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 (%sdirty)\n", page->mapping->host,
 179              page, page->index, PageDirty(page) ? "" : "not ");
 180
 181         /* Can we release the page from the cache? */
 182         if (!ceph_release_fscache_page(page, g))
 183                 return 0;
 184
 185         return !PagePrivate(page);
 186 }
 187
 188 /*
 189  * read a single page, without unlocking it.
 190  */
 191 static int readpage_nounlock(struct file *filp, struct page *page)
 192 {
 193         struct inode *inode = file_inode(filp);
 194         struct ceph_inode_info *ci = ceph_inode(inode);
 195         struct ceph_osd_client *osdc =
 196                 &ceph_inode_to_client(inode)->client->osdc;
 197         int err = 0;
 198         u64 off = page_offset(page);
 199         u64 len = PAGE_SIZE;
 200
 201         if (off >= i_size_read(inode)) {
 202                 zero_user_segment(page, 0, PAGE_SIZE);
 203                 SetPageUptodate(page);
 204                 return 0;
 205         }
 206
 207         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 208                 /*
 209                  * Uptodate inline data should have been added
 210                  * into page cache while getting Fcr caps.
 211                  */
 212                 if (off == 0)
 213                         return -EINVAL;
 214                 zero_user_segment(page, 0, PAGE_SIZE);
 215                 SetPageUptodate(page);
 216                 return 0;
 217         }
 218
 219         err = ceph_readpage_from_fscache(inode, page);
 220         if (err == 0)
 221                 goto out;
 222
 223         dout("readpage inode %p file %p page %p index %lu\n",
 224              inode, filp, page, page->index);
 225         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 226                                   off, &len,
 227                                   ci->i_truncate_seq, ci->i_truncate_size,
 228                                   &page, 1, 0);
 229         if (err == -ENOENT)
 230                 err = 0;
 231         if (err < 0) {
 232                 SetPageError(page);
 233                 ceph_fscache_readpage_cancel(inode, page);
 234                 goto out;
 235         }
 236         if (err < PAGE_SIZE)
 237                 /* zero fill remainder of page */
 238                 zero_user_segment(page, err, PAGE_SIZE);
 239         else
 240                 flush_dcache_page(page);
 241
 242         SetPageUptodate(page);
 243         ceph_readpage_to_fscache(inode, page);
 244
 245 out:
 246         return err < 0 ? err : 0;
 247 }
 248
 249 static int ceph_readpage(struct file *filp, struct page *page)
 250 {
 251         int r = readpage_nounlock(filp, page);
 252         unlock_page(page);
 253         return r;
 254 }
 255
 256 /*
 257  * Finish an async read(ahead) op.
 258  */
 259 static void finish_read(struct ceph_osd_request *req)
 260 {
 261         struct inode *inode = req->r_inode;
 262         struct ceph_osd_data *osd_data;
 263         int rc = req->r_result <= 0 ? req->r_result : 0;
 264         int bytes = req->r_result >= 0 ? req->r_result : 0;
 265         int num_pages;
 266         int i;
 267
 268         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 269
 270         /* unlock all pages, zeroing any data we didn't read */
 271         osd_data = osd_req_op_extent_osd_data(req, 0);
 272         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 273         num_pages = calc_pages_for((u64)osd_data->alignment,
 274                                         (u64)osd_data->length);
 275         for (i = 0; i < num_pages; i++) {
 276                 struct page *page = osd_data->pages[i];
 277
 278                 if (rc < 0 && rc != -ENOENT) {
 279                         ceph_fscache_readpage_cancel(inode, page);
 280                         goto unlock;
 281                 }
 282                 if (bytes < (int)PAGE_SIZE) {
 283                         /* zero (remainder of) page */
 284                         int s = bytes < 0 ? 0 : bytes;
 285                         zero_user_segment(page, s, PAGE_SIZE);
 286                 }
 287                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 288                      page->index);
 289                 flush_dcache_page(page);
 290                 SetPageUptodate(page);
 291                 ceph_readpage_to_fscache(inode, page);
 292 unlock:
 293                 unlock_page(page);
 294                 put_page(page);
 295                 bytes -= PAGE_SIZE;
 296         }
 297         kfree(osd_data->pages);
 298 }
 299
 300 /*
 301  * start an async read(ahead) operation.  return nr_pages we submitted
 302  * a read for on success, or negative error code.
 303  */
 304 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 305 {
 306         struct ceph_osd_client *osdc =
 307                 &ceph_inode_to_client(inode)->client->osdc;
 308         struct ceph_inode_info *ci = ceph_inode(inode);
 309         struct page *page = list_entry(page_list->prev, struct page, lru);
 310         struct ceph_vino vino;
 311         struct ceph_osd_request *req;
 312         u64 off;
 313         u64 len;
 314         int i;
 315         struct page **pages;
 316         pgoff_t next_index;
 317         int nr_pages = 0;
 318         int ret;
 319
 320         off = (u64) page_offset(page);
 321
 322         /* count pages */
 323         next_index = page->index;
 324         list_for_each_entry_reverse(page, page_list, lru) {
 325                 if (page->index != next_index)
 326                         break;
 327                 nr_pages++;
 328                 next_index++;
 329                 if (max && nr_pages == max)
 330                         break;
 331         }
 332         len = nr_pages << PAGE_SHIFT;
 333         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 334              off, len);
 335         vino = ceph_vino(inode);
 336         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 337                                     0, 1, CEPH_OSD_OP_READ,
 338                                     CEPH_OSD_FLAG_READ, NULL,
 339                                     ci->i_truncate_seq, ci->i_truncate_size,
 340                                     false);
 341         if (IS_ERR(req))
 342                 return PTR_ERR(req);
 343
 344         /* build page vector */
 345         nr_pages = calc_pages_for(0, len);
 346         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 347         ret = -ENOMEM;
 348         if (!pages)
 349                 goto out;
 350         for (i = 0; i < nr_pages; ++i) {
 351                 page = list_entry(page_list->prev, struct page, lru);
 352                 BUG_ON(PageLocked(page));
 353                 list_del(&page->lru);
 354
 355                 dout("start_read %p adding %p idx %lu\n", inode, page,
 356                      page->index);
 357                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 358                                           GFP_KERNEL)) {
 359                         ceph_fscache_uncache_page(inode, page);
 360                         put_page(page);
 361                         dout("start_read %p add_to_page_cache failed %p\n",
 362                              inode, page);
 363                         nr_pages = i;
 364                         if (nr_pages > 0) {
 365                                 len = nr_pages << PAGE_SHIFT;
 366                                 break;
 367                         }
 368                         goto out_pages;
 369                 }
 370                 pages[i] = page;
 371         }
 372         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 373         req->r_callback = finish_read;
 374         req->r_inode = inode;
 375
 376         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 377         ret = ceph_osdc_start_request(osdc, req, false);
 378         if (ret < 0)
 379                 goto out_pages;
 380         ceph_osdc_put_request(req);
 381         return nr_pages;
 382
 383 out_pages:
 384         for (i = 0; i < nr_pages; ++i) {
 385                 ceph_fscache_readpage_cancel(inode, pages[i]);
 386                 unlock_page(pages[i]);
 387         }
 388         ceph_put_page_vector(pages, nr_pages, false);
 389 out:
 390         ceph_osdc_put_request(req);
 391         return ret;
 392 }
 393
 394
 395 /*
 396  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 397  * the caller (VM) cleans them up.
