fs/ceph/addr.c

   1 #include "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/pagevec.h>
   9 #include <linux/task_io_accounting_ops.h>
  10
  11 #include "super.h"
  12 #include "osd_client.h"
  13
  14 /*
  15  * Ceph address space ops.
  16  *
  17  * There are a few funny things going on here.
  18  *
  19  * The page->private field is used to reference a struct
  20  * ceph_snap_context for _every_ dirty page.  This indicates which
  21  * snapshot the page was logically dirtied in, and thus which snap
  22  * context needs to be associated with the osd write during writeback.
  23  *
  24  * Similarly, struct ceph_inode_info maintains a set of counters to
  25  * count dirty pages on the inode.  In the absense of snapshots,
  26  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  27  *
  28  * When a snapshot is taken (that is, when the client receives
  29  * notification that a snapshot was taken), each inode with caps and
  30  * with dirty pages (dirty pages implies there is a cap) gets a new
  31  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  32  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  33  * moved to capsnap->dirty. (Unless a sync write is currently in
  34  * progress.  In that case, the capsnap is said to be "pending", new
  35  * writes cannot start, and the capsnap isn't "finalized" until the
  36  * write completes (or fails) and a final size/mtime for the inode for
  37  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  38  *
  39  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  40  * we look for the first capsnap in i_cap_snaps and write out pages in
  41  * that snap context _only_.  Then we move on to the next capsnap,
  42  * eventually reaching the "live" or "head" context (i.e., pages that
  43  * are not yet snapped) and are writing the most recently dirtied
  44  * pages.
  45  *
  46  * Invalidate and so forth must take care to ensure the dirty page
  47  * accounting is preserved.
  48  */
  49
  50
  51 /*
  52  * Dirty a page.  Optimistically adjust accounting, on the assumption
  53  * that we won't race with invalidate.  If we do, readjust.
  54  */
  55 static int ceph_set_page_dirty(struct page *page)
  56 {
  57         struct address_space *mapping = page->mapping;
  58         struct inode *inode;
  59         struct ceph_inode_info *ci;
  60         int undo = 0;
  61         struct ceph_snap_context *snapc;
  62
  63         if (unlikely(!mapping))
  64                 return !TestSetPageDirty(page);
  65
  66         if (TestSetPageDirty(page)) {
  67                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  68                      mapping->host, page, page->index);
  69                 return 0;
  70         }
  71
  72         inode = mapping->host;
  73         ci = ceph_inode(inode);
  74
  75         /*
  76          * Note that we're grabbing a snapc ref here without holding
  77          * any locks!
  78          */
  79         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  80
  81         /* dirty the head */
  82         spin_lock(&inode->i_lock);
  83         if (ci->i_wrbuffer_ref_head == 0)
  84                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  85         ++ci->i_wrbuffer_ref_head;
  86         if (ci->i_wrbuffer_ref == 0)
  87                 igrab(inode);
  88         ++ci->i_wrbuffer_ref;
  89         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
  90              "snapc %p seq %lld (%d snaps)\n",
  91              mapping->host, page, page->index,
  92              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
  93              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
  94              snapc, snapc->seq, snapc->num_snaps);
  95         spin_unlock(&inode->i_lock);
  96
  97         /* now adjust page */
  98         spin_lock_irq(&mapping->tree_lock);
  99         if (page->mapping) {    /* Race with truncate? */
 100                 WARN_ON_ONCE(!PageUptodate(page));
 101
 102                 if (mapping_cap_account_dirty(mapping)) {
 103                         __inc_zone_page_state(page, NR_FILE_DIRTY);
 104                         __inc_bdi_stat(mapping->backing_dev_info,
 105                                         BDI_RECLAIMABLE);
 106                         task_io_account_write(PAGE_CACHE_SIZE);
 107                 }
 108                 radix_tree_tag_set(&mapping->page_tree,
 109                                 page_index(page), PAGECACHE_TAG_DIRTY);
 110
 111                 /*
 112                  * Reference snap context in page->private.  Also set
 113                  * PagePrivate so that we get invalidatepage callback.
 114                  */
 115                 page->private = (unsigned long)snapc;
 116                 SetPagePrivate(page);
 117         } else {
 118                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 119                 undo = 1;
 120         }
 121
 122         spin_unlock_irq(&mapping->tree_lock);
 123
 124         if (undo)
 125                 /* whoops, we failed to dirty the page */
 126                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 127
 128         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 129
 130         BUG_ON(!PageDirty(page));
 131         return 1;
 132 }
 133
 134 /*
 135  * If we are truncating the full page (i.e. offset == 0), adjust the
 136  * dirty page counters appropriately.  Only called if there is private
 137  * data on the page.
