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 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  51 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  52         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  53          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  54
  55
  56
  57 /*
  58  * Dirty a page.  Optimistically adjust accounting, on the assumption
  59  * that we won't race with invalidate.  If we do, readjust.
  60  */
  61 static int ceph_set_page_dirty(struct page *page)
  62 {
  63         struct address_space *mapping = page->mapping;
  64         struct inode *inode;
  65         struct ceph_inode_info *ci;
  66         int undo = 0;
  67         struct ceph_snap_context *snapc;
  68
  69         if (unlikely(!mapping))
  70                 return !TestSetPageDirty(page);
  71
  72         if (TestSetPageDirty(page)) {
  73                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  74                      mapping->host, page, page->index);
  75                 return 0;
  76         }
  77
  78         inode = mapping->host;
  79         ci = ceph_inode(inode);
  80
  81         /*
  82          * Note that we're grabbing a snapc ref here without holding
  83          * any locks!
  84          */
  85         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  86
  87         /* dirty the head */
  88         spin_lock(&inode->i_lock);
  89         if (ci->i_wrbuffer_ref_head == 0)
  90                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  91         ++ci->i_wrbuffer_ref_head;
  92         if (ci->i_wrbuffer_ref == 0)
  93                 igrab(inode);
  94         ++ci->i_wrbuffer_ref;
  95         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
  96              "snapc %p seq %lld (%d snaps)\n",
  97              mapping->host, page, page->index,
  98              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
  99              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 100              snapc, snapc->seq, snapc->num_snaps);
 101         spin_unlock(&inode->i_lock);
 102
 103         /* now adjust page */
 104         spin_lock_irq(&mapping->tree_lock);
 105         if (page->mapping) {    /* Race with truncate? */
 106                 WARN_ON_ONCE(!PageUptodate(page));
 107
 108                 if (mapping_cap_account_dirty(mapping)) {
 109                         __inc_zone_page_state(page, NR_FILE_DIRTY);
 110                         __inc_bdi_stat(mapping->backing_dev_info,
 111                                         BDI_RECLAIMABLE);
 112                         task_io_account_write(PAGE_CACHE_SIZE);
 113                 }
 114                 radix_tree_tag_set(&mapping->page_tree,
 115                                 page_index(page), PAGECACHE_TAG_DIRTY);
 116
 117                 /*
 118                  * Reference snap context in page->private.  Also set
 119                  * PagePrivate so that we get invalidatepage callback.
 120                  */
 121                 page->private = (unsigned long)snapc;
 122                 SetPagePrivate(page);
 123         } else {
 124                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 125                 undo = 1;
 126         }
 127
 128         spin_unlock_irq(&mapping->tree_lock);
 129
 130         if (undo)
 131                 /* whoops, we failed to dirty the page */
 132                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 133
 134         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 135
 136         BUG_ON(!PageDirty(page));
 137         return 1;
 138 }
 139
 140 /*
 141  * If we are truncating the full page (i.e. offset == 0), adjust the
 142  * dirty page counters appropriately.  Only called if there is private
 143  * data on the page.
 144  */
 145 static void ceph_invalidatepage(struct page *page, unsigned long offset)
 146 {
 147         struct inode *inode = page->mapping->host;
 148         struct ceph_inode_info *ci;
 149         struct ceph_snap_context *snapc = (void *)page->private;
 150
 151         BUG_ON(!PageLocked(page));
 152         BUG_ON(!page->private);
 153         BUG_ON(!PagePrivate(page));
 154         BUG_ON(!page->mapping);
 155
 156         /*
 157          * We can get non-dirty pages here due to races between
 158          * set_page_dirty and truncate_complete_page; just spit out a
 159          * warning, in case we end up with accounting problems later.
 160          */
 161         if (!PageDirty(page))
 162                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 163
 164         if (offset == 0)
 165                 ClearPageChecked(page);
 166
 167         ci = ceph_inode(inode);
 168         if (offset == 0) {
 169                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 170                      inode, page, page->index, offset);
 171                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 172                 ceph_put_snap_context(snapc);
 173                 page->private = 0;
 174                 ClearPagePrivate(page);
 175         } else {
 176                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
 177                      inode, page, page->index);
 178         }
 179 }
 180
 181 /* just a sanity check */
 182 static int ceph_releasepage(struct page *page, gfp_t g)
 183 {
 184         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 185         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 186         WARN_ON(PageDirty(page));
 187         WARN_ON(page->private);
 188         WARN_ON(PagePrivate(page));
 189         return 0;
 190 }
 191
 192 /*
 193  * read a single page, without unlocking it.
