fs/orangefs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * (C) 2001 Clemson University and The University of Chicago
   4  * Copyright 2018 Omnibond Systems, L.L.C.
   5  *
   6  * See COPYING in top-level directory.
   7  */
   8
   9 /*
  10  *  Linux VFS inode operations.
  11  */
  12
  13 #include <linux/bvec.h>
  14 #include "protocol.h"
  15 #include "orangefs-kernel.h"
  16 #include "orangefs-bufmap.h"
  17
  18 static int orangefs_writepage_locked(struct page *page,
  19     struct writeback_control *wbc)
  20 {
  21         struct inode *inode = page->mapping->host;
  22         struct orangefs_write_range *wr = NULL;
  23         struct iov_iter iter;
  24         struct bio_vec bv;
  25         size_t len, wlen;
  26         ssize_t ret;
  27         loff_t off;
  28
  29         set_page_writeback(page);
  30
  31         len = i_size_read(inode);
  32         if (PagePrivate(page)) {
  33                 wr = (struct orangefs_write_range *)page_private(page);
  34                 WARN_ON(wr->pos >= len);
  35                 off = wr->pos;
  36                 if (off + wr->len > len)
  37                         wlen = len - off;
  38                 else
  39                         wlen = wr->len;
  40         } else {
  41                 WARN_ON(1);
  42                 off = page_offset(page);
  43                 if (off + PAGE_SIZE > len)
  44                         wlen = len - off;
  45                 else
  46                         wlen = PAGE_SIZE;
  47         }
  48         /* Should've been handled in orangefs_invalidatepage. */
  49         WARN_ON(off == len || off + wlen > len);
  50
  51         bv.bv_page = page;
  52         bv.bv_len = wlen;
  53         bv.bv_offset = off % PAGE_SIZE;
  54         WARN_ON(wlen == 0);
  55         iov_iter_bvec(&iter, WRITE, &bv, 1, wlen);
  56
  57         ret = wait_for_direct_io(ORANGEFS_IO_WRITE, inode, &off, &iter, wlen,
  58             len, wr, NULL, NULL);
  59         if (ret < 0) {
  60                 SetPageError(page);
  61                 mapping_set_error(page->mapping, ret);
  62         } else {
  63                 ret = 0;
  64         }
  65         if (wr) {
  66                 kfree(wr);
  67                 set_page_private(page, 0);
  68                 ClearPagePrivate(page);
  69                 put_page(page);
  70         }
  71         return ret;
  72 }
  73
  74 static int orangefs_writepage(struct page *page, struct writeback_control *wbc)
  75 {
  76         int ret;
  77         ret = orangefs_writepage_locked(page, wbc);
  78         unlock_page(page);
  79         end_page_writeback(page);
  80         return ret;
  81 }
  82
  83 struct orangefs_writepages {
  84         loff_t off;
  85         size_t len;
  86         kuid_t uid;
  87         kgid_t gid;
  88         int maxpages;
  89         int npages;
  90         struct page **pages;
  91         struct bio_vec *bv;
  92 };
  93
  94 static int orangefs_writepages_work(struct orangefs_writepages *ow,
  95     struct writeback_control *wbc)
  96 {
  97         struct inode *inode = ow->pages[0]->mapping->host;
  98         struct orangefs_write_range *wrp, wr;
  99         struct iov_iter iter;
 100         ssize_t ret;
 101         size_t len;
 102         loff_t off;
 103         int i;
 104
 105         len = i_size_read(inode);
 106
 107         for (i = 0; i < ow->npages; i++) {
 108                 set_page_writeback(ow->pages[i]);
 109                 ow->bv[i].bv_page = ow->pages[i];
 110                 ow->bv[i].bv_len = min(page_offset(ow->pages[i]) + PAGE_SIZE,
 111                     ow->off + ow->len) -
 112                     max(ow->off, page_offset(ow->pages[i]));
 113                 if (i == 0)
 114                         ow->bv[i].bv_offset = ow->off -
 115                             page_offset(ow->pages[i]);
 116                 else
 117                         ow->bv[i].bv_offset = 0;
 118         }
 119         iov_iter_bvec(&iter, WRITE, ow->bv, ow->npages, ow->len);
 120
 121         WARN_ON(ow->off >= len);
 122         if (ow->off + ow->len > len)
 123                 ow->len = len - ow->off;
 124
 125         off = ow->off;
 126         wr.uid = ow->uid;
 127         wr.gid = ow->gid;
 128         ret = wait_for_direct_io(ORANGEFS_IO_WRITE, inode, &off, &iter, ow->len,
 129             0, &wr, NULL, NULL);
 130         if (ret < 0) {
 131                 for (i = 0; i < ow->npages; i++) {
 132                         SetPageError(ow->pages[i]);
 133                         mapping_set_error(ow->pages[i]->mapping, ret);
 134                         if (PagePrivate(ow->pages[i])) {
 135                                 wrp = (struct orangefs_write_range *)
 136                                     page_private(ow->pages[i]);
 137                                 ClearPagePrivate(ow->pages[i]);
 138                                 put_page(ow->pages[i]);
 139                                 kfree(wrp);
 140                         }
 141                         end_page_writeback(ow->pages[i]);