 138  */
 139 static void ceph_invalidatepage(struct page *page, unsigned long offset)
 140 {
 141         struct inode *inode = page->mapping->host;
 142         struct ceph_inode_info *ci;
 143         struct ceph_snap_context *snapc = (void *)page->private;
 144
 145         BUG_ON(!PageLocked(page));
 146         BUG_ON(!page->private);
 147         BUG_ON(!PagePrivate(page));
 148         BUG_ON(!page->mapping);
 149
 150         /*
 151          * We can get non-dirty pages here due to races between
 152          * set_page_dirty and truncate_complete_page; just spit out a
 153          * warning, in case we end up with accounting problems later.
 154          */
 155         if (!PageDirty(page))
 156                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 157
 158         if (offset == 0)
 159                 ClearPageChecked(page);
 160
 161         ci = ceph_inode(inode);
 162         if (offset == 0) {
 163                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 164                      inode, page, page->index, offset);
 165                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 166                 ceph_put_snap_context(snapc);
 167                 page->private = 0;
 168                 ClearPagePrivate(page);
 169         } else {
 170                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
 171                      inode, page, page->index);
 172         }
 173 }
 174
 175 /* just a sanity check */
 176 static int ceph_releasepage(struct page *page, gfp_t g)
 177 {
 178         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 179         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 180         WARN_ON(PageDirty(page));
 181         WARN_ON(page->private);
 182         WARN_ON(PagePrivate(page));
 183         return 0;
 184 }
 185
 186 /*
 187  * read a single page, without unlocking it.
 188  */
 189 static int readpage_nounlock(struct file *filp, struct page *page)
 190 {
 191         struct inode *inode = filp->f_dentry->d_inode;
 192         struct ceph_inode_info *ci = ceph_inode(inode);
 193         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 194         int err = 0;
 195         u64 len = PAGE_CACHE_SIZE;
 196
 197         dout("readpage inode %p file %p page %p index %lu\n",
 198              inode, filp, page, page->index);
 199         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 200                                   page->index << PAGE_CACHE_SHIFT, &len,
 201                                   ci->i_truncate_seq, ci->i_truncate_size,
 202                                   &page, 1);
 203         if (err == -ENOENT)
 204                 err = 0;
 205         if (err < 0) {
 206                 SetPageError(page);
 207                 goto out;
 208         } else if (err < PAGE_CACHE_SIZE) {
 209                 /* zero fill remainder of page */
 210                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 211         }
 212         SetPageUptodate(page);
 213
 214 out:
 215         return err < 0 ? err : 0;
 216 }
 217
 218 static int ceph_readpage(struct file *filp, struct page *page)
 219 {
 220         int r = readpage_nounlock(filp, page);
 221         unlock_page(page);
 222         return r;
 223 }
 224
 225 /*
 226  * Build a vector of contiguous pages from the provided page list.
 227  */
 228 static struct page **page_vector_from_list(struct list_head *page_list,
 229                                            unsigned *nr_pages)
 230 {
 231         struct page **pages;
 232         struct page *page;
 233         int next_index, contig_pages = 0;
 234
 235         /* build page vector */
 236         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 237         if (!pages)
 238                 return ERR_PTR(-ENOMEM);
 239
 240         BUG_ON(list_empty(page_list));
 241         next_index = list_entry(page_list->prev, struct page, lru)->index;
 242         list_for_each_entry_reverse(page, page_list, lru) {
 243                 if (page->index == next_index) {
 244                         dout("readpages page %d %p\n", contig_pages, page);
 245                         pages[contig_pages] = page;
 246                         contig_pages++;
 247                         next_index++;
 248                 } else {
 249                         break;
 250                 }
 251         }
 252         *nr_pages = contig_pages;
 253         return pages;
 254 }
 255
 256 /*
 257  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 258  * the caller (VM) cleans them up.
