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