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         /*
  91          * Note that we're grabbing a snapc ref here without holding
  92          * any locks!
  93          */
  94         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         if (ci->i_head_snapc == NULL)
  99                 ci->i_head_snapc = ceph_get_snap_context(snapc);
 100         ++ci->i_wrbuffer_ref_head;
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         ret = __set_page_dirty_nobuffers(page);
 121         WARN_ON(!PageLocked(page));
 122         WARN_ON(!page->mapping);
 123
 124         return ret;
 125 }
 126
 127 /*
 128  * If we are truncating the full page (i.e. offset == 0), adjust the
 129  * dirty page counters appropriately.  Only called if there is private
 130  * data on the page.
 131  */
 132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 133                                 unsigned int length)
 134 {
 135         struct inode *inode;
 136         struct ceph_inode_info *ci;
 137         struct ceph_snap_context *snapc = page_snap_context(page);
 138
 139         inode = page->mapping->host;
 140         ci = ceph_inode(inode);
 141
 142         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 143                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 144                      inode, page, page->index, offset, length);
 145                 return;
 146         }
 147
 148         ceph_invalidate_fscache_page(inode, page);
 149
 150         if (!PagePrivate(page))
 151                 return;
 152
 153         /*
 154          * We can get non-dirty pages here due to races between
 155          * set_page_dirty and truncate_complete_page; just spit out a
 156          * warning, in case we end up with accounting problems later.
 157          */
 158         if (!PageDirty(page))
 159                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 160
 161         ClearPageChecked(page);
 162
 163         dout("%p invalidatepage %p idx %lu full dirty page\n",
 164              inode, page, page->index);
 165
 166         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 167         ceph_put_snap_context(snapc);
 168         page->private = 0;
 169         ClearPagePrivate(page);
 170 }
 171
 172 static int ceph_releasepage(struct page *page, gfp_t g)
 173 {
 174         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 175         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 176         WARN_ON(PageDirty(page));
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * read a single page, without unlocking it.
 187  */
 188 static int readpage_nounlock(struct file *filp, struct page *page)
 189 {
 190         struct inode *inode = file_inode(filp);
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_osd_client *osdc =
 193                 &ceph_inode_to_client(inode)->client->osdc;
 194         int err = 0;
 195         u64 len = PAGE_CACHE_SIZE;
 196
 197         err = ceph_readpage_from_fscache(inode, page);
 198
 199         if (err == 0)
 200                 goto out;
 201
 202         dout("readpage inode %p file %p page %p index %lu\n",
 203              inode, filp, page, page->index);
 204         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 205                                   (u64) page_offset(page), &len,
 206                                   ci->i_truncate_seq, ci->i_truncate_size,
 207                                   &page, 1, 0);
 208         if (err == -ENOENT)
 209                 err = 0;
 210         if (err < 0) {
 211                 SetPageError(page);
 212                 ceph_fscache_readpage_cancel(inode, page);
 213                 goto out;
 214         }
 215         if (err < PAGE_CACHE_SIZE)
 216                 /* zero fill remainder of page */
 217                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 218         else
 219                 flush_dcache_page(page);
 220
 221         SetPageUptodate(page);
 222         ceph_readpage_to_fscache(inode, page);
 223
 224 out:
 225         return err < 0 ? err : 0;
 226 }
 227
 228 static int ceph_readpage(struct file *filp, struct page *page)
 229 {
 230         int r = readpage_nounlock(filp, page);
 231         unlock_page(page);
 232         return r;
 233 }
 234
 235 /*
 236  * Finish an async read(ahead) op.
 237  */
 238 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 239 {
 240         struct inode *inode = req->r_inode;
 241         struct ceph_osd_data *osd_data;
 242         int rc = req->r_result;
 243         int bytes = le32_to_cpu(msg->hdr.data_len);
 244         int num_pages;
 245         int i;
 246
 247         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 248
 249         /* unlock all pages, zeroing any data we didn't read */
 250         osd_data = osd_req_op_extent_osd_data(req, 0);
 251         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 252         num_pages = calc_pages_for((u64)osd_data->alignment,
 253                                         (u64)osd_data->length);
 254         for (i = 0; i < num_pages; i++) {
 255                 struct page *page = osd_data->pages[i];
 256
 257                 if (rc < 0)
 258                         goto unlock;
 259                 if (bytes < (int)PAGE_CACHE_SIZE) {
 260                         /* zero (remainder of) page */
 261                         int s = bytes < 0 ? 0 : bytes;
 262                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 263                 }
 264                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 265                      page->index);
 266                 flush_dcache_page(page);
 267                 SetPageUptodate(page);
 268                 ceph_readpage_to_fscache(inode, page);
 269 unlock:
 270                 unlock_page(page);
 271                 page_cache_release(page);
 272                 bytes -= PAGE_CACHE_SIZE;
 273         }
 274         kfree(osd_data->pages);
 275 }
 276
 277 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 278 {
 279         int i;
 280
 281         for (i = 0; i < num_pages; i++)
 282                 unlock_page(pages[i]);
 283 }
 284
 285 /*
 286  * start an async read(ahead) operation.  return nr_pages we submitted
 287  * a read for on success, or negative error code.