 259  */
 260 static int ceph_readpages(struct file *file, struct address_space *mapping,
 261                           struct list_head *page_list, unsigned nr_pages)
 262 {
 263         struct inode *inode = file->f_dentry->d_inode;
 264         struct ceph_inode_info *ci = ceph_inode(inode);
 265         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 266         int rc = 0;
 267         struct page **pages;
 268         struct pagevec pvec;
 269         loff_t offset;
 270         u64 len;
 271
 272         dout("readpages %p file %p nr_pages %d\n",
 273              inode, file, nr_pages);
 274
 275         pages = page_vector_from_list(page_list, &nr_pages);
 276         if (IS_ERR(pages))
 277                 return PTR_ERR(pages);
 278
 279         /* guess read extent */
 280         offset = pages[0]->index << PAGE_CACHE_SHIFT;
 281         len = nr_pages << PAGE_CACHE_SHIFT;
 282         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 283                                  offset, &len,
 284                                  ci->i_truncate_seq, ci->i_truncate_size,
 285                                  pages, nr_pages);
 286         if (rc == -ENOENT)
 287                 rc = 0;
 288         if (rc < 0)
 289                 goto out;
 290
 291         /* set uptodate and add to lru in pagevec-sized chunks */
 292         pagevec_init(&pvec, 0);
 293         for (; !list_empty(page_list) && len > 0;
 294              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
 295                 struct page *page =
 296                         list_entry(page_list->prev, struct page, lru);
 297
 298                 list_del(&page->lru);
 299
 300                 if (rc < (int)PAGE_CACHE_SIZE) {
 301                         /* zero (remainder of) page */
 302                         int s = rc < 0 ? 0 : rc;
 303                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 304                 }
 305
 306                 if (add_to_page_cache(page, mapping, page->index, GFP_NOFS)) {
 307                         page_cache_release(page);
 308                         dout("readpages %p add_to_page_cache failed %p\n",
 309                              inode, page);
 310                         continue;
 311                 }
 312                 dout("readpages %p adding %p idx %lu\n", inode, page,
 313                      page->index);
 314                 flush_dcache_page(page);
 315                 SetPageUptodate(page);
 316                 unlock_page(page);
 317                 if (pagevec_add(&pvec, page) == 0)
 318                         pagevec_lru_add_file(&pvec);   /* add to lru */
 319         }
 320         pagevec_lru_add_file(&pvec);
 321         rc = 0;
 322
 323 out:
 324         kfree(pages);
 325         return rc;
 326 }
 327
 328 /*
 329  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 330  * only snap context we are allowed to write back.
 331  *
 332  * Caller holds i_lock.
 333  */
 334 static struct ceph_snap_context *__get_oldest_context(struct inode *inode,
 335                                                       u64 *snap_size)
 336 {
 337         struct ceph_inode_info *ci = ceph_inode(inode);
 338         struct ceph_snap_context *snapc = NULL;
 339         struct ceph_cap_snap *capsnap = NULL;
 340
 341         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 342                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 343                      capsnap->context, capsnap->dirty_pages);
 344                 if (capsnap->dirty_pages) {
 345                         snapc = ceph_get_snap_context(capsnap->context);
 346                         if (snap_size)
 347                                 *snap_size = capsnap->size;
 348                         break;
 349                 }
 350         }
 351         if (!snapc && ci->i_snap_realm) {
 352                 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
 353                 dout(" head snapc %p has %d dirty pages\n",
 354                      snapc, ci->i_wrbuffer_ref_head);
 355         }
 356         return snapc;
 357 }
 358
 359 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 360                                                     u64 *snap_size)
 361 {
 362         struct ceph_snap_context *snapc = NULL;
 363
 364         spin_lock(&inode->i_lock);
 365         snapc = __get_oldest_context(inode, snap_size);
 366         spin_unlock(&inode->i_lock);
 367         return snapc;
 368 }
 369
 370 /*
 371  * Write a single page, but leave the page locked.
 372  *
 373  * If we get a write error, set the page error bit, but still adjust the
 374  * dirty page accounting (i.e., page is no longer dirty).
 375  */
 376 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 377 {
 378         struct inode *inode;
 379         struct ceph_inode_info *ci;
 380         struct ceph_osd_client *osdc;
 381         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 382         int len = PAGE_CACHE_SIZE;
 383         loff_t i_size;
 384         int err = 0;
 385         struct ceph_snap_context *snapc;
 386         u64 snap_size = 0;
 387
 388         dout("writepage %p idx %lu\n", page, page->index);
 389
 390         if (!page->mapping || !page->mapping->host) {
 391                 dout("writepage %p - no mapping\n", page);
 392                 return -EFAULT;
 393         }
 394         inode = page->mapping->host;
 395         ci = ceph_inode(inode);
 396         osdc = &ceph_inode_to_client(inode)->osdc;
 397
 398         /* verify this is a writeable snap context */
 399         snapc = (void *)page->private;
 400         if (snapc == NULL) {
 401                 dout("writepage %p page %p not dirty?\n", inode, page);
 402                 goto out;
 403         }
 404         if (snapc != get_oldest_context(inode, &snap_size)) {
 405                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 406                      inode, page, (void *)page->private);
 407                 /* we should only noop if called by kswapd */
 408                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 409                 goto out;
 410         }
 411
 412         /* is this a partial page at end of file? */
 413         if (snap_size)
 414                 i_size = snap_size;
 415         else
 416                 i_size = i_size_read(inode);
 417         if (i_size < page_off + len)
 418                 len = i_size - page_off;
 419
 420         dout("writepage %p page %p index %lu on %llu~%u\n",
 421              inode, page, page->index, page_off, len);
 422
 423         set_page_writeback(page);
 424         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 425                                    &ci->i_layout, snapc,
 426                                    page_off, len,
 427                                    ci->i_truncate_seq, ci->i_truncate_size,
 428                                    &inode->i_mtime,
 429                                    &page, 1, 0, 0, true);
 430         if (err < 0) {
 431                 dout("writepage setting page/mapping error %d %p\n", err, page);
 432                 SetPageError(page);
 433                 mapping_set_error(&inode->i_data, err);
 434                 if (wbc)
 435                         wbc->pages_skipped++;
 436         } else {
 437                 dout("writepage cleaned page %p\n", page);
 438                 err = 0;  /* vfs expects us to return 0 */
 439         }
 440         page->private = 0;
 441         ClearPagePrivate(page);
 442         end_page_writeback(page);
 443         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 444         ceph_put_snap_context(snapc);
 445 out:
 446         return err;
 447 }
 448
 449 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 450 {
 451         int err = writepage_nounlock(page, wbc);
 452         unlock_page(page);
 453         return err;
 454 }
 455
 456
 457 /*
 458  * lame release_pages helper.  release_pages() isn't exported to
 459  * modules.