 194  */
 195 static int readpage_nounlock(struct file *filp, struct page *page)
 196 {
 197         struct inode *inode = filp->f_dentry->d_inode;
 198         struct ceph_inode_info *ci = ceph_inode(inode);
 199         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 200         int err = 0;
 201         u64 len = PAGE_CACHE_SIZE;
 202
 203         dout("readpage inode %p file %p page %p index %lu\n",
 204              inode, filp, page, page->index);
 205         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 206                                   page->index << PAGE_CACHE_SHIFT, &len,
 207                                   ci->i_truncate_seq, ci->i_truncate_size,
 208                                   &page, 1);
 209         if (err == -ENOENT)
 210                 err = 0;
 211         if (err < 0) {
 212                 SetPageError(page);
 213                 goto out;
 214         } else if (err < PAGE_CACHE_SIZE) {
 215                 /* zero fill remainder of page */
 216                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 217         }
 218         SetPageUptodate(page);
 219
 220 out:
 221         return err < 0 ? err : 0;
 222 }
 223
 224 static int ceph_readpage(struct file *filp, struct page *page)
 225 {
 226         int r = readpage_nounlock(filp, page);
 227         unlock_page(page);
 228         return r;
 229 }
 230
 231 /*
 232  * Build a vector of contiguous pages from the provided page list.
 233  */
 234 static struct page **page_vector_from_list(struct list_head *page_list,
 235                                            unsigned *nr_pages)
 236 {
 237         struct page **pages;
 238         struct page *page;
 239         int next_index, contig_pages = 0;
 240
 241         /* build page vector */
 242         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 243         if (!pages)
 244                 return ERR_PTR(-ENOMEM);
 245
 246         BUG_ON(list_empty(page_list));
 247         next_index = list_entry(page_list->prev, struct page, lru)->index;
 248         list_for_each_entry_reverse(page, page_list, lru) {
 249                 if (page->index == next_index) {
 250                         dout("readpages page %d %p\n", contig_pages, page);
 251                         pages[contig_pages] = page;
 252                         contig_pages++;
 253                         next_index++;
 254                 } else {
 255                         break;
 256                 }
 257         }
 258         *nr_pages = contig_pages;
 259         return pages;
 260 }
 261
 262 /*
 263  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 264  * the caller (VM) cleans them up.
 265  */
 266 static int ceph_readpages(struct file *file, struct address_space *mapping,
 267                           struct list_head *page_list, unsigned nr_pages)
 268 {
 269         struct inode *inode = file->f_dentry->d_inode;
 270         struct ceph_inode_info *ci = ceph_inode(inode);
 271         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 272         int rc = 0;
 273         struct page **pages;
 274         struct pagevec pvec;
 275         loff_t offset;
 276         u64 len;
 277
 278         dout("readpages %p file %p nr_pages %d\n",
 279              inode, file, nr_pages);
 280
 281         pages = page_vector_from_list(page_list, &nr_pages);
 282         if (IS_ERR(pages))
 283                 return PTR_ERR(pages);
 284
 285         /* guess read extent */
 286         offset = pages[0]->index << PAGE_CACHE_SHIFT;
 287         len = nr_pages << PAGE_CACHE_SHIFT;
 288         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 289                                  offset, &len,
 290                                  ci->i_truncate_seq, ci->i_truncate_size,
 291                                  pages, nr_pages);
 292         if (rc == -ENOENT)
 293                 rc = 0;
 294         if (rc < 0)
 295                 goto out;
 296
 297         /* set uptodate and add to lru in pagevec-sized chunks */
 298         pagevec_init(&pvec, 0);
 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(page, mapping, page->index, GFP_NOFS)) {
 313                         page_cache_release(page);
 314                         dout("readpages %p add_to_page_cache failed %p\n",
 315                              inode, page);
 316                         continue;
 317                 }
 318                 dout("readpages %p adding %p idx %lu\n", inode, page,
 319                      page->index);
 320                 flush_dcache_page(page);
 321                 SetPageUptodate(page);
 322                 unlock_page(page);
 323                 if (pagevec_add(&pvec, page) == 0)
 324                         pagevec_lru_add_file(&pvec);   /* add to lru */
 325         }
 326         pagevec_lru_add_file(&pvec);
 327         rc = 0;
 328
 329 out:
 330         kfree(pages);
 331         return rc;
 332 }
 333
 334 /*
 335  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 336  * only snap context we are allowed to write back.