 398  */
 399 static int ceph_readpages(struct file *file, struct address_space *mapping,
 400                           struct list_head *page_list, unsigned nr_pages)
 401 {
 402         struct inode *inode = file_inode(file);
 403         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 404         int rc = 0;
 405         int max = 0;
 406
 407         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 408                 return -EINVAL;
 409
 410         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 411                                          &nr_pages);
 412
 413         if (rc == 0)
 414                 goto out;
 415
 416         if (fsc->mount_options->rsize >= PAGE_SIZE)
 417                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 418                         >> PAGE_SHIFT;
 419
 420         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 421                 file, nr_pages,
 422              max);
 423         while (!list_empty(page_list)) {
 424                 rc = start_read(inode, page_list, max);
 425                 if (rc < 0)
 426                         goto out;
 427                 BUG_ON(rc == 0);
 428         }
 429 out:
 430         ceph_fscache_readpages_cancel(inode, page_list);
 431
 432         dout("readpages %p file %p ret %d\n", inode, file, rc);
 433         return rc;
 434 }
 435
 436 /*
 437  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 438  * only snap context we are allowed to write back.
 439  */
 440 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 441                                                     loff_t *snap_size)
 442 {
 443         struct ceph_inode_info *ci = ceph_inode(inode);
 444         struct ceph_snap_context *snapc = NULL;
 445         struct ceph_cap_snap *capsnap = NULL;
 446
 447         spin_lock(&ci->i_ceph_lock);
 448         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 449                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 450                      capsnap->context, capsnap->dirty_pages);
 451                 if (capsnap->dirty_pages) {
 452                         snapc = ceph_get_snap_context(capsnap->context);
 453                         if (snap_size)
 454                                 *snap_size = capsnap->size;
 455                         break;
 456                 }
 457         }
 458         if (!snapc && ci->i_wrbuffer_ref_head) {
 459                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 460                 dout(" head snapc %p has %d dirty pages\n",
 461                      snapc, ci->i_wrbuffer_ref_head);
 462         }
 463         spin_unlock(&ci->i_ceph_lock);
 464         return snapc;
 465 }
 466
 467 /*
 468  * Write a single page, but leave the page locked.
 469  *
 470  * If we get a write error, set the page error bit, but still adjust the
 471  * dirty page accounting (i.e., page is no longer dirty).
 472  */
 473 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 474 {
 475         struct inode *inode;
 476         struct ceph_inode_info *ci;
 477         struct ceph_fs_client *fsc;
 478         struct ceph_osd_client *osdc;
 479         struct ceph_snap_context *snapc, *oldest;
 480         loff_t page_off = page_offset(page);
 481         loff_t snap_size = -1;
 482         long writeback_stat;
 483         u64 truncate_size;
 484         u32 truncate_seq;
 485         int err = 0, len = PAGE_SIZE;
 486
 487         dout("writepage %p idx %lu\n", page, page->index);
 488
 489         if (!page->mapping || !page->mapping->host) {
 490                 dout("writepage %p - no mapping\n", page);
 491                 return -EFAULT;
 492         }
 493         inode = page->mapping->host;
 494         ci = ceph_inode(inode);
 495         fsc = ceph_inode_to_client(inode);
 496         osdc = &fsc->client->osdc;
 497
 498         /* verify this is a writeable snap context */
 499         snapc = page_snap_context(page);
 500         if (snapc == NULL) {
 501                 dout("writepage %p page %p not dirty?\n", inode, page);
 502                 goto out;
 503         }
 504         oldest = get_oldest_context(inode, &snap_size);
 505         if (snapc->seq > oldest->seq) {
 506                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 507                      inode, page, snapc);
 508                 /* we should only noop if called by kswapd */
 509                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 510                 ceph_put_snap_context(oldest);
 511                 goto out;
 512         }
 513         ceph_put_snap_context(oldest);
 514
 515         spin_lock(&ci->i_ceph_lock);
 516         truncate_seq = ci->i_truncate_seq;
 517         truncate_size = ci->i_truncate_size;
 518         if (snap_size == -1)
 519                 snap_size = i_size_read(inode);
 520         spin_unlock(&ci->i_ceph_lock);
 521
 522         /* is this a partial page at end of file? */
 523         if (page_off >= snap_size) {
 524                 dout("%p page eof %llu\n", page, snap_size);
 525                 goto out;
 526         }
 527         if (snap_size < page_off + len)
 528                 len = snap_size - page_off;
 529
 530         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 531              inode, page, page->index, page_off, len, snapc);
 532
 533         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 534         if (writeback_stat >
 535             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 536                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 537
 538         set_page_writeback(page);
 539         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 540                                    &ci->i_layout, snapc,
 541                                    page_off, len,
 542                                    truncate_seq, truncate_size,
 543                                    &inode->i_mtime, &page, 1);
 544         if (err < 0) {
 545                 struct writeback_control tmp_wbc;
 546                 if (!wbc)
 547                         wbc = &tmp_wbc;
 548                 if (err == -ERESTARTSYS) {
 549                         /* killed by SIGKILL */
 550                         dout("writepage interrupted page %p\n", page);
 551                         redirty_page_for_writepage(wbc, page);
 552                         end_page_writeback(page);
 553                         goto out;
 554                 }
 555                 dout("writepage setting page/mapping error %d %p\n",
 556                      err, page);
 557                 SetPageError(page);
 558                 mapping_set_error(&inode->i_data, err);
 559                 wbc->pages_skipped++;
 560         } else {
 561                 dout("writepage cleaned page %p\n", page);
 562                 err = 0;  /* vfs expects us to return 0 */
 563         }
 564         page->private = 0;
 565         ClearPagePrivate(page);
 566         end_page_writeback(page);
 567         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 568         ceph_put_snap_context(snapc);  /* page's reference */
 569 out:
 570         return err;
 571 }
 572
 573 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 574 {
 575         int err;
 576         struct inode *inode = page->mapping->host;
 577         BUG_ON(!inode);
 578         ihold(inode);
 579         err = writepage_nounlock(page, wbc);
 580         if (err == -ERESTARTSYS) {
 581                 /* direct memory reclaimer was killed by SIGKILL. return 0
 582                  * to prevent caller from setting mapping/page error */
 583                 err = 0;
 584         }
 585         unlock_page(page);
 586         iput(inode);
 587         return err;
 588 }
 589
 590 /*
 591  * lame release_pages helper.  release_pages() isn't exported to
 592  * modules.
 593  */
 594 static void ceph_release_pages(struct page **pages, int num)
 595 {
 596         struct pagevec pvec;
 597         int i;
 598
 599         pagevec_init(&pvec, 0);
 600         for (i = 0; i < num; i++) {
 601                 if (pagevec_add(&pvec, pages[i]) == 0)
 602                         pagevec_release(&pvec);
 603         }
 604         pagevec_release(&pvec);
 605 }
 606
 607 /*
 608  * async writeback completion handler.
 609  *
 610  * If we get an error, set the mapping error bit, but not the individual
 611  * page error bits.