 460  */
 461 static void ceph_release_pages(struct page **pages, int num)
 462 {
 463         struct pagevec pvec;
 464         int i;
 465
 466         pagevec_init(&pvec, 0);
 467         for (i = 0; i < num; i++) {
 468                 if (pagevec_add(&pvec, pages[i]) == 0)
 469                         pagevec_release(&pvec);
 470         }
 471         pagevec_release(&pvec);
 472 }
 473
 474
 475 /*
 476  * async writeback completion handler.
 477  *
 478  * If we get an error, set the mapping error bit, but not the individual
 479  * page error bits.
 480  */
 481 static void writepages_finish(struct ceph_osd_request *req,
 482                               struct ceph_msg *msg)
 483 {
 484         struct inode *inode = req->r_inode;
 485         struct ceph_osd_reply_head *replyhead;
 486         struct ceph_osd_op *op;
 487         struct ceph_inode_info *ci = ceph_inode(inode);
 488         unsigned wrote;
 489         loff_t offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 490         struct page *page;
 491         int i;
 492         struct ceph_snap_context *snapc = req->r_snapc;
 493         struct address_space *mapping = inode->i_mapping;
 494         struct writeback_control *wbc = req->r_wbc;
 495         __s32 rc = -EIO;
 496         u64 bytes = 0;
 497
 498         /* parse reply */
 499         replyhead = msg->front.iov_base;
 500         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 501         op = (void *)(replyhead + 1);
 502         rc = le32_to_cpu(replyhead->result);
 503         bytes = le64_to_cpu(op->extent.length);
 504
 505         if (rc >= 0) {
 506                 wrote = (bytes + (offset & ~PAGE_CACHE_MASK) + ~PAGE_CACHE_MASK)
 507                         >> PAGE_CACHE_SHIFT;
 508                 WARN_ON(wrote != req->r_num_pages);
 509         } else {
 510                 wrote = 0;
 511                 mapping_set_error(mapping, rc);
 512         }
 513         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 514              inode, rc, bytes, wrote);
 515
 516         /* clean all pages */
 517         for (i = 0; i < req->r_num_pages; i++) {
 518                 page = req->r_pages[i];
 519                 BUG_ON(!page);
 520                 WARN_ON(!PageUptodate(page));
 521
 522                 if (i >= wrote) {
 523                         dout("inode %p skipping page %p\n", inode, page);
 524                         wbc->pages_skipped++;
 525                 }
 526                 page->private = 0;
 527                 ClearPagePrivate(page);
 528                 ceph_put_snap_context(snapc);
 529                 dout("unlocking %d %p\n", i, page);
 530                 end_page_writeback(page);
 531                 unlock_page(page);
 532         }
 533         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 534         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 535
 536         ceph_release_pages(req->r_pages, req->r_num_pages);
 537         if (req->r_pages_from_pool)
 538                 mempool_free(req->r_pages,
 539                              ceph_client(inode->i_sb)->wb_pagevec_pool);
 540         else
 541                 kfree(req->r_pages);
 542         ceph_osdc_put_request(req);
 543 }
 544
 545 /*
 546  * allocate a page vec, either directly, or if necessary, via a the
 547  * mempool.  we avoid the mempool if we can because req->r_num_pages
 548  * may be less than the maximum write size.
