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 <linux/fileattr.h>
  15 #include "protocol.h"
  16 #include "orangefs-kernel.h"
  17 #include "orangefs-bufmap.h"
  18
  19 static int orangefs_writepage_locked(struct page *page,
  20     struct writeback_control *wbc)
  21 {
  22         struct inode *inode = page->mapping->host;
  23         struct orangefs_write_range *wr = NULL;
  24         struct iov_iter iter;
  25         struct bio_vec bv;
  26         size_t len, wlen;
  27         ssize_t ret;
  28         loff_t off;
  29
  30         set_page_writeback(page);
  31
  32         len = i_size_read(inode);
  33         if (PagePrivate(page)) {
  34                 wr = (struct orangefs_write_range *)page_private(page);
  35                 WARN_ON(wr->pos >= len);
  36                 off = wr->pos;
  37                 if (off + wr->len > len)
  38                         wlen = len - off;
  39                 else
  40                         wlen = wr->len;
  41         } else {
  42                 WARN_ON(1);
  43                 off = page_offset(page);
  44                 if (off + PAGE_SIZE > len)
  45                         wlen = len - off;
  46                 else
  47                         wlen = PAGE_SIZE;
  48         }
  49         /* Should've been handled in orangefs_invalidatepage. */
  50         WARN_ON(off == len || off + wlen > len);
  51
  52         bv.bv_page = page;
  53         bv.bv_len = wlen;
  54         bv.bv_offset = off % PAGE_SIZE;
  55         WARN_ON(wlen == 0);
  56         iov_iter_bvec(&iter, WRITE, &bv, 1, wlen);
  57
  58         ret = wait_for_direct_io(ORANGEFS_IO_WRITE, inode, &off, &iter, wlen,
  59             len, wr, NULL, NULL);
  60         if (ret < 0) {
  61                 SetPageError(page);
  62                 mapping_set_error(page->mapping, ret);
  63         } else {
  64                 ret = 0;
  65         }
  66         kfree(detach_page_private(page));
  67         return ret;
  68 }
  69
  70 static int orangefs_writepage(struct page *page, struct writeback_control *wbc)
  71 {
  72         int ret;
  73         ret = orangefs_writepage_locked(page, wbc);
  74         unlock_page(page);
  75         end_page_writeback(page);
  76         return ret;
  77 }
  78
  79 struct orangefs_writepages {
  80         loff_t off;
  81         size_t len;
  82         kuid_t uid;
  83         kgid_t gid;
  84         int maxpages;
  85         int npages;
  86         struct page **pages;
  87         struct bio_vec *bv;
  88 };
  89
  90 static int orangefs_writepages_work(struct orangefs_writepages *ow,
  91     struct writeback_control *wbc)
  92 {
  93         struct inode *inode = ow->pages[0]->mapping->host;
  94         struct orangefs_write_range *wrp, wr;
  95         struct iov_iter iter;
  96         ssize_t ret;
  97         size_t len;
  98         loff_t off;
  99         int i;
 100
 101         len = i_size_read(inode);
 102
 103         for (i = 0; i < ow->npages; i++) {
 104                 set_page_writeback(ow->pages[i]);
 105                 ow->bv[i].bv_page = ow->pages[i];
 106                 ow->bv[i].bv_len = min(page_offset(ow->pages[i]) + PAGE_SIZE,
 107                     ow->off + ow->len) -
 108                     max(ow->off, page_offset(ow->pages[i]));
 109                 if (i == 0)
 110                         ow->bv[i].bv_offset = ow->off -
 111                             page_offset(ow->pages[i]);
 112                 else
 113                         ow->bv[i].bv_offset = 0;
 114         }
 115         iov_iter_bvec(&iter, WRITE, ow->bv, ow->npages, ow->len);
 116
 117         WARN_ON(ow->off >= len);
 118         if (ow->off + ow->len > len)
 119                 ow->len = len - ow->off;
 120
 121         off = ow->off;
 122         wr.uid = ow->uid;
 123         wr.gid = ow->gid;
 124         ret = wait_for_direct_io(ORANGEFS_IO_WRITE, inode, &off, &iter, ow->len,
 125             0, &wr, NULL, NULL);
 126         if (ret < 0) {
 127                 for (i = 0; i < ow->npages; i++) {
 128                         SetPageError(ow->pages[i]);
 129                         mapping_set_error(ow->pages[i]->mapping, ret);
 130                         if (PagePrivate(ow->pages[i])) {
 131                                 wrp = (struct orangefs_write_range *)
 132                                     page_private(ow->pages[i]);
 133                                 ClearPagePrivate(ow->pages[i]);
 134                                 put_page(ow->pages[i]);
 135                                 kfree(wrp);
 136                         }
 137                         end_page_writeback(ow->pages[i]);
 138                         unlock_page(ow->pages[i]);
 139                 }
 140         } else {
 141                 ret = 0;
 142                 for (i = 0; i < ow->npages; i++) {
 143                         if (PagePrivate(ow->pages[i])) {
 144                                 wrp = (struct orangefs_write_range *)
 145                                     page_private(ow->pages[i]);
 146                                 ClearPagePrivate(ow->pages[i]);
 147                                 put_page(ow->pages[i]);
 148                                 kfree(wrp);
 149                         }
 150                         end_page_writeback(ow->pages[i]);
 151                         unlock_page(ow->pages[i]);
 152                 }
 153         }
 154         return ret;
 155 }
 156
 157 static int orangefs_writepages_callback(struct page *page,
 158     struct writeback_control *wbc, void *data)
 159 {
 160         struct orangefs_writepages *ow = data;
 161         struct orangefs_write_range *wr;
 162         int ret;
 163
 164         if (!PagePrivate(page)) {
 165                 unlock_page(page);
