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/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "osd_client.h"
  14
  15 /*
  16  * Ceph address space ops.
  17  *
  18  * There are a few funny things going on here.
  19  *
  20  * The page->private field is used to reference a struct
  21  * ceph_snap_context for _every_ dirty page.  This indicates which
  22  * snapshot the page was logically dirtied in, and thus which snap
  23  * context needs to be associated with the osd write during writeback.
  24  *
  25  * Similarly, struct ceph_inode_info maintains a set of counters to
  26  * count dirty pages on the inode.  In the absense of snapshots,
  27  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  28  *
  29  * When a snapshot is taken (that is, when the client receives
  30  * notification that a snapshot was taken), each inode with caps and
  31  * with dirty pages (dirty pages implies there is a cap) gets a new
  32  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  33  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  34  * moved to capsnap->dirty. (Unless a sync write is currently in
  35  * progress.  In that case, the capsnap is said to be "pending", new
  36  * writes cannot start, and the capsnap isn't "finalized" until the
  37  * write completes (or fails) and a final size/mtime for the inode for
  38  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  39  *
  40  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  41  * we look for the first capsnap in i_cap_snaps and write out pages in
  42  * that snap context _only_.  Then we move on to the next capsnap,
  43  * eventually reaching the "live" or "head" context (i.e., pages that
  44  * are not yet snapped) and are writing the most recently dirtied
  45  * pages.
  46  *
  47  * Invalidate and so forth must take care to ensure the dirty page
  48  * accounting is preserved.
  49  */
  50
  51 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  52 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  53         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  54          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  55
  56
  57
  58 /*
  59  * Dirty a page.  Optimistically adjust accounting, on the assumption
  60  * that we won't race with invalidate.  If we do, readjust.
  61  */
  62 static int ceph_set_page_dirty(struct page *page)
  63 {
  64         struct address_space *mapping = page->mapping;
  65         struct inode *inode;
  66         struct ceph_inode_info *ci;
  67         int undo = 0;
  68         struct ceph_snap_context *snapc;
  69
  70         if (unlikely(!mapping))
  71                 return !TestSetPageDirty(page);
  72
  73         if (TestSetPageDirty(page)) {
  74                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  75                      mapping->host, page, page->index);
  76                 return 0;
  77         }
  78
  79         inode = mapping->host;
  80         ci = ceph_inode(inode);
  81
  82         /*
  83          * Note that we're grabbing a snapc ref here without holding
  84          * any locks!
  85          */
  86         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  87
  88         /* dirty the head */
  89         spin_lock(&inode->i_lock);
  90         if (ci->i_wrbuffer_ref_head == 0)
  91                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  92         ++ci->i_wrbuffer_ref_head;
  93         if (ci->i_wrbuffer_ref == 0)
  94                 igrab(inode);
  95         ++ci->i_wrbuffer_ref;
  96         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
  97              "snapc %p seq %lld (%d snaps)\n",
  98              mapping->host, page, page->index,
  99              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 100              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 101              snapc, snapc->seq, snapc->num_snaps);
 102         spin_unlock(&inode->i_lock);
 103
 104         /* now adjust page */
 105         spin_lock_irq(&mapping->tree_lock);
 106         if (page->mapping) {    /* Race with truncate? */
 107                 WARN_ON_ONCE(!PageUptodate(page));
 108
 109                 if (mapping_cap_account_dirty(mapping)) {
 110                         __inc_zone_page_state(page, NR_FILE_DIRTY);
 111                         __inc_bdi_stat(mapping->backing_dev_info,
 112                                         BDI_RECLAIMABLE);
 113                         task_io_account_write(PAGE_CACHE_SIZE);
 114                 }
 115                 radix_tree_tag_set(&mapping->page_tree,
 116                                 page_index(page), PAGECACHE_TAG_DIRTY);
 117
 118                 /*
 119                  * Reference snap context in page->private.  Also set
 120                  * PagePrivate so that we get invalidatepage callback.
 121                  */
 122                 page->private = (unsigned long)snapc;
 123                 SetPagePrivate(page);
 124         } else {
 125                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 126                 undo = 1;
 127         }
 128
 129         spin_unlock_irq(&mapping->tree_lock);
 130
 131         if (undo)
 132                 /* whoops, we failed to dirty the page */
 133                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 134
 135         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 136
 137         BUG_ON(!PageDirty(page));
 138         return 1;
 139 }
 140
 141 /*
 142  * If we are truncating the full page (i.e. offset == 0), adjust the
 143  * dirty page counters appropriately.  Only called if there is private
 144  * data on the page.
 145  */
 146 static void ceph_invalidatepage(struct page *page, unsigned long offset)
 147 {
 148         struct inode *inode;
 149         struct ceph_inode_info *ci;
 150         struct ceph_snap_context *snapc = (void *)page->private;
 151
 152         BUG_ON(!PageLocked(page));
 153         BUG_ON(!page->private);
 154         BUG_ON(!PagePrivate(page));
 155         BUG_ON(!page->mapping);
 156
 157         inode = page->mapping->host;
 158
 159         /*
 160          * We can get non-dirty pages here due to races between
 161          * set_page_dirty and truncate_complete_page; just spit out a
 162          * warning, in case we end up with accounting problems later.
