lib/iov_iter.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 #include <crypto/hash.h>
   3 #include <linux/export.h>
   4 #include <linux/bvec.h>
   5 #include <linux/fault-inject-usercopy.h>
   6 #include <linux/uio.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/highmem.h>
   9 #include <linux/slab.h>
  10 #include <linux/vmalloc.h>
  11 #include <linux/splice.h>
  12 #include <linux/compat.h>
  13 #include <net/checksum.h>
  14 #include <linux/scatterlist.h>
  15 #include <linux/instrumented.h>
  16
  17 #define PIPE_PARANOIA /* for now */
  18
  19 /* covers ubuf and kbuf alike */
  20 #define iterate_buf(i, n, base, len, off, __p, STEP) {          \
  21         size_t __maybe_unused off = 0;                          \
  22         len = n;                                                \
  23         base = __p + i->iov_offset;                             \
  24         len -= (STEP);                                          \
  25         i->iov_offset += len;                                   \
  26         n = len;                                                \
  27 }
  28
  29 /* covers iovec and kvec alike */
  30 #define iterate_iovec(i, n, base, len, off, __p, STEP) {        \
  31         size_t off = 0;                                         \
  32         size_t skip = i->iov_offset;                            \
  33         do {                                                    \
  34                 len = min(n, __p->iov_len - skip);              \
  35                 if (likely(len)) {                              \
  36                         base = __p->iov_base + skip;            \
  37                         len -= (STEP);                          \
  38                         off += len;                             \
  39                         skip += len;                            \
  40                         n -= len;                               \
  41                         if (skip < __p->iov_len)                \
  42                                 break;                          \
  43                 }                                               \
  44                 __p++;                                          \
  45                 skip = 0;                                       \
  46         } while (n);                                            \
  47         i->iov_offset = skip;                                   \
  48         n = off;                                                \
  49 }
  50
  51 #define iterate_bvec(i, n, base, len, off, p, STEP) {           \
  52         size_t off = 0;                                         \
  53         unsigned skip = i->iov_offset;                          \
  54         while (n) {                                             \
  55                 unsigned offset = p->bv_offset + skip;          \
  56                 unsigned left;                                  \
  57                 void *kaddr = kmap_local_page(p->bv_page +      \
  58                                         offset / PAGE_SIZE);    \
  59                 base = kaddr + offset % PAGE_SIZE;              \
  60                 len = min(min(n, (size_t)(p->bv_len - skip)),   \
  61                      (size_t)(PAGE_SIZE - offset % PAGE_SIZE)); \
  62                 left = (STEP);                                  \
  63                 kunmap_local(kaddr);                            \
  64                 len -= left;                                    \
  65                 off += len;                                     \
  66                 skip += len;                                    \
  67                 if (skip == p->bv_len) {                        \
  68                         skip = 0;                               \
  69                         p++;                                    \
  70                 }                                               \
  71                 n -= len;                                       \
  72                 if (left)                                       \
  73                         break;                                  \
  74         }                                                       \
  75         i->iov_offset = skip;                                   \
  76         n = off;                                                \
  77 }
  78
  79 #define iterate_xarray(i, n, base, len, __off, STEP) {          \
  80         __label__ __out;                                        \
  81         size_t __off = 0;                                       \
  82         struct folio *folio;                                    \
  83         loff_t start = i->xarray_start + i->iov_offset;         \
  84         pgoff_t index = start / PAGE_SIZE;                      \
  85         XA_STATE(xas, i->xarray, index);                        \
  86                                                                 \
  87         len = PAGE_SIZE - offset_in_page(start);                \
  88         rcu_read_lock();                                        \
  89         xas_for_each(&xas, folio, ULONG_MAX) {                  \
  90                 unsigned left;                                  \
  91                 size_t offset;                                  \
  92                 if (xas_retry(&xas, folio))                     \
  93                         continue;                               \
  94                 if (WARN_ON(xa_is_value(folio)))                \
  95                         break;                                  \
  96                 if (WARN_ON(folio_test_hugetlb(folio)))         \
  97                         break;                                  \
  98                 offset = offset_in_folio(folio, start + __off); \
  99                 while (offset < folio_size(folio)) {            \
 100                         base = kmap_local_folio(folio, offset); \
 101                         len = min(n, len);                      \
 102                         left = (STEP);                          \
 103                         kunmap_local(base);                     \
 104                         len -= left;                            \
 105                         __off += len;                           \
 106                         n -= len;                               \
 107                         if (left || n == 0)                     \
 108                                 goto __out;                     \
 109                         offset += len;                          \
 110                         len = PAGE_SIZE;                        \
 111                 }                                               \
 112         }                                                       \
 113 __out:                                                          \
 114         rcu_read_unlock();                                      \
 115         i->iov_offset += __off;                                 \
 116         n = __off;                                              \
 117 }
 118
 119 #define __iterate_and_advance(i, n, base, len, off, I, K) {     \
 120         if (unlikely(i->count < n))                             \
 121                 n = i->count;                                   \
 122         if (likely(n)) {                                        \
 123                 if (likely(iter_is_ubuf(i))) {                  \
 124                         void __user *base;                      \
 125                         size_t len;                             \
 126                         iterate_buf(i, n, base, len, off,       \
 127                                                 i->ubuf, (I))   \
 128                 } else if (likely(iter_is_iovec(i))) {          \
 129                         const struct iovec *iov = iter_iov(i);  \
 130                         void __user *base;                      \
 131                         size_t len;                             \
 132                         iterate_iovec(i, n, base, len, off,     \
 133                                                 iov, (I))       \
 134                         i->nr_segs -= iov - iter_iov(i);        \
 135                         i->__iov = iov;                         \
 136                 } else if (iov_iter_is_bvec(i)) {               \
 137                         const struct bio_vec *bvec = i->bvec;   \
 138                         void *base;                             \
 139                         size_t len;                             \
 140                         iterate_bvec(i, n, base, len, off,      \
 141                                                 bvec, (K))      \
 142                         i->nr_segs -= bvec - i->bvec;           \
 143                         i->bvec = bvec;                         \
 144                 } else if (iov_iter_is_kvec(i)) {               \
 145                         const struct kvec *kvec = i->kvec;      \
 146                         void *base;                             \
 147                         size_t len;                             \
 148                         iterate_iovec(i, n, base, len, off,     \
 149                                                 kvec, (K))      \
 150                         i->nr_segs -= kvec - i->kvec;           \
 151                         i->kvec = kvec;                         \
 152                 } else if (iov_iter_is_xarray(i)) {             \
 153                         void *base;                             \
 154                         size_t len;                             \
 155                         iterate_xarray(i, n, base, len, off,    \
 156                                                         (K))    \
 157                 }                                               \
 158                 i->count -= n;                                  \
 159         }                                                       \
 160 }
 161 #define iterate_and_advance(i, n, base, len, off, I, K) \
 162         __iterate_and_advance(i, n, base, len, off, I, ((void)(K),0))
 163
 164 static int copyout(void __user *to, const void *from, size_t n)
 165 {
 166         if (should_fail_usercopy())
 167                 return n;
 168         if (access_ok(to, n)) {
 169                 instrument_copy_to_user(to, from, n);
 170                 n = raw_copy_to_user(to, from, n);
 171         }
 172         return n;
 173 }
 174
 175 static int copyin(void *to, const void __user *from, size_t n)
 176 {
 177         size_t res = n;
 178
 179         if (should_fail_usercopy())
 180                 return n;
 181         if (access_ok(from, n)) {
 182                 instrument_copy_from_user_before(to, from, n);
 183                 res = raw_copy_from_user(to, from, n);
 184                 instrument_copy_from_user_after(to, from, n, res);
 185         }
 186         return res;
 187 }
 188
 189 #ifdef PIPE_PARANOIA
 190 static bool sanity(const struct iov_iter *i)
 191 {
 192         struct pipe_inode_info *pipe = i->pipe;
 193         unsigned int p_head = pipe->head;
 194         unsigned int p_tail = pipe->tail;
 195         unsigned int p_occupancy = pipe_occupancy(p_head, p_tail);
 196         unsigned int i_head = i->head;
 197         unsigned int idx;
 198
 199         if (i->last_offset) {
 200                 struct pipe_buffer *p;
 201                 if (unlikely(p_occupancy == 0))
 202                         goto Bad;       // pipe must be non-empty
 203                 if (unlikely(i_head != p_head - 1))
 204                         goto Bad;       // must be at the last buffer...
 205
 206                 p = pipe_buf(pipe, i_head);
 207                 if (unlikely(p->offset + p->len != abs(i->last_offset)))
 208                         goto Bad;       // ... at the end of segment
 209         } else {
 210                 if (i_head != p_head)
 211                         goto Bad;       // must be right after the last buffer
 212         }
 213         return true;
 214 Bad:
 215         printk(KERN_ERR "idx = %d, offset = %d\n", i_head, i->last_offset);
 216         printk(KERN_ERR "head = %d, tail = %d, buffers = %d\n",
 217                         p_head, p_tail, pipe->ring_size);
 218         for (idx = 0; idx < pipe->ring_size; idx++)
 219                 printk(KERN_ERR "[%p %p %d %d]\n",
 220                         pipe->bufs[idx].ops,
 221                         pipe->bufs[idx].page,
 222                         pipe->bufs[idx].offset,
 223                         pipe->bufs[idx].len);
 224         WARN_ON(1);
 225         return false;
 226 }
 227 #else
 228 #define sanity(i) true
 229 #endif
 230
 231 static struct page *push_anon(struct pipe_inode_info *pipe, unsigned size)
 232 {
 233         struct page *page = alloc_page(GFP_USER);
 234         if (page) {
 235                 struct pipe_buffer *buf = pipe_buf(pipe, pipe->head++);
 236                 *buf = (struct pipe_buffer) {
 237                         .ops = &default_pipe_buf_ops,
 238                         .page = page,
 239                         .offset = 0,
 240                         .len = size
 241                 };
 242         }
 243         return page;
 244 }
 245
 246 static void push_page(struct pipe_inode_info *pipe, struct page *page,
 247                         unsigned int offset, unsigned int size)
 248 {
 249         struct pipe_buffer *buf = pipe_buf(pipe, pipe->head++);
 250         *buf = (struct pipe_buffer) {
 251                 .ops = &page_cache_pipe_buf_ops,
 252                 .page = page,
 253                 .offset = offset,
 254                 .len = size
 255         };
 256         get_page(page);
 257 }
 258
 259 static inline int last_offset(const struct pipe_buffer *buf)
 260 {
 261         if (buf->ops == &default_pipe_buf_ops)
 262                 return buf->len;        // buf->offset is 0 for those
 263         else
 264                 return -(buf->offset + buf->len);
 265 }
 266
 267 static struct page *append_pipe(struct iov_iter *i, size_t size,
 268                                 unsigned int *off)
 269 {
 270         struct pipe_inode_info *pipe = i->pipe;
 271         int offset = i->last_offset;
 272         struct pipe_buffer *buf;
 273         struct page *page;
 274
 275         if (offset > 0 && offset < PAGE_SIZE) {
 276                 // some space in the last buffer; add to it
 277                 buf = pipe_buf(pipe, pipe->head - 1);
 278                 size = min_t(size_t, size, PAGE_SIZE - offset);
 279                 buf->len += size;
 280                 i->last_offset += size;
 281                 i->count -= size;
 282                 *off = offset;
 283                 return buf->page;
 284         }
 285         // OK, we need a new buffer
 286         *off = 0;
 287         size = min_t(size_t, size, PAGE_SIZE);
 288         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 289                 return NULL;
 290         page = push_anon(pipe, size);
 291         if (!page)
 292                 return NULL;
 293         i->head = pipe->head - 1;
 294         i->last_offset = size;
 295         i->count -= size;
 296         return page;
 297 }
 298
 299 static size_t copy_page_to_iter_pipe(struct page *page, size_t offset, size_t bytes,
 300                          struct iov_iter *i)
 301 {
 302         struct pipe_inode_info *pipe = i->pipe;
 303         unsigned int head = pipe->head;
 304
 305         if (unlikely(bytes > i->count))
 306                 bytes = i->count;
 307
 308         if (unlikely(!bytes))
 309                 return 0;
 310
 311         if (!sanity(i))
 312                 return 0;
 313
 314         if (offset && i->last_offset == -offset) { // could we merge it?