 288  */
 289 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 290 {
 291         struct ceph_osd_client *osdc =
 292                 &ceph_inode_to_client(inode)->client->osdc;
 293         struct ceph_inode_info *ci = ceph_inode(inode);
 294         struct page *page = list_entry(page_list->prev, struct page, lru);
 295         struct ceph_vino vino;
 296         struct ceph_osd_request *req;
 297         u64 off;
 298         u64 len;
 299         int i;
 300         struct page **pages;
 301         pgoff_t next_index;
 302         int nr_pages = 0;
 303         int ret;
 304
 305         off = (u64) page_offset(page);
 306
 307         /* count pages */
 308         next_index = page->index;
 309         list_for_each_entry_reverse(page, page_list, lru) {
 310                 if (page->index != next_index)
 311                         break;
 312                 nr_pages++;
 313                 next_index++;
 314                 if (max && nr_pages == max)
 315                         break;
 316         }
 317         len = nr_pages << PAGE_CACHE_SHIFT;
 318         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 319              off, len);
 320         vino = ceph_vino(inode);
 321         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 322                                     1, CEPH_OSD_OP_READ,
 323                                     CEPH_OSD_FLAG_READ, NULL,
 324                                     ci->i_truncate_seq, ci->i_truncate_size,
 325                                     false);
 326         if (IS_ERR(req))
 327                 return PTR_ERR(req);
 328
 329         /* build page vector */
 330         nr_pages = calc_pages_for(0, len);
 331         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 332         ret = -ENOMEM;
 333         if (!pages)
 334                 goto out;
 335         for (i = 0; i < nr_pages; ++i) {
 336                 page = list_entry(page_list->prev, struct page, lru);
 337                 BUG_ON(PageLocked(page));
 338                 list_del(&page->lru);
 339
 340                 dout("start_read %p adding %p idx %lu\n", inode, page,
 341                      page->index);
 342                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 343                                           GFP_NOFS)) {
 344                         ceph_fscache_uncache_page(inode, page);
 345                         page_cache_release(page);
 346                         dout("start_read %p add_to_page_cache failed %p\n",
 347                              inode, page);
 348                         nr_pages = i;
 349                         goto out_pages;
 350                 }
 351                 pages[i] = page;
 352         }
 353         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 354         req->r_callback = finish_read;
 355         req->r_inode = inode;
 356
 357         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 358
 359         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 360         ret = ceph_osdc_start_request(osdc, req, false);
 361         if (ret < 0)
 362                 goto out_pages;
 363         ceph_osdc_put_request(req);
 364         return nr_pages;
 365
 366 out_pages:
 367         ceph_unlock_page_vector(pages, nr_pages);
 368         ceph_release_page_vector(pages, nr_pages);
 369 out:
 370         ceph_osdc_put_request(req);
 371         return ret;
 372 }
 373
 374
 375 /*
 376  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 377  * the caller (VM) cleans them up.
 378  */
 379 static int ceph_readpages(struct file *file, struct address_space *mapping,
 380                           struct list_head *page_list, unsigned nr_pages)
 381 {
 382         struct inode *inode = file_inode(file);
 383         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 384         int rc = 0;
 385         int max = 0;
 386
 387         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 388                                          &nr_pages);
 389
 390         if (rc == 0)
 391                 goto out;
 392
 393         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 394                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 395                         >> PAGE_SHIFT;
 396
 397         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 398                 file, nr_pages,
 399              max);
 400         while (!list_empty(page_list)) {
 401                 rc = start_read(inode, page_list, max);
 402                 if (rc < 0)
 403                         goto out;
 404                 BUG_ON(rc == 0);
 405         }
 406 out:
 407         ceph_fscache_readpages_cancel(inode, page_list);
 408
 409         dout("readpages %p file %p ret %d\n", inode, file, rc);
 410         return rc;
 411 }
 412
 413 /*
 414  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 415  * only snap context we are allowed to write back.
 416  */
 417 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 418                                                     u64 *snap_size)
 419 {
 420         struct ceph_inode_info *ci = ceph_inode(inode);
 421         struct ceph_snap_context *snapc = NULL;
 422         struct ceph_cap_snap *capsnap = NULL;
 423
 424         spin_lock(&ci->i_ceph_lock);
 425         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 426                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 427                      capsnap->context, capsnap->dirty_pages);
 428                 if (capsnap->dirty_pages) {
 429                         snapc = ceph_get_snap_context(capsnap->context);
 430                         if (snap_size)
 431                                 *snap_size = capsnap->size;
 432                         break;
 433                 }
 434         }
 435         if (!snapc && ci->i_wrbuffer_ref_head) {
 436                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 437                 dout(" head snapc %p has %d dirty pages\n",
 438                      snapc, ci->i_wrbuffer_ref_head);
 439         }
 440         spin_unlock(&ci->i_ceph_lock);
 441         return snapc;
 442 }
 443
 444 /*
 445  * Write a single page, but leave the page locked.
