }
 718         i->count -= orig_sz;
 719 }
 720
 721 void iov_iter_advance(struct iov_iter *i, size_t size)
 722 {
 723         if (unlikely(i->type & ITER_PIPE)) {
 724                 pipe_advance(i, size);
 725                 return;
 726         }
 727         iterate_and_advance(i, size, v, 0, 0, 0)
 728 }
 729 EXPORT_SYMBOL(iov_iter_advance);
 730
 731 /*
 732  * Return the count of just the current iov_iter segment.
 733  */
 734 size_t iov_iter_single_seg_count(const struct iov_iter *i)
 735 {
 736         if (unlikely(i->type & ITER_PIPE))
 737                 return i->count;        // it is a silly place, anyway
 738         if (i->nr_segs == 1)
 739                 return i->count;
 740         else if (i->type & ITER_BVEC)
 741                 return min(i->count, i->bvec->bv_len - i->iov_offset);
 742         else
 743                 return min(i->count, i->iov->iov_len - i->iov_offset);
 744 }
 745 EXPORT_SYMBOL(iov_iter_single_seg_count);
 746
 747 void iov_iter_kvec(struct iov_iter *i, int direction,
 748                         const struct kvec *kvec, unsigned long nr_segs,
 749                         size_t count)
 750 {
 751         BUG_ON(!(direction & ITER_KVEC));
 752         i->type = direction;
 753         i->kvec = kvec;
 754         i->nr_segs = nr_segs;
 755         i->iov_offset = 0;
 756         i->count = count;
 757 }
 758 EXPORT_SYMBOL(iov_iter_kvec);
 759
 760 void iov_iter_bvec(struct iov_iter *i, int direction,
 761                         const struct bio_vec *bvec, unsigned long nr_segs,
 762                         size_t count)
 763 {
 764         BUG_ON(!(direction & ITER_BVEC));
 765         i->type = direction;
 766         i->bvec = bvec;
 767         i->nr_segs = nr_segs;
 768         i->iov_offset = 0;
 769         i->count = count;
 770 }
 771 EXPORT_SYMBOL(iov_iter_bvec);
 772
 773 void iov_iter_pipe(struct iov_iter *i, int direction,
 774                         struct pipe_inode_info *pipe,
 775                         size_t count)
 776 {
 777         BUG_ON(direction != ITER_PIPE);
 778         i->type = direction;
 779         i->pipe = pipe;
 780         i->idx = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 781         i->iov_offset = 0;
 782         i->count = count;
 783 }
 784 EXPORT_SYMBOL(iov_iter_pipe);
 785
 786 unsigned long iov_iter_alignment(const struct iov_iter *i)
 787 {
 788         unsigned long res = 0;
 789         size_t size = i->count;
 790
 791         if (!size)
 792                 return 0;
 793
 794         if (unlikely(i->type & ITER_PIPE)) {
 795                 if (i->iov_offset && allocated(&i->pipe->bufs[i->idx]))
 796                         return size | i->iov_offset;
 797                 return size;
 798         }
 799         iterate_all_kinds(i, size, v,
 800                 (res |= (unsigned long)v.iov_base | v.iov_len, 0),
 801                 res |= v.bv_offset | v.bv_len,
 802                 res |= (unsigned long)v.iov_base | v.iov_len
 803         )
 804         return res;
 805 }
 806 EXPORT_SYMBOL(iov_iter_alignment);
 807
 808 unsigned long iov_iter_gap_alignment(const struct iov_iter *i)
 809 {
 810         unsigned long res = 0;
 811         size_t size = i->count;
 812         if (!size)
 813                 return 0;
 814
 815         if (unlikely(i->type & ITER_PIPE)) {
 816                 WARN_ON(1);
 817                 return ~0U;
 818         }
 819
 820         iterate_all_kinds(i, size, v,
 821                 (res |= (!res ? 0 : (unsigned long)v.iov_base) |
 822                         (size != v.iov_len ? size : 0), 0),
 823                 (res |= (!res ? 0 : (unsigned long)v.bv_offset) |
 824                         (size != v.bv_len ? size : 0)),
 825                 (res |= (!res ? 0 : (unsigned long)v.iov_base) |
 826                         (size != v.iov_len ? size : 0))
 827                 );
 828                 return res;
 829 }
 830 EXPORT_SYMBOL(iov_iter_gap_alignment);
 831
 832 static inline size_t __pipe_get_pages(struct iov_iter *i,