 315                 struct pipe_buffer *buf = pipe_buf(pipe, head - 1);
 316                 if (buf->page == page) {
 317                         buf->len += bytes;
 318                         i->last_offset -= bytes;
 319                         i->count -= bytes;
 320                         return bytes;
 321                 }
 322         }
 323         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 324                 return 0;
 325
 326         push_page(pipe, page, offset, bytes);
 327         i->last_offset = -(offset + bytes);
 328         i->head = head;
 329         i->count -= bytes;
 330         return bytes;
 331 }
 332
 333 /*
 334  * fault_in_iov_iter_readable - fault in iov iterator for reading
 335  * @i: iterator
 336  * @size: maximum length
 337  *
 338  * Fault in one or more iovecs of the given iov_iter, to a maximum length of
 339  * @size.  For each iovec, fault in each page that constitutes the iovec.
 340  *
 341  * Returns the number of bytes not faulted in (like copy_to_user() and
 342  * copy_from_user()).
 343  *
 344  * Always returns 0 for non-userspace iterators.
 345  */
 346 size_t fault_in_iov_iter_readable(const struct iov_iter *i, size_t size)
 347 {
 348         if (iter_is_ubuf(i)) {
 349                 size_t n = min(size, iov_iter_count(i));
 350                 n -= fault_in_readable(i->ubuf + i->iov_offset, n);
 351                 return size - n;
 352         } else if (iter_is_iovec(i)) {
 353                 size_t count = min(size, iov_iter_count(i));
 354                 const struct iovec *p;
 355                 size_t skip;
 356
 357                 size -= count;
 358                 for (p = iter_iov(i), skip = i->iov_offset; count; p++, skip = 0) {
 359                         size_t len = min(count, p->iov_len - skip);
 360                         size_t ret;
 361
 362                         if (unlikely(!len))
 363                                 continue;
 364                         ret = fault_in_readable(p->iov_base + skip, len);
 365                         count -= len - ret;
 366                         if (ret)
 367                                 break;
 368                 }
 369                 return count + size;
 370         }
 371         return 0;
 372 }
 373 EXPORT_SYMBOL(fault_in_iov_iter_readable);
 374
 375 /*
 376  * fault_in_iov_iter_writeable - fault in iov iterator for writing
 377  * @i: iterator
 378  * @size: maximum length
 379  *
 380  * Faults in the iterator using get_user_pages(), i.e., without triggering
 381  * hardware page faults.  This is primarily useful when we already know that
 382  * some or all of the pages in @i aren't in memory.
 383  *
 384  * Returns the number of bytes not faulted in, like copy_to_user() and
 385  * copy_from_user().
 386  *
 387  * Always returns 0 for non-user-space iterators.
 388  */
 389 size_t fault_in_iov_iter_writeable(const struct iov_iter *i, size_t size)
 390 {
 391         if (iter_is_ubuf(i)) {
 392                 size_t n = min(size, iov_iter_count(i));
 393                 n -= fault_in_safe_writeable(i->ubuf + i->iov_offset, n);
 394                 return size - n;
 395         } else if (iter_is_iovec(i)) {
 396                 size_t count = min(size, iov_iter_count(i));
 397                 const struct iovec *p;
 398                 size_t skip;
 399
 400                 size -= count;
 401                 for (p = iter_iov(i), skip = i->iov_offset; count; p++, skip = 0) {
 402                         size_t len = min(count, p->iov_len - skip);
 403                         size_t ret;
 404
 405                         if (unlikely(!len))
 406                                 continue;
 407                         ret = fault_in_safe_writeable(p->iov_base + skip, len);
 408                         count -= len - ret;
 409                         if (ret)
 410                                 break;
 411                 }
 412                 return count + size;
 413         }
 414         return 0;
 415 }
 416 EXPORT_SYMBOL(fault_in_iov_iter_writeable);
 417
 418 void iov_iter_init(struct iov_iter *i, unsigned int direction,
 419                         const struct iovec *iov, unsigned long nr_segs,
 420                         size_t count)
 421 {
 422         WARN_ON(direction & ~(READ | WRITE));
 423         *i = (struct iov_iter) {
 424                 .iter_type = ITER_IOVEC,
 425                 .nofault = false,
 426                 .user_backed = true,
 427                 .data_source = direction,
 428                 .__iov = iov,
 429                 .nr_segs = nr_segs,
 430                 .iov_offset = 0,
 431                 .count = count
 432         };
 433 }
 434 EXPORT_SYMBOL(iov_iter_init);
 435
 436 // returns the offset in partial buffer (if any)
 437 static inline unsigned int pipe_npages(const struct iov_iter *i, int *npages)
 438 {
 439         struct pipe_inode_info *pipe = i->pipe;
 440         int used = pipe->head - pipe->tail;
 441         int off = i->last_offset;
 442
 443         *npages = max((int)pipe->max_usage - used, 0);
 444
 445         if (off > 0 && off < PAGE_SIZE) { // anon and not full
 446                 (*npages)++;
 447                 return off;
 448         }
 449         return 0;
 450 }
 451
 452 static size_t copy_pipe_to_iter(const void *addr, size_t bytes,
 453                                 struct iov_iter *i)
 454 {
 455         unsigned int off, chunk;
 456
 457         if (unlikely(bytes > i->count))
 458                 bytes = i->count;
 459         if (unlikely(!bytes))
 460                 return 0;
 461
 462         if (!sanity(i))
 463                 return 0;
 464
 465         for (size_t n = bytes; n; n -= chunk) {
 466                 struct page *page = append_pipe(i, n, &off);
 467                 chunk = min_t(size_t, n, PAGE_SIZE - off);
 468                 if (!page)
 469                         return bytes - n;
 470                 memcpy_to_page(page, off, addr, chunk);
 471                 addr += chunk;
 472         }
 473         return bytes;
 474 }
 475
 476 static __wsum csum_and_memcpy(void *to, const void *from, size_t len,
 477                               __wsum sum, size_t off)
 478 {
 479         __wsum next = csum_partial_copy_nocheck(from, to, len);
 480         return csum_block_add(sum, next, off);
 481 }
 482
 483 static size_t csum_and_copy_to_pipe_iter(const void *addr, size_t bytes,
 484                                          struct iov_iter *i, __wsum *sump)
 485 {
 486         __wsum sum = *sump;
 487         size_t off = 0;
 488         unsigned int chunk, r;
 489
 490         if (unlikely(bytes > i->count))
 491                 bytes = i->count;
 492         if (unlikely(!bytes))
 493                 return 0;
 494
 495         if (!sanity(i))
 496                 return 0;
 497
 498         while (bytes) {
 499                 struct page *page = append_pipe(i, bytes, &r);
 500                 char *p;
 501
 502                 if (!page)
 503                         break;
 504                 chunk = min_t(size_t, bytes, PAGE_SIZE - r);
 505                 p = kmap_local_page(page);
 506                 sum = csum_and_memcpy(p + r, addr + off, chunk, sum, off);
 507                 kunmap_local(p);
 508                 off += chunk;
 509                 bytes -= chunk;
 510         }
 511         *sump = sum;
 512         return off;
 513 }
 514
 515 size_t _copy_to_iter(const void *addr, size_t bytes, struct iov_iter *i)
 516 {
 517         if (WARN_ON_ONCE(i->data_source))
 518                 return 0;
 519         if (unlikely(iov_iter_is_pipe(i)))
 520                 return copy_pipe_to_iter(addr, bytes, i);
 521         if (user_backed_iter(i))
 522                 might_fault();
 523         iterate_and_advance(i, bytes, base, len, off,
 524                 copyout(base, addr + off, len),
 525                 memcpy(base, addr + off, len)
 526         )
 527
 528         return bytes;
 529 }
 530 EXPORT_SYMBOL(_copy_to_iter);
 531
 532 #ifdef CONFIG_ARCH_HAS_COPY_MC
 533 static int copyout_mc(void __user *to, const void *from, size_t n)
 534 {
 535         if (access_ok(to, n)) {
 536                 instrument_copy_to_user(to, from, n);
 537                 n = copy_mc_to_user((__force void *) to, from, n);
 538         }
 539         return n;
 540 }
 541
 542 static size_t copy_mc_pipe_to_iter(const void *addr, size_t bytes,
 543                                 struct iov_iter *i)
 544 {
 545         size_t xfer = 0;
 546         unsigned int off, chunk;
 547
 548         if (unlikely(bytes > i->count))
 549                 bytes = i->count;
 550         if (unlikely(!bytes))
 551                 return 0;
 552
 553         if (!sanity(i))
 554                 return 0;
 555
 556         while (bytes) {
 557                 struct page *page = append_pipe(i, bytes, &off);
 558                 unsigned long rem;
 559                 char *p;
 560
 561                 if (!page)
 562                         break;
 563                 chunk = min_t(size_t, bytes, PAGE_SIZE - off);
 564                 p = kmap_local_page(page);
 565                 rem = copy_mc_to_kernel(p + off, addr + xfer, chunk);
 566                 chunk -= rem;
 567                 kunmap_local(p);
 568                 xfer += chunk;
 569                 bytes -= chunk;
 570                 if (rem) {
 571                         iov_iter_revert(i, rem);
 572                         break;
 573                 }
 574         }
 575         return xfer;
 576 }
 577
 578 /**
 579  * _copy_mc_to_iter - copy to iter with source memory error exception handling
 580  * @addr: source kernel address
 581  * @bytes: total transfer length
 582  * @i: destination iterator
 583  *
 584  * The pmem driver deploys this for the dax operation
 585  * (dax_copy_to_iter()) for dax reads (bypass page-cache and the
 586  * block-layer). Upon #MC read(2) aborts and returns EIO or the bytes
 587  * successfully copied.