 142                         unlock_page(ow->pages[i]);
 143                 }
 144         } else {
 145                 ret = 0;
 146                 for (i = 0; i < ow->npages; i++) {
 147                         if (PagePrivate(ow->pages[i])) {
 148                                 wrp = (struct orangefs_write_range *)
 149                                     page_private(ow->pages[i]);
 150                                 ClearPagePrivate(ow->pages[i]);
 151                                 put_page(ow->pages[i]);
 152                                 kfree(wrp);
 153                         }
 154                         end_page_writeback(ow->pages[i]);
 155                         unlock_page(ow->pages[i]);
 156                 }
 157         }
 158         return ret;
 159 }
 160
 161 static int orangefs_writepages_callback(struct page *page,
 162     struct writeback_control *wbc, void *data)
 163 {
 164         struct orangefs_writepages *ow = data;
 165         struct orangefs_write_range *wr;
 166         int ret;
 167
 168         if (!PagePrivate(page)) {
 169                 unlock_page(page);
 170                 /* It's not private so there's nothing to write, right? */
 171                 printk("writepages_callback not private!\n");
 172                 BUG();
 173                 return 0;
 174         }
 175         wr = (struct orangefs_write_range *)page_private(page);
 176
 177         ret = -1;
 178         if (ow->npages == 0) {
 179                 ow->off = wr->pos;
 180                 ow->len = wr->len;
 181                 ow->uid = wr->uid;
 182                 ow->gid = wr->gid;
 183                 ow->pages[ow->npages++] = page;
 184                 ret = 0;
 185                 goto done;
 186         }
 187         if (!uid_eq(ow->uid, wr->uid) || !gid_eq(ow->gid, wr->gid)) {
 188                 orangefs_writepages_work(ow, wbc);
 189                 ow->npages = 0;
 190                 ret = -1;
 191                 goto done;
 192         }
 193         if (ow->off + ow->len == wr->pos) {
 194                 ow->len += wr->len;
 195                 ow->pages[ow->npages++] = page;
 196                 ret = 0;
 197                 goto done;
 198         }
 199 done:
 200         if (ret == -1) {
 201                 if (ow->npages) {
 202                         orangefs_writepages_work(ow, wbc);
 203                         ow->npages = 0;
 204                 }
 205                 ret = orangefs_writepage_locked(page, wbc);
 206                 mapping_set_error(page->mapping, ret);
 207                 unlock_page(page);
 208                 end_page_writeback(page);
 209         } else {
 210                 if (ow->npages == ow->maxpages) {
 211                         orangefs_writepages_work(ow, wbc);
 212                         ow->npages = 0;
 213                 }
 214         }
 215         return ret;
 216 }
 217
 218 static int orangefs_writepages(struct address_space *mapping,
 219     struct writeback_control *wbc)
 220 {
 221         struct orangefs_writepages *ow;
 222         struct blk_plug plug;
 223         int ret;
 224         ow = kzalloc(sizeof(struct orangefs_writepages), GFP_KERNEL);
 225         if (!ow)
 226                 return -ENOMEM;
 227         ow->maxpages = orangefs_bufmap_size_query()/PAGE_SIZE;
 228         ow->pages = kcalloc(ow->maxpages, sizeof(struct page *), GFP_KERNEL);
 229         if (!ow->pages) {
 230                 kfree(ow);
 231                 return -ENOMEM;
 232         }
 233         ow->bv = kcalloc(ow->maxpages, sizeof(struct bio_vec), GFP_KERNEL);
 234         if (!ow->bv) {
 235                 kfree(ow->pages);
 236                 kfree(ow);
 237                 return -ENOMEM;
 238         }
 239         blk_start_plug(&plug);
 240         ret = write_cache_pages(mapping, wbc, orangefs_writepages_callback, ow);
 241         if (ow->npages)
 242                 ret = orangefs_writepages_work(ow, wbc);
 243         blk_finish_plug(&plug);
 244         kfree(ow->pages);
 245         kfree(ow->bv);
 246         kfree(ow);
 247         return ret;
 248 }
 249
 250 static int orangefs_launder_page(struct page *);
 251
 252 static int orangefs_readpage(struct file *file, struct page *page)
 253 {
 254         struct inode *inode = page->mapping->host;
 255         struct iov_iter iter;
 256         struct bio_vec bv;
 257         ssize_t ret;
 258         loff_t off; /* offset into this page */
 259         pgoff_t index; /* which page */
 260         struct page *next_page;
 261         char *kaddr;
 262         loff_t read_size;
 263         int buffer_index = -1; /* orangefs shared memory slot */
 264         int slot_index;   /* index into slot */
 265         int remaining;
 266
 267         /*
 268          * Get up to this many bytes from Orangefs at a time and try
 269          * to fill them into the page cache at once. Tests with dd made
 270          * this seem like a reasonable static number, if there was
 271          * interest perhaps this number could be made setable through
 272          * sysfs...