 612  */
 613 static void writepages_finish(struct ceph_osd_request *req)
 614 {
 615         struct inode *inode = req->r_inode;
 616         struct ceph_inode_info *ci = ceph_inode(inode);
 617         struct ceph_osd_data *osd_data;
 618         struct page *page;
 619         int num_pages, total_pages = 0;
 620         int i, j;
 621         int rc = req->r_result;
 622         struct ceph_snap_context *snapc = req->r_snapc;
 623         struct address_space *mapping = inode->i_mapping;
 624         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 625         bool remove_page;
 626
 627         dout("writepages_finish %p rc %d\n", inode, rc);
 628         if (rc < 0)
 629                 mapping_set_error(mapping, rc);
 630
 631         /*
 632          * We lost the cache cap, need to truncate the page before
 633          * it is unlocked, otherwise we'd truncate it later in the
 634          * page truncation thread, possibly losing some data that
 635          * raced its way in
 636          */
 637         remove_page = !(ceph_caps_issued(ci) &
 638                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 639
 640         /* clean all pages */
 641         for (i = 0; i < req->r_num_ops; i++) {
 642                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 643                         break;
 644
 645                 osd_data = osd_req_op_extent_osd_data(req, i);
 646                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 647                 num_pages = calc_pages_for((u64)osd_data->alignment,
 648                                            (u64)osd_data->length);
 649                 total_pages += num_pages;
 650                 for (j = 0; j < num_pages; j++) {
 651                         page = osd_data->pages[j];
 652                         BUG_ON(!page);
 653                         WARN_ON(!PageUptodate(page));
 654
 655                         if (atomic_long_dec_return(&fsc->writeback_count) <
 656                              CONGESTION_OFF_THRESH(
 657                                         fsc->mount_options->congestion_kb))
 658                                 clear_bdi_congested(&fsc->backing_dev_info,
 659                                                     BLK_RW_ASYNC);
 660
 661                         if (rc < 0)
 662                                 SetPageError(page);
 663
 664                         ceph_put_snap_context(page_snap_context(page));
 665                         page->private = 0;
 666                         ClearPagePrivate(page);
 667                         dout("unlocking %p\n", page);
 668                         end_page_writeback(page);
 669
 670                         if (remove_page)
 671                                 generic_error_remove_page(inode->i_mapping,
 672                                                           page);
 673
 674                         unlock_page(page);
 675                 }
 676                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 677                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 678
 679                 ceph_release_pages(osd_data->pages, num_pages);
 680         }
 681
 682         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 683
 684         osd_data = osd_req_op_extent_osd_data(req, 0);
 685         if (osd_data->pages_from_pool)
 686                 mempool_free(osd_data->pages,
 687                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 688         else
 689                 kfree(osd_data->pages);
 690         ceph_osdc_put_request(req);
 691 }
 692
 693 /*
 694  * initiate async writeback
 695  */
 696 static int ceph_writepages_start(struct address_space *mapping,
 697                                  struct writeback_control *wbc)
 698 {
 699         struct inode *inode = mapping->host;
 700         struct ceph_inode_info *ci = ceph_inode(inode);
 701         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 702         struct ceph_vino vino = ceph_vino(inode);
 703         pgoff_t index, start, end;
 704         int range_whole = 0;
 705         int should_loop = 1;
 706         pgoff_t max_pages = 0, max_pages_ever = 0;
 707         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 708         struct pagevec pvec;
 709         int done = 0;
 710         int rc = 0;
 711         unsigned wsize = 1 << inode->i_blkbits;
 712         struct ceph_osd_request *req = NULL;
 713         int do_sync = 0;
 714         loff_t snap_size, i_size;
 715         u64 truncate_size;
 716         u32 truncate_seq;
 717
 718         /*
 719          * Include a 'sync' in the OSD request if this is a data
 720          * integrity write (e.g., O_SYNC write or fsync()), or if our
 721          * cap is being revoked.
 722          */
 723         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 724                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 725                 do_sync = 1;
 726         dout("writepages_start %p dosync=%d (mode=%s)\n",
 727              inode, do_sync,
 728              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 729              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 730
 731         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 732                 if (ci->i_wrbuffer_ref > 0) {
 733                         pr_warn_ratelimited(
 734                                 "writepage_start %p %lld forced umount\n",
 735                                 inode, ceph_ino(inode));
 736                 }
 737                 mapping_set_error(mapping, -EIO);
 738                 return -EIO; /* we're in a forced umount, don't write! */
 739         }
 740         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 741                 wsize = fsc->mount_options->wsize;
 742         if (wsize < PAGE_SIZE)
 743                 wsize = PAGE_SIZE;
 744         max_pages_ever = wsize >> PAGE_SHIFT;
 745
 746         pagevec_init(&pvec, 0);
 747
 748         /* where to start/end? */
 749         if (wbc->range_cyclic) {
 750                 start = mapping->writeback_index; /* Start from prev offset */
 751                 end = -1;
 752                 dout(" cyclic, start at %lu\n", start);
 753         } else {
 754                 start = wbc->range_start >> PAGE_SHIFT;
 755                 end = wbc->range_end >> PAGE_SHIFT;
 756                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 757                         range_whole = 1;
 758                 should_loop = 0;
 759                 dout(" not cyclic, %lu to %lu\n", start, end);
 760         }
 761         index = start;
 762
 763 retry:
 764         /* find oldest snap context with dirty data */
 765         ceph_put_snap_context(snapc);
 766         snap_size = -1;
 767         snapc = get_oldest_context(inode, &snap_size);
 768         if (!snapc) {
 769                 /* hmm, why does writepages get called when there
 770                    is no dirty data? */
 771                 dout(" no snap context with dirty data?\n");
 772                 goto out;
 773         }
 774         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 775              snapc, snapc->seq, snapc->num_snaps);
 776
 777         spin_lock(&ci->i_ceph_lock);
 778         truncate_seq = ci->i_truncate_seq;
 779         truncate_size = ci->i_truncate_size;
 780         i_size = i_size_read(inode);
 781         spin_unlock(&ci->i_ceph_lock);
 782
 783         if (last_snapc && snapc != last_snapc) {
 784                 /* if we switched to a newer snapc, restart our scan at the
 785                  * start of the original file range. */
 786                 dout("  snapc differs from last pass, restarting at %lu\n",
 787                      index);
 788                 index = start;
 789         }
 790         last_snapc = snapc;
 791
 792         while (!done && index <= end) {
 793                 unsigned i;
 794                 int first;
 795                 pgoff_t strip_unit_end = 0;
 796                 int num_ops = 0, op_idx;
 797                 int pvec_pages, locked_pages = 0;
 798                 struct page **pages = NULL, **data_pages;
 799                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 800                 struct page *page;
 801                 int want;
 802                 u64 offset = 0, len = 0;
 803
 804                 max_pages = max_pages_ever;
 805
 806 get_more_pages:
 807                 first = -1;
 808                 want = min(end - index,
 809                            min((pgoff_t)PAGEVEC_SIZE,
 810                                max_pages - (pgoff_t)locked_pages) - 1)
 811                         + 1;
 812                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 813                                                 PAGECACHE_TAG_DIRTY,
 814                                                 want);
 815                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 816                 if (!pvec_pages && !locked_pages)
 817                         break;
 818                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 819                         page = pvec.pages[i];
 820                         dout("? %p idx %lu\n", page, page->index);
 821                         if (locked_pages == 0)
 822                                 lock_page(page);  /* first page */
 823                         else if (!trylock_page(page))
 824                                 break;
 825
 826                         /* only dirty pages, or our accounting breaks */
 827                         if (unlikely(!PageDirty(page)) ||
 828                             unlikely(page->mapping != mapping)) {
