fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /* dirty the head */
  91         spin_lock(&ci->i_ceph_lock);
  92         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  93         if (__ceph_have_pending_cap_snap(ci)) {
  94                 struct ceph_cap_snap *capsnap =
  95                                 list_last_entry(&ci->i_cap_snaps,
  96                                                 struct ceph_cap_snap,
  97                                                 ci_item);
  98                 snapc = ceph_get_snap_context(capsnap->context);
  99                 capsnap->dirty_pages++;
 100         } else {
 101                 BUG_ON(!ci->i_head_snapc);
 102                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 103                 ++ci->i_wrbuffer_ref_head;
 104         }
 105         if (ci->i_wrbuffer_ref == 0)
 106                 ihold(inode);
 107         ++ci->i_wrbuffer_ref;
 108         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 109              "snapc %p seq %lld (%d snaps)\n",
 110              mapping->host, page, page->index,
 111              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 112              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 113              snapc, snapc->seq, snapc->num_snaps);
 114         spin_unlock(&ci->i_ceph_lock);
 115
 116         /*
 117          * Reference snap context in page->private.  Also set
 118          * PagePrivate so that we get invalidatepage callback.
 119          */
 120         BUG_ON(PagePrivate(page));
 121         page->private = (unsigned long)snapc;
 122         SetPagePrivate(page);
 123
 124         ret = __set_page_dirty_nobuffers(page);
 125         WARN_ON(!PageLocked(page));
 126         WARN_ON(!page->mapping);
 127
 128         return ret;
 129 }
 130
 131 /*
 132  * If we are truncating the full page (i.e. offset == 0), adjust the
 133  * dirty page counters appropriately.  Only called if there is private
 134  * data on the page.
 135  */
 136 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 137                                 unsigned int length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc = page_snap_context(page);
 142
 143         inode = page->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != PAGE_SIZE) {
 147                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 148                      inode, page, page->index, offset, length);
 149                 return;
 150         }
 151
 152         ceph_invalidate_fscache_page(inode, page);
 153
 154         if (!PagePrivate(page))
 155                 return;
 156
 157         /*
 158          * We can get non-dirty pages here due to races between
 159          * set_page_dirty and truncate_complete_page; just spit out a
 160          * warning, in case we end up with accounting problems later.
 161          */
 162         if (!PageDirty(page))
 163                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 164
 165         ClearPageChecked(page);
 166
 167         dout("%p invalidatepage %p idx %lu full dirty page\n",
 168              inode, page, page->index);
 169
 170         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 171         ceph_put_snap_context(snapc);
 172         page->private = 0;
 173         ClearPagePrivate(page);
 174 }
 175
 176 static int ceph_releasepage(struct page *page, gfp_t g)
 177 {
 178         dout("%p releasepage %p idx %lu\n", page->mapping->host,
 179              page, page->index);
 180         WARN_ON(PageDirty(page));
 181
 182         /* Can we release the page from the cache? */
 183         if (!ceph_release_fscache_page(page, g))
 184                 return 0;
 185
 186         return !PagePrivate(page);
 187 }
 188
 189 /*
 190  * read a single page, without unlocking it.
 191  */
 192 static int readpage_nounlock(struct file *filp, struct page *page)
 193 {
 194         struct inode *inode = file_inode(filp);
 195         struct ceph_inode_info *ci = ceph_inode(inode);
 196         struct ceph_osd_client *osdc =
 197                 &ceph_inode_to_client(inode)->client->osdc;
 198         int err = 0;
 199         u64 off = page_offset(page);
 200         u64 len = PAGE_SIZE;
 201
 202         if (off >= i_size_read(inode)) {
 203                 zero_user_segment(page, 0, PAGE_SIZE);
 204                 SetPageUptodate(page);
 205                 return 0;
 206         }
 207
 208         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 209                 /*
 210                  * Uptodate inline data should have been added
 211                  * into page cache while getting Fcr caps.
 212                  */
 213                 if (off == 0)
 214                         return -EINVAL;
 215                 zero_user_segment(page, 0, PAGE_SIZE);
 216                 SetPageUptodate(page);
 217                 return 0;
 218         }
 219
 220         err = ceph_readpage_from_fscache(inode, page);
 221         if (err == 0)
 222                 goto out;
 223
 224         dout("readpage inode %p file %p page %p index %lu\n",
 225              inode, filp, page, page->index);
 226         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 227                                   off, &len,
 228                                   ci->i_truncate_seq, ci->i_truncate_size,
 229                                   &page, 1, 0);
 230         if (err == -ENOENT)
 231                 err = 0;
 232         if (err < 0) {
 233                 SetPageError(page);
 234                 ceph_fscache_readpage_cancel(inode, page);
 235                 goto out;
 236         }
 237         if (err < PAGE_SIZE)
 238                 /* zero fill remainder of page */
 239                 zero_user_segment(page, err, PAGE_SIZE);
 240         else
 241                 flush_dcache_page(page);
 242
 243         SetPageUptodate(page);
 244         ceph_readpage_to_fscache(inode, page);
 245
 246 out:
 247         return err < 0 ? err : 0;
 248 }
 249
 250 static int ceph_readpage(struct file *filp, struct page *page)
 251 {
 252         int r = readpage_nounlock(filp, page);
 253         unlock_page(page);
 254         return r;
 255 }
 256
 257 /*
 258  * Finish an async read(ahead) op.
 259  */
 260 static void finish_read(struct ceph_osd_request *req)
 261 {
 262         struct inode *inode = req->r_inode;
 263         struct ceph_osd_data *osd_data;
 264         int rc = req->r_result <= 0 ? req->r_result : 0;
 265         int bytes = req->r_result >= 0 ? req->r_result : 0;
 266         int num_pages;
 267         int i;
 268
 269         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 270
 271         /* unlock all pages, zeroing any data we didn't read */
 272         osd_data = osd_req_op_extent_osd_data(req, 0);
 273         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 274         num_pages = calc_pages_for((u64)osd_data->alignment,
 275                                         (u64)osd_data->length);
 276         for (i = 0; i < num_pages; i++) {
 277                 struct page *page = osd_data->pages[i];
 278
 279                 if (rc < 0 && rc != -ENOENT) {
 280                         ceph_fscache_readpage_cancel(inode, page);
 281                         goto unlock;
 282                 }
 283                 if (bytes < (int)PAGE_SIZE) {
 284                         /* zero (remainder of) page */
 285                         int s = bytes < 0 ? 0 : bytes;
 286                         zero_user_segment(page, s, PAGE_SIZE);
 287                 }
 288                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 289                      page->index);
 290                 flush_dcache_page(page);
 291                 SetPageUptodate(page);
 292                 ceph_readpage_to_fscache(inode, page);
 293 unlock:
 294                 unlock_page(page);
 295                 put_page(page);
 296                 bytes -= PAGE_SIZE;
 297         }
 298         kfree(osd_data->pages);
 299 }
 300
 301 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 302 {
 303         int i;
 304
 305         for (i = 0; i < num_pages; i++)
 306                 unlock_page(pages[i]);
 307 }
 308
 309 /*
 310  * start an async read(ahead) operation.  return nr_pages we submitted
 311  * a read for on success, or negative error code.