 166                 /* It's not private so there's nothing to write, right? */
 167                 printk("writepages_callback not private!\n");
 168                 BUG();
 169                 return 0;
 170         }
 171         wr = (struct orangefs_write_range *)page_private(page);
 172
 173         ret = -1;
 174         if (ow->npages == 0) {
 175                 ow->off = wr->pos;
 176                 ow->len = wr->len;
 177                 ow->uid = wr->uid;
 178                 ow->gid = wr->gid;
 179                 ow->pages[ow->npages++] = page;
 180                 ret = 0;
 181                 goto done;
 182         }
 183         if (!uid_eq(ow->uid, wr->uid) || !gid_eq(ow->gid, wr->gid)) {
 184                 orangefs_writepages_work(ow, wbc);
 185                 ow->npages = 0;
 186                 ret = -1;
 187                 goto done;
 188         }
 189         if (ow->off + ow->len == wr->pos) {
 190                 ow->len += wr->len;
 191                 ow->pages[ow->npages++] = page;
 192                 ret = 0;
 193                 goto done;
 194         }
 195 done:
 196         if (ret == -1) {
 197                 if (ow->npages) {
 198                         orangefs_writepages_work(ow, wbc);
 199                         ow->npages = 0;
 200                 }
 201                 ret = orangefs_writepage_locked(page, wbc);
 202                 mapping_set_error(page->mapping, ret);
 203                 unlock_page(page);
 204                 end_page_writeback(page);
 205         } else {
 206                 if (ow->npages == ow->maxpages) {
 207                         orangefs_writepages_work(ow, wbc);
 208                         ow->npages = 0;
 209                 }
 210         }
 211         return ret;
 212 }
 213
 214 static int orangefs_writepages(struct address_space *mapping,
 215     struct writeback_control *wbc)
 216 {
 217         struct orangefs_writepages *ow;
 218         struct blk_plug plug;
 219         int ret;
 220         ow = kzalloc(sizeof(struct orangefs_writepages), GFP_KERNEL);
 221         if (!ow)
 222                 return -ENOMEM;
 223         ow->maxpages = orangefs_bufmap_size_query()/PAGE_SIZE;
 224         ow->pages = kcalloc(ow->maxpages, sizeof(struct page *), GFP_KERNEL);
 225         if (!ow->pages) {
 226                 kfree(ow);
 227                 return -ENOMEM;
 228         }
 229         ow->bv = kcalloc(ow->maxpages, sizeof(struct bio_vec), GFP_KERNEL);
 230         if (!ow->bv) {
 231                 kfree(ow->pages);
 232                 kfree(ow);
 233                 return -ENOMEM;
 234         }
 235         blk_start_plug(&plug);
 236         ret = write_cache_pages(mapping, wbc, orangefs_writepages_callback, ow);
 237         if (ow->npages)
 238                 ret = orangefs_writepages_work(ow, wbc);
 239         blk_finish_plug(&plug);
 240         kfree(ow->pages);
 241         kfree(ow->bv);
 242         kfree(ow);
 243         return ret;
 244 }
 245
 246 static int orangefs_launder_page(struct page *);
 247
 248 static int orangefs_readpage(struct file *file, struct page *page)
 249 {
 250         struct inode *inode = page->mapping->host;
 251         struct iov_iter iter;
 252         struct bio_vec bv;
 253         ssize_t ret;
 254         loff_t off; /* offset into this page */
 255         pgoff_t index; /* which page */
 256         struct page *next_page;
 257         char *kaddr;
 258         loff_t read_size;
 259         int buffer_index = -1; /* orangefs shared memory slot */
 260         int slot_index;   /* index into slot */
 261         int remaining;
 262
 263         /*
 264          * Get up to this many bytes from Orangefs at a time and try
 265          * to fill them into the page cache at once. Tests with dd made
 266          * this seem like a reasonable static number, if there was
 267          * interest perhaps this number could be made setable through
 268          * sysfs...
 269          */
 270         read_size = 524288;
 271
 272         if (PageDirty(page))
 273                 orangefs_launder_page(page);
 274
 275         off = page_offset(page);
 276         index = off >> PAGE_SHIFT;
 277         bv.bv_page = page;
 278         bv.bv_len = PAGE_SIZE;
 279         bv.bv_offset = 0;
 280         iov_iter_bvec(&iter, READ, &bv, 1, PAGE_SIZE);
 281
 282         ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, &off, &iter,
 283             read_size, inode->i_size, NULL, &buffer_index, file);
 284         remaining = ret;
 285         /* this will only zero remaining unread portions of the page data */
 286         iov_iter_zero(~0U, &iter);
 287         /* takes care of potential aliasing */
 288         flush_dcache_page(page);
 289         if (ret < 0) {
 290                 SetPageError(page);
 291                 unlock_page(page);
 292                 goto out;
 293         } else {
 294                 SetPageUptodate(page);
 295                 if (PageError(page))
 296                         ClearPageError(page);
 297                 ret = 0;
 298         }
 299         /* unlock the page after the ->readpage() routine completes */
 300         unlock_page(page);
 301
 302         if (remaining > PAGE_SIZE) {
 303                 slot_index = 0;
 304                 while ((remaining - PAGE_SIZE) >= PAGE_SIZE) {
 305                         remaining -= PAGE_SIZE;
 306                         /*
 307                          * It is an optimization to try and fill more than one
 308                          * page... by now we've already gotten the single
 309                          * page we were after, if stuff doesn't seem to
 310                          * be going our way at this point just return
 311                          * and hope for the best.