 588  *
 589  * The main differences between this and typical _copy_to_iter().
 590  *
 591  * * Typical tail/residue handling after a fault retries the copy
 592  *   byte-by-byte until the fault happens again. Re-triggering machine
 593  *   checks is potentially fatal so the implementation uses source
 594  *   alignment and poison alignment assumptions to avoid re-triggering
 595  *   hardware exceptions.
 596  *
 597  * * ITER_KVEC, ITER_PIPE, and ITER_BVEC can return short copies.
 598  *   Compare to copy_to_iter() where only ITER_IOVEC attempts might return
 599  *   a short copy.
 600  *
 601  * Return: number of bytes copied (may be %0)
 602  */
 603 size_t _copy_mc_to_iter(const void *addr, size_t bytes, struct iov_iter *i)
 604 {
 605         if (WARN_ON_ONCE(i->data_source))
 606                 return 0;
 607         if (unlikely(iov_iter_is_pipe(i)))
 608                 return copy_mc_pipe_to_iter(addr, bytes, i);
 609         if (user_backed_iter(i))
 610                 might_fault();
 611         __iterate_and_advance(i, bytes, base, len, off,
 612                 copyout_mc(base, addr + off, len),
 613                 copy_mc_to_kernel(base, addr + off, len)
 614         )
 615
 616         return bytes;
 617 }
 618 EXPORT_SYMBOL_GPL(_copy_mc_to_iter);
 619 #endif /* CONFIG_ARCH_HAS_COPY_MC */
 620
 621 size_t _copy_from_iter(void *addr, size_t bytes, struct iov_iter *i)
 622 {
 623         if (WARN_ON_ONCE(!i->data_source))
 624                 return 0;
 625
 626         if (user_backed_iter(i))
 627                 might_fault();
 628         iterate_and_advance(i, bytes, base, len, off,
 629                 copyin(addr + off, base, len),
 630                 memcpy(addr + off, base, len)
 631         )
 632
 633         return bytes;
 634 }
 635 EXPORT_SYMBOL(_copy_from_iter);
 636
 637 size_t _copy_from_iter_nocache(void *addr, size_t bytes, struct iov_iter *i)
 638 {
 639         if (WARN_ON_ONCE(!i->data_source))
 640                 return 0;
 641
 642         iterate_and_advance(i, bytes, base, len, off,
 643                 __copy_from_user_inatomic_nocache(addr + off, base, len),
 644                 memcpy(addr + off, base, len)
 645         )
 646
 647         return bytes;
 648 }
 649 EXPORT_SYMBOL(_copy_from_iter_nocache);
 650
 651 #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
 652 /**
 653  * _copy_from_iter_flushcache - write destination through cpu cache
 654  * @addr: destination kernel address
 655  * @bytes: total transfer length
 656  * @i: source iterator
 657  *
 658  * The pmem driver arranges for filesystem-dax to use this facility via
 659  * dax_copy_from_iter() for ensuring that writes to persistent memory
 660  * are flushed through the CPU cache. It is differentiated from
 661  * _copy_from_iter_nocache() in that guarantees all data is flushed for
 662  * all iterator types. The _copy_from_iter_nocache() only attempts to
 663  * bypass the cache for the ITER_IOVEC case, and on some archs may use
 664  * instructions that strand dirty-data in the cache.
 665  *
 666  * Return: number of bytes copied (may be %0)
 667  */
 668 size_t _copy_from_iter_flushcache(void *addr, size_t bytes, struct iov_iter *i)
 669 {
 670         if (WARN_ON_ONCE(!i->data_source))
 671                 return 0;
 672
 673         iterate_and_advance(i, bytes, base, len, off,
 674                 __copy_from_user_flushcache(addr + off, base, len),
 675                 memcpy_flushcache(addr + off, base, len)
 676         )
 677
 678         return bytes;
 679 }
 680 EXPORT_SYMBOL_GPL(_copy_from_iter_flushcache);
 681 #endif
 682
 683 static inline bool page_copy_sane(struct page *page, size_t offset, size_t n)
 684 {
 685         struct page *head;
 686         size_t v = n + offset;
 687
 688         /*
 689          * The general case needs to access the page order in order
 690          * to compute the page size.
 691          * However, we mostly deal with order-0 pages and thus can
 692          * avoid a possible cache line miss for requests that fit all
 693          * page orders.
 694          */
 695         if (n <= v && v <= PAGE_SIZE)
 696                 return true;
 697
 698         head = compound_head(page);
 699         v += (page - head) << PAGE_SHIFT;
 700
 701         if (WARN_ON(n > v || v > page_size(head)))
 702                 return false;
 703         return true;
 704 }
 705
 706 size_t copy_page_to_iter(struct page *page, size_t offset, size_t bytes,
 707                          struct iov_iter *i)
 708 {
 709         size_t res = 0;
 710         if (!page_copy_sane(page, offset, bytes))
 711                 return 0;
 712         if (WARN_ON_ONCE(i->data_source))
 713                 return 0;
 714         if (unlikely(iov_iter_is_pipe(i)))
 715                 return copy_page_to_iter_pipe(page, offset, bytes, i);
 716         page += offset / PAGE_SIZE; // first subpage
 717         offset %= PAGE_SIZE;
 718         while (1) {
 719                 void *kaddr = kmap_local_page(page);
 720                 size_t n = min(bytes, (size_t)PAGE_SIZE - offset);
 721                 n = _copy_to_iter(kaddr + offset, n, i);
 722                 kunmap_local(kaddr);
 723                 res += n;
 724                 bytes -= n;
 725                 if (!bytes || !n)
 726                         break;
 727                 offset += n;
 728                 if (offset == PAGE_SIZE) {
 729                         page++;
 730                         offset = 0;
 731                 }
 732         }
 733         return res;
 734 }
 735 EXPORT_SYMBOL(copy_page_to_iter);
 736
 737 size_t copy_page_from_iter(struct page *page, size_t offset, size_t bytes,
 738                          struct iov_iter *i)
 739 {
 740         size_t res = 0;
 741         if (!page_copy_sane(page, offset, bytes))
 742                 return 0;
 743         page += offset / PAGE_SIZE; // first subpage
 744         offset %= PAGE_SIZE;
 745         while (1) {
 746                 void *kaddr = kmap_local_page(page);
 747                 size_t n = min(bytes, (size_t)PAGE_SIZE - offset);
 748                 n = _copy_from_iter(kaddr + offset, n, i);
 749                 kunmap_local(kaddr);
 750                 res += n;
 751                 bytes -= n;
 752                 if (!bytes || !n)
 753                         break;
 754                 offset += n;
 755                 if (offset == PAGE_SIZE) {
 756                         page++;
 757                         offset = 0;
 758                 }
 759         }
 760         return res;
 761 }
 762 EXPORT_SYMBOL(copy_page_from_iter);
 763
 764 static size_t pipe_zero(size_t bytes, struct iov_iter *i)
 765 {
 766         unsigned int chunk, off;
 767
 768         if (unlikely(bytes > i->count))
 769                 bytes = i->count;
 770         if (unlikely(!bytes))
 771                 return 0;
 772
 773         if (!sanity(i))
 774                 return 0;
 775
 776         for (size_t n = bytes; n; n -= chunk) {
 777                 struct page *page = append_pipe(i, n, &off);
 778                 char *p;
 779
 780                 if (!page)
 781                         return bytes - n;
 782                 chunk = min_t(size_t, n, PAGE_SIZE - off);
 783                 p = kmap_local_page(page);
 784                 memset(p + off, 0, chunk);
 785                 kunmap_local(p);
 786         }
 787         return bytes;
 788 }
 789
 790 size_t iov_iter_zero(size_t bytes, struct iov_iter *i)
 791 {
 792         if (unlikely(iov_iter_is_pipe(i)))
 793                 return pipe_zero(bytes, i);
 794         iterate_and_advance(i, bytes, base, len, count,
 795                 clear_user(base, len),
 796                 memset(base, 0, len)
 797         )
 798
 799         return bytes;
 800 }
 801 EXPORT_SYMBOL(iov_iter_zero);
 802
 803 size_t copy_page_from_iter_atomic(struct page *page, unsigned offset, size_t bytes,
 804                                   struct iov_iter *i)
 805 {
 806         char *kaddr = kmap_atomic(page), *p = kaddr + offset;
 807         if (!page_copy_sane(page, offset, bytes)) {
 808                 kunmap_atomic(kaddr);
 809                 return 0;
 810         }
 811         if (WARN_ON_ONCE(!i->data_source)) {
 812                 kunmap_atomic(kaddr);
 813                 return 0;
 814         }
 815         iterate_and_advance(i, bytes, base, len, off,
 816                 copyin(p + off, base, len),
 817                 memcpy(p + off, base, len)
 818         )
 819         kunmap_atomic(kaddr);
 820         return bytes;
 821 }
 822 EXPORT_SYMBOL(copy_page_from_iter_atomic);
 823
 824 static void pipe_advance(struct iov_iter *i, size_t size)
 825 {
 826         struct pipe_inode_info *pipe = i->pipe;
 827         int off = i->last_offset;
 828
 829         if (!off && !size) {
 830                 pipe_discard_from(pipe, i->start_head); // discard everything
 831                 return;
 832         }
 833         i->count -= size;
 834         while (1) {
 835                 struct pipe_buffer *buf = pipe_buf(pipe, i->head);
 836                 if (off) /* make it relative to the beginning of buffer */
 837                         size += abs(off) - buf->offset;
 838                 if (size <= buf->len) {
 839                         buf->len = size;
 840                         i->last_offset = last_offset(buf);
 841                         break;
 842                 }
 843                 size -= buf->len;
 844                 i->head++;
 845                 off = 0;
 846         }
 847         pipe_discard_from(pipe, i->head + 1); // discard everything past this one
 848 }
 849
 850 static void iov_iter_bvec_advance(struct iov_iter *i, size_t size)
 851 {
 852         const struct bio_vec *bvec, *end;
 853
 854         if (!i->count)
 855                 return;
 856         i->count -= size;
 857
 858         size += i->iov_offset;
 859
 860         for (bvec = i->bvec, end = bvec + i->nr_segs; bvec < end; bvec++) {
 861                 if (likely(size < bvec->bv_len))
 862                         break;
 863                 size -= bvec->bv_len;
 864         }
 865         i->iov_offset = size;
 866         i->nr_segs -= bvec - i->bvec;
 867         i->bvec = bvec;
 868 }
 869
 870 static void iov_iter_iovec_advance(struct iov_iter *i, size_t size)
 871 {
 872         const struct iovec *iov, *end;