 273          */
 274         read_size = 524288;
 275
 276         if (PageDirty(page))
 277                 orangefs_launder_page(page);
 278
 279         off = page_offset(page);
 280         index = off >> PAGE_SHIFT;
 281         bv.bv_page = page;
 282         bv.bv_len = PAGE_SIZE;
 283         bv.bv_offset = 0;
 284         iov_iter_bvec(&iter, READ, &bv, 1, PAGE_SIZE);
 285
 286         ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, &off, &iter,
 287             read_size, inode->i_size, NULL, &buffer_index, file);
 288         remaining = ret;
 289         /* this will only zero remaining unread portions of the page data */
 290         iov_iter_zero(~0U, &iter);
 291         /* takes care of potential aliasing */
 292         flush_dcache_page(page);
 293         if (ret < 0) {
 294                 SetPageError(page);
 295                 unlock_page(page);
 296                 goto out;
 297         } else {
 298                 SetPageUptodate(page);
 299                 if (PageError(page))
 300                         ClearPageError(page);
 301                 ret = 0;
 302         }
 303         /* unlock the page after the ->readpage() routine completes */
 304         unlock_page(page);
 305
 306         if (remaining > PAGE_SIZE) {
 307                 slot_index = 0;
 308                 while ((remaining - PAGE_SIZE) >= PAGE_SIZE) {
 309                         remaining -= PAGE_SIZE;
 310                         /*
 311                          * It is an optimization to try and fill more than one
 312                          * page... by now we've already gotten the single
 313                          * page we were after, if stuff doesn't seem to
 314                          * be going our way at this point just return
 315                          * and hope for the best.
 316                          *
 317                          * If we look for pages and they're already there is
 318                          * one reason to give up, and if they're not there
 319                          * and we can't create them is another reason.
 320                          */
 321
 322                         index++;
 323                         slot_index++;
 324                         next_page = find_get_page(inode->i_mapping, index);
 325                         if (next_page) {
 326                                 gossip_debug(GOSSIP_FILE_DEBUG,
 327                                         "%s: found next page, quitting\n",
 328                                         __func__);
 329                                 put_page(next_page);
 330                                 goto out;
 331                         }
 332                         next_page = find_or_create_page(inode->i_mapping,
 333                                                         index,
 334                                                         GFP_KERNEL);
 335                         /*
 336                          * I've never hit this, leave it as a printk for
 337                          * now so it will be obvious.
 338                          */
 339                         if (!next_page) {
 340                                 printk("%s: can't create next page, quitting\n",
 341                                         __func__);
 342                                 goto out;
 343                         }
 344                         kaddr = kmap_atomic(next_page);
 345                         orangefs_bufmap_page_fill(kaddr,
 346                                                 buffer_index,
 347                                                 slot_index);
 348                         kunmap_atomic(kaddr);
 349                         SetPageUptodate(next_page);
 350                         unlock_page(next_page);
 351                         put_page(next_page);
 352                 }
 353         }
 354
 355 out:
 356         if (buffer_index != -1)
 357                 orangefs_bufmap_put(buffer_index);
 358         return ret;
 359 }
 360
 361 static int orangefs_write_begin(struct file *file,
 362     struct address_space *mapping,
 363     loff_t pos, unsigned len, unsigned flags, struct page **pagep,
 364     void **fsdata)
 365 {
 366         struct orangefs_write_range *wr;
 367         struct page *page;
 368         pgoff_t index;
 369         int ret;
 370
 371         index = pos >> PAGE_SHIFT;
 372
 373         page = grab_cache_page_write_begin(mapping, index, flags);
 374         if (!page)
 375                 return -ENOMEM;
 376
 377         *pagep = page;
 378
 379         if (PageDirty(page) && !PagePrivate(page)) {
 380                 /*
 381                  * Should be impossible.  If it happens, launder the page
 382                  * since we don't know what's dirty.  This will WARN in
 383                  * orangefs_writepage_locked.
 384                  */
 385                 ret = orangefs_launder_page(page);
 386                 if (ret)
 387                         return ret;
 388         }
 389         if (PagePrivate(page)) {
 390                 struct orangefs_write_range *wr;
 391                 wr = (struct orangefs_write_range *)page_private(page);
 392                 if (wr->pos + wr->len == pos &&
 393                     uid_eq(wr->uid, current_fsuid()) &&
 394                     gid_eq(wr->gid, current_fsgid())) {
 395                         wr->len += len;
 396                         goto okay;
 397                 } else {
 398                         ret = orangefs_launder_page(page);
 399                         if (ret)
 400                                 return ret;
 401                 }
 402         }
 403
 404         wr = kmalloc(sizeof *wr, GFP_KERNEL);
 405         if (!wr)
 406                 return -ENOMEM;
 407
 408         wr->pos = pos;
 409         wr->len = len;
 410         wr->uid = current_fsuid();
 411         wr->gid = current_fsgid();
 412         SetPagePrivate(page);
 413         set_page_private(page, (unsigned long)wr);
 414         get_page(page);
 415 okay:
 416         return 0;
 417 }
 418
 419 static int orangefs_write_end(struct file *file, struct address_space *mapping,
 420     loff_t pos, unsigned len, unsigned copied, struct page *page, void *fsdata)
 421 {
 422         struct inode *inode = page->mapping->host;
 423         loff_t last_pos = pos + copied;
 424
 425         /*
 426          * No need to use i_size_read() here, the i_size
 427          * cannot change under us because we hold the i_mutex.