 446  *
 447  * If we get a write error, set the page error bit, but still adjust the
 448  * dirty page accounting (i.e., page is no longer dirty).
 449  */
 450 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 451 {
 452         struct inode *inode;
 453         struct ceph_inode_info *ci;
 454         struct ceph_fs_client *fsc;
 455         struct ceph_osd_client *osdc;
 456         struct ceph_snap_context *snapc, *oldest;
 457         loff_t page_off = page_offset(page);
 458         long writeback_stat;
 459         u64 truncate_size, snap_size = 0;
 460         u32 truncate_seq;
 461         int err = 0, len = PAGE_CACHE_SIZE;
 462
 463         dout("writepage %p idx %lu\n", page, page->index);
 464
 465         if (!page->mapping || !page->mapping->host) {
 466                 dout("writepage %p - no mapping\n", page);
 467                 return -EFAULT;
 468         }
 469         inode = page->mapping->host;
 470         ci = ceph_inode(inode);
 471         fsc = ceph_inode_to_client(inode);
 472         osdc = &fsc->client->osdc;
 473
 474         /* verify this is a writeable snap context */
 475         snapc = page_snap_context(page);
 476         if (snapc == NULL) {
 477                 dout("writepage %p page %p not dirty?\n", inode, page);
 478                 goto out;
 479         }
 480         oldest = get_oldest_context(inode, &snap_size);
 481         if (snapc->seq > oldest->seq) {
 482                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 483                      inode, page, snapc);
 484                 /* we should only noop if called by kswapd */
 485                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 486                 ceph_put_snap_context(oldest);
 487                 goto out;
 488         }
 489         ceph_put_snap_context(oldest);
 490
 491         spin_lock(&ci->i_ceph_lock);
 492         truncate_seq = ci->i_truncate_seq;
 493         truncate_size = ci->i_truncate_size;
 494         if (!snap_size)
 495                 snap_size = i_size_read(inode);
 496         spin_unlock(&ci->i_ceph_lock);
 497
 498         /* is this a partial page at end of file? */
 499         if (page_off >= snap_size) {
 500                 dout("%p page eof %llu\n", page, snap_size);
 501                 goto out;
 502         }
 503         if (snap_size < page_off + len)
 504                 len = snap_size - page_off;
 505
 506         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 507              inode, page, page->index, page_off, len, snapc);
 508
 509         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 510         if (writeback_stat >
 511             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 512                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 513
 514         ceph_readpage_to_fscache(inode, page);
 515
 516         set_page_writeback(page);
 517         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 518                                    &ci->i_layout, snapc,
 519                                    page_off, len,
 520                                    truncate_seq, truncate_size,
 521                                    &inode->i_mtime, &page, 1);
 522         if (err < 0) {
 523                 dout("writepage setting page/mapping error %d %p\n", err, page);
 524                 SetPageError(page);
 525                 mapping_set_error(&inode->i_data, err);
 526                 if (wbc)
 527                         wbc->pages_skipped++;
 528         } else {
 529                 dout("writepage cleaned page %p\n", page);
 530                 err = 0;  /* vfs expects us to return 0 */
 531         }
 532         page->private = 0;
 533         ClearPagePrivate(page);
 534         end_page_writeback(page);
 535         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 536         ceph_put_snap_context(snapc);  /* page's reference */
 537 out:
 538         return err;
 539 }
 540
 541 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 542 {
 543         int err;
 544         struct inode *inode = page->mapping->host;
 545         BUG_ON(!inode);
 546         ihold(inode);
 547         err = writepage_nounlock(page, wbc);
 548         unlock_page(page);
 549         iput(inode);
 550         return err;
 551 }
 552
 553
 554 /*
 555  * lame release_pages helper.  release_pages() isn't exported to
 556  * modules.
 557  */
 558 static void ceph_release_pages(struct page **pages, int num)
 559 {
 560         struct pagevec pvec;
 561         int i;
 562
 563         pagevec_init(&pvec, 0);
 564         for (i = 0; i < num; i++) {
 565                 if (pagevec_add(&pvec, pages[i]) == 0)
 566                         pagevec_release(&pvec);
 567         }
 568         pagevec_release(&pvec);
 569 }
 570
 571 /*
 572  * async writeback completion handler.
 573  *
 574  * If we get an error, set the mapping error bit, but not the individual
 575  * page error bits.