 873
 874         if (!i->count)
 875                 return;
 876         i->count -= size;
 877
 878         size += i->iov_offset; // from beginning of current segment
 879         for (iov = iter_iov(i), end = iov + i->nr_segs; iov < end; iov++) {
 880                 if (likely(size < iov->iov_len))
 881                         break;
 882                 size -= iov->iov_len;
 883         }
 884         i->iov_offset = size;
 885         i->nr_segs -= iov - iter_iov(i);
 886         i->__iov = iov;
 887 }
 888
 889 void iov_iter_advance(struct iov_iter *i, size_t size)
 890 {
 891         if (unlikely(i->count < size))
 892                 size = i->count;
 893         if (likely(iter_is_ubuf(i)) || unlikely(iov_iter_is_xarray(i))) {
 894                 i->iov_offset += size;
 895                 i->count -= size;
 896         } else if (likely(iter_is_iovec(i) || iov_iter_is_kvec(i))) {
 897                 /* iovec and kvec have identical layouts */
 898                 iov_iter_iovec_advance(i, size);
 899         } else if (iov_iter_is_bvec(i)) {
 900                 iov_iter_bvec_advance(i, size);
 901         } else if (iov_iter_is_pipe(i)) {
 902                 pipe_advance(i, size);
 903         } else if (iov_iter_is_discard(i)) {
 904                 i->count -= size;
 905         }
 906 }
 907 EXPORT_SYMBOL(iov_iter_advance);
 908
 909 void iov_iter_revert(struct iov_iter *i, size_t unroll)
 910 {
 911         if (!unroll)
 912                 return;
 913         if (WARN_ON(unroll > MAX_RW_COUNT))
 914                 return;
 915         i->count += unroll;
 916         if (unlikely(iov_iter_is_pipe(i))) {
 917                 struct pipe_inode_info *pipe = i->pipe;
 918                 unsigned int head = pipe->head;
 919
 920                 while (head > i->start_head) {
 921                         struct pipe_buffer *b = pipe_buf(pipe, --head);
 922                         if (unroll < b->len) {
 923                                 b->len -= unroll;
 924                                 i->last_offset = last_offset(b);
 925                                 i->head = head;
 926                                 return;
 927                         }
 928                         unroll -= b->len;
 929                         pipe_buf_release(pipe, b);
 930                         pipe->head--;
 931                 }
 932                 i->last_offset = 0;
 933                 i->head = head;
 934                 return;
 935         }
 936         if (unlikely(iov_iter_is_discard(i)))
 937                 return;
 938         if (unroll <= i->iov_offset) {
 939                 i->iov_offset -= unroll;
 940                 return;
 941         }
 942         unroll -= i->iov_offset;
 943         if (iov_iter_is_xarray(i) || iter_is_ubuf(i)) {
 944                 BUG(); /* We should never go beyond the start of the specified
 945                         * range since we might then be straying into pages that
 946                         * aren't pinned.
 947                         */
 948         } else if (iov_iter_is_bvec(i)) {
 949                 const struct bio_vec *bvec = i->bvec;
 950                 while (1) {
 951                         size_t n = (--bvec)->bv_len;
 952                         i->nr_segs++;
 953                         if (unroll <= n) {
 954                                 i->bvec = bvec;
 955                                 i->iov_offset = n - unroll;
 956                                 return;
 957                         }
 958                         unroll -= n;
 959                 }
 960         } else { /* same logics for iovec and kvec */
 961                 const struct iovec *iov = iter_iov(i);
 962                 while (1) {
 963                         size_t n = (--iov)->iov_len;
 964                         i->nr_segs++;
 965                         if (unroll <= n) {
 966                                 i->__iov = iov;
 967                                 i->iov_offset = n - unroll;
 968                                 return;
 969                         }
 970                         unroll -= n;
 971                 }
 972         }
 973 }
 974 EXPORT_SYMBOL(iov_iter_revert);
 975
 976 /*
 977  * Return the count of just the current iov_iter segment.
 978  */
 979 size_t iov_iter_single_seg_count(const struct iov_iter *i)
 980 {
 981         if (i->nr_segs > 1) {
 982                 if (likely(iter_is_iovec(i) || iov_iter_is_kvec(i)))
 983                         return min(i->count, iter_iov(i)->iov_len - i->iov_offset);
 984                 if (iov_iter_is_bvec(i))
 985                         return min(i->count, i->bvec->bv_len - i->iov_offset);
 986         }
 987         return i->count;
 988 }
 989 EXPORT_SYMBOL(iov_iter_single_seg_count);
 990
 991 void iov_iter_kvec(struct iov_iter *i, unsigned int direction,
 992                         const struct kvec *kvec, unsigned long nr_segs,
 993                         size_t count)
 994 {
 995         WARN_ON(direction & ~(READ | WRITE));
 996         *i = (struct iov_iter){
 997                 .iter_type = ITER_KVEC,
 998                 .data_source = direction,
 999                 .kvec = kvec,
1000                 .nr_segs = nr_segs,
1001                 .iov_offset = 0,
1002                 .count = count
1003         };
1004 }
1005 EXPORT_SYMBOL(iov_iter_kvec);
1006
1007 void iov_iter_bvec(struct iov_iter *i, unsigned int direction,
1008                         const struct bio_vec *bvec, unsigned long nr_segs,
1009                         size_t count)
1010 {
1011         WARN_ON(direction & ~(READ | WRITE));
1012         *i = (struct iov_iter){
1013                 .iter_type = ITER_BVEC,
1014                 .data_source = direction,
1015                 .bvec = bvec,
1016                 .nr_segs = nr_segs,
1017                 .iov_offset = 0,
1018                 .count = count
1019         };
1020 }
1021 EXPORT_SYMBOL(iov_iter_bvec);
1022
1023 void iov_iter_pipe(struct iov_iter *i, unsigned int direction,
1024                         struct pipe_inode_info *pipe,
1025                         size_t count)
1026 {
1027         BUG_ON(direction != READ);
1028         WARN_ON(pipe_full(pipe->head, pipe->tail, pipe->ring_size));
1029         *i = (struct iov_iter){
1030                 .iter_type = ITER_PIPE,
1031                 .data_source = false,
1032                 .pipe = pipe,
1033                 .head = pipe->head,
1034                 .start_head = pipe->head,
1035                 .last_offset = 0,
1036                 .count = count
1037         };
1038 }
1039 EXPORT_SYMBOL(iov_iter_pipe);
1040
1041 /**
1042  * iov_iter_xarray - Initialise an I/O iterator to use the pages in an xarray
1043  * @i: The iterator to initialise.
1044  * @direction: The direction of the transfer.
1045  * @xarray: The xarray to access.
1046  * @start: The start file position.
1047  * @count: The size of the I/O buffer in bytes.
1048  *
1049  * Set up an I/O iterator to either draw data out of the pages attached to an
1050  * inode or to inject data into those pages.  The pages *must* be prevented
1051  * from evaporation, either by taking a ref on them or locking them by the
1052  * caller.
1053  */
1054 void iov_iter_xarray(struct iov_iter *i, unsigned int direction,
1055                      struct xarray *xarray, loff_t start, size_t count)
1056 {
1057         BUG_ON(direction & ~1);
1058         *i = (struct iov_iter) {
1059                 .iter_type = ITER_XARRAY,
1060                 .data_source = direction,
1061                 .xarray = xarray,
1062                 .xarray_start = start,
1063                 .count = count,
1064                 .iov_offset = 0
1065         };
1066 }
1067 EXPORT_SYMBOL(iov_iter_xarray);
1068
1069 /**
1070  * iov_iter_discard - Initialise an I/O iterator that discards data
1071  * @i: The iterator to initialise.
1072  * @direction: The direction of the transfer.
1073  * @count: The size of the I/O buffer in bytes.
1074  *
1075  * Set up an I/O iterator that just discards everything that's written to it.
1076  * It's only available as a READ iterator.
1077  */
1078 void iov_iter_discard(struct iov_iter *i, unsigned int direction, size_t count)
1079 {
1080         BUG_ON(direction != READ);
1081         *i = (struct iov_iter){
1082                 .iter_type = ITER_DISCARD,
1083                 .data_source = false,
1084                 .count = count,
1085                 .iov_offset = 0
1086         };
1087 }
1088 EXPORT_SYMBOL(iov_iter_discard);
1089
1090 static bool iov_iter_aligned_iovec(const struct iov_iter *i, unsigned addr_mask,
1091                                    unsigned len_mask)
1092 {
1093         size_t size = i->count;
1094         size_t skip = i->iov_offset;
1095         unsigned k;
1096
1097         for (k = 0; k < i->nr_segs; k++, skip = 0) {
1098                 const struct iovec *iov = iter_iov(i) + k;
1099                 size_t len = iov->iov_len - skip;
1100
1101                 if (len > size)
1102                         len = size;
1103                 if (len & len_mask)
1104                         return false;
1105                 if ((unsigned long)(iov->iov_base + skip) & addr_mask)
1106                         return false;
1107
1108                 size -= len;
1109                 if (!size)
1110                         break;
1111         }
1112         return true;
1113 }
1114
1115 static bool iov_iter_aligned_bvec(const struct iov_iter *i, unsigned addr_mask,
1116                                   unsigned len_mask)
1117 {
1118         size_t size = i->count;
1119         unsigned skip = i->iov_offset;
1120         unsigned k;
1121
1122         for (k = 0; k < i->nr_segs; k++, skip = 0) {
1123                 size_t len = i->bvec[k].bv_len - skip;
1124
1125                 if (len > size)
1126                         len = size;
1127                 if (len & len_mask)
1128                         return false;
1129                 if ((unsigned long)(i->bvec[k].bv_offset + skip) & addr_mask)
1130                         return false;
1131
1132                 size -= len;
1133                 if (!size)
1134                         break;
1135         }
1136         return true;
1137 }
1138
1139 /**
1140  * iov_iter_is_aligned() - Check if the addresses and lengths of each segments
1141  *      are aligned to the parameters.