 428          */
 429         if (last_pos > inode->i_size)
 430                 i_size_write(inode, last_pos);
 431
 432         /* zero the stale part of the page if we did a short copy */
 433         if (!PageUptodate(page)) {
 434                 unsigned from = pos & (PAGE_SIZE - 1);
 435                 if (copied < len) {
 436                         zero_user(page, from + copied, len - copied);
 437                 }
 438                 /* Set fully written pages uptodate. */
 439                 if (pos == page_offset(page) &&
 440                     (len == PAGE_SIZE || pos + len == inode->i_size)) {
 441                         zero_user_segment(page, from + copied, PAGE_SIZE);
 442                         SetPageUptodate(page);
 443                 }
 444         }
 445
 446         set_page_dirty(page);
 447         unlock_page(page);
 448         put_page(page);
 449
 450         mark_inode_dirty_sync(file_inode(file));
 451         return copied;
 452 }
 453
 454 static void orangefs_invalidatepage(struct page *page,
 455                                  unsigned int offset,
 456                                  unsigned int length)
 457 {
 458         struct orangefs_write_range *wr;
 459         wr = (struct orangefs_write_range *)page_private(page);
 460
 461         if (offset == 0 && length == PAGE_SIZE) {
 462                 kfree((struct orangefs_write_range *)page_private(page));
 463                 set_page_private(page, 0);
 464                 ClearPagePrivate(page);
 465                 put_page(page);
 466                 return;
 467         /* write range entirely within invalidate range (or equal) */
 468         } else if (page_offset(page) + offset <= wr->pos &&
 469             wr->pos + wr->len <= page_offset(page) + offset + length) {
 470                 kfree((struct orangefs_write_range *)page_private(page));
 471                 set_page_private(page, 0);
 472                 ClearPagePrivate(page);
 473                 put_page(page);
 474                 /* XXX is this right? only caller in fs */
 475                 cancel_dirty_page(page);
 476                 return;
 477         /* invalidate range chops off end of write range */
 478         } else if (wr->pos < page_offset(page) + offset &&
 479             wr->pos + wr->len <= page_offset(page) + offset + length &&
 480              page_offset(page) + offset < wr->pos + wr->len) {
 481                 size_t x;
 482                 x = wr->pos + wr->len - (page_offset(page) + offset);
 483                 WARN_ON(x > wr->len);
 484                 wr->len -= x;
 485                 wr->uid = current_fsuid();
 486                 wr->gid = current_fsgid();
 487         /* invalidate range chops off beginning of write range */
 488         } else if (page_offset(page) + offset <= wr->pos &&
 489             page_offset(page) + offset + length < wr->pos + wr->len &&
 490             wr->pos < page_offset(page) + offset + length) {
 491                 size_t x;
 492                 x = page_offset(page) + offset + length - wr->pos;
 493                 WARN_ON(x > wr->len);
 494                 wr->pos += x;
 495                 wr->len -= x;
 496                 wr->uid = current_fsuid();
 497                 wr->gid = current_fsgid();
 498         /* invalidate range entirely within write range (punch hole) */
 499         } else if (wr->pos < page_offset(page) + offset &&
 500             page_offset(page) + offset + length < wr->pos + wr->len) {
 501                 /* XXX what do we do here... should not WARN_ON */
 502                 WARN_ON(1);
 503                 /* punch hole */
 504                 /*
 505                  * should we just ignore this and write it out anyway?
 506                  * it hardly makes sense
 507                  */
 508                 return;
 509         /* non-overlapping ranges */
 510         } else {
 511                 /* WARN if they do overlap */
 512                 if (!((page_offset(page) + offset + length <= wr->pos) ^
 513                     (wr->pos + wr->len <= page_offset(page) + offset))) {
 514                         WARN_ON(1);
 515                         printk("invalidate range offset %llu length %u\n",
 516                             page_offset(page) + offset, length);
 517                         printk("write range offset %llu length %zu\n",
 518                             wr->pos, wr->len);
 519                 }
 520                 return;
 521         }
 522
 523         /*
 524          * Above there are returns where wr is freed or where we WARN.
 525          * Thus the following runs if wr was modified above.
 526          */
 527
 528         orangefs_launder_page(page);
 529 }
 530
 531 static int orangefs_releasepage(struct page *page, gfp_t foo)
 532 {
 533         return !PagePrivate(page);
 534 }
 535
 536 static void orangefs_freepage(struct page *page)
 537 {
 538         if (PagePrivate(page)) {
 539                 kfree((struct orangefs_write_range *)page_private(page));
 540                 set_page_private(page, 0);
 541                 ClearPagePrivate(page);
 542                 put_page(page);
 543         }
 544 }
 545
 546 static int orangefs_launder_page(struct page *page)
 547 {
 548         int r = 0;
 549         struct writeback_control wbc = {
 550                 .sync_mode = WB_SYNC_ALL,
 551                 .nr_to_write = 0,
 552         };
 553         wait_on_page_writeback(page);
 554         if (clear_page_dirty_for_io(page)) {
 555                 r = orangefs_writepage_locked(page, &wbc);
 556                 end_page_writeback(page);
 557         }
 558         return r;
 559 }
 560
 561 static ssize_t orangefs_direct_IO(struct kiocb *iocb,
 562                                   struct iov_iter *iter)
 563 {
 564         /*
 565          * Comment from original do_readv_writev:
 566          * Common entry point for read/write/readv/writev
 567          * This function will dispatch it to either the direct I/O
 568          * or buffered I/O path depending on the mount options and/or
 569          * augmented/extended metadata attached to the file.
 570          * Note: File extended attributes override any mount options.
 571          */
 572         struct file *file = iocb->ki_filp;
 573         loff_t pos = iocb->ki_pos;
 574         enum ORANGEFS_io_type type = iov_iter_rw(iter) == WRITE ?