 163          */
 164         if (!PageDirty(page))
 165                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 166
 167         if (offset == 0)
 168                 ClearPageChecked(page);
 169
 170         ci = ceph_inode(inode);
 171         if (offset == 0) {
 172                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 173                      inode, page, page->index, offset);
 174                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 175                 ceph_put_snap_context(snapc);
 176                 page->private = 0;
 177                 ClearPagePrivate(page);
 178         } else {
 179                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
 180                      inode, page, page->index);
 181         }
 182 }
 183
 184 /* just a sanity check */
 185 static int ceph_releasepage(struct page *page, gfp_t g)
 186 {
 187         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 188         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 189         WARN_ON(PageDirty(page));
 190         WARN_ON(page->private);
 191         WARN_ON(PagePrivate(page));
 192         return 0;
 193 }
 194
 195 /*
 196  * read a single page, without unlocking it.
 197  */
 198 static int readpage_nounlock(struct file *filp, struct page *page)
 199 {
 200         struct inode *inode = filp->f_dentry->d_inode;
 201         struct ceph_inode_info *ci = ceph_inode(inode);
 202         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 203         int err = 0;
 204         u64 len = PAGE_CACHE_SIZE;
 205
 206         dout("readpage inode %p file %p page %p index %lu\n",
 207              inode, filp, page, page->index);
 208         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 209                                   page->index << PAGE_CACHE_SHIFT, &len,
 210                                   ci->i_truncate_seq, ci->i_truncate_size,
 211                                   &page, 1);
 212         if (err == -ENOENT)
 213                 err = 0;
 214         if (err < 0) {
 215                 SetPageError(page);
 216                 goto out;
 217         } else if (err < PAGE_CACHE_SIZE) {
 218                 /* zero fill remainder of page */
 219                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 220         }
 221         SetPageUptodate(page);
 222
 223 out:
 224         return err < 0 ? err : 0;
 225 }
 226
 227 static int ceph_readpage(struct file *filp, struct page *page)
 228 {
 229         int r = readpage_nounlock(filp, page);
 230         unlock_page(page);
 231         return r;
 232 }
 233
 234 /*
 235  * Build a vector of contiguous pages from the provided page list.
 236  */
 237 static struct page **page_vector_from_list(struct list_head *page_list,
 238                                            unsigned *nr_pages)
 239 {
 240         struct page **pages;
 241         struct page *page;
 242         int next_index, contig_pages = 0;
 243
 244         /* build page vector */
 245         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 246         if (!pages)
 247                 return ERR_PTR(-ENOMEM);
 248
 249         BUG_ON(list_empty(page_list));
 250         next_index = list_entry(page_list->prev, struct page, lru)->index;
 251         list_for_each_entry_reverse(page, page_list, lru) {
 252                 if (page->index == next_index) {
 253                         dout("readpages page %d %p\n", contig_pages, page);
 254                         pages[contig_pages] = page;
 255                         contig_pages++;
 256                         next_index++;
 257                 } else {
 258                         break;
 259                 }
 260         }
 261         *nr_pages = contig_pages;
 262         return pages;
 263 }
 264
 265 /*
 266  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 267  * the caller (VM) cleans them up.
 268  */
 269 static int ceph_readpages(struct file *file, struct address_space *mapping,
 270                           struct list_head *page_list, unsigned nr_pages)
 271 {
 272         struct inode *inode = file->f_dentry->d_inode;
 273         struct ceph_inode_info *ci = ceph_inode(inode);
 274         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 275         int rc = 0;
 276         struct page **pages;
 277         loff_t offset;
 278         u64 len;
 279
 280         dout("readpages %p file %p nr_pages %d\n",
 281              inode, file, nr_pages);
 282
 283         pages = page_vector_from_list(page_list, &nr_pages);
 284         if (IS_ERR(pages))
 285                 return PTR_ERR(pages);
 286
 287         /* guess read extent */
 288         offset = pages[0]->index << PAGE_CACHE_SHIFT;
 289         len = nr_pages << PAGE_CACHE_SHIFT;
 290         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 291                                  offset, &len,
 292                                  ci->i_truncate_seq, ci->i_truncate_size,
 293                                  pages, nr_pages);
 294         if (rc == -ENOENT)
 295                 rc = 0;
 296         if (rc < 0)
 297                 goto out;
 298
 299         for (; !list_empty(page_list) && len > 0;
 300              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
 301                 struct page *page =
 302                         list_entry(page_list->prev, struct page, lru);
 303
 304                 list_del(&page->lru);
 305
 306                 if (rc < (int)PAGE_CACHE_SIZE) {
 307                         /* zero (remainder of) page */
 308                         int s = rc < 0 ? 0 : rc;
 309                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 310                 }
 311
 312                 if (add_to_page_cache_lru(page, mapping, page->index,
 313                                           GFP_NOFS)) {
 314                         page_cache_release(page);
 315                         dout("readpages %p add_to_page_cache failed %p\n",
 316                              inode, page);
 317                         continue;
 318                 }
 319                 dout("readpages %p adding %p idx %lu\n", inode, page,
 320                      page->index);
 321                 flush_dcache_page(page);
 322                 SetPageUptodate(page);
 323                 unlock_page(page);
 324                 page_cache_release(page);
 325         }
 326         rc = 0;
 327
 328 out:
 329         kfree(pages);
 330         return rc;
 331 }
 332
 333 /*
 334  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 335  * only snap context we are allowed to write back.