 337  *
 338  * Caller holds i_lock.
 339  */
 340 static struct ceph_snap_context *__get_oldest_context(struct inode *inode,
 341                                                       u64 *snap_size)
 342 {
 343         struct ceph_inode_info *ci = ceph_inode(inode);
 344         struct ceph_snap_context *snapc = NULL;
 345         struct ceph_cap_snap *capsnap = NULL;
 346
 347         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 348                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 349                      capsnap->context, capsnap->dirty_pages);
 350                 if (capsnap->dirty_pages) {
 351                         snapc = ceph_get_snap_context(capsnap->context);
 352                         if (snap_size)
 353                                 *snap_size = capsnap->size;
 354                         break;
 355                 }
 356         }
 357         if (!snapc && ci->i_snap_realm) {
 358                 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
 359                 dout(" head snapc %p has %d dirty pages\n",
 360                      snapc, ci->i_wrbuffer_ref_head);
 361         }
 362         return snapc;
 363 }
 364
 365 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 366                                                     u64 *snap_size)
 367 {
 368         struct ceph_snap_context *snapc = NULL;
 369
 370         spin_lock(&inode->i_lock);
 371         snapc = __get_oldest_context(inode, snap_size);
 372         spin_unlock(&inode->i_lock);
 373         return snapc;
 374 }
 375
 376 /*
 377  * Write a single page, but leave the page locked.
 378  *
 379  * If we get a write error, set the page error bit, but still adjust the
 380  * dirty page accounting (i.e., page is no longer dirty).
 381  */
 382 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 383 {
 384         struct inode *inode;
 385         struct ceph_inode_info *ci;
 386         struct ceph_client *client;
 387         struct ceph_osd_client *osdc;
 388         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 389         int len = PAGE_CACHE_SIZE;
 390         loff_t i_size;
 391         int err = 0;
 392         struct ceph_snap_context *snapc;
 393         u64 snap_size = 0;
 394         long writeback_stat;
 395
 396         dout("writepage %p idx %lu\n", page, page->index);
 397
 398         if (!page->mapping || !page->mapping->host) {
 399                 dout("writepage %p - no mapping\n", page);
 400                 return -EFAULT;
 401         }
 402         inode = page->mapping->host;
 403         ci = ceph_inode(inode);
 404         client = ceph_inode_to_client(inode);
 405         osdc = &client->osdc;
 406
 407         /* verify this is a writeable snap context */
 408         snapc = (void *)page->private;
 409         if (snapc == NULL) {
 410                 dout("writepage %p page %p not dirty?\n", inode, page);
 411                 goto out;
 412         }
 413         if (snapc != get_oldest_context(inode, &snap_size)) {
 414                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 415                      inode, page, (void *)page->private);
 416                 /* we should only noop if called by kswapd */
 417                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 418                 goto out;
 419         }
 420
 421         /* is this a partial page at end of file? */
 422         if (snap_size)
 423                 i_size = snap_size;
 424         else
 425                 i_size = i_size_read(inode);
 426         if (i_size < page_off + len)
 427                 len = i_size - page_off;
 428
 429         dout("writepage %p page %p index %lu on %llu~%u\n",
 430              inode, page, page->index, page_off, len);
 431
 432         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 433         if (writeback_stat >
 434             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
 435                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 436
 437         set_page_writeback(page);
 438         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 439                                    &ci->i_layout, snapc,
 440                                    page_off, len,
 441                                    ci->i_truncate_seq, ci->i_truncate_size,
 442                                    &inode->i_mtime,
 443                                    &page, 1, 0, 0, true);
 444         if (err < 0) {
 445                 dout("writepage setting page/mapping error %d %p\n", err, page);
 446                 SetPageError(page);
 447                 mapping_set_error(&inode->i_data, err);
 448                 if (wbc)
 449                         wbc->pages_skipped++;
 450         } else {
 451                 dout("writepage cleaned page %p\n", page);
 452                 err = 0;  /* vfs expects us to return 0 */
 453         }
 454         page->private = 0;
 455         ClearPagePrivate(page);
 456         end_page_writeback(page);
 457         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 458         ceph_put_snap_context(snapc);
 459 out:
 460         return err;
 461 }
 462
 463 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 464 {
 465         int err;
 466         struct inode *inode = page->mapping->host;
 467         BUG_ON(!inode);
 468         igrab(inode);
 469         err = writepage_nounlock(page, wbc);
 470         unlock_page(page);
 471         iput(inode);
 472         return err;
 473 }
 474
 475
 476 /*
 477  * lame release_pages helper.  release_pages() isn't exported to
 478  * modules.