 312                          *
 313                          * If we look for pages and they're already there is
 314                          * one reason to give up, and if they're not there
 315                          * and we can't create them is another reason.
 316                          */
 317
 318                         index++;
 319                         slot_index++;
 320                         next_page = find_get_page(inode->i_mapping, index);
 321                         if (next_page) {
 322                                 gossip_debug(GOSSIP_FILE_DEBUG,
 323                                         "%s: found next page, quitting\n",
 324                                         __func__);
 325                                 put_page(next_page);
 326                                 goto out;
 327                         }
 328                         next_page = find_or_create_page(inode->i_mapping,
 329                                                         index,
 330                                                         GFP_KERNEL);
 331                         /*
 332                          * I've never hit this, leave it as a printk for
 333                          * now so it will be obvious.
 334                          */
 335                         if (!next_page) {
 336                                 printk("%s: can't create next page, quitting\n",
 337                                         __func__);
 338                                 goto out;
 339                         }
 340                         kaddr = kmap_atomic(next_page);
 341                         orangefs_bufmap_page_fill(kaddr,
 342                                                 buffer_index,
 343                                                 slot_index);
 344                         kunmap_atomic(kaddr);
 345                         SetPageUptodate(next_page);
 346                         unlock_page(next_page);
 347                         put_page(next_page);
 348                 }
 349         }
 350
 351 out:
 352         if (buffer_index != -1)
 353                 orangefs_bufmap_put(buffer_index);
 354         return ret;
 355 }
 356
 357 static int orangefs_write_begin(struct file *file,
 358     struct address_space *mapping,
 359     loff_t pos, unsigned len, unsigned flags, struct page **pagep,
 360     void **fsdata)
 361 {
 362         struct orangefs_write_range *wr;
 363         struct page *page;
 364         pgoff_t index;
 365         int ret;
 366
 367         index = pos >> PAGE_SHIFT;
 368
 369         page = grab_cache_page_write_begin(mapping, index, flags);
 370         if (!page)
 371                 return -ENOMEM;
 372
 373         *pagep = page;
 374
 375         if (PageDirty(page) && !PagePrivate(page)) {
 376                 /*
 377                  * Should be impossible.  If it happens, launder the page
 378                  * since we don't know what's dirty.  This will WARN in
 379                  * orangefs_writepage_locked.
 380                  */
 381                 ret = orangefs_launder_page(page);
 382                 if (ret)
 383                         return ret;
 384         }
 385         if (PagePrivate(page)) {
 386                 struct orangefs_write_range *wr;
 387                 wr = (struct orangefs_write_range *)page_private(page);
 388                 if (wr->pos + wr->len == pos &&
 389                     uid_eq(wr->uid, current_fsuid()) &&
 390                     gid_eq(wr->gid, current_fsgid())) {
 391                         wr->len += len;
 392                         goto okay;
 393                 } else {
 394                         ret = orangefs_launder_page(page);
 395                         if (ret)
 396                                 return ret;
 397                 }
 398         }
 399
 400         wr = kmalloc(sizeof *wr, GFP_KERNEL);
 401         if (!wr)
 402                 return -ENOMEM;
 403
 404         wr->pos = pos;
 405         wr->len = len;
 406         wr->uid = current_fsuid();
 407         wr->gid = current_fsgid();
 408         attach_page_private(page, wr);
 409 okay:
 410         return 0;
 411 }
 412
 413 static int orangefs_write_end(struct file *file, struct address_space *mapping,
 414     loff_t pos, unsigned len, unsigned copied, struct page *page, void *fsdata)
 415 {
 416         struct inode *inode = page->mapping->host;
 417         loff_t last_pos = pos + copied;
 418
 419         /*
 420          * No need to use i_size_read() here, the i_size
 421          * cannot change under us because we hold the i_mutex.
 422          */
 423         if (last_pos > inode->i_size)
 424                 i_size_write(inode, last_pos);
 425
 426         /* zero the stale part of the page if we did a short copy */
 427         if (!PageUptodate(page)) {
 428                 unsigned from = pos & (PAGE_SIZE - 1);
 429                 if (copied < len) {
 430                         zero_user(page, from + copied, len - copied);
 431                 }
 432                 /* Set fully written pages uptodate. */
 433                 if (pos == page_offset(page) &&
 434                     (len == PAGE_SIZE || pos + len == inode->i_size)) {
 435                         zero_user_segment(page, from + copied, PAGE_SIZE);
 436                         SetPageUptodate(page);
 437                 }
 438         }
 439
 440         set_page_dirty(page);
 441         unlock_page(page);
 442         put_page(page);
 443
 444         mark_inode_dirty_sync(file_inode(file));
 445         return copied;
 446 }
 447
 448 static void orangefs_invalidatepage(struct page *page,
 449                                  unsigned int offset,
 450                                  unsigned int length)
 451 {
 452         struct orangefs_write_range *wr;
 453         wr = (struct orangefs_write_range *)page_private(page);
 454
 455         if (offset == 0 && length == PAGE_SIZE) {
 456                 kfree(detach_page_private(page));
 457                 return;
 458         /* write range entirely within invalidate range (or equal) */
 459         } else if (page_offset(page) + offset <= wr->pos &&
 460             wr->pos + wr->len <= page_offset(page) + offset + length) {
 461                 kfree(detach_page_private(page));
 462                 /* XXX is this right? only caller in fs */
 463                 cancel_dirty_page(page);
 464                 return;
 465         /* invalidate range chops off end of write range */
 466         } else if (wr->pos < page_offset(page) + offset &&
 467             wr->pos + wr->len <= page_offset(page) + offset + length &&
 468              page_offset(page) + offset < wr->pos + wr->len) {
 469                 size_t x;
 470                 x = wr->pos + wr->len - (page_offset(page) + offset);
 471                 WARN_ON(x > wr->len);
 472                 wr->len -= x;
 473                 wr->uid = current_fsuid();
 474                 wr->gid = current_fsgid();
 475         /* invalidate range chops off beginning of write range */
 476         } else if (page_offset(page) + offset <= wr->pos &&
 477             page_offset(page) + offset + length < wr->pos + wr->len &&
 478             wr->pos < page_offset(page) + offset + length) {
 479                 size_t x;
 480                 x = page_offset(page) + offset + length - wr->pos;
 481                 WARN_ON(x > wr->len);
 482                 wr->pos += x;
 483                 wr->len -= x;
 484                 wr->uid = current_fsuid();
 485                 wr->gid = current_fsgid();
 486         /* invalidate range entirely within write range (punch hole) */
 487         } else if (wr->pos < page_offset(page) + offset &&
 488             page_offset(page) + offset + length < wr->pos + wr->len) {
 489                 /* XXX what do we do here... should not WARN_ON */
 490                 WARN_ON(1);
 491                 /* punch hole */
 492                 /*
 493                  * should we just ignore this and write it out anyway?