 336  */
 337 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 338                                                     u64 *snap_size)
 339 {
 340         struct ceph_inode_info *ci = ceph_inode(inode);
 341         struct ceph_snap_context *snapc = NULL;
 342         struct ceph_cap_snap *capsnap = NULL;
 343
 344         spin_lock(&inode->i_lock);
 345         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 346                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 347                      capsnap->context, capsnap->dirty_pages);
 348                 if (capsnap->dirty_pages) {
 349                         snapc = ceph_get_snap_context(capsnap->context);
 350                         if (snap_size)
 351                                 *snap_size = capsnap->size;
 352                         break;
 353                 }
 354         }
 355         if (!snapc && ci->i_head_snapc) {
 356                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 357                 dout(" head snapc %p has %d dirty pages\n",
 358                      snapc, ci->i_wrbuffer_ref_head);
 359         }
 360         spin_unlock(&inode->i_lock);
 361         return snapc;
 362 }
 363
 364 /*
 365  * Write a single page, but leave the page locked.
 366  *
 367  * If we get a write error, set the page error bit, but still adjust the
 368  * dirty page accounting (i.e., page is no longer dirty).
 369  */
 370 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 371 {
 372         struct inode *inode;
 373         struct ceph_inode_info *ci;
 374         struct ceph_client *client;
 375         struct ceph_osd_client *osdc;
 376         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 377         int len = PAGE_CACHE_SIZE;
 378         loff_t i_size;
 379         int err = 0;
 380         struct ceph_snap_context *snapc, *oldest;
 381         u64 snap_size = 0;
 382         long writeback_stat;
 383
 384         dout("writepage %p idx %lu\n", page, page->index);
 385
 386         if (!page->mapping || !page->mapping->host) {
 387                 dout("writepage %p - no mapping\n", page);
 388                 return -EFAULT;
 389         }
 390         inode = page->mapping->host;
 391         ci = ceph_inode(inode);
 392         client = ceph_inode_to_client(inode);
 393         osdc = &client->osdc;
 394
 395         /* verify this is a writeable snap context */
 396         snapc = (void *)page->private;
 397         if (snapc == NULL) {
 398                 dout("writepage %p page %p not dirty?\n", inode, page);
 399                 goto out;
 400         }
 401         oldest = get_oldest_context(inode, &snap_size);
 402         if (snapc->seq > oldest->seq) {
 403                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 404                      inode, page, (void *)page->private);
 405                 /* we should only noop if called by kswapd */
 406                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 407                 ceph_put_snap_context(oldest);
 408                 goto out;
 409         }
 410         ceph_put_snap_context(oldest);
 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         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 424         if (writeback_stat >
 425             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
 426                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 427
 428         set_page_writeback(page);
 429         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 430                                    &ci->i_layout, snapc,
 431                                    page_off, len,
 432                                    ci->i_truncate_seq, ci->i_truncate_size,
 433                                    &inode->i_mtime,
 434                                    &page, 1, 0, 0, true);
 435         if (err < 0) {
 436                 dout("writepage setting page/mapping error %d %p\n", err, page);
 437                 SetPageError(page);
 438                 mapping_set_error(&inode->i_data, err);
 439                 if (wbc)
 440                         wbc->pages_skipped++;
 441         } else {
 442                 dout("writepage cleaned page %p\n", page);
 443                 err = 0;  /* vfs expects us to return 0 */
 444         }
 445         page->private = 0;
 446         ClearPagePrivate(page);
 447         end_page_writeback(page);
 448         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 449         ceph_put_snap_context(snapc);  /* page's reference */
 450 out:
 451         return err;
 452 }
 453
 454 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 455 {
 456         int err;
 457         struct inode *inode = page->mapping->host;
 458         BUG_ON(!inode);
 459         igrab(inode);
 460         err = writepage_nounlock(page, wbc);
 461         unlock_page(page);
 462         iput(inode);
 463         return err;
 464 }
 465
 466
 467 /*
 468  * lame release_pages helper.  release_pages() isn't exported to
 469  * modules.