 479  */
 480 static void ceph_release_pages(struct page **pages, int num)
 481 {
 482         struct pagevec pvec;
 483         int i;
 484
 485         pagevec_init(&pvec, 0);
 486         for (i = 0; i < num; i++) {
 487                 if (pagevec_add(&pvec, pages[i]) == 0)
 488                         pagevec_release(&pvec);
 489         }
 490         pagevec_release(&pvec);
 491 }
 492
 493
 494 /*
 495  * async writeback completion handler.
 496  *
 497  * If we get an error, set the mapping error bit, but not the individual
 498  * page error bits.
 499  */
 500 static void writepages_finish(struct ceph_osd_request *req,
 501                               struct ceph_msg *msg)
 502 {
 503         struct inode *inode = req->r_inode;
 504         struct ceph_osd_reply_head *replyhead;
 505         struct ceph_osd_op *op;
 506         struct ceph_inode_info *ci = ceph_inode(inode);
 507         unsigned wrote;
 508         struct page *page;
 509         int i;
 510         struct ceph_snap_context *snapc = req->r_snapc;
 511         struct address_space *mapping = inode->i_mapping;
 512         struct writeback_control *wbc = req->r_wbc;
 513         __s32 rc = -EIO;
 514         u64 bytes = 0;
 515         struct ceph_client *client = ceph_inode_to_client(inode);
 516         long writeback_stat;
 517
 518         /* parse reply */
 519         replyhead = msg->front.iov_base;
 520         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 521         op = (void *)(replyhead + 1);
 522         rc = le32_to_cpu(replyhead->result);
 523         bytes = le64_to_cpu(op->extent.length);
 524
 525         if (rc >= 0) {
 526                 /*
 527                  * Assume we wrote the pages we originally sent.  The
 528                  * osd might reply with fewer pages if our writeback
 529                  * raced with a truncation and was adjusted at the osd,
 530                  * so don't believe the reply.
 531                  */
 532                 wrote = req->r_num_pages;
 533         } else {
 534                 wrote = 0;
 535                 mapping_set_error(mapping, rc);
 536         }
 537         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 538              inode, rc, bytes, wrote);
 539
 540         /* clean all pages */
 541         for (i = 0; i < req->r_num_pages; i++) {
 542                 page = req->r_pages[i];
 543                 BUG_ON(!page);
 544                 WARN_ON(!PageUptodate(page));
 545
 546                 writeback_stat =
 547                         atomic_long_dec_return(&client->writeback_count);
 548                 if (writeback_stat <
 549                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
 550                         clear_bdi_congested(&client->backing_dev_info,
 551                                             BLK_RW_ASYNC);
 552
 553                 if (i >= wrote) {
 554                         dout("inode %p skipping page %p\n", inode, page);
 555                         wbc->pages_skipped++;
 556                 }
 557                 page->private = 0;
 558                 ClearPagePrivate(page);
 559                 ceph_put_snap_context(snapc);
 560                 dout("unlocking %d %p\n", i, page);
 561                 end_page_writeback(page);
 562                 unlock_page(page);
 563         }
 564         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 565         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 566
 567         ceph_release_pages(req->r_pages, req->r_num_pages);
 568         if (req->r_pages_from_pool)
 569                 mempool_free(req->r_pages,
 570                              ceph_client(inode->i_sb)->wb_pagevec_pool);
 571         else
 572                 kfree(req->r_pages);
 573         ceph_osdc_put_request(req);
 574 }
 575
 576 /*
 577  * allocate a page vec, either directly, or if necessary, via a the
 578  * mempool.  we avoid the mempool if we can because req->r_num_pages
 579  * may be less than the maximum write size.