 494                  * it hardly makes sense
 495                  */
 496                 return;
 497         /* non-overlapping ranges */
 498         } else {
 499                 /* WARN if they do overlap */
 500                 if (!((page_offset(page) + offset + length <= wr->pos) ^
 501                     (wr->pos + wr->len <= page_offset(page) + offset))) {
 502                         WARN_ON(1);
 503                         printk("invalidate range offset %llu length %u\n",
 504                             page_offset(page) + offset, length);
 505                         printk("write range offset %llu length %zu\n",
 506                             wr->pos, wr->len);
 507                 }
 508                 return;
 509         }
 510
 511         /*
 512          * Above there are returns where wr is freed or where we WARN.
 513          * Thus the following runs if wr was modified above.
 514          */
 515
 516         orangefs_launder_page(page);
 517 }
 518
 519 static int orangefs_releasepage(struct page *page, gfp_t foo)
 520 {
 521         return !PagePrivate(page);
 522 }
 523
 524 static void orangefs_freepage(struct page *page)
 525 {
 526         kfree(detach_page_private(page));
 527 }
 528
 529 static int orangefs_launder_page(struct page *page)
 530 {
 531         int r = 0;
 532         struct writeback_control wbc = {
 533                 .sync_mode = WB_SYNC_ALL,
 534                 .nr_to_write = 0,
 535         };
 536         wait_on_page_writeback(page);
 537         if (clear_page_dirty_for_io(page)) {
 538                 r = orangefs_writepage_locked(page, &wbc);
 539                 end_page_writeback(page);
 540         }
 541         return r;
 542 }
 543
 544 static ssize_t orangefs_direct_IO(struct kiocb *iocb,
 545                                   struct iov_iter *iter)
 546 {
 547         /*
 548          * Comment from original do_readv_writev:
 549          * Common entry point for read/write/readv/writev
 550          * This function will dispatch it to either the direct I/O
 551          * or buffered I/O path depending on the mount options and/or
 552          * augmented/extended metadata attached to the file.
 553          * Note: File extended attributes override any mount options.
 554          */
 555         struct file *file = iocb->ki_filp;
 556         loff_t pos = iocb->ki_pos;
 557         enum ORANGEFS_io_type type = iov_iter_rw(iter) == WRITE ?
 558             ORANGEFS_IO_WRITE : ORANGEFS_IO_READ;
 559         loff_t *offset = &pos;
 560         struct inode *inode = file->f_mapping->host;
 561         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 562         struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
 563         size_t count = iov_iter_count(iter);
 564         ssize_t total_count = 0;
 565         ssize_t ret = -EINVAL;
 566         int i = 0;
 567
 568         gossip_debug(GOSSIP_FILE_DEBUG,
 569                 "%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
 570                 __func__,
 571                 handle,
 572                 (int)count);
 573
 574         if (type == ORANGEFS_IO_WRITE) {
 575                 gossip_debug(GOSSIP_FILE_DEBUG,
 576                              "%s(%pU): proceeding with offset : %llu, "
 577                              "size %d\n",
 578                              __func__,
 579                              handle,
 580                              llu(*offset),
 581                              (int)count);
 582         }
 583
 584         if (count == 0) {
 585                 ret = 0;
 586                 goto out;
 587         }
 588
 589         while (iov_iter_count(iter)) {
 590                 size_t each_count = iov_iter_count(iter);
 591                 size_t amt_complete;
 592                 i++;
 593
 594                 /* how much to transfer in this loop iteration */
 595                 if (each_count > orangefs_bufmap_size_query())
 596                         each_count = orangefs_bufmap_size_query();
 597
 598                 gossip_debug(GOSSIP_FILE_DEBUG,
 599                              "%s(%pU): size of each_count(%d)\n",
 600                              __func__,
 601                              handle,
 602                              (int)each_count);
 603                 gossip_debug(GOSSIP_FILE_DEBUG,
 604                              "%s(%pU): BEFORE wait_for_io: offset is %d\n",
 605                              __func__,
 606                              handle,
 607                              (int)*offset);
 608
 609                 ret = wait_for_direct_io(type, inode, offset, iter,
 610                                 each_count, 0, NULL, NULL, file);
 611                 gossip_debug(GOSSIP_FILE_DEBUG,
 612                              "%s(%pU): return from wait_for_io:%d\n",
 613                              __func__,
 614                              handle,