 576  */
 577 static void writepages_finish(struct ceph_osd_request *req,
 578                               struct ceph_msg *msg)
 579 {
 580         struct inode *inode = req->r_inode;
 581         struct ceph_inode_info *ci = ceph_inode(inode);
 582         struct ceph_osd_data *osd_data;
 583         unsigned wrote;
 584         struct page *page;
 585         int num_pages;
 586         int i;
 587         struct ceph_snap_context *snapc = req->r_snapc;
 588         struct address_space *mapping = inode->i_mapping;
 589         int rc = req->r_result;
 590         u64 bytes = req->r_ops[0].extent.length;
 591         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 592         long writeback_stat;
 593         unsigned issued = ceph_caps_issued(ci);
 594
 595         osd_data = osd_req_op_extent_osd_data(req, 0);
 596         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 597         num_pages = calc_pages_for((u64)osd_data->alignment,
 598                                         (u64)osd_data->length);
 599         if (rc >= 0) {
 600                 /*
 601                  * Assume we wrote the pages we originally sent.  The
 602                  * osd might reply with fewer pages if our writeback
 603                  * raced with a truncation and was adjusted at the osd,
 604                  * so don't believe the reply.
 605                  */
 606                 wrote = num_pages;
 607         } else {
 608                 wrote = 0;
 609                 mapping_set_error(mapping, rc);
 610         }
 611         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 612              inode, rc, bytes, wrote);
 613
 614         /* clean all pages */
 615         for (i = 0; i < num_pages; i++) {
 616                 page = osd_data->pages[i];
 617                 BUG_ON(!page);
 618                 WARN_ON(!PageUptodate(page));
 619
 620                 writeback_stat =
 621                         atomic_long_dec_return(&fsc->writeback_count);
 622                 if (writeback_stat <
 623                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 624                         clear_bdi_congested(&fsc->backing_dev_info,
 625                                             BLK_RW_ASYNC);
 626
 627                 ceph_put_snap_context(page_snap_context(page));
 628                 page->private = 0;
 629                 ClearPagePrivate(page);
 630                 dout("unlocking %d %p\n", i, page);
 631                 end_page_writeback(page);
 632
 633                 /*
 634                  * We lost the cache cap, need to truncate the page before
 635                  * it is unlocked, otherwise we'd truncate it later in the
 636                  * page truncation thread, possibly losing some data that
 637                  * raced its way in
 638                  */
 639                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 640                         generic_error_remove_page(inode->i_mapping, page);
 641
 642                 unlock_page(page);
 643         }
 644         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 645         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 646
 647         ceph_release_pages(osd_data->pages, num_pages);
 648         if (osd_data->pages_from_pool)
 649                 mempool_free(osd_data->pages,
 650                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 651         else
 652                 kfree(osd_data->pages);
 653         ceph_osdc_put_request(req);
 654 }
 655
 656 /*
 657  * initiate async writeback
 658  */
 659 static int ceph_writepages_start(struct address_space *mapping,
 660                                  struct writeback_control *wbc)
 661 {
 662         struct inode *inode = mapping->host;
 663         struct ceph_inode_info *ci = ceph_inode(inode);
 664         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 665         struct ceph_vino vino = ceph_vino(inode);
 666         pgoff_t index, start, end;
 667         int range_whole = 0;
 668         int should_loop = 1;
 669         pgoff_t max_pages = 0, max_pages_ever = 0;
 670         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 671         struct pagevec pvec;
 672         int done = 0;
 673         int rc = 0;
 674         unsigned wsize = 1 << inode->i_blkbits;
 675         struct ceph_osd_request *req = NULL;
 676         int do_sync;
 677         u64 truncate_size, snap_size;
 678         u32 truncate_seq;
 679
 680         /*
 681          * Include a 'sync' in the OSD request if this is a data
 682          * integrity write (e.g., O_SYNC write or fsync()), or if our
 683          * cap is being revoked.
 684          */
 685         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 686                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 687                 do_sync = 1;
 688         dout("writepages_start %p dosync=%d (mode=%s)\n",
 689              inode, do_sync,
 690              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 691              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 692
 693         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 694                 pr_warn("writepage_start %p on forced umount\n", inode);
 695                 return -EIO; /* we're in a forced umount, don't write! */
 696         }
 697         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 698                 wsize = fsc->mount_options->wsize;
 699         if (wsize < PAGE_CACHE_SIZE)
 700                 wsize = PAGE_CACHE_SIZE;
 701         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 702
 703         pagevec_init(&pvec, 0);
 704
 705         /* where to start/end? */
 706         if (wbc->range_cyclic) {
 707                 start = mapping->writeback_index; /* Start from prev offset */
 708                 end = -1;
 709                 dout(" cyclic, start at %lu\n", start);
 710         } else {
 711                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 712                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 713                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 714                         range_whole = 1;