 580  */
 581 static void alloc_page_vec(struct ceph_client *client,
 582                            struct ceph_osd_request *req)
 583 {
 584         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 585                                GFP_NOFS);
 586         if (!req->r_pages) {
 587                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
 588                 req->r_pages_from_pool = 1;
 589                 WARN_ON(!req->r_pages);
 590         }
 591 }
 592
 593 /*
 594  * initiate async writeback
 595  */
 596 static int ceph_writepages_start(struct address_space *mapping,
 597                                  struct writeback_control *wbc)
 598 {
 599         struct inode *inode = mapping->host;
 600         struct backing_dev_info *bdi = mapping->backing_dev_info;
 601         struct ceph_inode_info *ci = ceph_inode(inode);
 602         struct ceph_client *client;
 603         pgoff_t index, start, end;
 604         int range_whole = 0;
 605         int should_loop = 1;
 606         pgoff_t max_pages = 0, max_pages_ever = 0;
 607         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL;
 608         struct pagevec pvec;
 609         int done = 0;
 610         int rc = 0;
 611         unsigned wsize = 1 << inode->i_blkbits;
 612         struct ceph_osd_request *req = NULL;
 613         int do_sync;
 614         u64 snap_size = 0;
 615
 616         /*
 617          * Include a 'sync' in the OSD request if this is a data
 618          * integrity write (e.g., O_SYNC write or fsync()), or if our
 619          * cap is being revoked.
 620          */
 621         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 622         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 623                 do_sync = 1;
 624         dout("writepages_start %p dosync=%d (mode=%s)\n",
 625              inode, do_sync,
 626              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 627              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 628
 629         client = ceph_inode_to_client(inode);
 630         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
 631                 pr_warning("writepage_start %p on forced umount\n", inode);
 632                 return -EIO; /* we're in a forced umount, don't write! */
 633         }
 634         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
 635                 wsize = client->mount_args->wsize;
 636         if (wsize < PAGE_CACHE_SIZE)
 637                 wsize = PAGE_CACHE_SIZE;
 638         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 639
 640         pagevec_init(&pvec, 0);
 641
 642         /* ?? */
 643         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 644                 dout(" writepages congested\n");
 645                 wbc->encountered_congestion = 1;
 646                 goto out_final;
 647         }
 648
 649         /* where to start/end? */
 650         if (wbc->range_cyclic) {
 651                 start = mapping->writeback_index; /* Start from prev offset */
 652                 end = -1;
 653                 dout(" cyclic, start at %lu\n", start);
 654         } else {
 655                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 656                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 657                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 658                         range_whole = 1;
 659                 should_loop = 0;
 660                 dout(" not cyclic, %lu to %lu\n", start, end);
 661         }
 662         index = start;
 663
 664 retry:
 665         /* find oldest snap context with dirty data */
 666         ceph_put_snap_context(snapc);
 667         snapc = get_oldest_context(inode, &snap_size);
 668         if (!snapc) {
 669                 /* hmm, why does writepages get called when there
 670                    is no dirty data? */
 671                 dout(" no snap context with dirty data?\n");
 672                 goto out;
 673         }
 674         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 675              snapc, snapc->seq, snapc->num_snaps);
 676         if (last_snapc && snapc != last_snapc) {
 677                 /* if we switched to a newer snapc, restart our scan at the
 678                  * start of the original file range. */
 679                 dout("  snapc differs from last pass, restarting at %lu\n",
 680                      index);
 681                 index = start;
 682         }
 683         last_snapc = snapc;
 684
 685         while (!done && index <= end) {
 686                 unsigned i;
 687                 int first;
 688                 pgoff_t next;
 689                 int pvec_pages, locked_pages;
 690                 struct page *page;
 691                 int want;
 692                 u64 offset, len;
 693                 struct ceph_osd_request_head *reqhead;
 694                 struct ceph_osd_op *op;
 695                 long writeback_stat;
 696
 697                 next = 0;
 698                 locked_pages = 0;
 699                 max_pages = max_pages_ever;
 700
 701 get_more_pages:
 702                 first = -1;
 703                 want = min(end - index,
 704                            min((pgoff_t)PAGEVEC_SIZE,
 705                                max_pages - (pgoff_t)locked_pages) - 1)
 706                         + 1;
 707                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 708                                                 PAGECACHE_TAG_DIRTY,
 709                                                 want);
 710                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 711                 if (!pvec_pages && !locked_pages)