 575             ORANGEFS_IO_WRITE : ORANGEFS_IO_READ;
 576         loff_t *offset = &pos;
 577         struct inode *inode = file->f_mapping->host;
 578         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 579         struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
 580         size_t count = iov_iter_count(iter);
 581         ssize_t total_count = 0;
 582         ssize_t ret = -EINVAL;
 583         int i = 0;
 584
 585         gossip_debug(GOSSIP_FILE_DEBUG,
 586                 "%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
 587                 __func__,
 588                 handle,
 589                 (int)count);
 590
 591         if (type == ORANGEFS_IO_WRITE) {
 592                 gossip_debug(GOSSIP_FILE_DEBUG,
 593                              "%s(%pU): proceeding with offset : %llu, "
 594                              "size %d\n",
 595                              __func__,
 596                              handle,
 597                              llu(*offset),
 598                              (int)count);
 599         }
 600
 601         if (count == 0) {
 602                 ret = 0;
 603                 goto out;
 604         }
 605
 606         while (iov_iter_count(iter)) {
 607                 size_t each_count = iov_iter_count(iter);
 608                 size_t amt_complete;
 609                 i++;
 610
 611                 /* how much to transfer in this loop iteration */
 612                 if (each_count > orangefs_bufmap_size_query())
 613                         each_count = orangefs_bufmap_size_query();
 614
 615                 gossip_debug(GOSSIP_FILE_DEBUG,
 616                              "%s(%pU): size of each_count(%d)\n",
 617                              __func__,
 618                              handle,
 619                              (int)each_count);
 620                 gossip_debug(GOSSIP_FILE_DEBUG,
 621                              "%s(%pU): BEFORE wait_for_io: offset is %d\n",
 622                              __func__,
 623                              handle,
 624                              (int)*offset);
 625
 626                 ret = wait_for_direct_io(type, inode, offset, iter,
 627                                 each_count, 0, NULL, NULL, file);
 628                 gossip_debug(GOSSIP_FILE_DEBUG,
 629                              "%s(%pU): return from wait_for_io:%d\n",
 630                              __func__,
 631                              handle,
 632                              (int)ret);
 633
 634                 if (ret < 0)
 635                         goto out;
 636
 637                 *offset += ret;
 638                 total_count += ret;
 639                 amt_complete = ret;
 640
 641                 gossip_debug(GOSSIP_FILE_DEBUG,
 642                              "%s(%pU): AFTER wait_for_io: offset is %d\n",
 643                              __func__,
 644                              handle,
 645                              (int)*offset);
 646
 647                 /*
 648                  * if we got a short I/O operations,
 649                  * fall out and return what we got so far
 650                  */
 651                 if (amt_complete < each_count)
 652                         break;
 653         } /*end while */
 654
 655 out:
 656         if (total_count > 0)
 657                 ret = total_count;
 658         if (ret > 0) {
 659                 if (type == ORANGEFS_IO_READ) {
 660                         file_accessed(file);
 661                 } else {
 662                         file_update_time(file);
 663                         if (*offset > i_size_read(inode))
 664                                 i_size_write(inode, *offset);
 665                 }
 666         }
 667
 668         gossip_debug(GOSSIP_FILE_DEBUG,
 669                      "%s(%pU): Value(%d) returned.\n",
 670                      __func__,
 671                      handle,
 672                      (int)ret);
 673
 674         return ret;
 675 }
 676
 677 /** ORANGEFS2 implementation of address space operations */
 678 static const struct address_space_operations orangefs_address_operations = {
 679         .writepage = orangefs_writepage,
 680         .readpage = orangefs_readpage,
 681         .writepages = orangefs_writepages,
 682         .set_page_dirty = __set_page_dirty_nobuffers,
 683         .write_begin = orangefs_write_begin,
 684         .write_end = orangefs_write_end,
 685         .invalidatepage = orangefs_invalidatepage,
 686         .releasepage = orangefs_releasepage,
 687         .freepage = orangefs_freepage,
 688         .launder_page = orangefs_launder_page,
 689         .direct_IO = orangefs_direct_IO,
 690 };
 691
 692 vm_fault_t orangefs_page_mkwrite(struct vm_fault *vmf)
 693 {
 694         struct page *page = vmf->page;
 695         struct inode *inode = file_inode(vmf->vma->vm_file);
 696         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 697         unsigned long *bitlock = &orangefs_inode->bitlock;
 698         vm_fault_t ret;
 699         struct orangefs_write_range *wr;
 700
 701         sb_start_pagefault(inode->i_sb);
 702
 703         if (wait_on_bit(bitlock, 1, TASK_KILLABLE)) {
 704                 ret = VM_FAULT_RETRY;
 705                 goto out;
 706         }
 707
 708         lock_page(page);
 709         if (PageDirty(page) && !PagePrivate(page)) {
 710                 /*
 711                  * Should be impossible.  If it happens, launder the page
 712                  * since we don't know what's dirty.  This will WARN in
 713                  * orangefs_writepage_locked.