 715                 should_loop = 0;
 716                 dout(" not cyclic, %lu to %lu\n", start, end);
 717         }
 718         index = start;
 719
 720 retry:
 721         /* find oldest snap context with dirty data */
 722         ceph_put_snap_context(snapc);
 723         snap_size = 0;
 724         snapc = get_oldest_context(inode, &snap_size);
 725         if (!snapc) {
 726                 /* hmm, why does writepages get called when there
 727                    is no dirty data? */
 728                 dout(" no snap context with dirty data?\n");
 729                 goto out;
 730         }
 731         if (snap_size == 0)
 732                 snap_size = i_size_read(inode);
 733         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 734              snapc, snapc->seq, snapc->num_snaps);
 735
 736         spin_lock(&ci->i_ceph_lock);
 737         truncate_seq = ci->i_truncate_seq;
 738         truncate_size = ci->i_truncate_size;
 739         if (!snap_size)
 740                 snap_size = i_size_read(inode);
 741         spin_unlock(&ci->i_ceph_lock);
 742
 743         if (last_snapc && snapc != last_snapc) {
 744                 /* if we switched to a newer snapc, restart our scan at the
 745                  * start of the original file range. */
 746                 dout("  snapc differs from last pass, restarting at %lu\n",
 747                      index);
 748                 index = start;
 749         }
 750         last_snapc = snapc;
 751
 752         while (!done && index <= end) {
 753                 int num_ops = do_sync ? 2 : 1;
 754                 unsigned i;
 755                 int first;
 756                 pgoff_t next;
 757                 int pvec_pages, locked_pages;
 758                 struct page **pages = NULL;
 759                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 760                 struct page *page;
 761                 int want;
 762                 u64 offset, len;
 763                 long writeback_stat;
 764
 765                 next = 0;
 766                 locked_pages = 0;
 767                 max_pages = max_pages_ever;
 768
 769 get_more_pages:
 770                 first = -1;
 771                 want = min(end - index,
 772                            min((pgoff_t)PAGEVEC_SIZE,
 773                                max_pages - (pgoff_t)locked_pages) - 1)
 774                         + 1;
 775                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 776                                                 PAGECACHE_TAG_DIRTY,
 777                                                 want);
 778                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 779                 if (!pvec_pages && !locked_pages)
 780                         break;
 781                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 782                         page = pvec.pages[i];
 783                         dout("? %p idx %lu\n", page, page->index);
 784                         if (locked_pages == 0)
 785                                 lock_page(page);  /* first page */
 786                         else if (!trylock_page(page))
 787                                 break;
 788
 789                         /* only dirty pages, or our accounting breaks */
 790                         if (unlikely(!PageDirty(page)) ||
 791                             unlikely(page->mapping != mapping)) {
 792                                 dout("!dirty or !mapping %p\n", page);
 793                                 unlock_page(page);
 794                                 break;
 795                         }
 796                         if (!wbc->range_cyclic && page->index > end) {
 797                                 dout("end of range %p\n", page);
 798                                 done = 1;
 799                                 unlock_page(page);
 800                                 break;
 801                         }
 802                         if (next && (page->index != next)) {
 803                                 dout("not consecutive %p\n", page);
 804                                 unlock_page(page);
 805                                 break;
 806                         }
 807                         if (wbc->sync_mode != WB_SYNC_NONE) {
 808                                 dout("waiting on writeback %p\n", page);
 809                                 wait_on_page_writeback(page);
 810                         }
 811                         if (page_offset(page) >= snap_size) {
 812                                 dout("%p page eof %llu\n", page, snap_size);
 813                                 done = 1;
 814                                 unlock_page(page);
 815                                 break;
 816                         }
 817                         if (PageWriteback(page)) {
 818                                 dout("%p under writeback\n", page);
 819                                 unlock_page(page);
 820                                 break;
 821                         }
 822
 823                         /* only if matching snap context */
 824                         pgsnapc = page_snap_context(page);
 825                         if (pgsnapc->seq > snapc->seq) {
 826                                 dout("page snapc %p %lld > oldest %p %lld\n",
 827                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 828                                 unlock_page(page);
 829                                 if (!locked_pages)
 830                                         continue; /* keep looking for snap */
 831                                 break;
 832                         }
 833
 834                         if (!clear_page_dirty_for_io(page)) {
 835                                 dout("%p !clear_page_dirty_for_io\n", page);
 836                                 unlock_page(page);
 837                                 break;
 838                         }
 839
 840                         /*
 841                          * We have something to write.  If this is
 842                          * the first locked page this time through,
 843                          * allocate an osd request and a page array
 844                          * that it will use.