 712                         break;
 713                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 714                         page = pvec.pages[i];
 715                         dout("? %p idx %lu\n", page, page->index);
 716                         if (locked_pages == 0)
 717                                 lock_page(page);  /* first page */
 718                         else if (!trylock_page(page))
 719                                 break;
 720
 721                         /* only dirty pages, or our accounting breaks */
 722                         if (unlikely(!PageDirty(page)) ||
 723                             unlikely(page->mapping != mapping)) {
 724                                 dout("!dirty or !mapping %p\n", page);
 725                                 unlock_page(page);
 726                                 break;
 727                         }
 728                         if (!wbc->range_cyclic && page->index > end) {
 729                                 dout("end of range %p\n", page);
 730                                 done = 1;
 731                                 unlock_page(page);
 732                                 break;
 733                         }
 734                         if (next && (page->index != next)) {
 735                                 dout("not consecutive %p\n", page);
 736                                 unlock_page(page);
 737                                 break;
 738                         }
 739                         if (wbc->sync_mode != WB_SYNC_NONE) {
 740                                 dout("waiting on writeback %p\n", page);
 741                                 wait_on_page_writeback(page);
 742                         }
 743                         if ((snap_size && page_offset(page) > snap_size) ||
 744                             (!snap_size &&
 745                              page_offset(page) > i_size_read(inode))) {
 746                                 dout("%p page eof %llu\n", page, snap_size ?
 747                                      snap_size : i_size_read(inode));
 748                                 done = 1;
 749                                 unlock_page(page);
 750                                 break;
 751                         }
 752                         if (PageWriteback(page)) {
 753                                 dout("%p under writeback\n", page);
 754                                 unlock_page(page);
 755                                 break;
 756                         }
 757
 758                         /* only if matching snap context */
 759                         if (snapc != (void *)page->private) {
 760                                 dout("page snapc %p != oldest %p\n",
 761                                      (void *)page->private, snapc);
 762                                 unlock_page(page);
 763                                 if (!locked_pages)
 764                                         continue; /* keep looking for snap */
 765                                 break;
 766                         }
 767
 768                         if (!clear_page_dirty_for_io(page)) {
 769                                 dout("%p !clear_page_dirty_for_io\n", page);
 770                                 unlock_page(page);
 771                                 break;
 772                         }
 773
 774                         /* ok */
 775                         if (locked_pages == 0) {
 776                                 /* prepare async write request */
 777                                 offset = page->index << PAGE_CACHE_SHIFT;
 778                                 len = wsize;
 779                                 req = ceph_osdc_new_request(&client->osdc,
 780                                             &ci->i_layout,
 781                                             ceph_vino(inode),
 782                                             offset, &len,
 783                                             CEPH_OSD_OP_WRITE,
 784                                             CEPH_OSD_FLAG_WRITE |
 785                                                     CEPH_OSD_FLAG_ONDISK,
 786                                             snapc, do_sync,
 787                                             ci->i_truncate_seq,
 788                                             ci->i_truncate_size,
 789                                             &inode->i_mtime, true, 1);
 790                                 max_pages = req->r_num_pages;
 791
 792                                 alloc_page_vec(client, req);
 793                                 req->r_callback = writepages_finish;
 794                                 req->r_inode = inode;
 795                                 req->r_wbc = wbc;
 796                         }
 797
 798                         /* note position of first page in pvec */
 799                         if (first < 0)
 800                                 first = i;
 801                         dout("%p will write page %p idx %lu\n",
 802                              inode, page, page->index);
 803
 804                         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 805                         if (writeback_stat > CONGESTION_ON_THRESH(client->mount_args->congestion_kb)) {
 806                                 set_bdi_congested(&client->backing_dev_info, 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         return !oldest || snapc->seq <= oldest->seq;
 904 }
 905
 906 /*
 907  * We are only allowed to write into/dirty the page if the page is
 908  * clean, or already dirty within the same snap context.