 833                                 size_t maxsize,
 834                                 struct page **pages,
 835                                 int idx,
 836                                 size_t *start)
 837 {
 838         struct pipe_inode_info *pipe = i->pipe;
 839         ssize_t n = push_pipe(i, maxsize, &idx, start);
 840         if (!n)
 841                 return -EFAULT;
 842
 843         maxsize = n;
 844         n += *start;
 845         while (n > 0) {
 846                 get_page(*pages++ = pipe->bufs[idx].page);
 847                 idx = next_idx(idx, pipe);
 848                 n -= PAGE_SIZE;
 849         }
 850
 851         return maxsize;
 852 }
 853
 854 static ssize_t pipe_get_pages(struct iov_iter *i,
 855                    struct page **pages, size_t maxsize, unsigned maxpages,
 856                    size_t *start)
 857 {
 858         unsigned npages;
 859         size_t capacity;
 860         int idx;
 861
 862         if (!sanity(i))
 863                 return -EFAULT;
 864
 865         data_start(i, &idx, start);
 866         /* some of this one + all after this one */
 867         npages = ((i->pipe->curbuf - idx - 1) & (i->pipe->buffers - 1)) + 1;
 868         capacity = min(npages,maxpages) * PAGE_SIZE - *start;
 869
 870         return __pipe_get_pages(i, min(maxsize, capacity), pages, idx, start);
 871 }
 872
 873 ssize_t iov_iter_get_pages(struct iov_iter *i,
 874                    struct page **pages, size_t maxsize, unsigned maxpages,
 875                    size_t *start)
 876 {
 877         if (maxsize > i->count)
 878                 maxsize = i->count;
 879
 880         if (!maxsize)
 881                 return 0;
 882
 883         if (unlikely(i->type & ITER_PIPE))
 884                 return pipe_get_pages(i, pages, maxsize, maxpages, start);
 885         iterate_all_kinds(i, maxsize, v, ({
 886                 unsigned long addr = (unsigned long)v.iov_base;
 887                 size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1));
 888                 int n;
 889                 int res;
 890
 891                 if (len > maxpages * PAGE_SIZE)
 892                         len = maxpages * PAGE_SIZE;
 893                 addr &= ~(PAGE_SIZE - 1);
 894                 n = DIV_ROUND_UP(len, PAGE_SIZE);
 895                 res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, pages);
 896                 if (unlikely(res < 0))
 897                         return res;
 898                 return (res == n ? len : res * PAGE_SIZE) - *start;
 899         0;}),({
 900                 /* can't be more than PAGE_SIZE */
 901                 *start = v.bv_offset;
 902                 get_page(*pages = v.bv_page);
 903                 return v.bv_len;
 904         }),({
 905                 return -EFAULT;
 906         })
 907         )
 908         return 0;
 909 }
 910 EXPORT_SYMBOL(iov_iter_get_pages);
 911
 912 static struct page **get_pages_array(size_t n)
 913 {
 914         struct page **p = kmalloc(n * sizeof(struct page *), GFP_KERNEL);
 915         if (!p)
 916                 p = vmalloc(n * sizeof(struct page *));
 917         return p;
 918 }
 919
 920 static ssize_t pipe_get_pages_alloc(struct iov_iter *i,
 921                    struct page ***pages, size_t maxsize,
 922                    size_t *start)
 923 {
 924         struct page **p;
 925         size_t n;
 926         int idx;
 927         int npages;
 928
 929         if (!sanity(i))
 930                 return -EFAULT;
 931
 932         data_start(i, &idx, start);
 933         /* some of this one + all after this one */
 934         npages = ((i->pipe->curbuf - idx - 1) & (i->pipe->buffers - 1)) + 1;
 935         n = npages * PAGE_SIZE - *start;
 936         if (maxsize > n)
 937                 maxsize = n;
 938         else
 939                 npages = DIV_ROUND_UP(maxsize + *start, PAGE_SIZE);
 940         p = get_pages_array(npages);
 941         if (!p)
 942                 return -ENOMEM;
 943         n = __pipe_get_pages(i, maxsize, p, idx, start);
 944         if (n > 0)
 945                 *pages = p;