 312  */
 313 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 314 {
 315         struct ceph_osd_client *osdc =
 316                 &ceph_inode_to_client(inode)->client->osdc;
 317         struct ceph_inode_info *ci = ceph_inode(inode);
 318         struct page *page = list_entry(page_list->prev, struct page, lru);
 319         struct ceph_vino vino;
 320         struct ceph_osd_request *req;
 321         u64 off;
 322         u64 len;
 323         int i;
 324         struct page **pages;
 325         pgoff_t next_index;
 326         int nr_pages = 0;
 327         int ret;
 328
 329         off = (u64) page_offset(page);
 330
 331         /* count pages */
 332         next_index = page->index;
 333         list_for_each_entry_reverse(page, page_list, lru) {
 334                 if (page->index != next_index)
 335                         break;
 336                 nr_pages++;
 337                 next_index++;
 338                 if (max && nr_pages == max)
 339                         break;
 340         }
 341         len = nr_pages << PAGE_SHIFT;
 342         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 343              off, len);
 344         vino = ceph_vino(inode);
 345         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 346                                     0, 1, CEPH_OSD_OP_READ,
 347                                     CEPH_OSD_FLAG_READ, NULL,
 348                                     ci->i_truncate_seq, ci->i_truncate_size,
 349                                     false);
 350         if (IS_ERR(req))
 351                 return PTR_ERR(req);
 352
 353         /* build page vector */
 354         nr_pages = calc_pages_for(0, len);
 355         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 356         ret = -ENOMEM;
 357         if (!pages)
 358                 goto out;
 359         for (i = 0; i < nr_pages; ++i) {
 360                 page = list_entry(page_list->prev, struct page, lru);
 361                 BUG_ON(PageLocked(page));
 362                 list_del(&page->lru);
 363
 364                 dout("start_read %p adding %p idx %lu\n", inode, page,
 365                      page->index);
 366                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 367                                           GFP_KERNEL)) {
 368                         ceph_fscache_uncache_page(inode, page);
 369                         put_page(page);
 370                         dout("start_read %p add_to_page_cache failed %p\n",
 371                              inode, page);
 372                         nr_pages = i;
 373                         goto out_pages;
 374                 }
 375                 pages[i] = page;
 376         }
 377         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 378         req->r_callback = finish_read;
 379         req->r_inode = inode;
 380
 381         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 382         ret = ceph_osdc_start_request(osdc, req, false);
 383         if (ret < 0)
 384                 goto out_pages;
 385         ceph_osdc_put_request(req);
 386         return nr_pages;
 387
 388 out_pages:
 389         ceph_unlock_page_vector(pages, nr_pages);
 390         ceph_release_page_vector(pages, nr_pages);
 391 out:
 392         ceph_osdc_put_request(req);
 393         return ret;
 394 }
 395
 396
 397 /*
 398  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 399  * the caller (VM) cleans them up.
 400  */
 401 static int ceph_readpages(struct file *file, struct address_space *mapping,
 402                           struct list_head *page_list, unsigned nr_pages)
 403 {
 404         struct inode *inode = file_inode(file);
 405         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 406         int rc = 0;
 407         int max = 0;
 408
 409         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 410                 return -EINVAL;
 411
 412         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 413                                          &nr_pages);
 414
 415         if (rc == 0)
 416                 goto out;
 417
 418         if (fsc->mount_options->rsize >= PAGE_SIZE)
 419                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 420                         >> PAGE_SHIFT;
 421
 422         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 423                 file, nr_pages,
 424              max);
 425         while (!list_empty(page_list)) {
 426                 rc = start_read(inode, page_list, max);
 427                 if (rc < 0)
 428                         goto out;
 429                 BUG_ON(rc == 0);
 430         }
 431 out:
 432         ceph_fscache_readpages_cancel(inode, page_list);
 433
 434         dout("readpages %p file %p ret %d\n", inode, file, rc);
 435         return rc;
 436 }
 437
 438 /*
 439  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 440  * only snap context we are allowed to write back.
 441  */
 442 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 443                                                     loff_t *snap_size)
 444 {
 445         struct ceph_inode_info *ci = ceph_inode(inode);
 446         struct ceph_snap_context *snapc = NULL;
 447         struct ceph_cap_snap *capsnap = NULL;
 448
 449         spin_lock(&ci->i_ceph_lock);
 450         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 451                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 452                      capsnap->context, capsnap->dirty_pages);
 453                 if (capsnap->dirty_pages) {
 454                         snapc = ceph_get_snap_context(capsnap->context);
 455                         if (snap_size)
 456                                 *snap_size = capsnap->size;
 457                         break;
 458                 }
 459         }
 460         if (!snapc && ci->i_wrbuffer_ref_head) {
 461                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 462                 dout(" head snapc %p has %d dirty pages\n",
 463                      snapc, ci->i_wrbuffer_ref_head);
 464         }
 465         spin_unlock(&ci->i_ceph_lock);
 466         return snapc;
 467 }
 468
 469 /*
 470  * Write a single page, but leave the page locked.
 471  *
 472  * If we get a write error, set the page error bit, but still adjust the
 473  * dirty page accounting (i.e., page is no longer dirty).
 474  */
 475 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 476 {
 477         struct inode *inode;
 478         struct ceph_inode_info *ci;
 479         struct ceph_fs_client *fsc;
 480         struct ceph_osd_client *osdc;
 481         struct ceph_snap_context *snapc, *oldest;
 482         loff_t page_off = page_offset(page);
 483         loff_t snap_size = -1;
 484         long writeback_stat;
 485         u64 truncate_size;
 486         u32 truncate_seq;
 487         int err = 0, len = PAGE_SIZE;
 488
 489         dout("writepage %p idx %lu\n", page, page->index);
 490
 491         if (!page->mapping || !page->mapping->host) {
 492                 dout("writepage %p - no mapping\n", page);
 493                 return -EFAULT;
 494         }
 495         inode = page->mapping->host;
 496         ci = ceph_inode(inode);
 497         fsc = ceph_inode_to_client(inode);
 498         osdc = &fsc->client->osdc;
 499
 500         /* verify this is a writeable snap context */
 501         snapc = page_snap_context(page);
 502         if (snapc == NULL) {
 503                 dout("writepage %p page %p not dirty?\n", inode, page);
 504                 goto out;
 505         }
 506         oldest = get_oldest_context(inode, &snap_size);
 507         if (snapc->seq > oldest->seq) {
 508                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 509                      inode, page, snapc);
 510                 /* we should only noop if called by kswapd */
 511                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 512                 ceph_put_snap_context(oldest);
 513                 goto out;
 514         }
 515         ceph_put_snap_context(oldest);
 516
 517         spin_lock(&ci->i_ceph_lock);
 518         truncate_seq = ci->i_truncate_seq;
 519         truncate_size = ci->i_truncate_size;
 520         if (snap_size == -1)
 521                 snap_size = i_size_read(inode);
 522         spin_unlock(&ci->i_ceph_lock);
 523
 524         /* is this a partial page at end of file? */
 525         if (page_off >= snap_size) {
 526                 dout("%p page eof %llu\n", page, snap_size);
 527                 goto out;
 528         }
 529         if (snap_size < page_off + len)
 530                 len = snap_size - page_off;
 531
 532         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 533              inode, page, page->index, page_off, len, snapc);
 534
 535         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 536         if (writeback_stat >
 537             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 538                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 539
 540         ceph_readpage_to_fscache(inode, page);
 541
 542         set_page_writeback(page);
 543         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 544                                    &ci->i_layout, snapc,
 545                                    page_off, len,
 546                                    truncate_seq, truncate_size,
 547                                    &inode->i_mtime, &page, 1);
 548         if (err < 0) {
 549                 struct writeback_control tmp_wbc;
 550                 if (!wbc)
 551                         wbc = &tmp_wbc;
 552                 if (err == -ERESTARTSYS) {
 553                         /* killed by SIGKILL */
 554                         dout("writepage interrupted page %p\n", page);
 555                         redirty_page_for_writepage(wbc, page);
 556                         end_page_writeback(page);
 557                         goto out;
 558                 }
 559                 dout("writepage setting page/mapping error %d %p\n",
 560                      err, page);
 561                 SetPageError(page);
 562                 mapping_set_error(&inode->i_data, err);
 563                 wbc->pages_skipped++;
 564         } else {
 565                 dout("writepage cleaned page %p\n", page);
 566                 err = 0;  /* vfs expects us to return 0 */
 567         }
 568         page->private = 0;
 569         ClearPagePrivate(page);
 570         end_page_writeback(page);
 571         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 572         ceph_put_snap_context(snapc);  /* page's reference */
 573 out:
 574         return err;
 575 }
 576
 577 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 578 {
 579         int err;
 580         struct inode *inode = page->mapping->host;
 581         BUG_ON(!inode);
 582         ihold(inode);
 583         err = writepage_nounlock(page, wbc);
 584         if (err == -ERESTARTSYS) {
 585                 /* direct memory reclaimer was killed by SIGKILL. return 0
 586                  * to prevent caller from setting mapping/page error */
 587                 err = 0;
 588         }
 589         unlock_page(page);
 590         iput(inode);
 591         return err;
 592 }
 593
 594 /*
 595  * lame release_pages helper.  release_pages() isn't exported to
 596  * modules.