 909  */
 910 static int ceph_write_begin(struct file *file, struct address_space *mapping,
 911                             loff_t pos, unsigned len, unsigned flags,
 912                             struct page **pagep, void **fsdata)
 913 {
 914         struct inode *inode = file->f_dentry->d_inode;
 915         struct ceph_inode_info *ci = ceph_inode(inode);
 916         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 917         struct page *page;
 918         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 919         loff_t page_off = pos & PAGE_CACHE_MASK;
 920         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 921         int end_in_page = pos_in_page + len;
 922         loff_t i_size;
 923         struct ceph_snap_context *snapc;
 924         int r;
 925
 926         /* get a page*/
 927 retry:
 928         page = grab_cache_page_write_begin(mapping, index, 0);
 929         if (!page)
 930                 return -ENOMEM;
 931         *pagep = page;
 932
 933         dout("write_begin file %p inode %p page %p %d~%d\n", file,
 934              inode, page, (int)pos, (int)len);
 935
 936 retry_locked:
 937         /* writepages currently holds page lock, but if we change that later, */
 938         wait_on_page_writeback(page);
 939
 940         /* check snap context */
 941         BUG_ON(!ci->i_snap_realm);
 942         down_read(&mdsc->snap_rwsem);
 943         BUG_ON(!ci->i_snap_realm->cached_context);
 944         if (page->private &&
 945             (void *)page->private != ci->i_snap_realm->cached_context) {
 946                 /*
 947                  * this page is already dirty in another (older) snap
 948                  * context!  is it writeable now?
 949                  */
 950                 snapc = get_oldest_context(inode, NULL);
 951                 up_read(&mdsc->snap_rwsem);
 952
 953                 if (snapc != (void *)page->private) {
 954                         dout(" page %p snapc %p not current or oldest\n",
 955                              page, (void *)page->private);
 956                         /*
 957                          * queue for writeback, and wait for snapc to
 958                          * be writeable or written
 959                          */
 960                         snapc = ceph_get_snap_context((void *)page->private);
 961                         unlock_page(page);
 962                         if (ceph_queue_writeback(inode))
 963                                 igrab(inode);
 964                         wait_event_interruptible(ci->i_cap_wq,
 965                                context_is_writeable_or_written(inode, snapc));
 966                         ceph_put_snap_context(snapc);
 967                         goto retry;
 968                 }
 969
 970                 /* yay, writeable, do it now (without dropping page lock) */
 971                 dout(" page %p snapc %p not current, but oldest\n",
 972                      page, snapc);
 973                 if (!clear_page_dirty_for_io(page))
 974                         goto retry_locked;
 975                 r = writepage_nounlock(page, NULL);
 976                 if (r < 0)
 977                         goto fail_nosnap;
 978                 goto retry_locked;
 979         }
 980
 981         if (PageUptodate(page)) {
 982                 dout(" page %p already uptodate\n", page);
 983                 return 0;
 984         }
 985
 986         /* full page? */
 987         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 988                 return 0;
 989
 990         /* past end of file? */
 991         i_size = inode->i_size;   /* caller holds i_mutex */
 992
 993         if (i_size + len > inode->i_sb->s_maxbytes) {
 994                 /* file is too big */
 995                 r = -EINVAL;
 996                 goto fail;
 997         }
 998
 999         if (page_off >= i_size ||
1000             (pos_in_page == 0 && (pos+len) >= i_size &&
1001              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1002                 dout(" zeroing %p 0 - %d and %d - %d\n",
1003                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1004                 zero_user_segments(page,
1005                                    0, pos_in_page,
1006                                    end_in_page, PAGE_CACHE_SIZE);
1007                 return 0;
1008         }
1009
1010         /* we need to read it. */
1011         up_read(&mdsc->snap_rwsem);
1012         r = readpage_nounlock(file, page);
1013         if (r < 0)
1014                 goto fail_nosnap;
1015         goto retry_locked;
1016
1017 fail:
1018         up_read(&mdsc->snap_rwsem);
1019 fail_nosnap:
1020         unlock_page(page);
1021         return r;
1022 }
1023
1024 /*
1025  * we don't do anything in here that simple_write_end doesn't do
1026  * except adjust dirty page accounting and drop read lock on
1027  * mdsc->snap_rwsem.