 470  */
 471 static void ceph_release_pages(struct page **pages, int num)
 472 {
 473         struct pagevec pvec;
 474         int i;
 475
 476         pagevec_init(&pvec, 0);
 477         for (i = 0; i < num; i++) {
 478                 if (pagevec_add(&pvec, pages[i]) == 0)
 479                         pagevec_release(&pvec);
 480         }
 481         pagevec_release(&pvec);
 482 }
 483
 484
 485 /*
 486  * async writeback completion handler.
 487  *
 488  * If we get an error, set the mapping error bit, but not the individual
 489  * page error bits.
 490  */
 491 static void writepages_finish(struct ceph_osd_request *req,
 492                               struct ceph_msg *msg)
 493 {
 494         struct inode *inode = req->r_inode;
 495         struct ceph_osd_reply_head *replyhead;
 496         struct ceph_osd_op *op;
 497         struct ceph_inode_info *ci = ceph_inode(inode);
 498         unsigned wrote;
 499         struct page *page;
 500         int i;
 501         struct ceph_snap_context *snapc = req->r_snapc;
 502         struct address_space *mapping = inode->i_mapping;
 503         __s32 rc = -EIO;
 504         u64 bytes = 0;
 505         struct ceph_client *client = ceph_inode_to_client(inode);
 506         long writeback_stat;
 507         unsigned issued = ceph_caps_issued(ci);
 508
 509         /* parse reply */
 510         replyhead = msg->front.iov_base;
 511         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 512         op = (void *)(replyhead + 1);
 513         rc = le32_to_cpu(replyhead->result);
 514         bytes = le64_to_cpu(op->extent.length);
 515
 516         if (rc >= 0) {
 517                 /*
 518                  * Assume we wrote the pages we originally sent.  The
 519                  * osd might reply with fewer pages if our writeback
 520                  * raced with a truncation and was adjusted at the osd,
 521                  * so don't believe the reply.
 522                  */
 523                 wrote = req->r_num_pages;
 524         } else {
 525                 wrote = 0;
 526                 mapping_set_error(mapping, rc);
 527         }
 528         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 529              inode, rc, bytes, wrote);
 530
 531         /* clean all pages */
 532         for (i = 0; i < req->r_num_pages; i++) {
 533                 page = req->r_pages[i];
 534                 BUG_ON(!page);
 535                 WARN_ON(!PageUptodate(page));
 536
 537                 writeback_stat =
 538                         atomic_long_dec_return(&client->writeback_count);
 539                 if (writeback_stat <
 540                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
 541                         clear_bdi_congested(&client->backing_dev_info,
 542                                             BLK_RW_ASYNC);
 543
 544                 ceph_put_snap_context((void *)page->private);
 545                 page->private = 0;
 546                 ClearPagePrivate(page);
 547                 dout("unlocking %d %p\n", i, page);
 548                 end_page_writeback(page);
 549
 550                 /*
 551                  * We lost the cache cap, need to truncate the page before
 552                  * it is unlocked, otherwise we'd truncate it later in the
 553                  * page truncation thread, possibly losing some data that
 554                  * raced its way in
 555                  */
 556                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 557                         generic_error_remove_page(inode->i_mapping, page);
 558
 559                 unlock_page(page);
 560         }
 561         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 562         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 563
 564         ceph_release_pages(req->r_pages, req->r_num_pages);
 565         if (req->r_pages_from_pool)
 566                 mempool_free(req->r_pages,
 567                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 568         else
 569                 kfree(req->r_pages);
 570         ceph_osdc_put_request(req);
 571 }
 572
 573 /*
 574  * allocate a page vec, either directly, or if necessary, via a the
 575  * mempool.  we avoid the mempool if we can because req->r_num_pages
 576  * may be less than the maximum write size.
 577  */
 578 static void alloc_page_vec(struct ceph_client *client,
 579                            struct ceph_osd_request *req)
 580 {
 581         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 582                                GFP_NOFS);
 583         if (!req->r_pages) {
 584                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
 585                 req->r_pages_from_pool = 1;
 586                 WARN_ON(!req->r_pages);
 587         }
 588 }
 589
 590 /*
 591  * initiate async writeback
 592  */
 593 static int ceph_writepages_start(struct address_space *mapping,
 594                                  struct writeback_control *wbc)
 595 {
 596         struct inode *inode = mapping->host;
 597         struct backing_dev_info *bdi = mapping->backing_dev_info;
 598         struct ceph_inode_info *ci = ceph_inode(inode);
 599         struct ceph_client *client;
 600         pgoff_t index, start, end;
 601         int range_whole = 0;
 602         int should_loop = 1;
 603         pgoff_t max_pages = 0, max_pages_ever = 0;
 604         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 605         struct pagevec pvec;
 606         int done = 0;
 607         int rc = 0;
 608         unsigned wsize = 1 << inode->i_blkbits;
 609         struct ceph_osd_request *req = NULL;
 610         int do_sync;
 611         u64 snap_size = 0;
 612
 613         /*
 614          * Include a 'sync' in the OSD request if this is a data
 615          * integrity write (e.g., O_SYNC write or fsync()), or if our
 616          * cap is being revoked.