1142  *
1143  * @i: &struct iov_iter to restore
1144  * @addr_mask: bit mask to check against the iov element's addresses
1145  * @len_mask: bit mask to check against the iov element's lengths
1146  *
1147  * Return: false if any addresses or lengths intersect with the provided masks
1148  */
1149 bool iov_iter_is_aligned(const struct iov_iter *i, unsigned addr_mask,
1150                          unsigned len_mask)
1151 {
1152         if (likely(iter_is_ubuf(i))) {
1153                 if (i->count & len_mask)
1154                         return false;
1155                 if ((unsigned long)(i->ubuf + i->iov_offset) & addr_mask)
1156                         return false;
1157                 return true;
1158         }
1159
1160         if (likely(iter_is_iovec(i) || iov_iter_is_kvec(i)))
1161                 return iov_iter_aligned_iovec(i, addr_mask, len_mask);
1162
1163         if (iov_iter_is_bvec(i))
1164                 return iov_iter_aligned_bvec(i, addr_mask, len_mask);
1165
1166         if (iov_iter_is_pipe(i)) {
1167                 size_t size = i->count;
1168
1169                 if (size & len_mask)
1170                         return false;
1171                 if (size && i->last_offset > 0) {
1172                         if (i->last_offset & addr_mask)
1173                                 return false;
1174                 }
1175
1176                 return true;
1177         }
1178
1179         if (iov_iter_is_xarray(i)) {
1180                 if (i->count & len_mask)
1181                         return false;
1182                 if ((i->xarray_start + i->iov_offset) & addr_mask)
1183                         return false;
1184         }
1185
1186         return true;
1187 }
1188 EXPORT_SYMBOL_GPL(iov_iter_is_aligned);
1189
1190 static unsigned long iov_iter_alignment_iovec(const struct iov_iter *i)
1191 {
1192         unsigned long res = 0;
1193         size_t size = i->count;
1194         size_t skip = i->iov_offset;
1195         unsigned k;
1196
1197         for (k = 0; k < i->nr_segs; k++, skip = 0) {
1198                 const struct iovec *iov = iter_iov(i) + k;
1199                 size_t len = iov->iov_len - skip;
1200                 if (len) {
1201                         res |= (unsigned long)iov->iov_base + skip;
1202                         if (len > size)
1203                                 len = size;
1204                         res |= len;
1205                         size -= len;
1206                         if (!size)
1207                                 break;
1208                 }
1209         }
1210         return res;
1211 }
1212
1213 static unsigned long iov_iter_alignment_bvec(const struct iov_iter *i)
1214 {
1215         unsigned res = 0;
1216         size_t size = i->count;
1217         unsigned skip = i->iov_offset;
1218         unsigned k;
1219
1220         for (k = 0; k < i->nr_segs; k++, skip = 0) {
1221                 size_t len = i->bvec[k].bv_len - skip;
1222                 res |= (unsigned long)i->bvec[k].bv_offset + skip;
1223                 if (len > size)
1224                         len = size;
1225                 res |= len;
1226                 size -= len;
1227                 if (!size)
1228                         break;
1229         }
1230         return res;
1231 }
1232
1233 unsigned long iov_iter_alignment(const struct iov_iter *i)
1234 {
1235         if (likely(iter_is_ubuf(i))) {
1236                 size_t size = i->count;
1237                 if (size)
1238                         return ((unsigned long)i->ubuf + i->iov_offset) | size;
1239                 return 0;
1240         }
1241
1242         /* iovec and kvec have identical layouts */
1243         if (likely(iter_is_iovec(i) || iov_iter_is_kvec(i)))
1244                 return iov_iter_alignment_iovec(i);
1245
1246         if (iov_iter_is_bvec(i))
1247                 return iov_iter_alignment_bvec(i);
1248
1249         if (iov_iter_is_pipe(i)) {
1250                 size_t size = i->count;
1251
1252                 if (size && i->last_offset > 0)
1253                         return size | i->last_offset;
1254                 return size;
1255         }
1256
1257         if (iov_iter_is_xarray(i))
1258                 return (i->xarray_start + i->iov_offset) | i->count;
1259
1260         return 0;
1261 }
1262 EXPORT_SYMBOL(iov_iter_alignment);
1263
1264 unsigned long iov_iter_gap_alignment(const struct iov_iter *i)
1265 {
1266         unsigned long res = 0;
1267         unsigned long v = 0;
1268         size_t size = i->count;
1269         unsigned k;
1270
1271         if (iter_is_ubuf(i))
1272                 return 0;
1273
1274         if (WARN_ON(!iter_is_iovec(i)))
1275                 return ~0U;
1276
1277         for (k = 0; k < i->nr_segs; k++) {
1278                 const struct iovec *iov = iter_iov(i) + k;
1279                 if (iov->iov_len) {
1280                         unsigned long base = (unsigned long)iov->iov_base;
1281                         if (v) // if not the first one
1282                                 res |= base | v; // this start | previous end
1283                         v = base + iov->iov_len;
1284                         if (size <= iov->iov_len)
1285                                 break;
1286                         size -= iov->iov_len;
1287                 }
1288         }
1289         return res;
1290 }
1291 EXPORT_SYMBOL(iov_iter_gap_alignment);
1292
1293 static int want_pages_array(struct page ***res, size_t size,
1294                             size_t start, unsigned int maxpages)
1295 {
1296         unsigned int count = DIV_ROUND_UP(size + start, PAGE_SIZE);
1297
1298         if (count > maxpages)
1299                 count = maxpages;
1300         WARN_ON(!count);        // caller should've prevented that
1301         if (!*res) {
1302                 *res = kvmalloc_array(count, sizeof(struct page *), GFP_KERNEL);
1303                 if (!*res)
1304                         return 0;
1305         }
1306         return count;
1307 }
1308
1309 static ssize_t pipe_get_pages(struct iov_iter *i,
1310                    struct page ***pages, size_t maxsize, unsigned maxpages,
1311                    size_t *start)
1312 {
1313         unsigned int npages, count, off, chunk;
1314         struct page **p;
1315         size_t left;
1316
1317         if (!sanity(i))
1318                 return -EFAULT;
1319
1320         *start = off = pipe_npages(i, &npages);
1321         if (!npages)
1322                 return -EFAULT;
1323         count = want_pages_array(pages, maxsize, off, min(npages, maxpages));
1324         if (!count)
1325                 return -ENOMEM;
1326         p = *pages;
1327         for (npages = 0, left = maxsize ; npages < count; npages++, left -= chunk) {
1328                 struct page *page = append_pipe(i, left, &off);
1329                 if (!page)
1330                         break;
1331                 chunk = min_t(size_t, left, PAGE_SIZE - off);
1332                 get_page(*p++ = page);
1333         }
1334         if (!npages)
1335                 return -EFAULT;
1336         return maxsize - left;
1337 }
1338
1339 static ssize_t iter_xarray_populate_pages(struct page **pages, struct xarray *xa,
1340                                           pgoff_t index, unsigned int nr_pages)
1341 {
1342         XA_STATE(xas, xa, index);
1343         struct page *page;
1344         unsigned int ret = 0;
1345
1346         rcu_read_lock();
1347         for (page = xas_load(&xas); page; page = xas_next(&xas)) {
1348                 if (xas_retry(&xas, page))
1349                         continue;
1350
1351                 /* Has the page moved or been split? */
1352                 if (unlikely(page != xas_reload(&xas))) {
1353                         xas_reset(&xas);
1354                         continue;
1355                 }
1356
1357                 pages[ret] = find_subpage(page, xas.xa_index);
1358                 get_page(pages[ret]);
1359                 if (++ret == nr_pages)
1360                         break;
1361         }
1362         rcu_read_unlock();
1363         return ret;
1364 }
1365
1366 static ssize_t iter_xarray_get_pages(struct iov_iter *i,
1367                                      struct page ***pages, size_t maxsize,
1368                                      unsigned maxpages, size_t *_start_offset)
1369 {
1370         unsigned nr, offset, count;
1371         pgoff_t index;
1372         loff_t pos;
1373
1374         pos = i->xarray_start + i->iov_offset;
1375         index = pos >> PAGE_SHIFT;
1376         offset = pos & ~PAGE_MASK;
1377         *_start_offset = offset;
1378
1379         count = want_pages_array(pages, maxsize, offset, maxpages);
1380         if (!count)
1381                 return -ENOMEM;
1382         nr = iter_xarray_populate_pages(*pages, i->xarray, index, count);
1383         if (nr == 0)
1384                 return 0;
1385
1386         maxsize = min_t(size_t, nr * PAGE_SIZE - offset, maxsize);
1387         i->iov_offset += maxsize;
1388         i->count -= maxsize;
1389         return maxsize;
1390 }
1391
1392 /* must be done on non-empty ITER_UBUF or ITER_IOVEC one */
1393 static unsigned long first_iovec_segment(const struct iov_iter *i, size_t *size)
1394 {
1395         size_t skip;
1396         long k;
1397
1398         if (iter_is_ubuf(i))
1399                 return (unsigned long)i->ubuf + i->iov_offset;
1400
1401         for (k = 0, skip = i->iov_offset; k < i->nr_segs; k++, skip = 0) {
1402                 const struct iovec *iov = iter_iov(i) + k;
1403                 size_t len = iov->iov_len - skip;
1404
1405                 if (unlikely(!len))
1406                         continue;
1407                 if (*size > len)
1408                         *size = len;
1409                 return (unsigned long)iov->iov_base + skip;
1410         }
1411         BUG(); // if it had been empty, we wouldn't get called
1412 }
1413
1414 /* must be done on non-empty ITER_BVEC one */
1415 static struct page *first_bvec_segment(const struct iov_iter *i,
1416                                        size_t *size, size_t *start)
1417 {
1418         struct page *page;
1419         size_t skip = i->iov_offset, len;
1420
1421         len = i->bvec->bv_len - skip;
1422         if (*size > len)
1423                 *size = len;
1424         skip += i->bvec->bv_offset;
1425         page = i->bvec->bv_page + skip / PAGE_SIZE;
1426         *start = skip % PAGE_SIZE;
1427         return page;
1428 }
1429
1430 static ssize_t __iov_iter_get_pages_alloc(struct iov_iter *i,
1431                    struct page ***pages, size_t maxsize,
1432                    unsigned int maxpages, size_t *start,
1433                    iov_iter_extraction_t extraction_flags)
1434 {
1435         unsigned int n, gup_flags = 0;
1436
1437         if (maxsize > i->count)
1438                 maxsize = i->count;
1439         if (!maxsize)
1440                 return 0;
1441         if (maxsize > MAX_RW_COUNT)
1442                 maxsize = MAX_RW_COUNT;
1443         if (extraction_flags & ITER_ALLOW_P2PDMA)
1444                 gup_flags |= FOLL_PCI_P2PDMA;
1445
1446         if (likely(user_backed_iter(i))) {
1447                 unsigned long addr;
1448                 int res;
1449
1450                 if (iov_iter_rw(i) != WRITE)
1451                         gup_flags |= FOLL_WRITE;
1452                 if (i->nofault)
1453                         gup_flags |= FOLL_NOFAULT;
1454
1455                 addr = first_iovec_segment(i, &maxsize);
1456                 *start = addr % PAGE_SIZE;
1457                 addr &= PAGE_MASK;
1458                 n = want_pages_array(pages, maxsize, *start, maxpages);
1459                 if (!n)
1460                         return -ENOMEM;
1461                 res = get_user_pages_fast(addr, n, gup_flags, *pages);
1462                 if (unlikely(res <= 0))
1463                         return res;
1464                 maxsize = min_t(size_t, maxsize, res * PAGE_SIZE - *start);
1465                 iov_iter_advance(i, maxsize);