 549  */
 550 static void alloc_page_vec(struct ceph_client *client,
 551                            struct ceph_osd_request *req)
 552 {
 553         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 554                                GFP_NOFS);
 555         if (!req->r_pages) {
 556                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
 557                 req->r_pages_from_pool = 1;
 558                 WARN_ON(!req->r_pages);
 559         }
 560 }
 561
 562 /*
 563  * initiate async writeback
 564  */
 565 static int ceph_writepages_start(struct address_space *mapping,
 566                                  struct writeback_control *wbc)
 567 {
 568         struct inode *inode = mapping->host;
 569         struct backing_dev_info *bdi = mapping->backing_dev_info;
 570         struct ceph_inode_info *ci = ceph_inode(inode);
 571         struct ceph_client *client = ceph_inode_to_client(inode);
 572         pgoff_t index, start, end;
 573         int range_whole = 0;
 574         int should_loop = 1;
 575         pgoff_t max_pages = 0, max_pages_ever = 0;
 576         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL;
 577         struct pagevec pvec;
 578         int done = 0;
 579         int rc = 0;
 580         unsigned wsize = 1 << inode->i_blkbits;
 581         struct ceph_osd_request *req = NULL;
 582         int do_sync;
 583         u64 snap_size = 0;
 584
 585         /*
 586          * Include a 'sync' in the OSD request if this is a data
 587          * integrity write (e.g., O_SYNC write or fsync()), or if our
 588          * cap is being revoked.
 589          */
 590         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 591         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 592                 do_sync = 1;
 593         dout("writepages_start %p dosync=%d (mode=%s)\n",
 594              inode, do_sync,
 595              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 596              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 597
 598         client = ceph_inode_to_client(inode);
 599         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
 600                 pr_warning("writepage_start %p on forced umount\n", inode);
 601                 return -EIO; /* we're in a forced umount, don't write! */
 602         }
 603         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
 604                 wsize = client->mount_args->wsize;
 605         if (wsize < PAGE_CACHE_SIZE)
 606                 wsize = PAGE_CACHE_SIZE;
 607         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 608
 609         pagevec_init(&pvec, 0);
 610
 611         /* ?? */
 612         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 613                 dout(" writepages congested\n");
 614                 wbc->encountered_congestion = 1;
 615                 goto out_final;
 616         }
 617
 618         /* where to start/end? */
 619         if (wbc->range_cyclic) {
 620                 start = mapping->writeback_index; /* Start from prev offset */
 621                 end = -1;
 622                 dout(" cyclic, start at %lu\n", start);
 623         } else {
 624                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 625                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 626                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 627                         range_whole = 1;
 628                 should_loop = 0;
 629                 dout(" not cyclic, %lu to %lu\n", start, end);
 630         }
 631         index = start;
 632
 633 retry:
 634         /* find oldest snap context with dirty data */
 635         ceph_put_snap_context(snapc);
 636         snapc = get_oldest_context(inode, &snap_size);
 637         if (!snapc) {
 638                 /* hmm, why does writepages get called when there
 639                    is no dirty data? */
 640                 dout(" no snap context with dirty data?\n");
 641                 goto out;
 642         }
 643         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 644              snapc, snapc->seq, snapc->num_snaps);
 645         if (last_snapc && snapc != last_snapc) {
 646                 /* if we switched to a newer snapc, restart our scan at the
 647                  * start of the original file range. */
 648                 dout("  snapc differs from last pass, restarting at %lu\n",
 649                      index);
 650                 index = start;
 651         }
 652         last_snapc = snapc;
 653
 654         while (!done && index <= end) {
 655                 unsigned i;
 656                 int first;
 657                 pgoff_t next;
 658                 int pvec_pages, locked_pages;
 659                 struct page *page;
 660                 int want;
 661                 u64 offset, len;
 662                 struct ceph_osd_request_head *reqhead;
 663                 struct ceph_osd_op *op;
 664
 665                 next = 0;
 666                 locked_pages = 0;
 667                 max_pages = max_pages_ever;
 668
 669 get_more_pages:
 670                 first = -1;
 671                 want = min(end - index,
 672                            min((pgoff_t)PAGEVEC_SIZE,
 673                                max_pages - (pgoff_t)locked_pages) - 1)
 674                         + 1;
 675                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 676                                                 PAGECACHE_TAG_DIRTY,
 677                                                 want);
 678                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 679                 if (!pvec_pages && !locked_pages)
 680                         break;
 681                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 682                         page = pvec.pages[i];
 683                         dout("? %p idx %lu\n", page, page->index);
 684                         if (locked_pages == 0)
 685                                 lock_page(page);  /* first page */
 686                         else if (!trylock_page(page))
 687                                 break;
 688
 689                         /* only dirty pages, or our accounting breaks */
 690                         if (unlikely(!PageDirty(page)) ||
 691                             unlikely(page->mapping != mapping)) {
 692                                 dout("!dirty or !mapping %p\n", page);
 693                                 unlock_page(page);
 694                                 break;
 695                         }
 696                         if (!wbc->range_cyclic && page->index > end) {
 697                                 dout("end of range %p\n", page);
 698                                 done = 1;
 699                                 unlock_page(page);
 700                                 break;
 701                         }
 702                         if (next && (page->index != next)) {
 703                                 dout("not consecutive %p\n", page);
 704                                 unlock_page(page);
 705                                 break;
 706                         }
 707                         if (wbc->sync_mode != WB_SYNC_NONE) {
 708                                 dout("waiting on writeback %p\n", page);
 709                                 wait_on_page_writeback(page);
 710                         }
 711                         if ((snap_size && page_offset(page) > snap_size) ||
 712                             (!snap_size &&
 713                              page_offset(page) > i_size_read(inode))) {
 714                                 dout("%p page eof %llu\n", page, snap_size ?