 615                              (int)ret);
 616
 617                 if (ret < 0)
 618                         goto out;
 619
 620                 *offset += ret;
 621                 total_count += ret;
 622                 amt_complete = ret;
 623
 624                 gossip_debug(GOSSIP_FILE_DEBUG,
 625                              "%s(%pU): AFTER wait_for_io: offset is %d\n",
 626                              __func__,
 627                              handle,
 628                              (int)*offset);
 629
 630                 /*
 631                  * if we got a short I/O operations,
 632                  * fall out and return what we got so far
 633                  */
 634                 if (amt_complete < each_count)
 635                         break;
 636         } /*end while */
 637
 638 out:
 639         if (total_count > 0)
 640                 ret = total_count;
 641         if (ret > 0) {
 642                 if (type == ORANGEFS_IO_READ) {
 643                         file_accessed(file);
 644                 } else {
 645                         file_update_time(file);
 646                         if (*offset > i_size_read(inode))
 647                                 i_size_write(inode, *offset);
 648                 }
 649         }
 650
 651         gossip_debug(GOSSIP_FILE_DEBUG,
 652                      "%s(%pU): Value(%d) returned.\n",
 653                      __func__,
 654                      handle,
 655                      (int)ret);
 656
 657         return ret;
 658 }
 659
 660 /** ORANGEFS2 implementation of address space operations */
 661 static const struct address_space_operations orangefs_address_operations = {
 662         .writepage = orangefs_writepage,
 663         .readpage = orangefs_readpage,
 664         .writepages = orangefs_writepages,
 665         .set_page_dirty = __set_page_dirty_nobuffers,
 666         .write_begin = orangefs_write_begin,
 667         .write_end = orangefs_write_end,
 668         .invalidatepage = orangefs_invalidatepage,
 669         .releasepage = orangefs_releasepage,
 670         .freepage = orangefs_freepage,
 671         .launder_page = orangefs_launder_page,
 672         .direct_IO = orangefs_direct_IO,
 673 };
 674
 675 vm_fault_t orangefs_page_mkwrite(struct vm_fault *vmf)
 676 {
 677         struct page *page = vmf->page;
 678         struct inode *inode = file_inode(vmf->vma->vm_file);
 679         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 680         unsigned long *bitlock = &orangefs_inode->bitlock;
 681         vm_fault_t ret;
 682         struct orangefs_write_range *wr;
 683
 684         sb_start_pagefault(inode->i_sb);
 685
 686         if (wait_on_bit(bitlock, 1, TASK_KILLABLE)) {
 687                 ret = VM_FAULT_RETRY;
 688                 goto out;
 689         }
 690
 691         lock_page(page);
 692         if (PageDirty(page) && !PagePrivate(page)) {
 693                 /*
 694                  * Should be impossible.  If it happens, launder the page
 695                  * since we don't know what's dirty.  This will WARN in
 696                  * orangefs_writepage_locked.
 697                  */
 698                 if (orangefs_launder_page(page)) {
 699                         ret = VM_FAULT_LOCKED|VM_FAULT_RETRY;
 700                         goto out;
 701                 }
 702         }
 703         if (PagePrivate(page)) {
 704                 wr = (struct orangefs_write_range *)page_private(page);
 705                 if (uid_eq(wr->uid, current_fsuid()) &&
 706                     gid_eq(wr->gid, current_fsgid())) {
 707                         wr->pos = page_offset(page);
 708                         wr->len = PAGE_SIZE;
 709                         goto okay;
 710                 } else {
 711                         if (orangefs_launder_page(page)) {
 712                                 ret = VM_FAULT_LOCKED|VM_FAULT_RETRY;
 713                                 goto out;
 714                         }
 715                 }
 716         }
 717         wr = kmalloc(sizeof *wr, GFP_KERNEL);
 718         if (!wr) {
 719                 ret = VM_FAULT_LOCKED|VM_FAULT_RETRY;
 720                 goto out;
 721         }
 722         wr->pos = page_offset(page);
 723         wr->len = PAGE_SIZE;
 724         wr->uid = current_fsuid();
 725         wr->gid = current_fsgid();
 726         attach_page_private(page, wr);
 727 okay:
 728
 729         file_update_time(vmf->vma->vm_file);
 730         if (page->mapping != inode->i_mapping) {
 731                 unlock_page(page);
 732                 ret = VM_FAULT_LOCKED|VM_FAULT_NOPAGE;
 733                 goto out;
 734         }
 735
 736         /*
 737          * We mark the page dirty already here so that when freeze is in
 738          * progress, we are guaranteed that writeback during freezing will
 739          * see the dirty page and writeprotect it again.