1028  */
1029 static int ceph_write_end(struct file *file, struct address_space *mapping,
1030                           loff_t pos, unsigned len, unsigned copied,
1031                           struct page *page, void *fsdata)
1032 {
1033         struct inode *inode = file->f_dentry->d_inode;
1034         struct ceph_client *client = ceph_inode_to_client(inode);
1035         struct ceph_mds_client *mdsc = &client->mdsc;
1036         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1037         int check_cap = 0;
1038
1039         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1040              inode, page, (int)pos, (int)copied, (int)len);
1041
1042         /* zero the stale part of the page if we did a short copy */
1043         if (copied < len)
1044                 zero_user_segment(page, from+copied, len);
1045
1046         /* did file size increase? */
1047         /* (no need for i_size_read(); we caller holds i_mutex */
1048         if (pos+copied > inode->i_size)
1049                 check_cap = ceph_inode_set_size(inode, pos+copied);
1050
1051         if (!PageUptodate(page))
1052                 SetPageUptodate(page);
1053
1054         set_page_dirty(page);
1055
1056         unlock_page(page);
1057         up_read(&mdsc->snap_rwsem);
1058         page_cache_release(page);
1059
1060         if (check_cap)
1061                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1062
1063         return copied;
1064 }
1065
1066 /*
1067  * we set .direct_IO to indicate direct io is supported, but since we
1068  * intercept O_DIRECT reads and writes early, this function should
1069  * never get called.
1070  */
1071 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1072                               const struct iovec *iov,
1073                               loff_t pos, unsigned long nr_segs)
1074 {
1075         WARN_ON(1);
1076         return -EINVAL;
1077 }
1078
1079 const struct address_space_operations ceph_aops = {
1080         .readpage = ceph_readpage,
1081         .readpages = ceph_readpages,
1082         .writepage = ceph_writepage,
1083         .writepages = ceph_writepages_start,
1084         .write_begin = ceph_write_begin,
1085         .write_end = ceph_write_end,
1086         .set_page_dirty = ceph_set_page_dirty,
1087         .invalidatepage = ceph_invalidatepage,
1088         .releasepage = ceph_releasepage,
1089         .direct_IO = ceph_direct_io,
1090 };
1091
1092
1093 /*
1094  * vm ops
1095  */
1096
1097 /*
1098  * Reuse write_begin here for simplicity.
1099  */
1100 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1101 {
1102         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1103         struct page *page = vmf->page;
1104         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1105         loff_t off = page->index << PAGE_CACHE_SHIFT;
1106         loff_t size, len;
1107         struct page *locked_page = NULL;
1108         void *fsdata = NULL;
1109         int ret;
1110
1111         size = i_size_read(inode);
1112         if (off + PAGE_CACHE_SIZE <= size)
1113                 len = PAGE_CACHE_SIZE;
1114         else
1115                 len = size & ~PAGE_CACHE_MASK;
1116
1117         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1118              off, len, page, page->index);
1119         ret = ceph_write_begin(vma->vm_file, inode->i_mapping, off, len, 0,
1120                                &locked_page, &fsdata);
1121         WARN_ON(page != locked_page);
1122         if (!ret) {
1123                 /*
1124                  * doing the following, instead of calling
1125                  * ceph_write_end. Note that we keep the
1126                  * page locked
1127                  */
1128                 set_page_dirty(page);
1129                 up_read(&mdsc->snap_rwsem);
1130                 page_cache_release(page);
1131                 ret = VM_FAULT_LOCKED;
1132         } else {
1133                 ret = VM_FAULT_SIGBUS;
1134         }
1135         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1136         return ret;
1137 }
1138
1139 static struct vm_operations_struct ceph_vmops = {
1140         .fault          = filemap_fault,
1141         .page_mkwrite   = ceph_page_mkwrite,
1142 };
1143
1144 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1145 {
1146         struct address_space *mapping = file->f_mapping;
1147
1148         if (!mapping->a_ops->readpage)
1149                 return -ENOEXEC;
1150         file_accessed(file);
1151         vma->vm_ops = &ceph_vmops;
1152         vma->vm_flags |= VM_CAN_NONLINEAR;
1153         return 0;
1154 }