 715                                      snap_size : i_size_read(inode));
 716                                 done = 1;
 717                                 unlock_page(page);
 718                                 break;
 719                         }
 720                         if (PageWriteback(page)) {
 721                                 dout("%p under writeback\n", page);
 722                                 unlock_page(page);
 723                                 break;
 724                         }
 725
 726                         /* only if matching snap context */
 727                         if (snapc != (void *)page->private) {
 728                                 dout("page snapc %p != oldest %p\n",
 729                                      (void *)page->private, snapc);
 730                                 unlock_page(page);
 731                                 if (!locked_pages)
 732                                         continue; /* keep looking for snap */
 733                                 break;
 734                         }
 735
 736                         if (!clear_page_dirty_for_io(page)) {
 737                                 dout("%p !clear_page_dirty_for_io\n", page);
 738                                 unlock_page(page);
 739                                 break;
 740                         }
 741
 742                         /* ok */
 743                         if (locked_pages == 0) {
 744                                 /* prepare async write request */
 745                                 offset = page->index << PAGE_CACHE_SHIFT;
 746                                 len = wsize;
 747                                 req = ceph_osdc_new_request(&client->osdc,
 748                                             &ci->i_layout,
 749                                             ceph_vino(inode),
 750                                             offset, &len,
 751                                             CEPH_OSD_OP_WRITE,
 752                                             CEPH_OSD_FLAG_WRITE |
 753                                                     CEPH_OSD_FLAG_ONDISK,
 754                                             snapc, do_sync,
 755                                             ci->i_truncate_seq,
 756                                             ci->i_truncate_size,
 757                                             &inode->i_mtime, true, 1);
 758                                 max_pages = req->r_num_pages;
 759
 760                                 alloc_page_vec(client, req);
 761                                 req->r_callback = writepages_finish;
 762                                 req->r_inode = inode;
 763                                 req->r_wbc = wbc;
 764                         }
 765
 766                         /* note position of first page in pvec */
 767                         if (first < 0)
 768                                 first = i;
 769                         dout("%p will write page %p idx %lu\n",
 770                              inode, page, page->index);
 771                         set_page_writeback(page);
 772                         req->r_pages[locked_pages] = page;
 773                         locked_pages++;
 774                         next = page->index + 1;
 775                 }
 776
 777                 /* did we get anything? */
 778                 if (!locked_pages)
 779                         goto release_pvec_pages;
 780                 if (i) {
 781                         int j;
 782                         BUG_ON(!locked_pages || first < 0);
 783
 784                         if (pvec_pages && i == pvec_pages &&
 785                             locked_pages < max_pages) {
 786                                 dout("reached end pvec, trying for more\n");
 787                                 pagevec_reinit(&pvec);
 788                                 goto get_more_pages;
 789                         }
 790
 791                         /* shift unused pages over in the pvec...  we
 792                          * will need to release them below. */
 793                         for (j = i; j < pvec_pages; j++) {
 794                                 dout(" pvec leftover page %p\n",
 795                                      pvec.pages[j]);
 796                                 pvec.pages[j-i+first] = pvec.pages[j];
 797                         }
 798                         pvec.nr -= i-first;
 799                 }
 800
 801                 /* submit the write */
 802                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 803                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 804                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 805                 dout("writepages got %d pages at %llu~%llu\n",
 806                      locked_pages, offset, len);
 807
 808                 /* revise final length, page count */
 809                 req->r_num_pages = locked_pages;
 810                 reqhead = req->r_request->front.iov_base;
 811                 op = (void *)(reqhead + 1);
 812                 op->extent.length = cpu_to_le64(len);
 813                 op->payload_len = cpu_to_le32(len);
 814                 req->r_request->hdr.data_len = cpu_to_le32(len);
 815
 816                 ceph_osdc_start_request(&client->osdc, req, true);
 817                 req = NULL;
 818
 819                 /* continue? */
 820                 index = next;
 821                 wbc->nr_to_write -= locked_pages;
 822                 if (wbc->nr_to_write <= 0)
 823                         done = 1;
 824
 825 release_pvec_pages:
 826                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 827                      pvec.nr ? pvec.pages[0] : NULL);
 828                 pagevec_release(&pvec);
 829
 830                 if (locked_pages && !done)
 831                         goto retry;
 832         }
 833
 834         if (should_loop && !done) {
 835                 /* more to do; loop back to beginning of file */
 836                 dout("writepages looping back to beginning of file\n");
 837                 should_loop = 0;
 838                 index = 0;
 839                 goto retry;
 840         }
 841
 842         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 843                 mapping->writeback_index = index;
 844
 845 out:
 846         if (req)
 847                 ceph_osdc_put_request(req);
 848         if (rc > 0)
 849                 rc = 0;  /* vfs expects us to return 0 */
 850         ceph_put_snap_context(snapc);
 851         dout("writepages done, rc = %d\n", rc);
 852 out_final:
 853         return rc;
 854 }
 855
 856
 857
 858 /*
 859  * See if a given @snapc is either writeable, or already written.