 597  */
 598 static void ceph_release_pages(struct page **pages, int num)
 599 {
 600         struct pagevec pvec;
 601         int i;
 602
 603         pagevec_init(&pvec, 0);
 604         for (i = 0; i < num; i++) {
 605                 if (pagevec_add(&pvec, pages[i]) == 0)
 606                         pagevec_release(&pvec);
 607         }
 608         pagevec_release(&pvec);
 609 }
 610
 611 /*
 612  * async writeback completion handler.
 613  *
 614  * If we get an error, set the mapping error bit, but not the individual
 615  * page error bits.
 616  */
 617 static void writepages_finish(struct ceph_osd_request *req)
 618 {
 619         struct inode *inode = req->r_inode;
 620         struct ceph_inode_info *ci = ceph_inode(inode);
 621         struct ceph_osd_data *osd_data;
 622         struct page *page;
 623         int num_pages, total_pages = 0;
 624         int i, j;
 625         int rc = req->r_result;
 626         struct ceph_snap_context *snapc = req->r_snapc;
 627         struct address_space *mapping = inode->i_mapping;
 628         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 629         bool remove_page;
 630
 631         dout("writepages_finish %p rc %d\n", inode, rc);
 632         if (rc < 0)
 633                 mapping_set_error(mapping, rc);
 634
 635         /*
 636          * We lost the cache cap, need to truncate the page before
 637          * it is unlocked, otherwise we'd truncate it later in the
 638          * page truncation thread, possibly losing some data that
 639          * raced its way in
 640          */
 641         remove_page = !(ceph_caps_issued(ci) &
 642                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 643
 644         /* clean all pages */
 645         for (i = 0; i < req->r_num_ops; i++) {
 646                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 647                         break;
 648
 649                 osd_data = osd_req_op_extent_osd_data(req, i);
 650                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 651                 num_pages = calc_pages_for((u64)osd_data->alignment,
 652                                            (u64)osd_data->length);
 653                 total_pages += num_pages;
 654                 for (j = 0; j < num_pages; j++) {
 655                         page = osd_data->pages[j];
 656                         BUG_ON(!page);
 657                         WARN_ON(!PageUptodate(page));
 658
 659                         if (atomic_long_dec_return(&fsc->writeback_count) <
 660                              CONGESTION_OFF_THRESH(
 661                                         fsc->mount_options->congestion_kb))
 662                                 clear_bdi_congested(&fsc->backing_dev_info,
 663                                                     BLK_RW_ASYNC);
 664
 665                         if (rc < 0)
 666                                 SetPageError(page);
 667
 668                         ceph_put_snap_context(page_snap_context(page));
 669                         page->private = 0;
 670                         ClearPagePrivate(page);
 671                         dout("unlocking %p\n", page);
 672                         end_page_writeback(page);
 673
 674                         if (remove_page)
 675                                 generic_error_remove_page(inode->i_mapping,
 676                                                           page);
 677
 678                         unlock_page(page);
 679                 }
 680                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 681                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 682
 683                 ceph_release_pages(osd_data->pages, num_pages);
 684         }
 685
 686         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 687
 688         osd_data = osd_req_op_extent_osd_data(req, 0);
 689         if (osd_data->pages_from_pool)
 690                 mempool_free(osd_data->pages,
 691                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 692         else
 693                 kfree(osd_data->pages);
 694         ceph_osdc_put_request(req);
 695 }
 696
 697 /*
 698  * initiate async writeback
 699  */
 700 static int ceph_writepages_start(struct address_space *mapping,
 701                                  struct writeback_control *wbc)
 702 {
 703         struct inode *inode = mapping->host;
 704         struct ceph_inode_info *ci = ceph_inode(inode);
 705         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 706         struct ceph_vino vino = ceph_vino(inode);
 707         pgoff_t index, start, end;
 708         int range_whole = 0;
 709         int should_loop = 1;
 710         pgoff_t max_pages = 0, max_pages_ever = 0;
 711         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 712         struct pagevec pvec;
 713         int done = 0;
 714         int rc = 0;
 715         unsigned wsize = 1 << inode->i_blkbits;
 716         struct ceph_osd_request *req = NULL;
 717         int do_sync = 0;
 718         loff_t snap_size, i_size;
 719         u64 truncate_size;
 720         u32 truncate_seq;
 721
 722         /*
 723          * Include a 'sync' in the OSD request if this is a data
 724          * integrity write (e.g., O_SYNC write or fsync()), or if our
 725          * cap is being revoked.