 829                                 dout("!dirty or !mapping %p\n", page);
 830                                 unlock_page(page);
 831                                 break;
 832                         }
 833                         if (!wbc->range_cyclic && page->index > end) {
 834                                 dout("end of range %p\n", page);
 835                                 done = 1;
 836                                 unlock_page(page);
 837                                 break;
 838                         }
 839                         if (strip_unit_end && (page->index > strip_unit_end)) {
 840                                 dout("end of strip unit %p\n", page);
 841                                 unlock_page(page);
 842                                 break;
 843                         }
 844                         if (wbc->sync_mode != WB_SYNC_NONE) {
 845                                 dout("waiting on writeback %p\n", page);
 846                                 wait_on_page_writeback(page);
 847                         }
 848                         if (page_offset(page) >=
 849                             (snap_size == -1 ? i_size : snap_size)) {
 850                                 dout("%p page eof %llu\n", page,
 851                                      (snap_size == -1 ? i_size : snap_size));
 852                                 done = 1;
 853                                 unlock_page(page);
 854                                 break;
 855                         }
 856                         if (PageWriteback(page)) {
 857                                 dout("%p under writeback\n", page);
 858                                 unlock_page(page);
 859                                 break;
 860                         }
 861
 862                         /* only if matching snap context */
 863                         pgsnapc = page_snap_context(page);
 864                         if (pgsnapc->seq > snapc->seq) {
 865                                 dout("page snapc %p %lld > oldest %p %lld\n",
 866                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 867                                 unlock_page(page);
 868                                 if (!locked_pages)
 869                                         continue; /* keep looking for snap */
 870                                 break;
 871                         }
 872
 873                         if (!clear_page_dirty_for_io(page)) {
 874                                 dout("%p !clear_page_dirty_for_io\n", page);
 875                                 unlock_page(page);
 876                                 break;
 877                         }
 878
 879                         /*
 880                          * We have something to write.  If this is
 881                          * the first locked page this time through,
 882                          * calculate max possinle write size and
 883                          * allocate a page array
 884                          */
 885                         if (locked_pages == 0) {
 886                                 u64 objnum;
 887                                 u64 objoff;
 888
 889                                 /* prepare async write request */
 890                                 offset = (u64)page_offset(page);
 891                                 len = wsize;
 892
 893                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 894                                                                 offset, len,
 895                                                                 &objnum, &objoff,
 896                                                                 &len);
 897                                 if (rc < 0) {
 898                                         unlock_page(page);
 899                                         break;
 900                                 }
 901
 902                                 num_ops = 1 + do_sync;
 903                                 strip_unit_end = page->index +
 904                                         ((len - 1) >> PAGE_SHIFT);
 905
 906                                 BUG_ON(pages);
 907                                 max_pages = calc_pages_for(0, (u64)len);
 908                                 pages = kmalloc(max_pages * sizeof (*pages),
 909                                                 GFP_NOFS);
 910                                 if (!pages) {
 911                                         pool = fsc->wb_pagevec_pool;
 912                                         pages = mempool_alloc(pool, GFP_NOFS);
 913                                         BUG_ON(!pages);
 914                                 }
 915
 916                                 len = 0;
 917                         } else if (page->index !=
 918                                    (offset + len) >> PAGE_SHIFT) {
 919                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 920                                                         CEPH_OSD_MAX_OPS)) {
 921                                         redirty_page_for_writepage(wbc, page);
 922                                         unlock_page(page);
 923                                         break;
 924                                 }
 925
 926                                 num_ops++;
 927                                 offset = (u64)page_offset(page);
 928                                 len = 0;
 929                         }
 930
 931                         /* note position of first page in pvec */
 932                         if (first < 0)
 933                                 first = i;
 934                         dout("%p will write page %p idx %lu\n",
 935                              inode, page, page->index);
 936
 937                         if (atomic_long_inc_return(&fsc->writeback_count) >
 938                             CONGESTION_ON_THRESH(
 939                                     fsc->mount_options->congestion_kb)) {
 940                                 set_bdi_congested(&fsc->backing_dev_info,
 941                                                   BLK_RW_ASYNC);
 942                         }
 943
 944                         pages[locked_pages] = page;
 945                         locked_pages++;
 946                         len += PAGE_SIZE;
 947                 }
 948
 949                 /* did we get anything? */
 950                 if (!locked_pages)
 951                         goto release_pvec_pages;
 952                 if (i) {
 953                         int j;
 954                         BUG_ON(!locked_pages || first < 0);
 955
 956                         if (pvec_pages && i == pvec_pages &&
 957                             locked_pages < max_pages) {
 958                                 dout("reached end pvec, trying for more\n");
 959                                 pagevec_reinit(&pvec);
 960                                 goto get_more_pages;
 961                         }
 962
 963                         /* shift unused pages over in the pvec...  we
 964                          * will need to release them below. */
 965                         for (j = i; j < pvec_pages; j++) {
 966                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
 967                                 pvec.pages[j-i+first] = pvec.pages[j];
 968                         }
 969                         pvec.nr -= i-first;
 970                 }
 971
 972 new_request:
 973                 offset = page_offset(pages[0]);
 974                 len = wsize;
 975
 976                 req = ceph_osdc_new_request(&fsc->client->osdc,
 977                                         &ci->i_layout, vino,
 978                                         offset, &len, 0, num_ops,
 979                                         CEPH_OSD_OP_WRITE,
 980                                         CEPH_OSD_FLAG_WRITE |
 981                                         CEPH_OSD_FLAG_ONDISK,
 982                                         snapc, truncate_seq,
 983                                         truncate_size, false);
 984                 if (IS_ERR(req)) {
 985                         req = ceph_osdc_new_request(&fsc->client->osdc,
 986                                                 &ci->i_layout, vino,
 987                                                 offset, &len, 0,
 988                                                 min(num_ops,
 989                                                     CEPH_OSD_SLAB_OPS),
 990                                                 CEPH_OSD_OP_WRITE,
 991                                                 CEPH_OSD_FLAG_WRITE |
 992                                                 CEPH_OSD_FLAG_ONDISK,
 993                                                 snapc, truncate_seq,
 994                                                 truncate_size, true);
 995                         BUG_ON(IS_ERR(req));
 996                 }
 997                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
 998                              PAGE_SIZE - offset);
 999
1000                 req->r_callback = writepages_finish;
1001                 req->r_inode = inode;
1002
1003                 /* Format the osd request message and submit the write */
1004                 len = 0;
1005                 data_pages = pages;
1006                 op_idx = 0;
1007                 for (i = 0; i < locked_pages; i++) {
1008                         u64 cur_offset = page_offset(pages[i]);
1009                         if (offset + len != cur_offset) {
1010                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1011                                         break;
1012                                 osd_req_op_extent_dup_last(req, op_idx,