 845                          */
 846                         if (locked_pages == 0) {
 847                                 BUG_ON(pages);
 848                                 /* prepare async write request */
 849                                 offset = (u64)page_offset(page);
 850                                 len = wsize;
 851                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 852                                                         &ci->i_layout, vino,
 853                                                         offset, &len, num_ops,
 854                                                         CEPH_OSD_OP_WRITE,
 855                                                         CEPH_OSD_FLAG_WRITE |
 856                                                         CEPH_OSD_FLAG_ONDISK,
 857                                                         snapc, truncate_seq,
 858                                                         truncate_size, true);
 859                                 if (IS_ERR(req)) {
 860                                         rc = PTR_ERR(req);
 861                                         unlock_page(page);
 862                                         break;
 863                                 }
 864
 865                                 req->r_callback = writepages_finish;
 866                                 req->r_inode = inode;
 867
 868                                 max_pages = calc_pages_for(0, (u64)len);
 869                                 pages = kmalloc(max_pages * sizeof (*pages),
 870                                                 GFP_NOFS);
 871                                 if (!pages) {
 872                                         pool = fsc->wb_pagevec_pool;
 873                                         pages = mempool_alloc(pool, GFP_NOFS);
 874                                         BUG_ON(!pages);
 875                                 }
 876                         }
 877
 878                         /* note position of first page in pvec */
 879                         if (first < 0)
 880                                 first = i;
 881                         dout("%p will write page %p idx %lu\n",
 882                              inode, page, page->index);
 883
 884                         writeback_stat =
 885                                atomic_long_inc_return(&fsc->writeback_count);
 886                         if (writeback_stat > CONGESTION_ON_THRESH(
 887                                     fsc->mount_options->congestion_kb)) {
 888                                 set_bdi_congested(&fsc->backing_dev_info,
 889                                                   BLK_RW_ASYNC);
 890                         }
 891
 892                         set_page_writeback(page);
 893                         pages[locked_pages] = page;
 894                         locked_pages++;
 895                         next = page->index + 1;
 896                 }
 897
 898                 /* did we get anything? */
 899                 if (!locked_pages)
 900                         goto release_pvec_pages;
 901                 if (i) {
 902                         int j;
 903                         BUG_ON(!locked_pages || first < 0);
 904
 905                         if (pvec_pages && i == pvec_pages &&
 906                             locked_pages < max_pages) {
 907                                 dout("reached end pvec, trying for more\n");
 908                                 pagevec_reinit(&pvec);
 909                                 goto get_more_pages;
 910                         }
 911
 912                         /* shift unused pages over in the pvec...  we
 913                          * will need to release them below. */
 914                         for (j = i; j < pvec_pages; j++) {
 915                                 dout(" pvec leftover page %p\n",
 916                                      pvec.pages[j]);
 917                                 pvec.pages[j-i+first] = pvec.pages[j];
 918                         }
 919                         pvec.nr -= i-first;
 920                 }
 921
 922                 /* Format the osd request message and submit the write */
 923
 924                 offset = page_offset(pages[0]);
 925                 len = min(snap_size - offset,
 926                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 927                 dout("writepages got %d pages at %llu~%llu\n",
 928                      locked_pages, offset, len);
 929
 930                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 931                                                         !!pool, false);
 932
 933                 pages = NULL;   /* request message now owns the pages array */
 934                 pool = NULL;
 935
 936                 /* Update the write op length in case we changed it */
 937
 938                 osd_req_op_extent_update(req, 0, len);
 939
 940                 vino = ceph_vino(inode);
 941                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 942                                         &inode->i_mtime);
 943
 944                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 945                 BUG_ON(rc);
 946                 req = NULL;
 947
 948                 /* continue? */
 949                 index = next;
 950                 wbc->nr_to_write -= locked_pages;
 951                 if (wbc->nr_to_write <= 0)
 952                         done = 1;
 953
 954 release_pvec_pages:
 955                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 956                      pvec.nr ? pvec.pages[0] : NULL);
 957                 pagevec_release(&pvec);
 958
 959                 if (locked_pages && !done)
 960                         goto retry;
 961         }
 962
 963         if (should_loop && !done) {
 964                 /* more to do; loop back to beginning of file */
 965                 dout("writepages looping back to beginning of file\n");
 966                 should_loop = 0;
 967                 index = 0;
 968                 goto retry;
 969         }
 970
 971         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 972                 mapping->writeback_index = index;
 973
 974 out:
 975         if (req)
 976                 ceph_osdc_put_request(req);
 977         ceph_put_snap_context(snapc);
 978         dout("writepages done, rc = %d\n", rc);
 979         return rc;
 980 }
 981
 982
 983
 984 /*
 985  * See if a given @snapc is either writeable, or already written.
 986  */
 987 static int context_is_writeable_or_written(struct inode *inode,
 988                                            struct ceph_snap_context *snapc)
 989 {
 990         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 991         int ret = !oldest || snapc->seq <= oldest->seq;
 992
 993         ceph_put_snap_context(oldest);
 994         return ret;
 995 }
 996
 997 /*
 998  * We are only allowed to write into/dirty the page if the page is
 999  * clean, or already dirty within the same snap context.
1000  *
1001  * called with page locked.
1002  * return success with page locked,
1003  * or any failure (incl -EAGAIN) with page unlocked.
1004  */
1005 static int ceph_update_writeable_page(struct file *file,
1006                             loff_t pos, unsigned len,
1007                             struct page *page)
1008 {
1009         struct inode *inode = file_inode(file);
1010         struct ceph_inode_info *ci = ceph_inode(inode);
1011         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1012         loff_t page_off = pos & PAGE_CACHE_MASK;
1013         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1014         int end_in_page = pos_in_page + len;
1015         loff_t i_size;
1016         int r;
1017         struct ceph_snap_context *snapc, *oldest;
1018
1019 retry_locked:
1020         /* writepages currently holds page lock, but if we change that later, */
1021         wait_on_page_writeback(page);
1022
1023         /* check snap context */
1024         BUG_ON(!ci->i_snap_realm);
1025         down_read(&mdsc->snap_rwsem);
1026         BUG_ON(!ci->i_snap_realm->cached_context);
1027         snapc = page_snap_context(page);
1028         if (snapc && snapc != ci->i_head_snapc) {
1029                 /*
1030                  * this page is already dirty in another (older) snap
1031                  * context!  is it writeable now?