1466                 return maxsize;
1467         }
1468         if (iov_iter_is_bvec(i)) {
1469                 struct page **p;
1470                 struct page *page;
1471
1472                 page = first_bvec_segment(i, &maxsize, start);
1473                 n = want_pages_array(pages, maxsize, *start, maxpages);
1474                 if (!n)
1475                         return -ENOMEM;
1476                 p = *pages;
1477                 for (int k = 0; k < n; k++)
1478                         get_page(p[k] = page + k);
1479                 maxsize = min_t(size_t, maxsize, n * PAGE_SIZE - *start);
1480                 i->count -= maxsize;
1481                 i->iov_offset += maxsize;
1482                 if (i->iov_offset == i->bvec->bv_len) {
1483                         i->iov_offset = 0;
1484                         i->bvec++;
1485                         i->nr_segs--;
1486                 }
1487                 return maxsize;
1488         }
1489         if (iov_iter_is_pipe(i))
1490                 return pipe_get_pages(i, pages, maxsize, maxpages, start);
1491         if (iov_iter_is_xarray(i))
1492                 return iter_xarray_get_pages(i, pages, maxsize, maxpages, start);
1493         return -EFAULT;
1494 }
1495
1496 ssize_t iov_iter_get_pages(struct iov_iter *i,
1497                    struct page **pages, size_t maxsize, unsigned maxpages,
1498                    size_t *start, iov_iter_extraction_t extraction_flags)
1499 {
1500         if (!maxpages)
1501                 return 0;
1502         BUG_ON(!pages);
1503
1504         return __iov_iter_get_pages_alloc(i, &pages, maxsize, maxpages,
1505                                           start, extraction_flags);
1506 }
1507 EXPORT_SYMBOL_GPL(iov_iter_get_pages);
1508
1509 ssize_t iov_iter_get_pages2(struct iov_iter *i, struct page **pages,
1510                 size_t maxsize, unsigned maxpages, size_t *start)
1511 {
1512         return iov_iter_get_pages(i, pages, maxsize, maxpages, start, 0);
1513 }
1514 EXPORT_SYMBOL(iov_iter_get_pages2);
1515
1516 ssize_t iov_iter_get_pages_alloc(struct iov_iter *i,
1517                    struct page ***pages, size_t maxsize,
1518                    size_t *start, iov_iter_extraction_t extraction_flags)
1519 {
1520         ssize_t len;
1521
1522         *pages = NULL;
1523
1524         len = __iov_iter_get_pages_alloc(i, pages, maxsize, ~0U, start,
1525                                          extraction_flags);
1526         if (len <= 0) {
1527                 kvfree(*pages);
1528                 *pages = NULL;
1529         }
1530         return len;
1531 }
1532 EXPORT_SYMBOL_GPL(iov_iter_get_pages_alloc);
1533
1534 ssize_t iov_iter_get_pages_alloc2(struct iov_iter *i,
1535                 struct page ***pages, size_t maxsize, size_t *start)
1536 {
1537         return iov_iter_get_pages_alloc(i, pages, maxsize, start, 0);
1538 }
1539 EXPORT_SYMBOL(iov_iter_get_pages_alloc2);
1540
1541 size_t csum_and_copy_from_iter(void *addr, size_t bytes, __wsum *csum,
1542                                struct iov_iter *i)
1543 {
1544         __wsum sum, next;
1545         sum = *csum;
1546         if (WARN_ON_ONCE(!i->data_source))
1547                 return 0;
1548
1549         iterate_and_advance(i, bytes, base, len, off, ({
1550                 next = csum_and_copy_from_user(base, addr + off, len);
1551                 sum = csum_block_add(sum, next, off);
1552                 next ? 0 : len;
1553         }), ({
1554                 sum = csum_and_memcpy(addr + off, base, len, sum, off);
1555         })
1556         )
1557         *csum = sum;
1558         return bytes;
1559 }
1560 EXPORT_SYMBOL(csum_and_copy_from_iter);
1561
1562 size_t csum_and_copy_to_iter(const void *addr, size_t bytes, void *_csstate,
1563                              struct iov_iter *i)
1564 {
1565         struct csum_state *csstate = _csstate;
1566         __wsum sum, next;
1567
1568         if (WARN_ON_ONCE(i->data_source))
1569                 return 0;
1570         if (unlikely(iov_iter_is_discard(i))) {
1571                 // can't use csum_memcpy() for that one - data is not copied
1572                 csstate->csum = csum_block_add(csstate->csum,
1573                                                csum_partial(addr, bytes, 0),
1574                                                csstate->off);
1575                 csstate->off += bytes;
1576                 return bytes;
1577         }
1578
1579         sum = csum_shift(csstate->csum, csstate->off);
1580         if (unlikely(iov_iter_is_pipe(i)))
1581                 bytes = csum_and_copy_to_pipe_iter(addr, bytes, i, &sum);
1582         else iterate_and_advance(i, bytes, base, len, off, ({
1583                 next = csum_and_copy_to_user(addr + off, base, len);
1584                 sum = csum_block_add(sum, next, off);
1585                 next ? 0 : len;
1586         }), ({
1587                 sum = csum_and_memcpy(base, addr + off, len, sum, off);
1588         })
1589         )
1590         csstate->csum = csum_shift(sum, csstate->off);
1591         csstate->off += bytes;
1592         return bytes;
1593 }
1594 EXPORT_SYMBOL(csum_and_copy_to_iter);
1595
1596 size_t hash_and_copy_to_iter(const void *addr, size_t bytes, void *hashp,
1597                 struct iov_iter *i)
1598 {
1599 #ifdef CONFIG_CRYPTO_HASH
1600         struct ahash_request *hash = hashp;
1601         struct scatterlist sg;
1602         size_t copied;
1603
1604         copied = copy_to_iter(addr, bytes, i);
1605         sg_init_one(&sg, addr, copied);
1606         ahash_request_set_crypt(hash, &sg, NULL, copied);
1607         crypto_ahash_update(hash);
1608         return copied;
1609 #else
1610         return 0;
1611 #endif
1612 }
1613 EXPORT_SYMBOL(hash_and_copy_to_iter);
1614
1615 static int iov_npages(const struct iov_iter *i, int maxpages)
1616 {
1617         size_t skip = i->iov_offset, size = i->count;
1618         const struct iovec *p;
1619         int npages = 0;
1620
1621         for (p = iter_iov(i); size; skip = 0, p++) {
1622                 unsigned offs = offset_in_page(p->iov_base + skip);
1623                 size_t len = min(p->iov_len - skip, size);
1624
1625                 if (len) {
1626                         size -= len;
1627                         npages += DIV_ROUND_UP(offs + len, PAGE_SIZE);
1628                         if (unlikely(npages > maxpages))
1629                                 return maxpages;
1630                 }
1631         }
1632         return npages;
1633 }
1634
1635 static int bvec_npages(const struct iov_iter *i, int maxpages)
1636 {
1637         size_t skip = i->iov_offset, size = i->count;
1638         const struct bio_vec *p;
1639         int npages = 0;
1640
1641         for (p = i->bvec; size; skip = 0, p++) {
1642                 unsigned offs = (p->bv_offset + skip) % PAGE_SIZE;
1643                 size_t len = min(p->bv_len - skip, size);
1644
1645                 size -= len;
1646                 npages += DIV_ROUND_UP(offs + len, PAGE_SIZE);
1647                 if (unlikely(npages > maxpages))
1648                         return maxpages;
1649         }
1650         return npages;
1651 }
1652
1653 int iov_iter_npages(const struct iov_iter *i, int maxpages)
1654 {
1655         if (unlikely(!i->count))
1656                 return 0;
1657         if (likely(iter_is_ubuf(i))) {
1658                 unsigned offs = offset_in_page(i->ubuf + i->iov_offset);
1659                 int npages = DIV_ROUND_UP(offs + i->count, PAGE_SIZE);
1660                 return min(npages, maxpages);
1661         }
1662         /* iovec and kvec have identical layouts */
1663         if (likely(iter_is_iovec(i) || iov_iter_is_kvec(i)))
1664                 return iov_npages(i, maxpages);
1665         if (iov_iter_is_bvec(i))
1666                 return bvec_npages(i, maxpages);
1667         if (iov_iter_is_pipe(i)) {
1668                 int npages;
1669
1670                 if (!sanity(i))
1671                         return 0;
1672
1673                 pipe_npages(i, &npages);
1674                 return min(npages, maxpages);
1675         }
1676         if (iov_iter_is_xarray(i)) {
1677                 unsigned offset = (i->xarray_start + i->iov_offset) % PAGE_SIZE;
1678                 int npages = DIV_ROUND_UP(offset + i->count, PAGE_SIZE);
1679                 return min(npages, maxpages);
1680         }
1681         return 0;
1682 }
1683 EXPORT_SYMBOL(iov_iter_npages);
1684
1685 const void *dup_iter(struct iov_iter *new, struct iov_iter *old, gfp_t flags)
1686 {
1687         *new = *old;
1688         if (unlikely(iov_iter_is_pipe(new))) {
1689                 WARN_ON(1);
1690                 return NULL;
1691         }
1692         if (iov_iter_is_bvec(new))
1693                 return new->bvec = kmemdup(new->bvec,
1694                                     new->nr_segs * sizeof(struct bio_vec),
1695                                     flags);
1696         else if (iov_iter_is_kvec(new) || iter_is_iovec(new))
1697                 /* iovec and kvec have identical layout */
1698                 return new->__iov = kmemdup(new->__iov,
1699                                    new->nr_segs * sizeof(struct iovec),
1700                                    flags);
1701         return NULL;
1702 }
1703 EXPORT_SYMBOL(dup_iter);
1704
1705 static __noclone int copy_compat_iovec_from_user(struct iovec *iov,
1706                 const struct iovec __user *uvec, unsigned long nr_segs)
1707 {
1708         const struct compat_iovec __user *uiov =
1709                 (const struct compat_iovec __user *)uvec;
1710         int ret = -EFAULT, i;
1711
1712         if (!user_access_begin(uiov, nr_segs * sizeof(*uiov)))
1713                 return -EFAULT;
1714
1715         for (i = 0; i < nr_segs; i++) {
1716                 compat_uptr_t buf;
1717                 compat_ssize_t len;
1718
1719                 unsafe_get_user(len, &uiov[i].iov_len, uaccess_end);
1720                 unsafe_get_user(buf, &uiov[i].iov_base, uaccess_end);
1721
1722                 /* check for compat_size_t not fitting in compat_ssize_t .. */
1723                 if (len < 0) {
1724                         ret = -EINVAL;
1725                         goto uaccess_end;
1726                 }
1727                 iov[i].iov_base = compat_ptr(buf);
1728                 iov[i].iov_len = len;
1729         }
1730
1731         ret = 0;
1732 uaccess_end:
1733         user_access_end();
1734         return ret;
1735 }
1736
1737 static int copy_iovec_from_user(struct iovec *iov,
1738                 const struct iovec __user *uiov, unsigned long nr_segs)
1739 {
1740         int ret = -EFAULT;
1741
1742         if (!user_access_begin(uiov, nr_segs * sizeof(*uiov)))
1743                 return -EFAULT;
1744
1745         do {
1746                 void __user *buf;
1747                 ssize_t len;
1748
1749                 unsafe_get_user(len, &uiov->iov_len, uaccess_end);
1750                 unsafe_get_user(buf, &uiov->iov_base, uaccess_end);
1751
1752                 /* check for size_t not fitting in ssize_t .. */
1753                 if (unlikely(len < 0)) {
1754                         ret = -EINVAL;
1755                         goto uaccess_end;
1756                 }
1757                 iov->iov_base = buf;
1758                 iov->iov_len = len;
1759
1760                 uiov++; iov++;
1761         } while (--nr_segs);
1762
1763         ret = 0;
1764 uaccess_end:
1765         user_access_end();
1766         return ret;
1767 }
1768
1769 struct iovec *iovec_from_user(const struct iovec __user *uvec,
1770                 unsigned long nr_segs, unsigned long fast_segs,
1771                 struct iovec *fast_iov, bool compat)
1772 {
1773         struct iovec *iov = fast_iov;
1774         int ret;
1775
1776         /*
1777          * SuS says "The readv() function *may* fail if the iovcnt argument was
1778          * less than or equal to 0, or greater than {IOV_MAX}.  Linux has
1779          * traditionally returned zero for zero segments, so...