 617          */
 618         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 619         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 620                 do_sync = 1;
 621         dout("writepages_start %p dosync=%d (mode=%s)\n",
 622              inode, do_sync,
 623              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 624              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 625
 626         client = ceph_inode_to_client(inode);
 627         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
 628                 pr_warning("writepage_start %p on forced umount\n", inode);
 629                 return -EIO; /* we're in a forced umount, don't write! */
 630         }
 631         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
 632                 wsize = client->mount_args->wsize;
 633         if (wsize < PAGE_CACHE_SIZE)
 634                 wsize = PAGE_CACHE_SIZE;
 635         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 636
 637         pagevec_init(&pvec, 0);
 638
 639         /* ?? */
 640         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 641                 dout(" writepages congested\n");
 642                 wbc->encountered_congestion = 1;
 643                 goto out_final;
 644         }
 645
 646         /* where to start/end? */
 647         if (wbc->range_cyclic) {
 648                 start = mapping->writeback_index; /* Start from prev offset */
 649                 end = -1;
 650                 dout(" cyclic, start at %lu\n", start);
 651         } else {
 652                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 653                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 654                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 655                         range_whole = 1;
 656                 should_loop = 0;
 657                 dout(" not cyclic, %lu to %lu\n", start, end);
 658         }
 659         index = start;
 660
 661 retry:
 662         /* find oldest snap context with dirty data */
 663         ceph_put_snap_context(snapc);
 664         snapc = get_oldest_context(inode, &snap_size);
 665         if (!snapc) {
 666                 /* hmm, why does writepages get called when there
 667                    is no dirty data? */
 668                 dout(" no snap context with dirty data?\n");
 669                 goto out;
 670         }
 671         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 672              snapc, snapc->seq, snapc->num_snaps);
 673         if (last_snapc && snapc != last_snapc) {
 674                 /* if we switched to a newer snapc, restart our scan at the
 675                  * start of the original file range. */
 676                 dout("  snapc differs from last pass, restarting at %lu\n",
 677                      index);
 678                 index = start;
 679         }
 680         last_snapc = snapc;
 681
 682         while (!done && index <= end) {
 683                 unsigned i;
 684                 int first;
 685                 pgoff_t next;
 686                 int pvec_pages, locked_pages;
 687                 struct page *page;
 688                 int want;
 689                 u64 offset, len;
 690                 struct ceph_osd_request_head *reqhead;
 691                 struct ceph_osd_op *op;
 692                 long writeback_stat;
 693
 694                 next = 0;
 695                 locked_pages = 0;
 696                 max_pages = max_pages_ever;
 697
 698 get_more_pages:
 699                 first = -1;
 700                 want = min(end - index,
 701                            min((pgoff_t)PAGEVEC_SIZE,
 702                                max_pages - (pgoff_t)locked_pages) - 1)
 703                         + 1;
 704                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 705                                                 PAGECACHE_TAG_DIRTY,
 706                                                 want);
 707                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 708                 if (!pvec_pages && !locked_pages)
 709                         break;
 710                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 711                         page = pvec.pages[i];
 712                         dout("? %p idx %lu\n", page, page->index);
 713                         if (locked_pages == 0)
 714                                 lock_page(page);  /* first page */
 715                         else if (!trylock_page(page))
 716                                 break;
 717
 718                         /* only dirty pages, or our accounting breaks */
 719                         if (unlikely(!PageDirty(page)) ||
 720                             unlikely(page->mapping != mapping)) {
 721                                 dout("!dirty or !mapping %p\n", page);
 722                                 unlock_page(page);
 723                                 break;
 724                         }
 725                         if (!wbc->range_cyclic && page->index > end) {
 726                                 dout("end of range %p\n", page);
 727                                 done = 1;
 728                                 unlock_page(page);
 729                                 break;
 730                         }
 731                         if (next && (page->index != next)) {
 732                                 dout("not consecutive %p\n", page);
 733                                 unlock_page(page);
 734                                 break;
 735                         }
 736                         if (wbc->sync_mode != WB_SYNC_NONE) {
 737                                 dout("waiting on writeback %p\n", page);
 738                                 wait_on_page_writeback(page);
 739                         }
 740                         if ((snap_size && page_offset(page) > snap_size) ||
 741                             (!snap_size &&
 742                              page_offset(page) > i_size_read(inode))) {
 743                                 dout("%p page eof %llu\n", page, snap_size ?