1032                  */
1033                 oldest = get_oldest_context(inode, NULL);
1034                 up_read(&mdsc->snap_rwsem);
1035
1036                 if (snapc->seq > oldest->seq) {
1037                         ceph_put_snap_context(oldest);
1038                         dout(" page %p snapc %p not current or oldest\n",
1039                              page, snapc);
1040                         /*
1041                          * queue for writeback, and wait for snapc to
1042                          * be writeable or written
1043                          */
1044                         snapc = ceph_get_snap_context(snapc);
1045                         unlock_page(page);
1046                         ceph_queue_writeback(inode);
1047                         r = wait_event_interruptible(ci->i_cap_wq,
1048                                context_is_writeable_or_written(inode, snapc));
1049                         ceph_put_snap_context(snapc);
1050                         if (r == -ERESTARTSYS)
1051                                 return r;
1052                         return -EAGAIN;
1053                 }
1054                 ceph_put_snap_context(oldest);
1055
1056                 /* yay, writeable, do it now (without dropping page lock) */
1057                 dout(" page %p snapc %p not current, but oldest\n",
1058                      page, snapc);
1059                 if (!clear_page_dirty_for_io(page))
1060                         goto retry_locked;
1061                 r = writepage_nounlock(page, NULL);
1062                 if (r < 0)
1063                         goto fail_nosnap;
1064                 goto retry_locked;
1065         }
1066
1067         if (PageUptodate(page)) {
1068                 dout(" page %p already uptodate\n", page);
1069                 return 0;
1070         }
1071
1072         /* full page? */
1073         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1074                 return 0;
1075
1076         /* past end of file? */
1077         i_size = inode->i_size;   /* caller holds i_mutex */
1078
1079         if (i_size + len > inode->i_sb->s_maxbytes) {
1080                 /* file is too big */
1081                 r = -EINVAL;
1082                 goto fail;
1083         }
1084
1085         if (page_off >= i_size ||
1086             (pos_in_page == 0 && (pos+len) >= i_size &&
1087              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1088                 dout(" zeroing %p 0 - %d and %d - %d\n",
1089                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1090                 zero_user_segments(page,
1091                                    0, pos_in_page,
1092                                    end_in_page, PAGE_CACHE_SIZE);
1093                 return 0;
1094         }
1095
1096         /* we need to read it. */
1097         up_read(&mdsc->snap_rwsem);
1098         r = readpage_nounlock(file, page);
1099         if (r < 0)
1100                 goto fail_nosnap;
1101         goto retry_locked;
1102
1103 fail:
1104         up_read(&mdsc->snap_rwsem);
1105 fail_nosnap:
1106         unlock_page(page);
1107         return r;
1108 }
1109
1110 /*
1111  * We are only allowed to write into/dirty the page if the page is
1112  * clean, or already dirty within the same snap context.
1113  */
1114 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1115                             loff_t pos, unsigned len, unsigned flags,
1116                             struct page **pagep, void **fsdata)
1117 {
1118         struct inode *inode = file_inode(file);
1119         struct page *page;
1120         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1121         int r;
1122
1123         do {
1124                 /* get a page */
1125                 page = grab_cache_page_write_begin(mapping, index, 0);
1126                 if (!page)
1127                         return -ENOMEM;
1128                 *pagep = page;
1129
1130                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1131                      inode, page, (int)pos, (int)len);
1132
1133                 r = ceph_update_writeable_page(file, pos, len, page);
1134         } while (r == -EAGAIN);
1135
1136         return r;
1137 }
1138
1139 /*
1140  * we don't do anything in here that simple_write_end doesn't do
1141  * except adjust dirty page accounting and drop read lock on
1142  * mdsc->snap_rwsem.
1143  */
1144 static int ceph_write_end(struct file *file, struct address_space *mapping,
1145                           loff_t pos, unsigned len, unsigned copied,
1146                           struct page *page, void *fsdata)
1147 {
1148         struct inode *inode = file_inode(file);
1149         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1150         struct ceph_mds_client *mdsc = fsc->mdsc;
1151         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1152         int check_cap = 0;
1153
1154         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1155              inode, page, (int)pos, (int)copied, (int)len);
1156
1157         /* zero the stale part of the page if we did a short copy */
1158         if (copied < len)
1159                 zero_user_segment(page, from+copied, len);
1160
1161         /* did file size increase? */
1162         /* (no need for i_size_read(); we caller holds i_mutex */
1163         if (pos+copied > inode->i_size)
1164                 check_cap = ceph_inode_set_size(inode, pos+copied);
1165
1166         if (!PageUptodate(page))
1167                 SetPageUptodate(page);
1168
1169         set_page_dirty(page);
1170
1171         unlock_page(page);
1172         up_read(&mdsc->snap_rwsem);
1173         page_cache_release(page);
1174
1175         if (check_cap)
1176                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1177
1178         return copied;
1179 }
1180
1181 /*
1182  * we set .direct_IO to indicate direct io is supported, but since we
1183  * intercept O_DIRECT reads and writes early, this function should
1184  * never get called.