1013                                                            cur_offset - offset);
1014                                 dout("writepages got pages at %llu~%llu\n",
1015                                      offset, len);
1016                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1017                                                         data_pages, len, 0,
1018                                                         !!pool, false);
1019                                 osd_req_op_extent_update(req, op_idx, len);
1020
1021                                 len = 0;
1022                                 offset = cur_offset;
1023                                 data_pages = pages + i;
1024                                 op_idx++;
1025                         }
1026
1027                         set_page_writeback(pages[i]);
1028                         len += PAGE_SIZE;
1029                 }
1030
1031                 if (snap_size != -1) {
1032                         len = min(len, snap_size - offset);
1033                 } else if (i == locked_pages) {
1034                         /* writepages_finish() clears writeback pages
1035                          * according to the data length, so make sure
1036                          * data length covers all locked pages */
1037                         u64 min_len = len + 1 - PAGE_SIZE;
1038                         len = min(len, (u64)i_size_read(inode) - offset);
1039                         len = max(len, min_len);
1040                 }
1041                 dout("writepages got pages at %llu~%llu\n", offset, len);
1042
1043                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1044                                                  0, !!pool, false);
1045                 osd_req_op_extent_update(req, op_idx, len);
1046
1047                 if (do_sync) {
1048                         op_idx++;
1049                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1050                 }
1051                 BUG_ON(op_idx + 1 != req->r_num_ops);
1052
1053                 pool = NULL;
1054                 if (i < locked_pages) {
1055                         BUG_ON(num_ops <= req->r_num_ops);
1056                         num_ops -= req->r_num_ops;
1057                         num_ops += do_sync;
1058                         locked_pages -= i;
1059
1060                         /* allocate new pages array for next request */
1061                         data_pages = pages;
1062                         pages = kmalloc(locked_pages * sizeof (*pages),
1063                                         GFP_NOFS);
1064                         if (!pages) {
1065                                 pool = fsc->wb_pagevec_pool;
1066                                 pages = mempool_alloc(pool, GFP_NOFS);
1067                                 BUG_ON(!pages);
1068                         }
1069                         memcpy(pages, data_pages + i,
1070                                locked_pages * sizeof(*pages));
1071                         memset(data_pages + i, 0,
1072                                locked_pages * sizeof(*pages));
1073                 } else {
1074                         BUG_ON(num_ops != req->r_num_ops);
1075                         index = pages[i - 1]->index + 1;
1076                         /* request message now owns the pages array */
1077                         pages = NULL;
1078                 }
1079
1080                 req->r_mtime = inode->i_mtime;
1081                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1082                 BUG_ON(rc);
1083                 req = NULL;
1084
1085                 wbc->nr_to_write -= i;
1086                 if (pages)
1087                         goto new_request;
1088
1089                 if (wbc->nr_to_write <= 0)
1090                         done = 1;
1091
1092 release_pvec_pages:
1093                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1094                      pvec.nr ? pvec.pages[0] : NULL);
1095                 pagevec_release(&pvec);
1096
1097                 if (locked_pages && !done)
1098                         goto retry;
1099         }
1100
1101         if (should_loop && !done) {
1102                 /* more to do; loop back to beginning of file */
1103                 dout("writepages looping back to beginning of file\n");
1104                 should_loop = 0;
1105                 index = 0;
1106                 goto retry;
1107         }
1108
1109         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1110                 mapping->writeback_index = index;
1111
1112 out:
1113         ceph_osdc_put_request(req);
1114         ceph_put_snap_context(snapc);
1115         dout("writepages done, rc = %d\n", rc);
1116         return rc;
1117 }
1118
1119
1120
1121 /*
1122  * See if a given @snapc is either writeable, or already written.
1123  */
1124 static int context_is_writeable_or_written(struct inode *inode,
1125                                            struct ceph_snap_context *snapc)
1126 {
1127         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1128         int ret = !oldest || snapc->seq <= oldest->seq;
1129
1130         ceph_put_snap_context(oldest);
1131         return ret;
1132 }
1133
1134 /*
1135  * We are only allowed to write into/dirty the page if the page is
1136  * clean, or already dirty within the same snap context.
1137  *
1138  * called with page locked.
1139  * return success with page locked,
1140  * or any failure (incl -EAGAIN) with page unlocked.
1141  */
1142 static int ceph_update_writeable_page(struct file *file,
1143                             loff_t pos, unsigned len,
1144                             struct page *page)
1145 {
1146         struct inode *inode = file_inode(file);
1147         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1148         struct ceph_inode_info *ci = ceph_inode(inode);
1149         loff_t page_off = pos & PAGE_MASK;
1150         int pos_in_page = pos & ~PAGE_MASK;
1151         int end_in_page = pos_in_page + len;
1152         loff_t i_size;
1153         int r;
1154         struct ceph_snap_context *snapc, *oldest;
1155
1156         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1157                 dout(" page %p forced umount\n", page);
1158                 unlock_page(page);
1159                 return -EIO;
1160         }
1161
1162 retry_locked:
1163         /* writepages currently holds page lock, but if we change that later, */
1164         wait_on_page_writeback(page);
1165
1166         snapc = page_snap_context(page);
1167         if (snapc && snapc != ci->i_head_snapc) {
1168                 /*
1169                  * this page is already dirty in another (older) snap
1170                  * context!  is it writeable now?
1171                  */
1172                 oldest = get_oldest_context(inode, NULL);
1173
1174                 if (snapc->seq > oldest->seq) {
1175                         ceph_put_snap_context(oldest);
1176                         dout(" page %p snapc %p not current or oldest\n",
1177                              page, snapc);
1178                         /*
1179                          * queue for writeback, and wait for snapc to
1180                          * be writeable or written
1181                          */
1182                         snapc = ceph_get_snap_context(snapc);
1183                         unlock_page(page);
1184                         ceph_queue_writeback(inode);
1185                         r = wait_event_killable(ci->i_cap_wq,
1186                                context_is_writeable_or_written(inode, snapc));
1187                         ceph_put_snap_context(snapc);
1188                         if (r == -ERESTARTSYS)
1189                                 return r;
1190                         return -EAGAIN;
1191                 }
1192                 ceph_put_snap_context(oldest);
1193
1194                 /* yay, writeable, do it now (without dropping page lock) */
1195                 dout(" page %p snapc %p not current, but oldest\n",
1196                      page, snapc);
1197                 if (!clear_page_dirty_for_io(page))
1198                         goto retry_locked;
1199                 r = writepage_nounlock(page, NULL);
1200                 if (r < 0)
1201                         goto fail_nosnap;
1202                 goto retry_locked;
1203         }
1204
1205         if (PageUptodate(page)) {
1206                 dout(" page %p already uptodate\n", page);
1207                 return 0;
1208         }
1209
1210         /* full page? */
1211         if (pos_in_page == 0 && len == PAGE_SIZE)
1212                 return 0;
1213
1214         /* past end of file? */
1215         i_size = i_size_read(inode);
1216
1217         if (page_off >= i_size ||
1218             (pos_in_page == 0 && (pos+len) >= i_size &&
1219              end_in_page - pos_in_page != PAGE_SIZE)) {
1220                 dout(" zeroing %p 0 - %d and %d - %d\n",
1221                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1222                 zero_user_segments(page,
1223                                    0, pos_in_page,
1224                                    end_in_page, PAGE_SIZE);
1225                 return 0;
1226         }
1227
1228         /* we need to read it. */
1229         r = readpage_nounlock(file, page);
1230         if (r < 0)
1231                 goto fail_nosnap;
1232         goto retry_locked;
1233 fail_nosnap:
1234         unlock_page(page);
1235         return r;
1236 }
1237
1238 /*
1239  * We are only allowed to write into/dirty the page if the page is
1240  * clean, or already dirty within the same snap context.