 744                                      snap_size : i_size_read(inode));
 745                                 done = 1;
 746                                 unlock_page(page);
 747                                 break;
 748                         }
 749                         if (PageWriteback(page)) {
 750                                 dout("%p under writeback\n", page);
 751                                 unlock_page(page);
 752                                 break;
 753                         }
 754
 755                         /* only if matching snap context */
 756                         pgsnapc = (void *)page->private;
 757                         if (pgsnapc->seq > snapc->seq) {
 758                                 dout("page snapc %p %lld > oldest %p %lld\n",
 759                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 760                                 unlock_page(page);
 761                                 if (!locked_pages)
 762                                         continue; /* keep looking for snap */
 763                                 break;
 764                         }
 765
 766                         if (!clear_page_dirty_for_io(page)) {
 767                                 dout("%p !clear_page_dirty_for_io\n", page);
 768                                 unlock_page(page);
 769                                 break;
 770                         }
 771
 772                         /* ok */
 773                         if (locked_pages == 0) {
 774                                 /* prepare async write request */
 775                                 offset = page->index << PAGE_CACHE_SHIFT;
 776                                 len = wsize;
 777                                 req = ceph_osdc_new_request(&client->osdc,
 778                                             &ci->i_layout,
 779                                             ceph_vino(inode),
 780                                             offset, &len,
 781                                             CEPH_OSD_OP_WRITE,
 782                                             CEPH_OSD_FLAG_WRITE |
 783                                                     CEPH_OSD_FLAG_ONDISK,
 784                                             snapc, do_sync,
 785                                             ci->i_truncate_seq,
 786                                             ci->i_truncate_size,
 787                                             &inode->i_mtime, true, 1);
 788                                 max_pages = req->r_num_pages;
 789
 790                                 alloc_page_vec(client, req);
 791                                 req->r_callback = writepages_finish;
 792                                 req->r_inode = inode;
 793                         }
 794
 795                         /* note position of first page in pvec */
 796                         if (first < 0)
 797                                 first = i;
 798                         dout("%p will write page %p idx %lu\n",
 799                              inode, page, page->index);
 800
 801                         writeback_stat =
 802                                atomic_long_inc_return(&client->writeback_count);
 803                         if (writeback_stat > CONGESTION_ON_THRESH(
 804                                     client->mount_args->congestion_kb)) {
 805                                 set_bdi_congested(&client->backing_dev_info,
 806                                                   BLK_RW_ASYNC);
 807                         }
 808
 809                         set_page_writeback(page);
 810                         req->r_pages[locked_pages] = page;
 811                         locked_pages++;
 812                         next = page->index + 1;
 813                 }
 814
 815                 /* did we get anything? */
 816                 if (!locked_pages)
 817                         goto release_pvec_pages;
 818                 if (i) {
 819                         int j;
 820                         BUG_ON(!locked_pages || first < 0);
 821
 822                         if (pvec_pages && i == pvec_pages &&
 823                             locked_pages < max_pages) {
 824                                 dout("reached end pvec, trying for more\n");
 825                                 pagevec_reinit(&pvec);
 826                                 goto get_more_pages;
 827                         }
 828
 829                         /* shift unused pages over in the pvec...  we
 830                          * will need to release them below. */
 831                         for (j = i; j < pvec_pages; j++) {
 832                                 dout(" pvec leftover page %p\n",
 833                                      pvec.pages[j]);
 834                                 pvec.pages[j-i+first] = pvec.pages[j];
 835                         }
 836                         pvec.nr -= i-first;
 837                 }
 838
 839                 /* submit the write */
 840                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 841                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 842                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 843                 dout("writepages got %d pages at %llu~%llu\n",
 844                      locked_pages, offset, len);
 845
 846                 /* revise final length, page count */
 847                 req->r_num_pages = locked_pages;
 848                 reqhead = req->r_request->front.iov_base;
 849                 op = (void *)(reqhead + 1);
 850                 op->extent.length = cpu_to_le64(len);
 851                 op->payload_len = cpu_to_le32(len);
 852                 req->r_request->hdr.data_len = cpu_to_le32(len);
 853
 854                 ceph_osdc_start_request(&client->osdc, req, true);
 855                 req = NULL;
 856
 857                 /* continue? */
 858                 index = next;
 859                 wbc->nr_to_write -= locked_pages;
 860                 if (wbc->nr_to_write <= 0)
 861                         done = 1;
 862
 863 release_pvec_pages:
 864                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 865                      pvec.nr ? pvec.pages[0] : NULL);
 866                 pagevec_release(&pvec);
 867
 868                 if (locked_pages && !done)
 869                         goto retry;
 870         }
 871
 872         if (should_loop && !done) {
 873                 /* more to do; loop back to beginning of file */
 874                 dout("writepages looping back to beginning of file\n");
 875                 should_loop = 0;
 876                 index = 0;
 877                 goto retry;
 878         }
 879
 880         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 881                 mapping->writeback_index = index;
 882
 883 out:
 884         if (req)
 885                 ceph_osdc_put_request(req);
 886         if (rc > 0)
 887                 rc = 0;  /* vfs expects us to return 0 */
 888         ceph_put_snap_context(snapc);
 889         dout("writepages done, rc = %d\n", rc);
 890 out_final:
 891         return rc;
 892 }
 893
 894
 895
 896 /*
 897  * See if a given @snapc is either writeable, or already written.