 714                  */
 715                 if (orangefs_launder_page(page)) {
 716                         ret = VM_FAULT_LOCKED|VM_FAULT_RETRY;
 717                         goto out;
 718                 }
 719         }
 720         if (PagePrivate(page)) {
 721                 wr = (struct orangefs_write_range *)page_private(page);
 722                 if (uid_eq(wr->uid, current_fsuid()) &&
 723                     gid_eq(wr->gid, current_fsgid())) {
 724                         wr->pos = page_offset(page);
 725                         wr->len = PAGE_SIZE;
 726                         goto okay;
 727                 } else {
 728                         if (orangefs_launder_page(page)) {
 729                                 ret = VM_FAULT_LOCKED|VM_FAULT_RETRY;
 730                                 goto out;
 731                         }
 732                 }
 733         }
 734         wr = kmalloc(sizeof *wr, GFP_KERNEL);
 735         if (!wr) {
 736                 ret = VM_FAULT_LOCKED|VM_FAULT_RETRY;
 737                 goto out;
 738         }
 739         wr->pos = page_offset(page);
 740         wr->len = PAGE_SIZE;
 741         wr->uid = current_fsuid();
 742         wr->gid = current_fsgid();
 743         SetPagePrivate(page);
 744         set_page_private(page, (unsigned long)wr);
 745         get_page(page);
 746 okay:
 747
 748         file_update_time(vmf->vma->vm_file);
 749         if (page->mapping != inode->i_mapping) {
 750                 unlock_page(page);
 751                 ret = VM_FAULT_LOCKED|VM_FAULT_NOPAGE;
 752                 goto out;
 753         }
 754
 755         /*
 756          * We mark the page dirty already here so that when freeze is in
 757          * progress, we are guaranteed that writeback during freezing will
 758          * see the dirty page and writeprotect it again.
 759          */
 760         set_page_dirty(page);
 761         wait_for_stable_page(page);
 762         ret = VM_FAULT_LOCKED;
 763 out:
 764         sb_end_pagefault(inode->i_sb);
 765         return ret;
 766 }
 767
 768 static int orangefs_setattr_size(struct inode *inode, struct iattr *iattr)
 769 {
 770         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 771         struct orangefs_kernel_op_s *new_op;
 772         loff_t orig_size;
 773         int ret = -EINVAL;
 774
 775         gossip_debug(GOSSIP_INODE_DEBUG,
 776                      "%s: %pU: Handle is %pU | fs_id %d | size is %llu\n",
 777                      __func__,
 778                      get_khandle_from_ino(inode),
 779                      &orangefs_inode->refn.khandle,
 780                      orangefs_inode->refn.fs_id,
 781                      iattr->ia_size);
 782
 783         /* Ensure that we have a up to date size, so we know if it changed. */
 784         ret = orangefs_inode_getattr(inode, ORANGEFS_GETATTR_SIZE);
 785         if (ret == -ESTALE)
 786                 ret = -EIO;
 787         if (ret) {
 788                 gossip_err("%s: orangefs_inode_getattr failed, ret:%d:.\n",
 789                     __func__, ret);
 790                 return ret;
 791         }
 792         orig_size = i_size_read(inode);
 793
 794         /* This is truncate_setsize in a different order. */
 795         truncate_pagecache(inode, iattr->ia_size);
 796         i_size_write(inode, iattr->ia_size);
 797         if (iattr->ia_size > orig_size)
 798                 pagecache_isize_extended(inode, orig_size, iattr->ia_size);
 799
 800         new_op = op_alloc(ORANGEFS_VFS_OP_TRUNCATE);
 801         if (!new_op)
 802                 return -ENOMEM;
 803
 804         new_op->upcall.req.truncate.refn = orangefs_inode->refn;
 805         new_op->upcall.req.truncate.size = (__s64) iattr->ia_size;
 806
 807         ret = service_operation(new_op,
 808                 __func__,
 809                 get_interruptible_flag(inode));
 810
 811         /*
 812          * the truncate has no downcall members to retrieve, but
 813          * the status value tells us if it went through ok or not
 814          */
 815         gossip_debug(GOSSIP_INODE_DEBUG, "%s: ret:%d:\n", __func__, ret);
 816
 817         op_release(new_op);
 818
 819         if (ret != 0)
 820                 return ret;
 821
 822         if (orig_size != i_size_read(inode))
 823                 iattr->ia_valid |= ATTR_CTIME | ATTR_MTIME;
 824
 825         return ret;
 826 }
 827
 828 int __orangefs_setattr(struct inode *inode, struct iattr *iattr)
 829 {
 830         int ret;
 831
 832         if (iattr->ia_valid & ATTR_MODE) {
 833                 if (iattr->ia_mode & (S_ISVTX)) {
 834                         if (is_root_handle(inode)) {
 835                                 /*
 836                                  * allow sticky bit to be set on root (since
 837                                  * it shows up that way by default anyhow),
 838                                  * but don't show it to the server
 839                                  */
 840                                 iattr->ia_mode -= S_ISVTX;
 841                         } else {
 842                                 gossip_debug(GOSSIP_UTILS_DEBUG,
 843                                              "User attempted to set sticky bit on non-root directory; returning EINVAL.\n");
 844                                 ret = -EINVAL;
 845                                 goto out;
 846                         }
 847                 }
 848                 if (iattr->ia_mode & (S_ISUID)) {
 849                         gossip_debug(GOSSIP_UTILS_DEBUG,
 850                                      "Attempting to set setuid bit (not supported); returning EINVAL.\n");
 851                         ret = -EINVAL;
 852                         goto out;
 853                 }
 854         }
 855
 856         if (iattr->ia_valid & ATTR_SIZE) {
 857                 ret = orangefs_setattr_size(inode, iattr);
 858                 if (ret)
 859                         goto out;
 860         }
 861
 862 again:
 863         spin_lock(&inode->i_lock);
 864         if (ORANGEFS_I(inode)->attr_valid) {
 865                 if (uid_eq(ORANGEFS_I(inode)->attr_uid, current_fsuid()) &&
 866                     gid_eq(ORANGEFS_I(inode)->attr_gid, current_fsgid())) {
 867                         ORANGEFS_I(inode)->attr_valid = iattr->ia_valid;
 868                 } else {
 869                         spin_unlock(&inode->i_lock);
 870                         write_inode_now(inode, 1);
 871                         goto again;
 872                 }
 873         } else {
 874                 ORANGEFS_I(inode)->attr_valid = iattr->ia_valid;
 875                 ORANGEFS_I(inode)->attr_uid = current_fsuid();
 876                 ORANGEFS_I(inode)->attr_gid = current_fsgid();
 877         }
 878         setattr_copy(inode, iattr);
 879         spin_unlock(&inode->i_lock);
 880         mark_inode_dirty(inode);
 881
 882         if (iattr->ia_valid & ATTR_MODE)
 883                 /* change mod on a file that has ACLs */
 884                 ret = posix_acl_chmod(inode, inode->i_mode);
 885
 886         ret = 0;
 887 out:
 888         return ret;
 889 }
 890
 891 /*
 892  * Change attributes of an object referenced by dentry.