 860  */
 861 static int context_is_writeable_or_written(struct inode *inode,
 862                                            struct ceph_snap_context *snapc)
 863 {
 864         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 865         return !oldest || snapc->seq <= oldest->seq;
 866 }
 867
 868 /*
 869  * We are only allowed to write into/dirty the page if the page is
 870  * clean, or already dirty within the same snap context.
 871  */
 872 static int ceph_write_begin(struct file *file, struct address_space *mapping,
 873                             loff_t pos, unsigned len, unsigned flags,
 874                             struct page **pagep, void **fsdata)
 875 {
 876         struct inode *inode = file->f_dentry->d_inode;
 877         struct ceph_inode_info *ci = ceph_inode(inode);
 878         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 879         struct page *page;
 880         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 881         loff_t page_off = pos & PAGE_CACHE_MASK;
 882         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 883         int end_in_page = pos_in_page + len;
 884         loff_t i_size;
 885         struct ceph_snap_context *snapc;
 886         int r;
 887
 888         /* get a page*/
 889 retry:
 890         page = grab_cache_page_write_begin(mapping, index, 0);
 891         if (!page)
 892                 return -ENOMEM;
 893         *pagep = page;
 894
 895         dout("write_begin file %p inode %p page %p %d~%d\n", file,
 896              inode, page, (int)pos, (int)len);
 897
 898 retry_locked:
 899         /* writepages currently holds page lock, but if we change that later, */
 900         wait_on_page_writeback(page);
 901
 902         /* check snap context */
 903         BUG_ON(!ci->i_snap_realm);
 904         down_read(&mdsc->snap_rwsem);
 905         BUG_ON(!ci->i_snap_realm->cached_context);
 906         if (page->private &&
 907             (void *)page->private != ci->i_snap_realm->cached_context) {
 908                 /*
 909                  * this page is already dirty in another (older) snap
 910                  * context!  is it writeable now?
 911                  */
 912                 snapc = get_oldest_context(inode, NULL);
 913                 up_read(&mdsc->snap_rwsem);
 914
 915                 if (snapc != (void *)page->private) {
 916                         dout(" page %p snapc %p not current or oldest\n",
 917                              page, (void *)page->private);
 918                         /*
 919                          * queue for writeback, and wait for snapc to
 920                          * be writeable or written
 921                          */
 922                         snapc = ceph_get_snap_context((void *)page->private);
 923                         unlock_page(page);
 924                         if (ceph_queue_writeback(inode))
 925                                 igrab(inode);
 926                         wait_event_interruptible(ci->i_cap_wq,
 927                                context_is_writeable_or_written(inode, snapc));
 928                         ceph_put_snap_context(snapc);
 929                         goto retry;
 930                 }
 931
 932                 /* yay, writeable, do it now (without dropping page lock) */
 933                 dout(" page %p snapc %p not current, but oldest\n",
 934                      page, snapc);
 935                 if (!clear_page_dirty_for_io(page))
 936                         goto retry_locked;
 937                 r = writepage_nounlock(page, NULL);
 938                 if (r < 0)
 939                         goto fail_nosnap;
 940                 goto retry_locked;
 941         }
 942
 943         if (PageUptodate(page)) {
 944                 dout(" page %p already uptodate\n", page);
 945                 return 0;
 946         }
 947
 948         /* full page? */
 949         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 950                 return 0;
 951
 952         /* past end of file? */
 953         i_size = inode->i_size;   /* caller holds i_mutex */
 954
 955         if (i_size + len > inode->i_sb->s_maxbytes) {
 956                 /* file is too big */
 957                 r = -EINVAL;
 958                 goto fail;
 959         }
 960
 961         if (page_off >= i_size ||
 962             (pos_in_page == 0 && (pos+len) >= i_size &&
 963              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
 964                 dout(" zeroing %p 0 - %d and %d - %d\n",
 965                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
 966                 zero_user_segments(page,
 967                                    0, pos_in_page,
 968                                    end_in_page, PAGE_CACHE_SIZE);
 969                 return 0;
 970         }
 971
 972         /* we need to read it. */
 973         up_read(&mdsc->snap_rwsem);
 974         r = readpage_nounlock(file, page);
 975         if (r < 0)
 976                 goto fail_nosnap;
 977         goto retry_locked;
 978
 979 fail:
 980         up_read(&mdsc->snap_rwsem);
 981 fail_nosnap:
 982         unlock_page(page);
 983         return r;
 984 }
 985
 986 /*
 987  * we don't do anything in here that simple_write_end doesn't do
 988  * except adjust dirty page accounting and drop read lock on
 989  * mdsc->snap_rwsem.