 898  */
 899 static int context_is_writeable_or_written(struct inode *inode,
 900                                            struct ceph_snap_context *snapc)
 901 {
 902         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 903         int ret = !oldest || snapc->seq <= oldest->seq;
 904
 905         ceph_put_snap_context(oldest);
 906         return ret;
 907 }
 908
 909 /*
 910  * We are only allowed to write into/dirty the page if the page is
 911  * clean, or already dirty within the same snap context.
 912  *
 913  * called with page locked.
 914  * return success with page locked,
 915  * or any failure (incl -EAGAIN) with page unlocked.
 916  */
 917 static int ceph_update_writeable_page(struct file *file,
 918                             loff_t pos, unsigned len,
 919                             struct page *page)
 920 {
 921         struct inode *inode = file->f_dentry->d_inode;
 922         struct ceph_inode_info *ci = ceph_inode(inode);
 923         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 924         loff_t page_off = pos & PAGE_CACHE_MASK;
 925         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 926         int end_in_page = pos_in_page + len;
 927         loff_t i_size;
 928         int r;
 929         struct ceph_snap_context *snapc, *oldest;
 930
 931 retry_locked:
 932         /* writepages currently holds page lock, but if we change that later, */
 933         wait_on_page_writeback(page);
 934
 935         /* check snap context */
 936         BUG_ON(!ci->i_snap_realm);
 937         down_read(&mdsc->snap_rwsem);
 938         BUG_ON(!ci->i_snap_realm->cached_context);
 939         snapc = (void *)page->private;
 940         if (snapc && snapc != ci->i_head_snapc) {
 941                 /*
 942                  * this page is already dirty in another (older) snap
 943                  * context!  is it writeable now?
 944                  */
 945                 oldest = get_oldest_context(inode, NULL);
 946                 up_read(&mdsc->snap_rwsem);
 947
 948                 if (snapc->seq > oldest->seq) {
 949                         ceph_put_snap_context(oldest);
 950                         dout(" page %p snapc %p not current or oldest\n",
 951                              page, snapc);
 952                         /*
 953                          * queue for writeback, and wait for snapc to
 954                          * be writeable or written
 955                          */
 956                         snapc = ceph_get_snap_context(snapc);
 957                         unlock_page(page);
 958                         ceph_queue_writeback(inode);
 959                         r = wait_event_interruptible(ci->i_cap_wq,
 960                                context_is_writeable_or_written(inode, snapc));
 961                         ceph_put_snap_context(snapc);
 962                         if (r == -ERESTARTSYS)
 963                                 return r;
 964                         return -EAGAIN;
 965                 }
 966                 ceph_put_snap_context(oldest);
 967
 968                 /* yay, writeable, do it now (without dropping page lock) */
 969                 dout(" page %p snapc %p not current, but oldest\n",
 970                      page, snapc);
 971                 if (!clear_page_dirty_for_io(page))
 972                         goto retry_locked;
 973                 r = writepage_nounlock(page, NULL);
 974                 if (r < 0)
 975                         goto fail_nosnap;
 976                 goto retry_locked;
 977         }
 978
 979         if (PageUptodate(page)) {
 980                 dout(" page %p already uptodate\n", page);
 981                 return 0;
 982         }
 983
 984         /* full page? */
 985         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 986                 return 0;
 987
 988         /* past end of file? */
 989         i_size = inode->i_size;   /* caller holds i_mutex */
 990
 991         if (i_size + len > inode->i_sb->s_maxbytes) {
 992                 /* file is too big */
 993                 r = -EINVAL;
 994                 goto fail;
 995         }
 996
 997         if (page_off >= i_size ||
 998             (pos_in_page == 0 && (pos+len) >= i_size &&
 999              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1000                 dout(" zeroing %p 0 - %d and %d - %d\n",
1001                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1002                 zero_user_segments(page,
1003                                    0, pos_in_page,
1004                                    end_in_page, PAGE_CACHE_SIZE);
1005                 return 0;
1006         }
1007
1008         /* we need to read it. */
1009         up_read(&mdsc->snap_rwsem);
1010         r = readpage_nounlock(file, page);
1011         if (r < 0)
1012                 goto fail_nosnap;
1013         goto retry_locked;
1014
1015 fail:
1016         up_read(&mdsc->snap_rwsem);
1017 fail_nosnap:
1018         unlock_page(page);
1019         return r;
1020 }
1021
1022 /*
1023  * We are only allowed to write into/dirty the page if the page is
1024  * clean, or already dirty within the same snap context.