 893  */
 894 int orangefs_setattr(struct dentry *dentry, struct iattr *iattr)
 895 {
 896         int ret;
 897         gossip_debug(GOSSIP_INODE_DEBUG, "__orangefs_setattr: called on %pd\n",
 898             dentry);
 899         ret = setattr_prepare(dentry, iattr);
 900         if (ret)
 901                 goto out;
 902         ret = __orangefs_setattr(d_inode(dentry), iattr);
 903         sync_inode_metadata(d_inode(dentry), 1);
 904 out:
 905         gossip_debug(GOSSIP_INODE_DEBUG, "orangefs_setattr: returning %d\n",
 906             ret);
 907         return ret;
 908 }
 909
 910 /*
 911  * Obtain attributes of an object given a dentry
 912  */
 913 int orangefs_getattr(const struct path *path, struct kstat *stat,
 914                      u32 request_mask, unsigned int flags)
 915 {
 916         int ret;
 917         struct inode *inode = path->dentry->d_inode;
 918
 919         gossip_debug(GOSSIP_INODE_DEBUG,
 920                      "orangefs_getattr: called on %pd mask %u\n",
 921                      path->dentry, request_mask);
 922
 923         ret = orangefs_inode_getattr(inode,
 924             request_mask & STATX_SIZE ? ORANGEFS_GETATTR_SIZE : 0);
 925         if (ret == 0) {
 926                 generic_fillattr(inode, stat);
 927
 928                 /* override block size reported to stat */
 929                 if (!(request_mask & STATX_SIZE))
 930                         stat->result_mask &= ~STATX_SIZE;
 931
 932                 stat->attributes_mask = STATX_ATTR_IMMUTABLE |
 933                     STATX_ATTR_APPEND;
 934                 if (inode->i_flags & S_IMMUTABLE)
 935                         stat->attributes |= STATX_ATTR_IMMUTABLE;
 936                 if (inode->i_flags & S_APPEND)
 937                         stat->attributes |= STATX_ATTR_APPEND;
 938         }
 939         return ret;
 940 }
 941
 942 int orangefs_permission(struct inode *inode, int mask)
 943 {
 944         int ret;
 945
 946         if (mask & MAY_NOT_BLOCK)
 947                 return -ECHILD;
 948
 949         gossip_debug(GOSSIP_INODE_DEBUG, "%s: refreshing\n", __func__);
 950
 951         /* Make sure the permission (and other common attrs) are up to date. */
 952         ret = orangefs_inode_getattr(inode, 0);
 953         if (ret < 0)
 954                 return ret;
 955
 956         return generic_permission(inode, mask);
 957 }
 958
 959 int orangefs_update_time(struct inode *inode, struct timespec64 *time, int flags)
 960 {
 961         struct iattr iattr;
 962         gossip_debug(GOSSIP_INODE_DEBUG, "orangefs_update_time: %pU\n",
 963             get_khandle_from_ino(inode));
 964         generic_update_time(inode, time, flags);
 965         memset(&iattr, 0, sizeof iattr);
 966         if (flags & S_ATIME)
 967                 iattr.ia_valid |= ATTR_ATIME;
 968         if (flags & S_CTIME)
 969                 iattr.ia_valid |= ATTR_CTIME;
 970         if (flags & S_MTIME)
 971                 iattr.ia_valid |= ATTR_MTIME;
 972         return __orangefs_setattr(inode, &iattr);
 973 }
 974
 975 /* ORANGEFS2 implementation of VFS inode operations for files */
 976 static const struct inode_operations orangefs_file_inode_operations = {
 977         .get_acl = orangefs_get_acl,
 978         .set_acl = orangefs_set_acl,
 979         .setattr = orangefs_setattr,
 980         .getattr = orangefs_getattr,
 981         .listxattr = orangefs_listxattr,
 982         .permission = orangefs_permission,
 983         .update_time = orangefs_update_time,
 984 };
 985
 986 static int orangefs_init_iops(struct inode *inode)
 987 {
 988         inode->i_mapping->a_ops = &orangefs_address_operations;
 989
 990         switch (inode->i_mode & S_IFMT) {
 991         case S_IFREG:
 992                 inode->i_op = &orangefs_file_inode_operations;
 993                 inode->i_fop = &orangefs_file_operations;
 994                 break;
 995         case S_IFLNK:
 996                 inode->i_op = &orangefs_symlink_inode_operations;
 997                 break;
 998         case S_IFDIR:
 999                 inode->i_op = &orangefs_dir_inode_operations;
1000                 inode->i_fop = &orangefs_dir_operations;
1001                 break;
1002         default:
1003                 gossip_debug(GOSSIP_INODE_DEBUG,
1004                              "%s: unsupported mode\n",
1005                              __func__);
1006                 return -EINVAL;
1007         }
1008
1009         return 0;
1010 }
1011
1012 /*
1013  * Given an ORANGEFS object identifier (fsid, handle), convert it into
1014  * a ino_t type that will be used as a hash-index from where the handle will
1015  * be searched for in the VFS hash table of inodes.