1780          */
1781         if (nr_segs == 0)
1782                 return iov;
1783         if (nr_segs > UIO_MAXIOV)
1784                 return ERR_PTR(-EINVAL);
1785         if (nr_segs > fast_segs) {
1786                 iov = kmalloc_array(nr_segs, sizeof(struct iovec), GFP_KERNEL);
1787                 if (!iov)
1788                         return ERR_PTR(-ENOMEM);
1789         }
1790
1791         if (unlikely(compat))
1792                 ret = copy_compat_iovec_from_user(iov, uvec, nr_segs);
1793         else
1794                 ret = copy_iovec_from_user(iov, uvec, nr_segs);
1795         if (ret) {
1796                 if (iov != fast_iov)
1797                         kfree(iov);
1798                 return ERR_PTR(ret);
1799         }
1800
1801         return iov;
1802 }
1803
1804 /*
1805  * Single segment iovec supplied by the user, import it as ITER_UBUF.
1806  */
1807 static ssize_t __import_iovec_ubuf(int type, const struct iovec __user *uvec,
1808                                    struct iovec **iovp, struct iov_iter *i,
1809                                    bool compat)
1810 {
1811         struct iovec *iov = *iovp;
1812         ssize_t ret;
1813
1814         if (compat)
1815                 ret = copy_compat_iovec_from_user(iov, uvec, 1);
1816         else
1817                 ret = copy_iovec_from_user(iov, uvec, 1);
1818         if (unlikely(ret))
1819                 return ret;
1820
1821         ret = import_ubuf(type, iov->iov_base, iov->iov_len, i);
1822         if (unlikely(ret))
1823                 return ret;
1824         *iovp = NULL;
1825         return i->count;
1826 }
1827
1828 ssize_t __import_iovec(int type, const struct iovec __user *uvec,
1829                  unsigned nr_segs, unsigned fast_segs, struct iovec **iovp,
1830                  struct iov_iter *i, bool compat)
1831 {
1832         ssize_t total_len = 0;
1833         unsigned long seg;
1834         struct iovec *iov;
1835
1836         if (nr_segs == 1)
1837                 return __import_iovec_ubuf(type, uvec, iovp, i, compat);
1838
1839         iov = iovec_from_user(uvec, nr_segs, fast_segs, *iovp, compat);
1840         if (IS_ERR(iov)) {
1841                 *iovp = NULL;
1842                 return PTR_ERR(iov);
1843         }
1844
1845         /*
1846          * According to the Single Unix Specification we should return EINVAL if
1847          * an element length is < 0 when cast to ssize_t or if the total length
1848          * would overflow the ssize_t return value of the system call.
1849          *
1850          * Linux caps all read/write calls to MAX_RW_COUNT, and avoids the
1851          * overflow case.
1852          */
1853         for (seg = 0; seg < nr_segs; seg++) {
1854                 ssize_t len = (ssize_t)iov[seg].iov_len;
1855
1856                 if (!access_ok(iov[seg].iov_base, len)) {
1857                         if (iov != *iovp)
1858                                 kfree(iov);
1859                         *iovp = NULL;
1860                         return -EFAULT;
1861                 }
1862
1863                 if (len > MAX_RW_COUNT - total_len) {
1864                         len = MAX_RW_COUNT - total_len;
1865                         iov[seg].iov_len = len;
1866                 }
1867                 total_len += len;
1868         }
1869
1870         iov_iter_init(i, type, iov, nr_segs, total_len);
1871         if (iov == *iovp)
1872                 *iovp = NULL;
1873         else
1874                 *iovp = iov;
1875         return total_len;
1876 }
1877
1878 /**
1879  * import_iovec() - Copy an array of &struct iovec from userspace
1880  *     into the kernel, check that it is valid, and initialize a new
1881  *     &struct iov_iter iterator to access it.
1882  *
1883  * @type: One of %READ or %WRITE.
1884  * @uvec: Pointer to the userspace array.
1885  * @nr_segs: Number of elements in userspace array.
1886  * @fast_segs: Number of elements in @iov.
1887  * @iovp: (input and output parameter) Pointer to pointer to (usually small
1888  *     on-stack) kernel array.
1889  * @i: Pointer to iterator that will be initialized on success.
1890  *
1891  * If the array pointed to by *@iov is large enough to hold all @nr_segs,
1892  * then this function places %NULL in *@iov on return. Otherwise, a new
1893  * array will be allocated and the result placed in *@iov. This means that
1894  * the caller may call kfree() on *@iov regardless of whether the small
1895  * on-stack array was used or not (and regardless of whether this function
1896  * returns an error or not).
1897  *
1898  * Return: Negative error code on error, bytes imported on success
1899  */
1900 ssize_t import_iovec(int type, const struct iovec __user *uvec,
1901                  unsigned nr_segs, unsigned fast_segs,
1902                  struct iovec **iovp, struct iov_iter *i)
1903 {
1904         return __import_iovec(type, uvec, nr_segs, fast_segs, iovp, i,
1905                               in_compat_syscall());
1906 }
1907 EXPORT_SYMBOL(import_iovec);
1908
1909 int import_single_range(int rw, void __user *buf, size_t len,
1910                  struct iovec *iov, struct iov_iter *i)
1911 {
1912         if (len > MAX_RW_COUNT)
1913                 len = MAX_RW_COUNT;
1914         if (unlikely(!access_ok(buf, len)))
1915                 return -EFAULT;
1916
1917         iov_iter_ubuf(i, rw, buf, len);
1918         return 0;
1919 }
1920 EXPORT_SYMBOL(import_single_range);
1921
1922 int import_ubuf(int rw, void __user *buf, size_t len, struct iov_iter *i)
1923 {
1924         if (len > MAX_RW_COUNT)
1925                 len = MAX_RW_COUNT;
1926         if (unlikely(!access_ok(buf, len)))
1927                 return -EFAULT;
1928
1929         iov_iter_ubuf(i, rw, buf, len);
1930         return 0;
1931 }
1932
1933 /**
1934  * iov_iter_restore() - Restore a &struct iov_iter to the same state as when
1935  *     iov_iter_save_state() was called.
1936  *
1937  * @i: &struct iov_iter to restore
1938  * @state: state to restore from
1939  *
1940  * Used after iov_iter_save_state() to bring restore @i, if operations may
1941  * have advanced it.
1942  *
1943  * Note: only works on ITER_IOVEC, ITER_BVEC, and ITER_KVEC
1944  */
1945 void iov_iter_restore(struct iov_iter *i, struct iov_iter_state *state)
1946 {
1947         if (WARN_ON_ONCE(!iov_iter_is_bvec(i) && !iter_is_iovec(i) &&
1948                          !iter_is_ubuf(i)) && !iov_iter_is_kvec(i))
1949                 return;
1950         i->iov_offset = state->iov_offset;
1951         i->count = state->count;
1952         if (iter_is_ubuf(i))
1953                 return;
1954         /*
1955          * For the *vec iters, nr_segs + iov is constant - if we increment
1956          * the vec, then we also decrement the nr_segs count. Hence we don't
1957          * need to track both of these, just one is enough and we can deduct
1958          * the other from that. ITER_KVEC and ITER_IOVEC are the same struct
1959          * size, so we can just increment the iov pointer as they are unionzed.
1960          * ITER_BVEC _may_ be the same size on some archs, but on others it is
1961          * not. Be safe and handle it separately.
1962          */
1963         BUILD_BUG_ON(sizeof(struct iovec) != sizeof(struct kvec));
1964         if (iov_iter_is_bvec(i))
1965                 i->bvec -= state->nr_segs - i->nr_segs;
1966         else
1967                 i->__iov -= state->nr_segs - i->nr_segs;
1968         i->nr_segs = state->nr_segs;
1969 }
1970
1971 /*
1972  * Extract a list of contiguous pages from an ITER_XARRAY iterator.  This does not
1973  * get references on the pages, nor does it get a pin on them.