1025  */
1026 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1027                             loff_t pos, unsigned len, unsigned flags,
1028                             struct page **pagep, void **fsdata)
1029 {
1030         struct inode *inode = file->f_dentry->d_inode;
1031         struct page *page;
1032         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1033         int r;
1034
1035         do {
1036                 /* get a page */
1037                 page = grab_cache_page_write_begin(mapping, index, 0);
1038                 if (!page)
1039                         return -ENOMEM;
1040                 *pagep = page;
1041
1042                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1043                      inode, page, (int)pos, (int)len);
1044
1045                 r = ceph_update_writeable_page(file, pos, len, page);
1046         } while (r == -EAGAIN);
1047
1048         return r;
1049 }
1050
1051 /*
1052  * we don't do anything in here that simple_write_end doesn't do
1053  * except adjust dirty page accounting and drop read lock on
1054  * mdsc->snap_rwsem.
1055  */
1056 static int ceph_write_end(struct file *file, struct address_space *mapping,
1057                           loff_t pos, unsigned len, unsigned copied,
1058                           struct page *page, void *fsdata)
1059 {
1060         struct inode *inode = file->f_dentry->d_inode;
1061         struct ceph_client *client = ceph_inode_to_client(inode);
1062         struct ceph_mds_client *mdsc = &client->mdsc;
1063         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1064         int check_cap = 0;
1065
1066         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1067              inode, page, (int)pos, (int)copied, (int)len);
1068
1069         /* zero the stale part of the page if we did a short copy */
1070         if (copied < len)
1071                 zero_user_segment(page, from+copied, len);
1072
1073         /* did file size increase? */
1074         /* (no need for i_size_read(); we caller holds i_mutex */
1075         if (pos+copied > inode->i_size)
1076                 check_cap = ceph_inode_set_size(inode, pos+copied);
1077
1078         if (!PageUptodate(page))
1079                 SetPageUptodate(page);
1080
1081         set_page_dirty(page);
1082
1083         unlock_page(page);
1084         up_read(&mdsc->snap_rwsem);
1085         page_cache_release(page);
1086
1087         if (check_cap)
1088                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1089
1090         return copied;
1091 }
1092
1093 /*
1094  * we set .direct_IO to indicate direct io is supported, but since we
1095  * intercept O_DIRECT reads and writes early, this function should
1096  * never get called.
1097  */
1098 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1099                               const struct iovec *iov,
1100                               loff_t pos, unsigned long nr_segs)
1101 {
1102         WARN_ON(1);
1103         return -EINVAL;
1104 }
1105
1106 const struct address_space_operations ceph_aops = {
1107         .readpage = ceph_readpage,
1108         .readpages = ceph_readpages,
1109         .writepage = ceph_writepage,
1110         .writepages = ceph_writepages_start,
1111         .write_begin = ceph_write_begin,
1112         .write_end = ceph_write_end,
1113         .set_page_dirty = ceph_set_page_dirty,
1114         .invalidatepage = ceph_invalidatepage,
1115         .releasepage = ceph_releasepage,
1116         .direct_IO = ceph_direct_io,
1117 };
1118
1119
1120 /*
1121  * vm ops
1122  */
1123
1124 /*
1125  * Reuse write_begin here for simplicity.
1126  */
1127 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1128 {
1129         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1130         struct page *page = vmf->page;
1131         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1132         loff_t off = page->index << PAGE_CACHE_SHIFT;
1133         loff_t size, len;
1134         int ret;
1135
1136         size = i_size_read(inode);
1137         if (off + PAGE_CACHE_SIZE <= size)
1138                 len = PAGE_CACHE_SIZE;
1139         else
1140                 len = size & ~PAGE_CACHE_MASK;
1141
1142         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1143              off, len, page, page->index);
1144
1145         lock_page(page);
1146
1147         ret = VM_FAULT_NOPAGE;
1148         if ((off > size) ||
1149             (page->mapping != inode->i_mapping))
1150                 goto out;
1151
1152         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1153         if (ret == 0) {
1154                 /* success.  we'll keep the page locked. */
1155                 set_page_dirty(page);
1156                 up_read(&mdsc->snap_rwsem);
1157                 ret = VM_FAULT_LOCKED;
1158         } else {
1159                 if (ret == -ENOMEM)
1160                         ret = VM_FAULT_OOM;
1161                 else
1162                         ret = VM_FAULT_SIGBUS;
1163         }
1164 out:
1165         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1166         if (ret != VM_FAULT_LOCKED)
1167                 unlock_page(page);
1168         return ret;
1169 }
1170
1171 static struct vm_operations_struct ceph_vmops = {
1172         .fault          = filemap_fault,
1173         .page_mkwrite   = ceph_page_mkwrite,
1174 };
1175
1176 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1177 {
1178         struct address_space *mapping = file->f_mapping;
1179
1180         if (!mapping->a_ops->readpage)
1181                 return -ENOEXEC;
1182         file_accessed(file);
1183         vma->vm_ops = &ceph_vmops;
1184         vma->vm_flags |= VM_CAN_NONLINEAR;
1185         return 0;
1186 }