 740          */
 741         set_page_dirty(page);
 742         wait_for_stable_page(page);
 743         ret = VM_FAULT_LOCKED;
 744 out:
 745         sb_end_pagefault(inode->i_sb);
 746         return ret;
 747 }
 748
 749 static int orangefs_setattr_size(struct inode *inode, struct iattr *iattr)
 750 {
 751         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 752         struct orangefs_kernel_op_s *new_op;
 753         loff_t orig_size;
 754         int ret = -EINVAL;
 755
 756         gossip_debug(GOSSIP_INODE_DEBUG,
 757                      "%s: %pU: Handle is %pU | fs_id %d | size is %llu\n",
 758                      __func__,
 759                      get_khandle_from_ino(inode),
 760                      &orangefs_inode->refn.khandle,
 761                      orangefs_inode->refn.fs_id,
 762                      iattr->ia_size);
 763
 764         /* Ensure that we have a up to date size, so we know if it changed. */
 765         ret = orangefs_inode_getattr(inode, ORANGEFS_GETATTR_SIZE);
 766         if (ret == -ESTALE)
 767                 ret = -EIO;
 768         if (ret) {
 769                 gossip_err("%s: orangefs_inode_getattr failed, ret:%d:.\n",
 770                     __func__, ret);
 771                 return ret;
 772         }
 773         orig_size = i_size_read(inode);
 774
 775         /* This is truncate_setsize in a different order. */
 776         truncate_pagecache(inode, iattr->ia_size);
 777         i_size_write(inode, iattr->ia_size);
 778         if (iattr->ia_size > orig_size)
 779                 pagecache_isize_extended(inode, orig_size, iattr->ia_size);
 780
 781         new_op = op_alloc(ORANGEFS_VFS_OP_TRUNCATE);
 782         if (!new_op)
 783                 return -ENOMEM;
 784
 785         new_op->upcall.req.truncate.refn = orangefs_inode->refn;
 786         new_op->upcall.req.truncate.size = (__s64) iattr->ia_size;
 787
 788         ret = service_operation(new_op,
 789                 __func__,
 790                 get_interruptible_flag(inode));
 791
 792         /*
 793          * the truncate has no downcall members to retrieve, but
 794          * the status value tells us if it went through ok or not
 795          */
 796         gossip_debug(GOSSIP_INODE_DEBUG, "%s: ret:%d:\n", __func__, ret);
 797
 798         op_release(new_op);
 799
 800         if (ret != 0)
 801                 return ret;
 802
 803         if (orig_size != i_size_read(inode))
 804                 iattr->ia_valid |= ATTR_CTIME | ATTR_MTIME;
 805
 806         return ret;
 807 }
 808
 809 int __orangefs_setattr(struct inode *inode, struct iattr *iattr)
 810 {
 811         int ret;
 812
 813         if (iattr->ia_valid & ATTR_MODE) {
 814                 if (iattr->ia_mode & (S_ISVTX)) {
 815                         if (is_root_handle(inode)) {
 816                                 /*
 817                                  * allow sticky bit to be set on root (since
 818                                  * it shows up that way by default anyhow),
 819                                  * but don't show it to the server
 820                                  */
 821                                 iattr->ia_mode -= S_ISVTX;
 822                         } else {
 823                                 gossip_debug(GOSSIP_UTILS_DEBUG,
 824                                              "User attempted to set sticky bit on non-root directory; returning EINVAL.\n");
 825                                 ret = -EINVAL;
 826                                 goto out;
 827                         }
 828                 }
 829                 if (iattr->ia_mode & (S_ISUID)) {
 830                         gossip_debug(GOSSIP_UTILS_DEBUG,
 831                                      "Attempting to set setuid bit (not supported); returning EINVAL.\n");
 832                         ret = -EINVAL;
 833                         goto out;
 834                 }
 835         }
 836
 837         if (iattr->ia_valid & ATTR_SIZE) {
 838                 ret = orangefs_setattr_size(inode, iattr);
 839                 if (ret)
 840                         goto out;
 841         }
 842
 843 again:
 844         spin_lock(&inode->i_lock);
 845         if (ORANGEFS_I(inode)->attr_valid) {
 846                 if (uid_eq(ORANGEFS_I(inode)->attr_uid, current_fsuid()) &&
 847                     gid_eq(ORANGEFS_I(inode)->attr_gid, current_fsgid())) {
 848                         ORANGEFS_I(inode)->attr_valid = iattr->ia_valid;
 849                 } else {
 850                         spin_unlock(&inode->i_lock);
 851                         write_inode_now(inode, 1);
 852                         goto again;
 853                 }
 854         } else {
 855                 ORANGEFS_I(inode)->attr_valid = iattr->ia_valid;
 856                 ORANGEFS_I(inode)->attr_uid = current_fsuid();
 857                 ORANGEFS_I(inode)->attr_gid = current_fsgid();
 858         }
 859         setattr_copy(&init_user_ns, inode, iattr);
 860         spin_unlock(&inode->i_lock);
 861         mark_inode_dirty(inode);
 862
 863         if (iattr->ia_valid & ATTR_MODE)
 864                 /* change mod on a file that has ACLs */
 865                 ret = posix_acl_chmod(&init_user_ns, inode, inode->i_mode);
 866
 867         ret = 0;
 868 out:
 869         return ret;
 870 }
 871
 872 /*
 873  * Change attributes of an object referenced by dentry.