 726          */
 727         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 728                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 729                 do_sync = 1;
 730         dout("writepages_start %p dosync=%d (mode=%s)\n",
 731              inode, do_sync,
 732              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 733              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 734
 735         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 736                 if (ci->i_wrbuffer_ref > 0) {
 737                         pr_warn_ratelimited(
 738                                 "writepage_start %p %lld forced umount\n",
 739                                 inode, ceph_ino(inode));
 740                 }
 741                 mapping_set_error(mapping, -EIO);
 742                 return -EIO; /* we're in a forced umount, don't write! */
 743         }
 744         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 745                 wsize = fsc->mount_options->wsize;
 746         if (wsize < PAGE_SIZE)
 747                 wsize = PAGE_SIZE;
 748         max_pages_ever = wsize >> PAGE_SHIFT;
 749
 750         pagevec_init(&pvec, 0);
 751
 752         /* where to start/end? */
 753         if (wbc->range_cyclic) {
 754                 start = mapping->writeback_index; /* Start from prev offset */
 755                 end = -1;
 756                 dout(" cyclic, start at %lu\n", start);
 757         } else {
 758                 start = wbc->range_start >> PAGE_SHIFT;
 759                 end = wbc->range_end >> PAGE_SHIFT;
 760                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 761                         range_whole = 1;
 762                 should_loop = 0;
 763                 dout(" not cyclic, %lu to %lu\n", start, end);
 764         }
 765         index = start;
 766
 767 retry:
 768         /* find oldest snap context with dirty data */
 769         ceph_put_snap_context(snapc);
 770         snap_size = -1;
 771         snapc = get_oldest_context(inode, &snap_size);
 772         if (!snapc) {
 773                 /* hmm, why does writepages get called when there
 774                    is no dirty data? */
 775                 dout(" no snap context with dirty data?\n");
 776                 goto out;
 777         }
 778         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 779              snapc, snapc->seq, snapc->num_snaps);
 780
 781         spin_lock(&ci->i_ceph_lock);
 782         truncate_seq = ci->i_truncate_seq;
 783         truncate_size = ci->i_truncate_size;
 784         i_size = i_size_read(inode);
 785         spin_unlock(&ci->i_ceph_lock);
 786
 787         if (last_snapc && snapc != last_snapc) {
 788                 /* if we switched to a newer snapc, restart our scan at the
 789                  * start of the original file range. */
 790                 dout("  snapc differs from last pass, restarting at %lu\n",
 791                      index);
 792                 index = start;
 793         }
 794         last_snapc = snapc;
 795
 796         while (!done && index <= end) {
 797                 unsigned i;
 798                 int first;
 799                 pgoff_t strip_unit_end = 0;
 800                 int num_ops = 0, op_idx;
 801                 int pvec_pages, locked_pages = 0;
 802                 struct page **pages = NULL, **data_pages;
 803                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 804                 struct page *page;
 805                 int want;
 806                 u64 offset = 0, len = 0;
 807
 808                 max_pages = max_pages_ever;
 809
 810 get_more_pages:
 811                 first = -1;
 812                 want = min(end - index,
 813                            min((pgoff_t)PAGEVEC_SIZE,
 814                                max_pages - (pgoff_t)locked_pages) - 1)
 815                         + 1;
 816                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 817                                                 PAGECACHE_TAG_DIRTY,
 818                                                 want);
 819                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 820                 if (!pvec_pages && !locked_pages)
 821                         break;
 822                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 823                         page = pvec.pages[i];
 824                         dout("? %p idx %lu\n", page, page->index);
 825                         if (locked_pages == 0)
 826                                 lock_page(page);  /* first page */
 827                         else if (!trylock_page(page))
 828                                 break;
 829
 830                         /* only dirty pages, or our accounting breaks */
 831                         if (unlikely(!PageDirty(page)) ||
 832                             unlikely(page->mapping != mapping)) {
 833                                 dout("!dirty or !mapping %p\n", page);
 834                                 unlock_page(page);
 835                                 break;
 836                         }
 837                         if (!wbc->range_cyclic && page->index > end) {
 838                                 dout("end of range %p\n", page);
 839                                 done = 1;
 840                                 unlock_page(page);
 841                                 break;
 842                         }
 843                         if (strip_unit_end && (page->index > strip_unit_end)) {
 844                                 dout("end of strip unit %p\n", page);
 845                                 unlock_page(page);
 846                                 break;
 847                         }
 848                         if (wbc->sync_mode != WB_SYNC_NONE) {
 849                                 dout("waiting on writeback %p\n", page);
 850                                 wait_on_page_writeback(page);
 851                         }
 852                         if (page_offset(page) >=
 853                             (snap_size == -1 ? i_size : snap_size)) {
 854                                 dout("%p page eof %llu\n", page,
 855                                      (snap_size == -1 ? i_size : snap_size));
 856                                 done = 1;
 857                                 unlock_page(page);
 858                                 break;
 859                         }
 860                         if (PageWriteback(page)) {
 861                                 dout("%p under writeback\n", page);
 862                                 unlock_page(page);
 863                                 break;
 864                         }
 865
 866                         /* only if matching snap context */
 867                         pgsnapc = page_snap_context(page);
 868                         if (pgsnapc->seq > snapc->seq) {
 869                                 dout("page snapc %p %lld > oldest %p %lld\n",
 870                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 871                                 unlock_page(page);
 872                                 if (!locked_pages)
 873                                         continue; /* keep looking for snap */
 874                                 break;
 875                         }
 876
 877                         if (!clear_page_dirty_for_io(page)) {
 878                                 dout("%p !clear_page_dirty_for_io\n", page);
 879                                 unlock_page(page);
 880                                 break;
 881                         }
 882
 883                         /*
 884                          * We have something to write.  If this is
 885                          * the first locked page this time through,
 886                          * calculate max possinle write size and
 887                          * allocate a page array
 888                          */
 889                         if (locked_pages == 0) {
 890                                 u64 objnum;
 891                                 u64 objoff;
 892
 893                                 /* prepare async write request */
 894                                 offset = (u64)page_offset(page);
 895                                 len = wsize;
 896
 897                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 898                                                                 offset, len,
 899                                                                 &objnum, &objoff,
 900                                                                 &len);
 901                                 if (rc < 0) {
 902                                         unlock_page(page);
 903                                         break;
 904                                 }
 905
 906                                 num_ops = 1 + do_sync;
 907                                 strip_unit_end = page->index +
 908                                         ((len - 1) >> PAGE_SHIFT);
 909
 910                                 BUG_ON(pages);
 911                                 max_pages = calc_pages_for(0, (u64)len);
 912                                 pages = kmalloc(max_pages * sizeof (*pages),
 913                                                 GFP_NOFS);
 914                                 if (!pages) {
 915                                         pool = fsc->wb_pagevec_pool;
 916                                         pages = mempool_alloc(pool, GFP_NOFS);
 917                                         BUG_ON(!pages);
 918                                 }
 919
 920                                 len = 0;
 921                         } else if (page->index !=
 922                                    (offset + len) >> PAGE_SHIFT) {
 923                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 924                                                         CEPH_OSD_MAX_OPS)) {
 925                                         redirty_page_for_writepage(wbc, page);
 926                                         unlock_page(page);
 927                                         break;
 928                                 }
 929
 930                                 num_ops++;
 931                                 offset = (u64)page_offset(page);
 932                                 len = 0;
 933                         }
 934
 935                         /* note position of first page in pvec */
 936                         if (first < 0)
 937                                 first = i;
 938                         dout("%p will write page %p idx %lu\n",
 939                              inode, page, page->index);
 940