1241  */
1242 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1243                             loff_t pos, unsigned len, unsigned flags,
1244                             struct page **pagep, void **fsdata)
1245 {
1246         struct inode *inode = file_inode(file);
1247         struct page *page;
1248         pgoff_t index = pos >> PAGE_SHIFT;
1249         int r;
1250
1251         do {
1252                 /* get a page */
1253                 page = grab_cache_page_write_begin(mapping, index, 0);
1254                 if (!page)
1255                         return -ENOMEM;
1256
1257                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1258                      inode, page, (int)pos, (int)len);
1259
1260                 r = ceph_update_writeable_page(file, pos, len, page);
1261                 if (r < 0)
1262                         put_page(page);
1263                 else
1264                         *pagep = page;
1265         } while (r == -EAGAIN);
1266
1267         return r;
1268 }
1269
1270 /*
1271  * we don't do anything in here that simple_write_end doesn't do
1272  * except adjust dirty page accounting
1273  */
1274 static int ceph_write_end(struct file *file, struct address_space *mapping,
1275                           loff_t pos, unsigned len, unsigned copied,
1276                           struct page *page, void *fsdata)
1277 {
1278         struct inode *inode = file_inode(file);
1279         unsigned from = pos & (PAGE_SIZE - 1);
1280         int check_cap = 0;
1281
1282         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1283              inode, page, (int)pos, (int)copied, (int)len);
1284
1285         /* zero the stale part of the page if we did a short copy */
1286         if (copied < len)
1287                 zero_user_segment(page, from+copied, len);
1288
1289         /* did file size increase? */
1290         if (pos+copied > i_size_read(inode))
1291                 check_cap = ceph_inode_set_size(inode, pos+copied);
1292
1293         if (!PageUptodate(page))
1294                 SetPageUptodate(page);
1295
1296         set_page_dirty(page);
1297
1298         unlock_page(page);
1299         put_page(page);
1300
1301         if (check_cap)
1302                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1303
1304         return copied;
1305 }
1306
1307 /*
1308  * we set .direct_IO to indicate direct io is supported, but since we
1309  * intercept O_DIRECT reads and writes early, this function should
1310  * never get called.
1311  */
1312 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1313 {
1314         WARN_ON(1);
1315         return -EINVAL;
1316 }
1317
1318 const struct address_space_operations ceph_aops = {
1319         .readpage = ceph_readpage,
1320         .readpages = ceph_readpages,
1321         .writepage = ceph_writepage,
1322         .writepages = ceph_writepages_start,
1323         .write_begin = ceph_write_begin,
1324         .write_end = ceph_write_end,
1325         .set_page_dirty = ceph_set_page_dirty,
1326         .invalidatepage = ceph_invalidatepage,
1327         .releasepage = ceph_releasepage,
1328         .direct_IO = ceph_direct_io,
1329 };
1330
1331 static void ceph_block_sigs(sigset_t *oldset)
1332 {
1333         sigset_t mask;
1334         siginitsetinv(&mask, sigmask(SIGKILL));
1335         sigprocmask(SIG_BLOCK, &mask, oldset);
1336 }
1337
1338 static void ceph_restore_sigs(sigset_t *oldset)
1339 {
1340         sigprocmask(SIG_SETMASK, oldset, NULL);
1341 }
1342
1343 /*
1344  * vm ops
1345  */
1346 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1347 {
1348         struct inode *inode = file_inode(vma->vm_file);
1349         struct ceph_inode_info *ci = ceph_inode(inode);
1350         struct ceph_file_info *fi = vma->vm_file->private_data;
1351         struct page *pinned_page = NULL;
1352         loff_t off = vmf->pgoff << PAGE_SHIFT;
1353         int want, got, ret;
1354         sigset_t oldset;
1355
1356         ceph_block_sigs(&oldset);
1357
1358         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1359              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1360         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1361                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1362         else
1363                 want = CEPH_CAP_FILE_CACHE;
1364
1365         got = 0;
1366         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1367         if (ret < 0)
1368                 goto out_restore;
1369
1370         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1371              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1372
1373         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1374             ci->i_inline_version == CEPH_INLINE_NONE)
1375                 ret = filemap_fault(vma, vmf);
1376         else
1377                 ret = -EAGAIN;
1378
1379         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1380              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1381         if (pinned_page)
1382                 put_page(pinned_page);
1383         ceph_put_cap_refs(ci, got);
1384
1385         if (ret != -EAGAIN)
1386                 goto out_restore;
1387
1388         /* read inline data */
1389         if (off >= PAGE_SIZE) {
1390                 /* does not support inline data > PAGE_SIZE */
1391                 ret = VM_FAULT_SIGBUS;
1392         } else {
1393                 int ret1;
1394                 struct address_space *mapping = inode->i_mapping;
1395                 struct page *page = find_or_create_page(mapping, 0,
1396                                                 mapping_gfp_constraint(mapping,
1397                                                 ~__GFP_FS));
1398                 if (!page) {
1399                         ret = VM_FAULT_OOM;
1400                         goto out_inline;
1401                 }
1402                 ret1 = __ceph_do_getattr(inode, page,
1403                                          CEPH_STAT_CAP_INLINE_DATA, true);
1404                 if (ret1 < 0 || off >= i_size_read(inode)) {
1405                         unlock_page(page);
1406                         put_page(page);
1407                         if (ret1 < 0)
1408                                 ret = ret1;
1409                         else
1410                                 ret = VM_FAULT_SIGBUS;
1411                         goto out_inline;
1412                 }
1413                 if (ret1 < PAGE_SIZE)
1414                         zero_user_segment(page, ret1, PAGE_SIZE);
1415                 else
1416                         flush_dcache_page(page);
1417                 SetPageUptodate(page);
1418                 vmf->page = page;
1419                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1420 out_inline:
1421                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1422                      inode, off, (size_t)PAGE_SIZE, ret);
1423         }
1424 out_restore:
1425         ceph_restore_sigs(&oldset);
1426         if (ret < 0)
1427                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1428
1429         return ret;
1430 }
1431
1432 /*
1433  * Reuse write_begin here for simplicity.