 946         else
 947                 kvfree(p);
 948         return n;
 949 }
 950
 951 ssize_t iov_iter_get_pages_alloc(struct iov_iter *i,
 952                    struct page ***pages, size_t maxsize,
 953                    size_t *start)
 954 {
 955         struct page **p;
 956
 957         if (maxsize > i->count)
 958                 maxsize = i->count;
 959
 960         if (!maxsize)
 961                 return 0;
 962
 963         if (unlikely(i->type & ITER_PIPE))
 964                 return pipe_get_pages_alloc(i, pages, maxsize, start);
 965         iterate_all_kinds(i, maxsize, v, ({
 966                 unsigned long addr = (unsigned long)v.iov_base;
 967                 size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1));
 968                 int n;
 969                 int res;
 970
 971                 addr &= ~(PAGE_SIZE - 1);
 972                 n = DIV_ROUND_UP(len, PAGE_SIZE);
 973                 p = get_pages_array(n);
 974                 if (!p)
 975                         return -ENOMEM;
 976                 res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, p);
 977                 if (unlikely(res < 0)) {
 978                         kvfree(p);
 979                         return res;
 980                 }
 981                 *pages = p;
 982                 return (res == n ? len : res * PAGE_SIZE) - *start;
 983         0;}),({
 984                 /* can't be more than PAGE_SIZE */
 985                 *start = v.bv_offset;
 986                 *pages = p = get_pages_array(1);
 987                 if (!p)
 988                         return -ENOMEM;
 989                 get_page(*p = v.bv_page);
 990                 return v.bv_len;
 991         }),({
 992                 return -EFAULT;
 993         })
 994         )
 995         return 0;
 996 }
 997 EXPORT_SYMBOL(iov_iter_get_pages_alloc);
 998
 999 size_t csum_and_copy_from_iter(void *addr, size_t bytes, __wsum *csum,
1000                                struct iov_iter *i)
1001 {
1002         char *to = addr;
1003         __wsum sum, next;
1004         size_t off = 0;
1005         sum = *csum;
1006         if (unlikely(i->type & ITER_PIPE)) {
1007                 WARN_ON(1);
1008                 return 0;
1009         }
1010         iterate_and_advance(i, bytes, v, ({
1011                 int err = 0;
1012                 next = csum_and_copy_from_user(v.iov_base,
1013                                                (to += v.iov_len) - v.iov_len,
1014                                                v.iov_len, 0, &err);
1015                 if (!err) {
1016                         sum = csum_block_add(sum, next, off);
1017                         off += v.iov_len;
1018                 }
1019                 err ? v.iov_len : 0;
1020         }), ({
1021                 char *p = kmap_atomic(v.bv_page);
1022                 next = csum_partial_copy_nocheck(p + v.bv_offset,
1023                                                  (to += v.bv_len) - v.bv_len,
1024                                                  v.bv_len, 0);
1025                 kunmap_atomic(p);
1026                 sum = csum_block_add(sum, next, off);
1027                 off += v.bv_len;
1028         }),({
1029                 next = csum_partial_copy_nocheck(v.iov_base,
1030                                                  (to += v.iov_len) - v.iov_len,
1031                                                  v.iov_len, 0);
1032                 sum = csum_block_add(sum, next, off);
1033                 off += v.iov_len;
1034         })
1035         )
1036         *csum = sum;
1037         return bytes;
1038 }
1039 EXPORT_SYMBOL(csum_and_copy_from_iter);
1040
1041 size_t csum_and_copy_to_iter(const void *addr, size_t bytes, __wsum *csum,
1042                              struct iov_iter *i)
1043 {
1044         const char *from = addr;
1045         __wsum sum, next;
1046         size_t off = 0;
1047         sum = *csum;
1048         if (unlikely(i->type & ITER_PIPE)) {
1049                 WARN_ON(1);     /* for now */