 941                         if (atomic_long_inc_return(&fsc->writeback_count) >
 942                             CONGESTION_ON_THRESH(
 943                                     fsc->mount_options->congestion_kb)) {
 944                                 set_bdi_congested(&fsc->backing_dev_info,
 945                                                   BLK_RW_ASYNC);
 946                         }
 947
 948                         pages[locked_pages] = page;
 949                         locked_pages++;
 950                         len += PAGE_SIZE;
 951                 }
 952
 953                 /* did we get anything? */
 954                 if (!locked_pages)
 955                         goto release_pvec_pages;
 956                 if (i) {
 957                         int j;
 958                         BUG_ON(!locked_pages || first < 0);
 959
 960                         if (pvec_pages && i == pvec_pages &&
 961                             locked_pages < max_pages) {
 962                                 dout("reached end pvec, trying for more\n");
 963                                 pagevec_reinit(&pvec);
 964                                 goto get_more_pages;
 965                         }
 966
 967                         /* shift unused pages over in the pvec...  we
 968                          * will need to release them below. */
 969                         for (j = i; j < pvec_pages; j++) {
 970                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
 971                                 pvec.pages[j-i+first] = pvec.pages[j];
 972                         }
 973                         pvec.nr -= i-first;
 974                 }
 975
 976 new_request:
 977                 offset = page_offset(pages[0]);
 978                 len = wsize;
 979
 980                 req = ceph_osdc_new_request(&fsc->client->osdc,
 981                                         &ci->i_layout, vino,
 982                                         offset, &len, 0, num_ops,
 983                                         CEPH_OSD_OP_WRITE,
 984                                         CEPH_OSD_FLAG_WRITE |
 985                                         CEPH_OSD_FLAG_ONDISK,
 986                                         snapc, truncate_seq,
 987                                         truncate_size, false);
 988                 if (IS_ERR(req)) {
 989                         req = ceph_osdc_new_request(&fsc->client->osdc,
 990                                                 &ci->i_layout, vino,
 991                                                 offset, &len, 0,
 992                                                 min(num_ops,
 993                                                     CEPH_OSD_SLAB_OPS),
 994                                                 CEPH_OSD_OP_WRITE,
 995                                                 CEPH_OSD_FLAG_WRITE |
 996                                                 CEPH_OSD_FLAG_ONDISK,
 997                                                 snapc, truncate_seq,
 998                                                 truncate_size, true);
 999                         BUG_ON(IS_ERR(req));
1000                 }
1001                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1002                              PAGE_SIZE - offset);
1003
1004                 req->r_callback = writepages_finish;
1005                 req->r_inode = inode;
1006
1007                 /* Format the osd request message and submit the write */
1008                 len = 0;
1009                 data_pages = pages;
1010                 op_idx = 0;
1011                 for (i = 0; i < locked_pages; i++) {
1012                         u64 cur_offset = page_offset(pages[i]);
1013                         if (offset + len != cur_offset) {
1014                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1015                                         break;
1016                                 osd_req_op_extent_dup_last(req, op_idx,
1017                                                            cur_offset - offset);
1018                                 dout("writepages got pages at %llu~%llu\n",
1019                                      offset, len);
1020                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1021                                                         data_pages, len, 0,
1022                                                         !!pool, false);
1023                                 osd_req_op_extent_update(req, op_idx, len);
1024
1025                                 len = 0;
1026                                 offset = cur_offset;
1027                                 data_pages = pages + i;
1028                                 op_idx++;
1029                         }
1030
1031                         set_page_writeback(pages[i]);
1032                         len += PAGE_SIZE;
1033                 }
1034
1035                 if (snap_size != -1) {
1036                         len = min(len, snap_size - offset);
1037                 } else if (i == locked_pages) {
1038                         /* writepages_finish() clears writeback pages
1039                          * according to the data length, so make sure
1040                          * data length covers all locked pages */
1041                         u64 min_len = len + 1 - PAGE_SIZE;
1042                         len = min(len, (u64)i_size_read(inode) - offset);
1043                         len = max(len, min_len);
1044                 }
1045                 dout("writepages got pages at %llu~%llu\n", offset, len);
1046
1047                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1048                                                  0, !!pool, false);
1049                 osd_req_op_extent_update(req, op_idx, len);
1050
1051                 if (do_sync) {
1052                         op_idx++;
1053                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1054                 }
1055                 BUG_ON(op_idx + 1 != req->r_num_ops);
1056
1057                 pool = NULL;
1058                 if (i < locked_pages) {
1059                         BUG_ON(num_ops <= req->r_num_ops);
1060                         num_ops -= req->r_num_ops;
1061                         num_ops += do_sync;
1062                         locked_pages -= i;
1063
1064                         /* allocate new pages array for next request */
1065                         data_pages = pages;
1066                         pages = kmalloc(locked_pages * sizeof (*pages),
1067                                         GFP_NOFS);
1068                         if (!pages) {
1069                                 pool = fsc->wb_pagevec_pool;
1070                                 pages = mempool_alloc(pool, GFP_NOFS);
1071                                 BUG_ON(!pages);
1072                         }
1073                         memcpy(pages, data_pages + i,
1074                                locked_pages * sizeof(*pages));
1075                         memset(data_pages + i, 0,
1076                                locked_pages * sizeof(*pages));
1077                 } else {
1078                         BUG_ON(num_ops != req->r_num_ops);
1079                         index = pages[i - 1]->index + 1;
1080                         /* request message now owns the pages array */
1081                         pages = NULL;
1082                 }
1083
1084                 req->r_mtime = inode->i_mtime;
1085                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1086                 BUG_ON(rc);
1087                 req = NULL;
1088
1089                 wbc->nr_to_write -= i;
1090                 if (pages)
1091                         goto new_request;
1092
1093                 if (wbc->nr_to_write <= 0)
1094                         done = 1;
1095
1096 release_pvec_pages:
1097                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1098                      pvec.nr ? pvec.pages[0] : NULL);
1099                 pagevec_release(&pvec);
1100
1101                 if (locked_pages && !done)
1102                         goto retry;
1103         }
1104
1105         if (should_loop && !done) {
1106                 /* more to do; loop back to beginning of file */
1107                 dout("writepages looping back to beginning of file\n");
1108                 should_loop = 0;
1109                 index = 0;
1110                 goto retry;
1111         }
1112
1113         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1114                 mapping->writeback_index = index;
1115
1116 out:
1117         ceph_osdc_put_request(req);
1118         ceph_put_snap_context(snapc);
1119         dout("writepages done, rc = %d\n", rc);
1120         return rc;
1121 }
1122
1123
1124
1125 /*
1126  * See if a given @snapc is either writeable, or already written.
1127  */
1128 static int context_is_writeable_or_written(struct inode *inode,
1129                                            struct ceph_snap_context *snapc)
1130 {
1131         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1132         int ret = !oldest || snapc->seq <= oldest->seq;
1133
1134         ceph_put_snap_context(oldest);
1135         return ret;
1136 }
1137
1138 /*
1139  * We are only allowed to write into/dirty the page if the page is
1140  * clean, or already dirty within the same snap context.
1141  *
1142  * called with page locked.
1143  * return success with page locked,
1144  * or any failure (incl -EAGAIN) with page unlocked.
1145  */
1146 static int ceph_update_writeable_page(struct file *file,
1147                             loff_t pos, unsigned len,
1148                             struct page *page)
1149 {
1150         struct inode *inode = file_inode(file);
1151         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1152         struct ceph_inode_info *ci = ceph_inode(inode);
1153         loff_t page_off = pos & PAGE_MASK;
1154         int pos_in_page = pos & ~PAGE_MASK;
1155         int end_in_page = pos_in_page + len;
1156         loff_t i_size;
1157         int r;
1158         struct ceph_snap_context *snapc, *oldest;
1159
1160         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1161                 dout(" page %p forced umount\n", page);
1162                 unlock_page(page);
1163                 return -EIO;
1164         }
1165
1166 retry_locked:
1167         /* writepages currently holds page lock, but if we change that later, */
1168         wait_on_page_writeback(page);
1169
1170         snapc = page_snap_context(page);
1171         if (snapc && snapc != ci->i_head_snapc) {
1172                 /*
1173                  * this page is already dirty in another (older) snap
1174                  * context!  is it writeable now?