1016  */
1017 static inline ino_t orangefs_handle_hash(struct orangefs_object_kref *ref)
1018 {
1019         if (!ref)
1020                 return 0;
1021         return orangefs_khandle_to_ino(&(ref->khandle));
1022 }
1023
1024 /*
1025  * Called to set up an inode from iget5_locked.
1026  */
1027 static int orangefs_set_inode(struct inode *inode, void *data)
1028 {
1029         struct orangefs_object_kref *ref = (struct orangefs_object_kref *) data;
1030         ORANGEFS_I(inode)->refn.fs_id = ref->fs_id;
1031         ORANGEFS_I(inode)->refn.khandle = ref->khandle;
1032         ORANGEFS_I(inode)->attr_valid = 0;
1033         hash_init(ORANGEFS_I(inode)->xattr_cache);
1034         ORANGEFS_I(inode)->mapping_time = jiffies - 1;
1035         ORANGEFS_I(inode)->bitlock = 0;
1036         return 0;
1037 }
1038
1039 /*
1040  * Called to determine if handles match.
1041  */
1042 static int orangefs_test_inode(struct inode *inode, void *data)
1043 {
1044         struct orangefs_object_kref *ref = (struct orangefs_object_kref *) data;
1045         struct orangefs_inode_s *orangefs_inode = NULL;
1046
1047         orangefs_inode = ORANGEFS_I(inode);
1048         /* test handles and fs_ids... */
1049         return (!ORANGEFS_khandle_cmp(&(orangefs_inode->refn.khandle),
1050                                 &(ref->khandle)) &&
1051                         orangefs_inode->refn.fs_id == ref->fs_id);
1052 }
1053
1054 /*
1055  * Front-end to lookup the inode-cache maintained by the VFS using the ORANGEFS
1056  * file handle.
1057  *
1058  * @sb: the file system super block instance.
1059  * @ref: The ORANGEFS object for which we are trying to locate an inode.
1060  */
1061 struct inode *orangefs_iget(struct super_block *sb,
1062                 struct orangefs_object_kref *ref)
1063 {
1064         struct inode *inode = NULL;
1065         unsigned long hash;
1066         int error;
1067
1068         hash = orangefs_handle_hash(ref);
1069         inode = iget5_locked(sb,
1070                         hash,
1071                         orangefs_test_inode,
1072                         orangefs_set_inode,
1073                         ref);
1074
1075         if (!inode)
1076                 return ERR_PTR(-ENOMEM);
1077
1078         if (!(inode->i_state & I_NEW))
1079                 return inode;
1080
1081         error = orangefs_inode_getattr(inode, ORANGEFS_GETATTR_NEW);
1082         if (error) {
1083                 iget_failed(inode);
1084                 return ERR_PTR(error);
1085         }
1086
1087         inode->i_ino = hash;    /* needed for stat etc */
1088         orangefs_init_iops(inode);
1089         unlock_new_inode(inode);
1090
1091         gossip_debug(GOSSIP_INODE_DEBUG,
1092                      "iget handle %pU, fsid %d hash %ld i_ino %lu\n",
1093                      &ref->khandle,
1094                      ref->fs_id,
1095                      hash,
1096                      inode->i_ino);
1097
1098         return inode;
1099 }
1100
1101 /*
1102  * Allocate an inode for a newly created file and insert it into the inode hash.
1103  */
1104 struct inode *orangefs_new_inode(struct super_block *sb, struct inode *dir,
1105                 int mode, dev_t dev, struct orangefs_object_kref *ref)
1106 {
1107         unsigned long hash = orangefs_handle_hash(ref);
1108         struct inode *inode;
1109         int error;
1110
1111         gossip_debug(GOSSIP_INODE_DEBUG,
1112                      "%s:(sb is %p | MAJOR(dev)=%u | MINOR(dev)=%u mode=%o)\n",
1113                      __func__,
1114                      sb,
1115                      MAJOR(dev),
1116                      MINOR(dev),
1117                      mode);
1118
1119         inode = new_inode(sb);
1120         if (!inode)
1121                 return ERR_PTR(-ENOMEM);
1122
1123         orangefs_set_inode(inode, ref);
1124         inode->i_ino = hash;    /* needed for stat etc */
1125
1126         error = orangefs_inode_getattr(inode, ORANGEFS_GETATTR_NEW);
1127         if (error)
1128                 goto out_iput;
1129
1130         orangefs_init_iops(inode);
1131         inode->i_rdev = dev;
1132
1133         error = insert_inode_locked4(inode, hash, orangefs_test_inode, ref);
1134         if (error < 0)
1135                 goto out_iput;
1136
1137         gossip_debug(GOSSIP_INODE_DEBUG,
1138                      "Initializing ACL's for inode %pU\n",
1139                      get_khandle_from_ino(inode));
1140         orangefs_init_acl(inode, dir);
1141         return inode;
1142
1143 out_iput:
1144         iput(inode);
1145         return ERR_PTR(error);
1146 }