1050                 return 0;
1051         }
1052         iterate_and_advance(i, bytes, v, ({
1053                 int err = 0;
1054                 next = csum_and_copy_to_user((from += v.iov_len) - v.iov_len,
1055                                              v.iov_base,
1056                                              v.iov_len, 0, &err);
1057                 if (!err) {
1058                         sum = csum_block_add(sum, next, off);
1059                         off += v.iov_len;
1060                 }
1061                 err ? v.iov_len : 0;
1062         }), ({
1063                 char *p = kmap_atomic(v.bv_page);
1064                 next = csum_partial_copy_nocheck((from += v.bv_len) - v.bv_len,
1065                                                  p + v.bv_offset,
1066                                                  v.bv_len, 0);
1067                 kunmap_atomic(p);
1068                 sum = csum_block_add(sum, next, off);
1069                 off += v.bv_len;
1070         }),({
1071                 next = csum_partial_copy_nocheck((from += v.iov_len) - v.iov_len,
1072                                                  v.iov_base,
1073                                                  v.iov_len, 0);
1074                 sum = csum_block_add(sum, next, off);
1075                 off += v.iov_len;
1076         })
1077         )
1078         *csum = sum;
1079         return bytes;
1080 }
1081 EXPORT_SYMBOL(csum_and_copy_to_iter);
1082
1083 int iov_iter_npages(const struct iov_iter *i, int maxpages)
1084 {
1085         size_t size = i->count;
1086         int npages = 0;
1087
1088         if (!size)
1089                 return 0;
1090
1091         if (unlikely(i->type & ITER_PIPE)) {
1092                 struct pipe_inode_info *pipe = i->pipe;
1093                 size_t off;
1094                 int idx;
1095
1096                 if (!sanity(i))
1097                         return 0;
1098
1099                 data_start(i, &idx, &off);
1100                 /* some of this one + all after this one */
1101                 npages = ((pipe->curbuf - idx - 1) & (pipe->buffers - 1)) + 1;
1102                 if (npages >= maxpages)
1103                         return maxpages;
1104         } else iterate_all_kinds(i, size, v, ({
1105                 unsigned long p = (unsigned long)v.iov_base;
1106                 npages += DIV_ROUND_UP(p + v.iov_len, PAGE_SIZE)
1107                         - p / PAGE_SIZE;
1108                 if (npages >= maxpages)
1109                         return maxpages;
1110         0;}),({
1111                 npages++;
1112                 if (npages >= maxpages)
1113                         return maxpages;
1114         }),({
1115                 unsigned long p = (unsigned long)v.iov_base;
1116                 npages += DIV_ROUND_UP(p + v.iov_len, PAGE_SIZE)
1117                         - p / PAGE_SIZE;
1118                 if (npages >= maxpages)
1119                         return maxpages;
1120         })
1121         )
1122         return npages;
1123 }
1124 EXPORT_SYMBOL(iov_iter_npages);
1125
1126 const void *dup_iter(struct iov_iter *new, struct iov_iter *old, gfp_t flags)
1127 {
1128         *new = *old;
1129         if (unlikely(new->type & ITER_PIPE)) {
1130                 WARN_ON(1);
1131                 return NULL;
1132         }
1133         if (new->type & ITER_BVEC)
1134                 return new->bvec = kmemdup(new->bvec,
1135                                     new->nr_segs * sizeof(struct bio_vec),
1136                                     flags);
1137         else
1138                 /* iovec and kvec have identical layout */
1139                 return new->iov = kmemdup(new->iov,
1140                                    new->nr_segs * sizeof(struct iovec),
1141                                    flags);
1142 }
1143 EXPORT_SYMBOL(dup_iter);
1144
1145 /**
1146  * import_iovec() - Copy an array of &struct iovec from userspace
1147  *     into the kernel, check that it is valid, and initialize a new
1148  *     &struct iov_iter iterator to access it.