1175                  */
1176                 oldest = get_oldest_context(inode, NULL);
1177
1178                 if (snapc->seq > oldest->seq) {
1179                         ceph_put_snap_context(oldest);
1180                         dout(" page %p snapc %p not current or oldest\n",
1181                              page, snapc);
1182                         /*
1183                          * queue for writeback, and wait for snapc to
1184                          * be writeable or written
1185                          */
1186                         snapc = ceph_get_snap_context(snapc);
1187                         unlock_page(page);
1188                         ceph_queue_writeback(inode);
1189                         r = wait_event_killable(ci->i_cap_wq,
1190                                context_is_writeable_or_written(inode, snapc));
1191                         ceph_put_snap_context(snapc);
1192                         if (r == -ERESTARTSYS)
1193                                 return r;
1194                         return -EAGAIN;
1195                 }
1196                 ceph_put_snap_context(oldest);
1197
1198                 /* yay, writeable, do it now (without dropping page lock) */
1199                 dout(" page %p snapc %p not current, but oldest\n",
1200                      page, snapc);
1201                 if (!clear_page_dirty_for_io(page))
1202                         goto retry_locked;
1203                 r = writepage_nounlock(page, NULL);
1204                 if (r < 0)
1205                         goto fail_nosnap;
1206                 goto retry_locked;
1207         }
1208
1209         if (PageUptodate(page)) {
1210                 dout(" page %p already uptodate\n", page);
1211                 return 0;
1212         }
1213
1214         /* full page? */
1215         if (pos_in_page == 0 && len == PAGE_SIZE)
1216                 return 0;
1217
1218         /* past end of file? */
1219         i_size = i_size_read(inode);
1220
1221         if (page_off >= i_size ||
1222             (pos_in_page == 0 && (pos+len) >= i_size &&
1223              end_in_page - pos_in_page != PAGE_SIZE)) {
1224                 dout(" zeroing %p 0 - %d and %d - %d\n",
1225                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1226                 zero_user_segments(page,
1227                                    0, pos_in_page,
1228                                    end_in_page, PAGE_SIZE);
1229                 return 0;
1230         }
1231
1232         /* we need to read it. */
1233         r = readpage_nounlock(file, page);
1234         if (r < 0)
1235                 goto fail_nosnap;
1236         goto retry_locked;
1237 fail_nosnap:
1238         unlock_page(page);
1239         return r;
1240 }
1241
1242 /*
1243  * We are only allowed to write into/dirty the page if the page is
1244  * clean, or already dirty within the same snap context.
1245  */
1246 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1247                             loff_t pos, unsigned len, unsigned flags,
1248                             struct page **pagep, void **fsdata)
1249 {
1250         struct inode *inode = file_inode(file);
1251         struct page *page;
1252         pgoff_t index = pos >> PAGE_SHIFT;
1253         int r;
1254
1255         do {
1256                 /* get a page */
1257                 page = grab_cache_page_write_begin(mapping, index, 0);
1258                 if (!page)
1259                         return -ENOMEM;
1260
1261                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1262                      inode, page, (int)pos, (int)len);
1263
1264                 r = ceph_update_writeable_page(file, pos, len, page);
1265                 if (r < 0)
1266                         put_page(page);
1267                 else
1268                         *pagep = page;
1269         } while (r == -EAGAIN);
1270
1271         return r;
1272 }
1273
1274 /*
1275  * we don't do anything in here that simple_write_end doesn't do
1276  * except adjust dirty page accounting
1277  */
1278 static int ceph_write_end(struct file *file, struct address_space *mapping,
1279                           loff_t pos, unsigned len, unsigned copied,
1280                           struct page *page, void *fsdata)
1281 {
1282         struct inode *inode = file_inode(file);
1283         unsigned from = pos & (PAGE_SIZE - 1);
1284         int check_cap = 0;
1285
1286         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1287              inode, page, (int)pos, (int)copied, (int)len);
1288
1289         /* zero the stale part of the page if we did a short copy */
1290         if (copied < len)
1291                 zero_user_segment(page, from+copied, len);
1292
1293         /* did file size increase? */
1294         if (pos+copied > i_size_read(inode))
1295                 check_cap = ceph_inode_set_size(inode, pos+copied);
1296
1297         if (!PageUptodate(page))
1298                 SetPageUptodate(page);
1299
1300         set_page_dirty(page);
1301
1302         unlock_page(page);
1303         put_page(page);
1304
1305         if (check_cap)
1306                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1307
1308         return copied;
1309 }
1310
1311 /*
1312  * we set .direct_IO to indicate direct io is supported, but since we
1313  * intercept O_DIRECT reads and writes early, this function should
1314  * never get called.
1315  */
1316 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1317 {
1318         WARN_ON(1);
1319         return -EINVAL;
1320 }
1321
1322 const struct address_space_operations ceph_aops = {
1323         .readpage = ceph_readpage,
1324         .readpages = ceph_readpages,
1325         .writepage = ceph_writepage,
1326         .writepages = ceph_writepages_start,
1327         .write_begin = ceph_write_begin,
1328         .write_end = ceph_write_end,
1329         .set_page_dirty = ceph_set_page_dirty,
1330         .invalidatepage = ceph_invalidatepage,
1331         .releasepage = ceph_releasepage,
1332         .direct_IO = ceph_direct_io,
1333 };
1334
1335 static void ceph_block_sigs(sigset_t *oldset)
1336 {
1337         sigset_t mask;
1338         siginitsetinv(&mask, sigmask(SIGKILL));
1339         sigprocmask(SIG_BLOCK, &mask, oldset);
1340 }
1341
1342 static void ceph_restore_sigs(sigset_t *oldset)
1343 {
1344         sigprocmask(SIG_SETMASK, oldset, NULL);
1345 }
1346
1347 /*
1348  * vm ops
1349  */
1350 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1351 {
1352         struct inode *inode = file_inode(vma->vm_file);
1353         struct ceph_inode_info *ci = ceph_inode(inode);
1354         struct ceph_file_info *fi = vma->vm_file->private_data;
1355         struct page *pinned_page = NULL;
1356         loff_t off = vmf->pgoff << PAGE_SHIFT;
1357         int want, got, ret;
1358         sigset_t oldset;
1359
1360         ceph_block_sigs(&oldset);
1361
1362         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1363              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1364         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1365                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1366         else
1367                 want = CEPH_CAP_FILE_CACHE;
1368
1369         got = 0;
1370         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1371         if (ret < 0)
1372                 goto out_restore;
1373
1374         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1375              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1376
1377         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1378             ci->i_inline_version == CEPH_INLINE_NONE)
1379                 ret = filemap_fault(vma, vmf);
1380         else
1381                 ret = -EAGAIN;
1382
1383         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1384              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1385         if (pinned_page)
1386                 put_page(pinned_page);
1387         ceph_put_cap_refs(ci, got);
1388
1389         if (ret != -EAGAIN)
1390                 goto out_restore;
1391
1392         /* read inline data */
1393         if (off >= PAGE_SIZE) {
1394                 /* does not support inline data > PAGE_SIZE */
1395                 ret = VM_FAULT_SIGBUS;
1396         } else {
1397                 int ret1;
1398                 struct address_space *mapping = inode->i_mapping;
1399                 struct page *page = find_or_create_page(mapping, 0,
1400                                                 mapping_gfp_constraint(mapping,
1401                                                 ~__GFP_FS));
1402                 if (!page) {
1403                         ret = VM_FAULT_OOM;
1404                         goto out_inline;
1405                 }
1406                 ret1 = __ceph_do_getattr(inode, page,
1407                                          CEPH_STAT_CAP_INLINE_DATA, true);
1408                 if (ret1 < 0 || off >= i_size_read(inode)) {
1409                         unlock_page(page);
1410                         put_page(page);
1411                         if (ret1 < 0)
1412                                 ret = ret1;
1413                         else
1414                                 ret = VM_FAULT_SIGBUS;
1415                         goto out_inline;
1416                 }
1417                 if (ret1 < PAGE_SIZE)
1418                         zero_user_segment(page, ret1, PAGE_SIZE);
1419                 else
1420                         flush_dcache_page(page);
1421                 SetPageUptodate(page);
1422                 vmf->page = page;
1423                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1424 out_inline:
1425                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1426                      inode, off, (size_t)PAGE_SIZE, ret);
1427         }
1428 out_restore:
1429         ceph_restore_sigs(&oldset);
1430         if (ret < 0)
1431                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1432
1433         return ret;
1434 }
1435
1436 /*
1437  * Reuse write_begin here for simplicity.