 874  */
 875 int orangefs_setattr(struct user_namespace *mnt_userns, struct dentry *dentry,
 876                      struct iattr *iattr)
 877 {
 878         int ret;
 879         gossip_debug(GOSSIP_INODE_DEBUG, "__orangefs_setattr: called on %pd\n",
 880             dentry);
 881         ret = setattr_prepare(&init_user_ns, dentry, iattr);
 882         if (ret)
 883                 goto out;
 884         ret = __orangefs_setattr(d_inode(dentry), iattr);
 885         sync_inode_metadata(d_inode(dentry), 1);
 886 out:
 887         gossip_debug(GOSSIP_INODE_DEBUG, "orangefs_setattr: returning %d\n",
 888             ret);
 889         return ret;
 890 }
 891
 892 /*
 893  * Obtain attributes of an object given a dentry
 894  */
 895 int orangefs_getattr(struct user_namespace *mnt_userns, const struct path *path,
 896                      struct kstat *stat, u32 request_mask, unsigned int flags)
 897 {
 898         int ret;
 899         struct inode *inode = path->dentry->d_inode;
 900
 901         gossip_debug(GOSSIP_INODE_DEBUG,
 902                      "orangefs_getattr: called on %pd mask %u\n",
 903                      path->dentry, request_mask);
 904
 905         ret = orangefs_inode_getattr(inode,
 906             request_mask & STATX_SIZE ? ORANGEFS_GETATTR_SIZE : 0);
 907         if (ret == 0) {
 908                 generic_fillattr(&init_user_ns, inode, stat);
 909
 910                 /* override block size reported to stat */
 911                 if (!(request_mask & STATX_SIZE))
 912                         stat->result_mask &= ~STATX_SIZE;
 913
 914                 stat->attributes_mask = STATX_ATTR_IMMUTABLE |
 915                     STATX_ATTR_APPEND;
 916                 if (inode->i_flags & S_IMMUTABLE)
 917                         stat->attributes |= STATX_ATTR_IMMUTABLE;
 918                 if (inode->i_flags & S_APPEND)
 919                         stat->attributes |= STATX_ATTR_APPEND;
 920         }
 921         return ret;
 922 }
 923
 924 int orangefs_permission(struct user_namespace *mnt_userns,
 925                         struct inode *inode, int mask)
 926 {
 927         int ret;
 928
 929         if (mask & MAY_NOT_BLOCK)
 930                 return -ECHILD;
 931
 932         gossip_debug(GOSSIP_INODE_DEBUG, "%s: refreshing\n", __func__);
 933
 934         /* Make sure the permission (and other common attrs) are up to date. */
 935         ret = orangefs_inode_getattr(inode, 0);
 936         if (ret < 0)
 937                 return ret;
 938
 939         return generic_permission(&init_user_ns, inode, mask);
 940 }
 941
 942 int orangefs_update_time(struct inode *inode, struct timespec64 *time, int flags)
 943 {
 944         struct iattr iattr;
 945         gossip_debug(GOSSIP_INODE_DEBUG, "orangefs_update_time: %pU\n",
 946             get_khandle_from_ino(inode));
 947         generic_update_time(inode, time, flags);
 948         memset(&iattr, 0, sizeof iattr);
 949         if (flags & S_ATIME)
 950                 iattr.ia_valid |= ATTR_ATIME;
 951         if (flags & S_CTIME)
 952                 iattr.ia_valid |= ATTR_CTIME;
 953         if (flags & S_MTIME)
 954                 iattr.ia_valid |= ATTR_MTIME;
 955         return __orangefs_setattr(inode, &iattr);
 956 }
 957
 958 static int orangefs_fileattr_get(struct dentry *dentry, struct fileattr *fa)
 959 {
 960         u64 val = 0;
 961         int ret;
 962
 963         gossip_debug(GOSSIP_FILE_DEBUG, "%s: called on %pd\n", __func__,
 964                      dentry);
 965
 966         ret = orangefs_inode_getxattr(d_inode(dentry),
 967                                       "user.pvfs2.meta_hint",
 968                                       &val, sizeof(val));
 969         if (ret < 0 && ret != -ENODATA)
 970                 return ret;
 971
 972         gossip_debug(GOSSIP_FILE_DEBUG, "%s: flags=%u\n", __func__, (u32) val);
 973
 974         fileattr_fill_flags(fa, val);
 975         return 0;
 976 }
 977
 978 static int orangefs_fileattr_set(struct user_namespace *mnt_userns,
 979                                  struct dentry *dentry, struct fileattr *fa)
 980 {
 981         u64 val = 0;
 982
 983         gossip_debug(GOSSIP_FILE_DEBUG, "%s: called on %pd\n", __func__,
 984                      dentry);
 985         /*
 986          * ORANGEFS_MIRROR_FL is set internally when the mirroring mode is
 987          * turned on for a file. The user is not allowed to turn on this bit,
 988          * but the bit is present if the user first gets the flags and then
 989          * updates the flags with some new settings. So, we ignore it in the
 990          * following edit. bligon.
 991          */
 992         if (fileattr_has_fsx(fa) ||
 993             (fa->flags & ~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL | ORANGEFS_MIRROR_FL))) {
 994                 gossip_err("%s: only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n",
 995                            __func__);
 996                 return -EOPNOTSUPP;
 997         }
 998         val = fa->flags;
 999         gossip_debug(GOSSIP_FILE_DEBUG, "%s: flags=%u\n", __func__, (u32) val);
1000         return orangefs_inode_setxattr(d_inode(dentry),
1001                                        "user.pvfs2.meta_hint",
1002                                        &val, sizeof(val), 0);
1003 }
1004
1005 /* ORANGEFS2 implementation of VFS inode operations for files */
1006 static const struct inode_operations orangefs_file_inode_operations = {
1007         .get_acl = orangefs_get_acl,
1008         .set_acl = orangefs_set_acl,
1009         .setattr = orangefs_setattr,
1010         .getattr = orangefs_getattr,
1011         .listxattr = orangefs_listxattr,
1012         .permission = orangefs_permission,
1013         .update_time = orangefs_update_time,
1014         .fileattr_get = orangefs_fileattr_get,
1015         .fileattr_set = orangefs_fileattr_set,
1016 };
1017
1018 static int orangefs_init_iops(struct inode *inode)
1019 {
1020         inode->i_mapping->a_ops = &orangefs_address_operations;
1021
1022         switch (inode->i_mode & S_IFMT) {
1023         case S_IFREG:
1024                 inode->i_op = &orangefs_file_inode_operations;
1025                 inode->i_fop = &orangefs_file_operations;
1026                 break;
1027         case S_IFLNK:
1028                 inode->i_op = &orangefs_symlink_inode_operations;
1029                 break;
1030         case S_IFDIR:
1031                 inode->i_op = &orangefs_dir_inode_operations;
1032                 inode->i_fop = &orangefs_dir_operations;
1033                 break;
1034         default:
1035                 gossip_debug(GOSSIP_INODE_DEBUG,
1036                              "%s: unsupported mode\n",
1037                              __func__);
1038                 return -EINVAL;
1039         }
1040
1041         return 0;
1042 }
1043
1044 /*
1045  * Given an ORANGEFS object identifier (fsid, handle), convert it into
1046  * a ino_t type that will be used as a hash-index from where the handle will
1047  * be searched for in the VFS hash table of inodes.