1434  */
1435 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1436 {
1437         struct inode *inode = file_inode(vma->vm_file);
1438         struct ceph_inode_info *ci = ceph_inode(inode);
1439         struct ceph_file_info *fi = vma->vm_file->private_data;
1440         struct ceph_cap_flush *prealloc_cf;
1441         struct page *page = vmf->page;
1442         loff_t off = page_offset(page);
1443         loff_t size = i_size_read(inode);
1444         size_t len;
1445         int want, got, ret;
1446         sigset_t oldset;
1447
1448         prealloc_cf = ceph_alloc_cap_flush();
1449         if (!prealloc_cf)
1450                 return VM_FAULT_OOM;
1451
1452         ceph_block_sigs(&oldset);
1453
1454         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1455                 struct page *locked_page = NULL;
1456                 if (off == 0) {
1457                         lock_page(page);
1458                         locked_page = page;
1459                 }
1460                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1461                 if (locked_page)
1462                         unlock_page(locked_page);
1463                 if (ret < 0)
1464                         goto out_free;
1465         }
1466
1467         if (off + PAGE_SIZE <= size)
1468                 len = PAGE_SIZE;
1469         else
1470                 len = size & ~PAGE_MASK;
1471
1472         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1473              inode, ceph_vinop(inode), off, len, size);
1474         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1475                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1476         else
1477                 want = CEPH_CAP_FILE_BUFFER;
1478
1479         got = 0;
1480         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1481                             &got, NULL);
1482         if (ret < 0)
1483                 goto out_free;
1484
1485         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1486              inode, off, len, ceph_cap_string(got));
1487
1488         /* Update time before taking page lock */
1489         file_update_time(vma->vm_file);
1490
1491         do {
1492                 lock_page(page);
1493
1494                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1495                         unlock_page(page);
1496                         ret = VM_FAULT_NOPAGE;
1497                         break;
1498                 }
1499
1500                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1501                 if (ret >= 0) {
1502                         /* success.  we'll keep the page locked. */
1503                         set_page_dirty(page);
1504                         ret = VM_FAULT_LOCKED;
1505                 }
1506         } while (ret == -EAGAIN);
1507
1508         if (ret == VM_FAULT_LOCKED ||
1509             ci->i_inline_version != CEPH_INLINE_NONE) {
1510                 int dirty;
1511                 spin_lock(&ci->i_ceph_lock);
1512                 ci->i_inline_version = CEPH_INLINE_NONE;
1513                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1514                                                &prealloc_cf);
1515                 spin_unlock(&ci->i_ceph_lock);
1516                 if (dirty)
1517                         __mark_inode_dirty(inode, dirty);
1518         }
1519
1520         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1521              inode, off, len, ceph_cap_string(got), ret);
1522         ceph_put_cap_refs(ci, got);
1523 out_free:
1524         ceph_restore_sigs(&oldset);
1525         ceph_free_cap_flush(prealloc_cf);
1526         if (ret < 0)
1527                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1528         return ret;
1529 }
1530
1531 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1532                            char *data, size_t len)
1533 {
1534         struct address_space *mapping = inode->i_mapping;
1535         struct page *page;
1536
1537         if (locked_page) {
1538                 page = locked_page;
1539         } else {
1540                 if (i_size_read(inode) == 0)
1541                         return;
1542                 page = find_or_create_page(mapping, 0,
1543                                            mapping_gfp_constraint(mapping,
1544                                            ~__GFP_FS));
1545                 if (!page)
1546                         return;
1547                 if (PageUptodate(page)) {
1548                         unlock_page(page);
1549                         put_page(page);
1550                         return;
1551                 }
1552         }
1553
1554         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1555              inode, ceph_vinop(inode), len, locked_page);
1556
1557         if (len > 0) {
1558                 void *kaddr = kmap_atomic(page);
1559                 memcpy(kaddr, data, len);
1560                 kunmap_atomic(kaddr);
1561         }
1562
1563         if (page != locked_page) {
1564                 if (len < PAGE_SIZE)
1565                         zero_user_segment(page, len, PAGE_SIZE);
1566                 else
1567                         flush_dcache_page(page);
1568
1569                 SetPageUptodate(page);
1570                 unlock_page(page);
1571                 put_page(page);
1572         }
1573 }
1574
1575 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1576 {
1577         struct inode *inode = file_inode(filp);
1578         struct ceph_inode_info *ci = ceph_inode(inode);
1579         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1580         struct ceph_osd_request *req;
1581         struct page *page = NULL;
1582         u64 len, inline_version;
1583         int err = 0;
1584         bool from_pagecache = false;
1585
1586         spin_lock(&ci->i_ceph_lock);
1587         inline_version = ci->i_inline_version;
1588         spin_unlock(&ci->i_ceph_lock);
1589
1590         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1591              inode, ceph_vinop(inode), inline_version);
1592
1593         if (inline_version == 1 || /* initial version, no data */
1594             inline_version == CEPH_INLINE_NONE)
1595                 goto out;
1596
1597         if (locked_page) {
1598                 page = locked_page;
1599                 WARN_ON(!PageUptodate(page));
1600         } else if (ceph_caps_issued(ci) &
1601                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1602                 page = find_get_page(inode->i_mapping, 0);
1603                 if (page) {
1604                         if (PageUptodate(page)) {
1605                                 from_pagecache = true;
1606                                 lock_page(page);
1607                         } else {
1608                                 put_page(page);
1609                                 page = NULL;
1610                         }
1611                 }
1612         }
1613
1614         if (page) {
1615                 len = i_size_read(inode);
1616                 if (len > PAGE_SIZE)
1617                         len = PAGE_SIZE;
1618         } else {
1619                 page = __page_cache_alloc(GFP_NOFS);
1620                 if (!page) {
1621                         err = -ENOMEM;
1622                         goto out;
1623                 }
1624                 err = __ceph_do_getattr(inode, page,
1625                                         CEPH_STAT_CAP_INLINE_DATA, true);
1626                 if (err < 0) {
1627                         /* no inline data */
1628                         if (err == -ENODATA)
1629                                 err = 0;
1630                         goto out;
1631                 }
1632                 len = err;
1633         }
1634
1635         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1636                                     ceph_vino(inode), 0, &len, 0, 1,
1637                                     CEPH_OSD_OP_CREATE,
1638                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1639                                     NULL, 0, 0, false);
1640         if (IS_ERR(req)) {
1641                 err = PTR_ERR(req);
1642                 goto out;
1643         }
1644
1645         req->r_mtime = inode->i_mtime;
1646         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1647         if (!err)
1648                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1649         ceph_osdc_put_request(req);
1650         if (err < 0)
1651                 goto out;
1652
1653         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1654                                     ceph_vino(inode), 0, &len, 1, 3,
1655                                     CEPH_OSD_OP_WRITE,
1656                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1657                                     NULL, ci->i_truncate_seq,
1658                                     ci->i_truncate_size, false);
1659         if (IS_ERR(req)) {
1660                 err = PTR_ERR(req);
1661                 goto out;
1662         }
1663
1664         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1665
1666         {
1667                 __le64 xattr_buf = cpu_to_le64(inline_version);
1668                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1669                                             "inline_version", &xattr_buf,
1670                                             sizeof(xattr_buf),
1671                                             CEPH_OSD_CMPXATTR_OP_GT,
1672                                             CEPH_OSD_CMPXATTR_MODE_U64);
1673                 if (err)
1674                         goto out_put;
1675         }
1676
1677         {
1678                 char xattr_buf[32];
1679                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1680                                          "%llu", inline_version);
1681                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1682                                             "inline_version",
1683                                             xattr_buf, xattr_len, 0, 0);
1684                 if (err)
1685                         goto out_put;
1686         }
1687
1688         req->r_mtime = inode->i_mtime;
1689         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1690         if (!err)
1691                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1692 out_put:
1693         ceph_osdc_put_request(req);
1694         if (err == -ECANCELED)
1695                 err = 0;
1696 out:
1697         if (page && page != locked_page) {
1698                 if (from_pagecache) {
1699                         unlock_page(page);
1700                         put_page(page);
1701                 } else
1702                         __free_pages(page, 0);
1703         }
1704
1705         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1706              inode, ceph_vinop(inode), inline_version, err);
1707         return err;
1708 }
1709
1710 static const struct vm_operations_struct ceph_vmops = {
1711         .fault          = ceph_filemap_fault,
1712         .page_mkwrite   = ceph_page_mkwrite,
1713 };
1714
1715 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1716 {
1717         struct address_space *mapping = file->f_mapping;
1718
1719         if (!mapping->a_ops->readpage)
1720                 return -ENOEXEC;
1721         file_accessed(file);
1722         vma->vm_ops = &ceph_vmops;
1723         return 0;
1724 }
1725
1726 enum {
1727         POOL_READ       = 1,
1728         POOL_WRITE      = 2,
1729 };
1730
1731 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1732                                 s64 pool, struct ceph_string *pool_ns)
1733 {
1734         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1735         struct ceph_mds_client *mdsc = fsc->mdsc;
1736         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1737         struct rb_node **p, *parent;
1738         struct ceph_pool_perm *perm;
1739         struct page **pages;
1740         size_t pool_ns_len;
1741         int err = 0, err2 = 0, have = 0;
1742
1743         down_read(&mdsc->pool_perm_rwsem);
1744         p = &mdsc->pool_perm_tree.rb_node;
1745         while (*p) {
1746                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1747                 if (pool < perm->pool)
1748                         p = &(*p)->rb_left;
1749                 else if (pool > perm->pool)
1750                         p = &(*p)->rb_right;
1751                 else {
1752                         int ret = ceph_compare_string(pool_ns,
1753                                                 perm->pool_ns,
1754                                                 perm->pool_ns_len);
1755                         if (ret < 0)
1756                                 p = &(*p)->rb_left;
1757                         else if (ret > 0)
1758                                 p = &(*p)->rb_right;
1759                         else {
1760                                 have = perm->perm;
1761                                 break;
1762                         }
1763                 }
1764         }
1765         up_read(&mdsc->pool_perm_rwsem);
1766         if (*p)
1767                 goto out;
1768
1769         if (pool_ns)
1770                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1771                      pool, (int)pool_ns->len, pool_ns->str);
1772         else
1773                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1774
1775         down_write(&mdsc->pool_perm_rwsem);
1776         p = &mdsc->pool_perm_tree.rb_node;
1777         parent = NULL;
1778         while (*p) {
1779                 parent = *p;
1780                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1781                 if (pool < perm->pool)
1782                         p = &(*p)->rb_left;
1783                 else if (pool > perm->pool)
1784                         p = &(*p)->rb_right;
1785                 else {
1786                         int ret = ceph_compare_string(pool_ns,
1787                                                 perm->pool_ns,
1788                                                 perm->pool_ns_len);
1789                         if (ret < 0)
1790                                 p = &(*p)->rb_left;
1791                         else if (ret > 0)
1792                                 p = &(*p)->rb_right;
1793                         else {
1794                                 have = perm->perm;
1795                                 break;
1796                         }
1797                 }
1798         }
1799         if (*p) {
1800                 up_write(&mdsc->pool_perm_rwsem);
1801                 goto out;
1802         }
1803
1804         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1805                                          1, false, GFP_NOFS);
1806         if (!rd_req) {
1807                 err = -ENOMEM;
1808                 goto out_unlock;
1809         }
1810
1811         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1812         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1813         rd_req->r_base_oloc.pool = pool;
1814         if (pool_ns)
1815                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1816         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1817
1818         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1819         if (err)
1820                 goto out_unlock;
1821
1822         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1823                                          1, false, GFP_NOFS);
1824         if (!wr_req) {
1825                 err = -ENOMEM;
1826                 goto out_unlock;
1827         }
1828
1829         wr_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ACK;
1830         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1831         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1832         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1833
1834         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1835         if (err)
1836                 goto out_unlock;
1837
1838         /* one page should be large enough for STAT data */
1839         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1840         if (IS_ERR(pages)) {
1841                 err = PTR_ERR(pages);
1842                 goto out_unlock;
1843         }
1844
1845         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1846                                      0, false, true);
1847         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1848
1849         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1850         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1851
1852         if (!err)
1853                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1854         if (!err2)
1855                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1856
1857         if (err >= 0 || err == -ENOENT)
1858                 have |= POOL_READ;
1859         else if (err != -EPERM)
1860                 goto out_unlock;
1861
1862         if (err2 == 0 || err2 == -EEXIST)
1863                 have |= POOL_WRITE;
1864         else if (err2 != -EPERM) {
1865                 err = err2;
1866                 goto out_unlock;
1867         }
1868
1869         pool_ns_len = pool_ns ? pool_ns->len : 0;
1870         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1871         if (!perm) {
1872                 err = -ENOMEM;
1873                 goto out_unlock;
1874         }
1875
1876         perm->pool = pool;
1877         perm->perm = have;
1878         perm->pool_ns_len = pool_ns_len;
1879         if (pool_ns_len > 0)
1880                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1881         perm->pool_ns[pool_ns_len] = 0;
1882
1883         rb_link_node(&perm->node, parent, p);
1884         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1885         err = 0;
1886 out_unlock:
1887         up_write(&mdsc->pool_perm_rwsem);
1888
1889         ceph_osdc_put_request(rd_req);
1890         ceph_osdc_put_request(wr_req);
1891 out:
1892         if (!err)
1893                 err = have;
1894         if (pool_ns)
1895                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1896                      pool, (int)pool_ns->len, pool_ns->str, err);
1897         else
1898                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1899         return err;
1900 }
1901
1902 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1903 {
1904         s64 pool;
1905         struct ceph_string *pool_ns;
1906         int ret, flags;
1907
1908         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1909                                 NOPOOLPERM))
1910                 return 0;
1911
1912         spin_lock(&ci->i_ceph_lock);
1913         flags = ci->i_ceph_flags;
1914         pool = ci->i_layout.pool_id;
1915         spin_unlock(&ci->i_ceph_lock);
1916 check:
1917         if (flags & CEPH_I_POOL_PERM) {
1918                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1919                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1920                              pool);
1921                         return -EPERM;
1922                 }
1923                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1924                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1925                              pool);
1926                         return -EPERM;
1927                 }
1928                 return 0;
1929         }
1930
1931         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1932         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1933         ceph_put_string(pool_ns);
1934         if (ret < 0)
1935                 return ret;
1936
1937         flags = CEPH_I_POOL_PERM;
1938         if (ret & POOL_READ)
1939                 flags |= CEPH_I_POOL_RD;
1940         if (ret & POOL_WRITE)
1941                 flags |= CEPH_I_POOL_WR;
1942
1943         spin_lock(&ci->i_ceph_lock);
1944         if (pool == ci->i_layout.pool_id &&
1945             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
1946                 ci->i_ceph_flags |= flags;
1947         } else {
1948                 pool = ci->i_layout.pool_id;
1949                 flags = ci->i_ceph_flags;
1950         }
1951         spin_unlock(&ci->i_ceph_lock);
1952         goto check;
1953 }
1954
1955 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1956 {
1957         struct ceph_pool_perm *perm;
1958         struct rb_node *n;
1959
1960         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1961                 n = rb_first(&mdsc->pool_perm_tree);
1962                 perm = rb_entry(n, struct ceph_pool_perm, node);
1963                 rb_erase(n, &mdsc->pool_perm_tree);
1964                 kfree(perm);
1965         }
1966 }