1974  */
1975 static ssize_t iov_iter_extract_xarray_pages(struct iov_iter *i,
1976                                              struct page ***pages, size_t maxsize,
1977                                              unsigned int maxpages,
1978                                              iov_iter_extraction_t extraction_flags,
1979                                              size_t *offset0)
1980 {
1981         struct page *page, **p;
1982         unsigned int nr = 0, offset;
1983         loff_t pos = i->xarray_start + i->iov_offset;
1984         pgoff_t index = pos >> PAGE_SHIFT;
1985         XA_STATE(xas, i->xarray, index);
1986
1987         offset = pos & ~PAGE_MASK;
1988         *offset0 = offset;
1989
1990         maxpages = want_pages_array(pages, maxsize, offset, maxpages);
1991         if (!maxpages)
1992                 return -ENOMEM;
1993         p = *pages;
1994
1995         rcu_read_lock();
1996         for (page = xas_load(&xas); page; page = xas_next(&xas)) {
1997                 if (xas_retry(&xas, page))
1998                         continue;
1999
2000                 /* Has the page moved or been split? */
2001                 if (unlikely(page != xas_reload(&xas))) {
2002                         xas_reset(&xas);
2003                         continue;
2004                 }
2005
2006                 p[nr++] = find_subpage(page, xas.xa_index);
2007                 if (nr == maxpages)
2008                         break;
2009         }
2010         rcu_read_unlock();
2011
2012         maxsize = min_t(size_t, nr * PAGE_SIZE - offset, maxsize);
2013         iov_iter_advance(i, maxsize);
2014         return maxsize;
2015 }
2016
2017 /*
2018  * Extract a list of contiguous pages from an ITER_BVEC iterator.  This does
2019  * not get references on the pages, nor does it get a pin on them.
2020  */
2021 static ssize_t iov_iter_extract_bvec_pages(struct iov_iter *i,
2022                                            struct page ***pages, size_t maxsize,
2023                                            unsigned int maxpages,
2024                                            iov_iter_extraction_t extraction_flags,
2025                                            size_t *offset0)
2026 {
2027         struct page **p, *page;
2028         size_t skip = i->iov_offset, offset;
2029         int k;
2030
2031         for (;;) {
2032                 if (i->nr_segs == 0)
2033                         return 0;
2034                 maxsize = min(maxsize, i->bvec->bv_len - skip);
2035                 if (maxsize)
2036                         break;
2037                 i->iov_offset = 0;
2038                 i->nr_segs--;
2039                 i->bvec++;
2040                 skip = 0;
2041         }
2042
2043         skip += i->bvec->bv_offset;
2044         page = i->bvec->bv_page + skip / PAGE_SIZE;
2045         offset = skip % PAGE_SIZE;
2046         *offset0 = offset;
2047
2048         maxpages = want_pages_array(pages, maxsize, offset, maxpages);
2049         if (!maxpages)
2050                 return -ENOMEM;
2051         p = *pages;
2052         for (k = 0; k < maxpages; k++)
2053                 p[k] = page + k;
2054
2055         maxsize = min_t(size_t, maxsize, maxpages * PAGE_SIZE - offset);
2056         iov_iter_advance(i, maxsize);
2057         return maxsize;
2058 }
2059
2060 /*
2061  * Extract a list of virtually contiguous pages from an ITER_KVEC iterator.
2062  * This does not get references on the pages, nor does it get a pin on them.
2063  */
2064 static ssize_t iov_iter_extract_kvec_pages(struct iov_iter *i,
2065                                            struct page ***pages, size_t maxsize,
2066                                            unsigned int maxpages,
2067                                            iov_iter_extraction_t extraction_flags,
2068                                            size_t *offset0)
2069 {
2070         struct page **p, *page;
2071         const void *kaddr;
2072         size_t skip = i->iov_offset, offset, len;
2073         int k;
2074
2075         for (;;) {
2076                 if (i->nr_segs == 0)
2077                         return 0;
2078                 maxsize = min(maxsize, i->kvec->iov_len - skip);
2079                 if (maxsize)
2080                         break;
2081                 i->iov_offset = 0;
2082                 i->nr_segs--;
2083                 i->kvec++;
2084                 skip = 0;
2085         }
2086
2087         kaddr = i->kvec->iov_base + skip;
2088         offset = (unsigned long)kaddr & ~PAGE_MASK;
2089         *offset0 = offset;
2090
2091         maxpages = want_pages_array(pages, maxsize, offset, maxpages);
2092         if (!maxpages)
2093                 return -ENOMEM;
2094         p = *pages;
2095
2096         kaddr -= offset;
2097         len = offset + maxsize;
2098         for (k = 0; k < maxpages; k++) {
2099                 size_t seg = min_t(size_t, len, PAGE_SIZE);
2100
2101                 if (is_vmalloc_or_module_addr(kaddr))
2102                         page = vmalloc_to_page(kaddr);
2103                 else
2104                         page = virt_to_page(kaddr);
2105
2106                 p[k] = page;
2107                 len -= seg;
2108                 kaddr += PAGE_SIZE;
2109         }
2110
2111         maxsize = min_t(size_t, maxsize, maxpages * PAGE_SIZE - offset);
2112         iov_iter_advance(i, maxsize);
2113         return maxsize;
2114 }
2115
2116 /*
2117  * Extract a list of contiguous pages from a user iterator and get a pin on
2118  * each of them.  This should only be used if the iterator is user-backed
2119  * (IOBUF/UBUF).
2120  *
2121  * It does not get refs on the pages, but the pages must be unpinned by the
2122  * caller once the transfer is complete.
2123  *
2124  * This is safe to be used where background IO/DMA *is* going to be modifying
2125  * the buffer; using a pin rather than a ref makes forces fork() to give the
2126  * child a copy of the page.
2127  */
2128 static ssize_t iov_iter_extract_user_pages(struct iov_iter *i,
2129                                            struct page ***pages,
2130                                            size_t maxsize,
2131                                            unsigned int maxpages,
2132                                            iov_iter_extraction_t extraction_flags,
2133                                            size_t *offset0)
2134 {
2135         unsigned long addr;
2136         unsigned int gup_flags = 0;
2137         size_t offset;
2138         int res;
2139
2140         if (i->data_source == ITER_DEST)
2141                 gup_flags |= FOLL_WRITE;
2142         if (extraction_flags & ITER_ALLOW_P2PDMA)
2143                 gup_flags |= FOLL_PCI_P2PDMA;
2144         if (i->nofault)
2145                 gup_flags |= FOLL_NOFAULT;
2146
2147         addr = first_iovec_segment(i, &maxsize);
2148         *offset0 = offset = addr % PAGE_SIZE;
2149         addr &= PAGE_MASK;
2150         maxpages = want_pages_array(pages, maxsize, offset, maxpages);
2151         if (!maxpages)
2152                 return -ENOMEM;
2153         res = pin_user_pages_fast(addr, maxpages, gup_flags, *pages);
2154         if (unlikely(res <= 0))
2155                 return res;
2156         maxsize = min_t(size_t, maxsize, res * PAGE_SIZE - offset);
2157         iov_iter_advance(i, maxsize);
2158         return maxsize;
2159 }
2160
2161 /**
2162  * iov_iter_extract_pages - Extract a list of contiguous pages from an iterator
2163  * @i: The iterator to extract from
2164  * @pages: Where to return the list of pages
2165  * @maxsize: The maximum amount of iterator to extract
2166  * @maxpages: The maximum size of the list of pages
2167  * @extraction_flags: Flags to qualify request
2168  * @offset0: Where to return the starting offset into (*@pages)[0]
2169  *
2170  * Extract a list of contiguous pages from the current point of the iterator,
2171  * advancing the iterator.  The maximum number of pages and the maximum amount
2172  * of page contents can be set.
2173  *
2174  * If *@pages is NULL, a page list will be allocated to the required size and
2175  * *@pages will be set to its base.  If *@pages is not NULL, it will be assumed
2176  * that the caller allocated a page list at least @maxpages in size and this
2177  * will be filled in.
2178  *
2179  * @extraction_flags can have ITER_ALLOW_P2PDMA set to request peer-to-peer DMA
2180  * be allowed on the pages extracted.
2181  *
2182  * The iov_iter_extract_will_pin() function can be used to query how cleanup
2183  * should be performed.
2184  *
2185  * Extra refs or pins on the pages may be obtained as follows:
2186  *
2187  *  (*) If the iterator is user-backed (ITER_IOVEC/ITER_UBUF), pins will be
2188  *      added to the pages, but refs will not be taken.
2189  *      iov_iter_extract_will_pin() will return true.
2190  *
2191  *  (*) If the iterator is ITER_KVEC, ITER_BVEC or ITER_XARRAY, the pages are
2192  *      merely listed; no extra refs or pins are obtained.
2193  *      iov_iter_extract_will_pin() will return 0.
2194  *
2195  * Note also:
2196  *
2197  *  (*) Use with ITER_DISCARD is not supported as that has no content.
2198  *
2199  * On success, the function sets *@pages to the new pagelist, if allocated, and
2200  * sets *offset0 to the offset into the first page.
2201  *
2202  * It may also return -ENOMEM and -EFAULT.
2203  */
2204 ssize_t iov_iter_extract_pages(struct iov_iter *i,
2205                                struct page ***pages,
2206                                size_t maxsize,
2207                                unsigned int maxpages,
2208                                iov_iter_extraction_t extraction_flags,
2209                                size_t *offset0)
2210 {
2211         maxsize = min_t(size_t, min_t(size_t, maxsize, i->count), MAX_RW_COUNT);
2212         if (!maxsize)
2213                 return 0;
2214
2215         if (likely(user_backed_iter(i)))
2216                 return iov_iter_extract_user_pages(i, pages, maxsize,
2217                                                    maxpages, extraction_flags,
2218                                                    offset0);
2219         if (iov_iter_is_kvec(i))
2220                 return iov_iter_extract_kvec_pages(i, pages, maxsize,
2221                                                    maxpages, extraction_flags,
2222                                                    offset0);
2223         if (iov_iter_is_bvec(i))
2224                 return iov_iter_extract_bvec_pages(i, pages, maxsize,
2225                                                    maxpages, extraction_flags,
2226                                                    offset0);
2227         if (iov_iter_is_xarray(i))
2228                 return iov_iter_extract_xarray_pages(i, pages, maxsize,
2229                                                      maxpages, extraction_flags,
2230                                                      offset0);
2231         return -EFAULT;
2232 }
2233 EXPORT_SYMBOL_GPL(iov_iter_extract_pages);