1438  */
1439 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1440 {
1441         struct inode *inode = file_inode(vma->vm_file);
1442         struct ceph_inode_info *ci = ceph_inode(inode);
1443         struct ceph_file_info *fi = vma->vm_file->private_data;
1444         struct ceph_cap_flush *prealloc_cf;
1445         struct page *page = vmf->page;
1446         loff_t off = page_offset(page);
1447         loff_t size = i_size_read(inode);
1448         size_t len;
1449         int want, got, ret;
1450         sigset_t oldset;
1451
1452         prealloc_cf = ceph_alloc_cap_flush();
1453         if (!prealloc_cf)
1454                 return VM_FAULT_OOM;
1455
1456         ceph_block_sigs(&oldset);
1457
1458         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1459                 struct page *locked_page = NULL;
1460                 if (off == 0) {
1461                         lock_page(page);
1462                         locked_page = page;
1463                 }
1464                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1465                 if (locked_page)
1466                         unlock_page(locked_page);
1467                 if (ret < 0)
1468                         goto out_free;
1469         }
1470
1471         if (off + PAGE_SIZE <= size)
1472                 len = PAGE_SIZE;
1473         else
1474                 len = size & ~PAGE_MASK;
1475
1476         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1477              inode, ceph_vinop(inode), off, len, size);
1478         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1479                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1480         else
1481                 want = CEPH_CAP_FILE_BUFFER;
1482
1483         got = 0;
1484         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1485                             &got, NULL);
1486         if (ret < 0)
1487                 goto out_free;
1488
1489         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1490              inode, off, len, ceph_cap_string(got));
1491
1492         /* Update time before taking page lock */
1493         file_update_time(vma->vm_file);
1494
1495         do {
1496                 lock_page(page);
1497
1498                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1499                         unlock_page(page);
1500                         ret = VM_FAULT_NOPAGE;
1501                         break;
1502                 }
1503
1504                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1505                 if (ret >= 0) {
1506                         /* success.  we'll keep the page locked. */
1507                         set_page_dirty(page);
1508                         ret = VM_FAULT_LOCKED;
1509                 }
1510         } while (ret == -EAGAIN);
1511
1512         if (ret == VM_FAULT_LOCKED ||
1513             ci->i_inline_version != CEPH_INLINE_NONE) {
1514                 int dirty;
1515                 spin_lock(&ci->i_ceph_lock);
1516                 ci->i_inline_version = CEPH_INLINE_NONE;
1517                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1518                                                &prealloc_cf);
1519                 spin_unlock(&ci->i_ceph_lock);
1520                 if (dirty)
1521                         __mark_inode_dirty(inode, dirty);
1522         }
1523
1524         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1525              inode, off, len, ceph_cap_string(got), ret);
1526         ceph_put_cap_refs(ci, got);
1527 out_free:
1528         ceph_restore_sigs(&oldset);
1529         ceph_free_cap_flush(prealloc_cf);
1530         if (ret < 0)
1531                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1532         return ret;
1533 }
1534
1535 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1536                            char *data, size_t len)
1537 {
1538         struct address_space *mapping = inode->i_mapping;
1539         struct page *page;
1540
1541         if (locked_page) {
1542                 page = locked_page;
1543         } else {
1544                 if (i_size_read(inode) == 0)
1545                         return;
1546                 page = find_or_create_page(mapping, 0,
1547                                            mapping_gfp_constraint(mapping,
1548                                            ~__GFP_FS));
1549                 if (!page)
1550                         return;
1551                 if (PageUptodate(page)) {
1552                         unlock_page(page);
1553                         put_page(page);
1554                         return;
1555                 }
1556         }
1557
1558         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1559              inode, ceph_vinop(inode), len, locked_page);
1560
1561         if (len > 0) {
1562                 void *kaddr = kmap_atomic(page);
1563                 memcpy(kaddr, data, len);
1564                 kunmap_atomic(kaddr);
1565         }
1566
1567         if (page != locked_page) {
1568                 if (len < PAGE_SIZE)
1569                         zero_user_segment(page, len, PAGE_SIZE);
1570                 else
1571                         flush_dcache_page(page);
1572
1573                 SetPageUptodate(page);
1574                 unlock_page(page);
1575                 put_page(page);
1576         }
1577 }
1578
1579 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1580 {
1581         struct inode *inode = file_inode(filp);
1582         struct ceph_inode_info *ci = ceph_inode(inode);
1583         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1584         struct ceph_osd_request *req;
1585         struct page *page = NULL;
1586         u64 len, inline_version;
1587         int err = 0;
1588         bool from_pagecache = false;
1589
1590         spin_lock(&ci->i_ceph_lock);
1591         inline_version = ci->i_inline_version;
1592         spin_unlock(&ci->i_ceph_lock);
1593
1594         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1595              inode, ceph_vinop(inode), inline_version);
1596
1597         if (inline_version == 1 || /* initial version, no data */
1598             inline_version == CEPH_INLINE_NONE)
1599                 goto out;
1600
1601         if (locked_page) {
1602                 page = locked_page;
1603                 WARN_ON(!PageUptodate(page));
1604         } else if (ceph_caps_issued(ci) &
1605                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1606                 page = find_get_page(inode->i_mapping, 0);
1607                 if (page) {
1608                         if (PageUptodate(page)) {
1609                                 from_pagecache = true;
1610                                 lock_page(page);
1611                         } else {
1612                                 put_page(page);
1613                                 page = NULL;
1614                         }
1615                 }
1616         }
1617
1618         if (page) {
1619                 len = i_size_read(inode);
1620                 if (len > PAGE_SIZE)
1621                         len = PAGE_SIZE;
1622         } else {
1623                 page = __page_cache_alloc(GFP_NOFS);
1624                 if (!page) {
1625                         err = -ENOMEM;
1626                         goto out;
1627                 }
1628                 err = __ceph_do_getattr(inode, page,
1629                                         CEPH_STAT_CAP_INLINE_DATA, true);
1630                 if (err < 0) {
1631                         /* no inline data */
1632                         if (err == -ENODATA)
1633                                 err = 0;
1634                         goto out;
1635                 }
1636                 len = err;
1637         }
1638
1639         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1640                                     ceph_vino(inode), 0, &len, 0, 1,
1641                                     CEPH_OSD_OP_CREATE,
1642                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1643                                     NULL, 0, 0, false);
1644         if (IS_ERR(req)) {
1645                 err = PTR_ERR(req);
1646                 goto out;
1647         }
1648
1649         req->r_mtime = inode->i_mtime;
1650         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1651         if (!err)
1652                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1653         ceph_osdc_put_request(req);
1654         if (err < 0)
1655                 goto out;
1656
1657         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1658                                     ceph_vino(inode), 0, &len, 1, 3,
1659                                     CEPH_OSD_OP_WRITE,
1660                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1661                                     NULL, ci->i_truncate_seq,
1662                                     ci->i_truncate_size, false);
1663         if (IS_ERR(req)) {
1664                 err = PTR_ERR(req);
1665                 goto out;
1666         }
1667
1668         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1669
1670         {
1671                 __le64 xattr_buf = cpu_to_le64(inline_version);
1672                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1673                                             "inline_version", &xattr_buf,
1674                                             sizeof(xattr_buf),