1048  */
1049 static inline ino_t orangefs_handle_hash(struct orangefs_object_kref *ref)
1050 {
1051         if (!ref)
1052                 return 0;
1053         return orangefs_khandle_to_ino(&(ref->khandle));
1054 }
1055
1056 /*
1057  * Called to set up an inode from iget5_locked.
1058  */
1059 static int orangefs_set_inode(struct inode *inode, void *data)
1060 {
1061         struct orangefs_object_kref *ref = (struct orangefs_object_kref *) data;
1062         ORANGEFS_I(inode)->refn.fs_id = ref->fs_id;
1063         ORANGEFS_I(inode)->refn.khandle = ref->khandle;
1064         ORANGEFS_I(inode)->attr_valid = 0;
1065         hash_init(ORANGEFS_I(inode)->xattr_cache);
1066         ORANGEFS_I(inode)->mapping_time = jiffies - 1;
1067         ORANGEFS_I(inode)->bitlock = 0;
1068         return 0;
1069 }
1070
1071 /*
1072  * Called to determine if handles match.
1073  */
1074 static int orangefs_test_inode(struct inode *inode, void *data)
1075 {
1076         struct orangefs_object_kref *ref = (struct orangefs_object_kref *) data;
1077         struct orangefs_inode_s *orangefs_inode = NULL;
1078
1079         orangefs_inode = ORANGEFS_I(inode);
1080         /* test handles and fs_ids... */
1081         return (!ORANGEFS_khandle_cmp(&(orangefs_inode->refn.khandle),
1082                                 &(ref->khandle)) &&
1083                         orangefs_inode->refn.fs_id == ref->fs_id);
1084 }
1085
1086 /*
1087  * Front-end to lookup the inode-cache maintained by the VFS using the ORANGEFS
1088  * file handle.
1089  *
1090  * @sb: the file system super block instance.
1091  * @ref: The ORANGEFS object for which we are trying to locate an inode.
1092  */
1093 struct inode *orangefs_iget(struct super_block *sb,
1094                 struct orangefs_object_kref *ref)
1095 {
1096         struct inode *inode = NULL;
1097         unsigned long hash;
1098         int error;
1099
1100         hash = orangefs_handle_hash(ref);
1101         inode = iget5_locked(sb,
1102                         hash,
1103                         orangefs_test_inode,
1104                         orangefs_set_inode,
1105                         ref);
1106
1107         if (!inode)
1108                 return ERR_PTR(-ENOMEM);
1109
1110         if (!(inode->i_state & I_NEW))
1111                 return inode;
1112
1113         error = orangefs_inode_getattr(inode, ORANGEFS_GETATTR_NEW);
1114         if (error) {
1115                 iget_failed(inode);
1116                 return ERR_PTR(error);
1117         }
1118
1119         inode->i_ino = hash;    /* needed for stat etc */
1120         orangefs_init_iops(inode);
1121         unlock_new_inode(inode);
1122
1123         gossip_debug(GOSSIP_INODE_DEBUG,
1124                      "iget handle %pU, fsid %d hash %ld i_ino %lu\n",
1125                      &ref->khandle,
1126                      ref->fs_id,
1127                      hash,
1128                      inode->i_ino);
1129
1130         return inode;
1131 }
1132
1133 /*
1134  * Allocate an inode for a newly created file and insert it into the inode hash.
1135  */
1136 struct inode *orangefs_new_inode(struct super_block *sb, struct inode *dir,
1137                 int mode, dev_t dev, struct orangefs_object_kref *ref)
1138 {
1139         unsigned long hash = orangefs_handle_hash(ref);
1140         struct inode *inode;
1141         int error;
1142
1143         gossip_debug(GOSSIP_INODE_DEBUG,
1144                      "%s:(sb is %p | MAJOR(dev)=%u | MINOR(dev)=%u mode=%o)\n",
1145                      __func__,
1146                      sb,
1147                      MAJOR(dev),
1148                      MINOR(dev),
1149                      mode);
1150
1151         inode = new_inode(sb);
1152         if (!inode)
1153                 return ERR_PTR(-ENOMEM);
1154
1155         orangefs_set_inode(inode, ref);
1156         inode->i_ino = hash;    /* needed for stat etc */
1157
1158         error = orangefs_inode_getattr(inode, ORANGEFS_GETATTR_NEW);
1159         if (error)
1160                 goto out_iput;
1161
1162         orangefs_init_iops(inode);
1163         inode->i_rdev = dev;
1164
1165         error = insert_inode_locked4(inode, hash, orangefs_test_inode, ref);
1166         if (error < 0)
1167                 goto out_iput;
1168
1169         gossip_debug(GOSSIP_INODE_DEBUG,
1170                      "Initializing ACL's for inode %pU\n",
1171                      get_khandle_from_ino(inode));
1172         orangefs_init_acl(inode, dir);
1173         return inode;
1174
1175 out_iput:
1176         iput(inode);
1177         return ERR_PTR(error);
1178 }