 990  */
 991 static int ceph_write_end(struct file *file, struct address_space *mapping,
 992                           loff_t pos, unsigned len, unsigned copied,
 993                           struct page *page, void *fsdata)
 994 {
 995         struct inode *inode = file->f_dentry->d_inode;
 996         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 997         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 998         int check_cap = 0;
 999
1000         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1001              inode, page, (int)pos, (int)copied, (int)len);
1002
1003         /* zero the stale part of the page if we did a short copy */
1004         if (copied < len)
1005                 zero_user_segment(page, from+copied, len);
1006
1007         /* did file size increase? */
1008         /* (no need for i_size_read(); we caller holds i_mutex */
1009         if (pos+copied > inode->i_size)
1010                 check_cap = ceph_inode_set_size(inode, pos+copied);
1011
1012         if (!PageUptodate(page))
1013                 SetPageUptodate(page);
1014
1015         set_page_dirty(page);
1016
1017         unlock_page(page);
1018         up_read(&mdsc->snap_rwsem);
1019         page_cache_release(page);
1020
1021         if (check_cap)
1022                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1023
1024         return copied;
1025 }
1026
1027 /*
1028  * we set .direct_IO to indicate direct io is supported, but since we
1029  * intercept O_DIRECT reads and writes early, this function should
1030  * never get called.
1031  */
1032 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1033                               const struct iovec *iov,
1034                               loff_t pos, unsigned long nr_segs)
1035 {
1036         WARN_ON(1);
1037         return -EINVAL;
1038 }
1039
1040 const struct address_space_operations ceph_aops = {
1041         .readpage = ceph_readpage,
1042         .readpages = ceph_readpages,
1043         .writepage = ceph_writepage,
1044         .writepages = ceph_writepages_start,
1045         .write_begin = ceph_write_begin,
1046         .write_end = ceph_write_end,
1047         .set_page_dirty = ceph_set_page_dirty,
1048         .invalidatepage = ceph_invalidatepage,
1049         .releasepage = ceph_releasepage,
1050         .direct_IO = ceph_direct_io,
1051 };
1052
1053
1054 /*
1055  * vm ops
1056  */
1057
1058 /*
1059  * Reuse write_begin here for simplicity.
1060  */
1061 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1062 {
1063         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1064         struct page *page = vmf->page;
1065         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1066         loff_t off = page->index << PAGE_CACHE_SHIFT;
1067         loff_t size, len;
1068         struct page *locked_page = NULL;
1069         void *fsdata = NULL;
1070         int ret;
1071
1072         size = i_size_read(inode);
1073         if (off + PAGE_CACHE_SIZE <= size)
1074                 len = PAGE_CACHE_SIZE;
1075         else
1076                 len = size & ~PAGE_CACHE_MASK;
1077
1078         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1079              off, len, page, page->index);
1080         ret = ceph_write_begin(vma->vm_file, inode->i_mapping, off, len, 0,
1081                                &locked_page, &fsdata);
1082         WARN_ON(page != locked_page);
1083         if (!ret) {
1084                 /*
1085                  * doing the following, instead of calling
1086                  * ceph_write_end. Note that we keep the
1087                  * page locked
1088                  */
1089                 set_page_dirty(page);
1090                 up_read(&mdsc->snap_rwsem);
1091                 page_cache_release(page);
1092                 ret = VM_FAULT_LOCKED;
1093         } else {
1094                 ret = VM_FAULT_SIGBUS;
1095         }
1096         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1097         return ret;
1098 }
1099
1100 static struct vm_operations_struct ceph_vmops = {
1101         .fault          = filemap_fault,
1102         .page_mkwrite   = ceph_page_mkwrite,
1103 };
1104
1105 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1106 {
1107         struct address_space *mapping = file->f_mapping;
1108
1109         if (!mapping->a_ops->readpage)
1110                 return -ENOEXEC;
1111         file_accessed(file);
1112         vma->vm_ops = &ceph_vmops;
1113         vma->vm_flags |= VM_CAN_NONLINEAR;
1114         return 0;
1115 }