1149  *
1150  * @type: One of %READ or %WRITE.
1151  * @uvector: Pointer to the userspace array.
1152  * @nr_segs: Number of elements in userspace array.
1153  * @fast_segs: Number of elements in @iov.
1154  * @iov: (input and output parameter) Pointer to pointer to (usually small
1155  *     on-stack) kernel array.
1156  * @i: Pointer to iterator that will be initialized on success.
1157  *
1158  * If the array pointed to by *@iov is large enough to hold all @nr_segs,
1159  * then this function places %NULL in *@iov on return. Otherwise, a new
1160  * array will be allocated and the result placed in *@iov. This means that
1161  * the caller may call kfree() on *@iov regardless of whether the small
1162  * on-stack array was used or not (and regardless of whether this function
1163  * returns an error or not).
1164  *
1165  * Return: 0 on success or negative error code on error.
1166  */
1167 int import_iovec(int type, const struct iovec __user * uvector,
1168                  unsigned nr_segs, unsigned fast_segs,
1169                  struct iovec **iov, struct iov_iter *i)
1170 {
1171         ssize_t n;
1172         struct iovec *p;
1173         n = rw_copy_check_uvector(type, uvector, nr_segs, fast_segs,
1174                                   *iov, &p);
1175         if (n < 0) {
1176                 if (p != *iov)
1177                         kfree(p);
1178                 *iov = NULL;
1179                 return n;
1180         }
1181         iov_iter_init(i, type, p, nr_segs, n);
1182         *iov = p == *iov ? NULL : p;
1183         return 0;
1184 }
1185 EXPORT_SYMBOL(import_iovec);
1186
1187 #ifdef CONFIG_COMPAT
1188 #include <linux/compat.h>
1189
1190 int compat_import_iovec(int type, const struct compat_iovec __user * uvector,
1191                  unsigned nr_segs, unsigned fast_segs,
1192                  struct iovec **iov, struct iov_iter *i)
1193 {
1194         ssize_t n;
1195         struct iovec *p;
1196         n = compat_rw_copy_check_uvector(type, uvector, nr_segs, fast_segs,
1197                                   *iov, &p);
1198         if (n < 0) {
1199                 if (p != *iov)
1200                         kfree(p);
1201                 *iov = NULL;
1202                 return n;
1203         }
1204         iov_iter_init(i, type, p, nr_segs, n);
1205         *iov = p == *iov ? NULL : p;
1206         return 0;
1207 }
1208 #endif
1209
1210 int import_single_range(int rw, void __user *buf, size_t len,
1211                  struct iovec *iov, struct iov_iter *i)
1212 {
1213         if (len > MAX_RW_COUNT)
1214                 len = MAX_RW_COUNT;
1215         if (unlikely(!access_ok(!rw, buf, len)))
1216                 return -EFAULT;
1217
1218         iov->iov_base = buf;
1219         iov->iov_len = len;
1220         iov_iter_init(i, rw, iov, 1, len);
1221         return 0;
1222 }
1223 EXPORT_SYMBOL(import_single_range);