1675                                             CEPH_OSD_CMPXATTR_OP_GT,
1676                                             CEPH_OSD_CMPXATTR_MODE_U64);
1677                 if (err)
1678                         goto out_put;
1679         }
1680
1681         {
1682                 char xattr_buf[32];
1683                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1684                                          "%llu", inline_version);
1685                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1686                                             "inline_version",
1687                                             xattr_buf, xattr_len, 0, 0);
1688                 if (err)
1689                         goto out_put;
1690         }
1691
1692         req->r_mtime = inode->i_mtime;
1693         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1694         if (!err)
1695                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1696 out_put:
1697         ceph_osdc_put_request(req);
1698         if (err == -ECANCELED)
1699                 err = 0;
1700 out:
1701         if (page && page != locked_page) {
1702                 if (from_pagecache) {
1703                         unlock_page(page);
1704                         put_page(page);
1705                 } else
1706                         __free_pages(page, 0);
1707         }
1708
1709         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1710              inode, ceph_vinop(inode), inline_version, err);
1711         return err;
1712 }
1713
1714 static const struct vm_operations_struct ceph_vmops = {
1715         .fault          = ceph_filemap_fault,
1716         .page_mkwrite   = ceph_page_mkwrite,
1717 };
1718
1719 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1720 {
1721         struct address_space *mapping = file->f_mapping;
1722
1723         if (!mapping->a_ops->readpage)
1724                 return -ENOEXEC;
1725         file_accessed(file);
1726         vma->vm_ops = &ceph_vmops;
1727         return 0;
1728 }
1729
1730 enum {
1731         POOL_READ       = 1,
1732         POOL_WRITE      = 2,
1733 };
1734
1735 static int __ceph_pool_perm_get(struct ceph_inode_info *ci, u32 pool)
1736 {
1737         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1738         struct ceph_mds_client *mdsc = fsc->mdsc;
1739         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1740         struct rb_node **p, *parent;
1741         struct ceph_pool_perm *perm;
1742         struct page **pages;
1743         int err = 0, err2 = 0, have = 0;
1744
1745         down_read(&mdsc->pool_perm_rwsem);
1746         p = &mdsc->pool_perm_tree.rb_node;
1747         while (*p) {
1748                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1749                 if (pool < perm->pool)
1750                         p = &(*p)->rb_left;
1751                 else if (pool > perm->pool)
1752                         p = &(*p)->rb_right;
1753                 else {
1754                         have = perm->perm;
1755                         break;
1756                 }
1757         }
1758         up_read(&mdsc->pool_perm_rwsem);
1759         if (*p)
1760                 goto out;
1761
1762         dout("__ceph_pool_perm_get pool %u no perm cached\n", pool);
1763
1764         down_write(&mdsc->pool_perm_rwsem);
1765         parent = NULL;
1766         while (*p) {
1767                 parent = *p;
1768                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1769                 if (pool < perm->pool)
1770                         p = &(*p)->rb_left;
1771                 else if (pool > perm->pool)
1772                         p = &(*p)->rb_right;
1773                 else {
1774                         have = perm->perm;
1775                         break;
1776                 }
1777         }
1778         if (*p) {
1779                 up_write(&mdsc->pool_perm_rwsem);
1780                 goto out;
1781         }
1782
1783         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1784                                          1, false, GFP_NOFS);
1785         if (!rd_req) {
1786                 err = -ENOMEM;
1787                 goto out_unlock;
1788         }
1789
1790         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1791         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1792         rd_req->r_base_oloc.pool = pool;
1793         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1794
1795         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1796         if (err)
1797                 goto out_unlock;
1798
1799         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1800                                          1, false, GFP_NOFS);
1801         if (!wr_req) {
1802                 err = -ENOMEM;
1803                 goto out_unlock;
1804         }
1805
1806         wr_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ACK;
1807         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1808         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1809         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1810
1811         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1812         if (err)
1813                 goto out_unlock;
1814
1815         /* one page should be large enough for STAT data */
1816         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1817         if (IS_ERR(pages)) {
1818                 err = PTR_ERR(pages);
1819                 goto out_unlock;
1820         }
1821
1822         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1823                                      0, false, true);
1824         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1825
1826         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1827         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1828
1829         if (!err)
1830                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1831         if (!err2)
1832                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1833
1834         if (err >= 0 || err == -ENOENT)
1835                 have |= POOL_READ;
1836         else if (err != -EPERM)
1837                 goto out_unlock;
1838
1839         if (err2 == 0 || err2 == -EEXIST)
1840                 have |= POOL_WRITE;
1841         else if (err2 != -EPERM) {
1842                 err = err2;
1843                 goto out_unlock;
1844         }
1845
1846         perm = kmalloc(sizeof(*perm), GFP_NOFS);
1847         if (!perm) {
1848                 err = -ENOMEM;
1849                 goto out_unlock;
1850         }
1851
1852         perm->pool = pool;
1853         perm->perm = have;
1854         rb_link_node(&perm->node, parent, p);
1855         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1856         err = 0;
1857 out_unlock:
1858         up_write(&mdsc->pool_perm_rwsem);
1859
1860         ceph_osdc_put_request(rd_req);
1861         ceph_osdc_put_request(wr_req);
1862 out:
1863         if (!err)
1864                 err = have;
1865         dout("__ceph_pool_perm_get pool %u result = %d\n", pool, err);
1866         return err;
1867 }
1868
1869 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1870 {
1871         u32 pool;
1872         int ret, flags;
1873
1874         /* does not support pool namespace yet */
1875         if (ci->i_pool_ns_len)
1876                 return -EIO;
1877
1878         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1879                                 NOPOOLPERM))
1880                 return 0;
1881
1882         spin_lock(&ci->i_ceph_lock);
1883         flags = ci->i_ceph_flags;
1884         pool = ceph_file_layout_pg_pool(ci->i_layout);
1885         spin_unlock(&ci->i_ceph_lock);
1886 check:
1887         if (flags & CEPH_I_POOL_PERM) {
1888                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1889                         dout("ceph_pool_perm_check pool %u no read perm\n",
1890                              pool);
1891                         return -EPERM;
1892                 }
1893                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1894                         dout("ceph_pool_perm_check pool %u no write perm\n",
1895                              pool);
1896                         return -EPERM;
1897                 }
1898                 return 0;
1899         }
1900
1901         ret = __ceph_pool_perm_get(ci, pool);
1902         if (ret < 0)
1903                 return ret;
1904
1905         flags = CEPH_I_POOL_PERM;
1906         if (ret & POOL_READ)
1907                 flags |= CEPH_I_POOL_RD;
1908         if (ret & POOL_WRITE)
1909                 flags |= CEPH_I_POOL_WR;
1910
1911         spin_lock(&ci->i_ceph_lock);
1912         if (pool == ceph_file_layout_pg_pool(ci->i_layout)) {
1913                 ci->i_ceph_flags = flags;
1914         } else {
1915                 pool = ceph_file_layout_pg_pool(ci->i_layout);
1916                 flags = ci->i_ceph_flags;
1917         }
1918         spin_unlock(&ci->i_ceph_lock);
1919         goto check;
1920 }
1921
1922 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1923 {
1924         struct ceph_pool_perm *perm;
1925         struct rb_node *n;
1926
1927         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1928                 n = rb_first(&mdsc->pool_perm_tree);
1929                 perm = rb_entry(n, struct ceph_pool_perm, node);
1930                 rb_erase(n, &mdsc->pool_perm_tree);
1931                 kfree(perm);
1932         }
1933 }