1185  */
1186 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1187                               struct iov_iter *iter,
1188                               loff_t pos)
1189 {
1190         WARN_ON(1);
1191         return -EINVAL;
1192 }
1193
1194 const struct address_space_operations ceph_aops = {
1195         .readpage = ceph_readpage,
1196         .readpages = ceph_readpages,
1197         .writepage = ceph_writepage,
1198         .writepages = ceph_writepages_start,
1199         .write_begin = ceph_write_begin,
1200         .write_end = ceph_write_end,
1201         .set_page_dirty = ceph_set_page_dirty,
1202         .invalidatepage = ceph_invalidatepage,
1203         .releasepage = ceph_releasepage,
1204         .direct_IO = ceph_direct_io,
1205 };
1206
1207
1208 /*
1209  * vm ops
1210  */
1211 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1212 {
1213         struct inode *inode = file_inode(vma->vm_file);
1214         struct ceph_inode_info *ci = ceph_inode(inode);
1215         struct ceph_file_info *fi = vma->vm_file->private_data;
1216         loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1217         int want, got, ret;
1218
1219         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1220              inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1221         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1222                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1223         else
1224                 want = CEPH_CAP_FILE_CACHE;
1225         while (1) {
1226                 got = 0;
1227                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
1228                 if (ret == 0)
1229                         break;
1230                 if (ret != -ERESTARTSYS) {
1231                         WARN_ON(1);
1232                         return VM_FAULT_SIGBUS;
1233                 }
1234         }
1235         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1236              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1237
1238         ret = filemap_fault(vma, vmf);
1239
1240         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1241              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1242         ceph_put_cap_refs(ci, got);
1243
1244         return ret;
1245 }
1246
1247 /*
1248  * Reuse write_begin here for simplicity.
1249  */
1250 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1251 {
1252         struct inode *inode = file_inode(vma->vm_file);
1253         struct ceph_inode_info *ci = ceph_inode(inode);
1254         struct ceph_file_info *fi = vma->vm_file->private_data;
1255         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1256         struct page *page = vmf->page;
1257         loff_t off = page_offset(page);
1258         loff_t size = i_size_read(inode);
1259         size_t len;
1260         int want, got, ret;
1261
1262         if (off + PAGE_CACHE_SIZE <= size)
1263                 len = PAGE_CACHE_SIZE;
1264         else
1265                 len = size & ~PAGE_CACHE_MASK;
1266
1267         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1268              inode, ceph_vinop(inode), off, len, size);
1269         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1270                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1271         else
1272                 want = CEPH_CAP_FILE_BUFFER;
1273         while (1) {
1274                 got = 0;
1275                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, off + len);
1276                 if (ret == 0)
1277                         break;
1278                 if (ret != -ERESTARTSYS) {
1279                         WARN_ON(1);
1280                         return VM_FAULT_SIGBUS;
1281                 }
1282         }
1283         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1284              inode, off, len, ceph_cap_string(got));
1285
1286         /* Update time before taking page lock */
1287         file_update_time(vma->vm_file);
1288
1289         lock_page(page);
1290
1291         ret = VM_FAULT_NOPAGE;
1292         if ((off > size) ||
1293             (page->mapping != inode->i_mapping))
1294                 goto out;
1295
1296         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1297         if (ret == 0) {
1298                 /* success.  we'll keep the page locked. */
1299                 set_page_dirty(page);
1300                 up_read(&mdsc->snap_rwsem);
1301                 ret = VM_FAULT_LOCKED;
1302         } else {
1303                 if (ret == -ENOMEM)
1304                         ret = VM_FAULT_OOM;
1305                 else
1306                         ret = VM_FAULT_SIGBUS;
1307         }
1308 out:
1309         if (ret != VM_FAULT_LOCKED) {
1310                 unlock_page(page);
1311         } else {
1312                 int dirty;
1313                 spin_lock(&ci->i_ceph_lock);
1314                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1315                 spin_unlock(&ci->i_ceph_lock);
1316                 if (dirty)
1317                         __mark_inode_dirty(inode, dirty);
1318         }
1319
1320         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1321              inode, off, len, ceph_cap_string(got), ret);
1322         ceph_put_cap_refs(ci, got);
1323
1324         return ret;
1325 }
1326
1327 static struct vm_operations_struct ceph_vmops = {
1328         .fault          = ceph_filemap_fault,
1329         .page_mkwrite   = ceph_page_mkwrite,
1330         .remap_pages    = generic_file_remap_pages,
1331 };
1332
1333 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1334 {
1335         struct address_space *mapping = file->f_mapping;
1336
1337         if (!mapping->a_ops->readpage)
1338                 return -ENOEXEC;
1339         file_accessed(file);
1340         vma->vm_ops = &ceph_vmops;
1341         return 0;
1342 }