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;
 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         inode = page->mapping->host;
 157
 158         /*
 159          * We can get non-dirty pages here due to races between
 160          * set_page_dirty and truncate_complete_page; just spit out a
 161          * warning, in case we end up with accounting problems later.
 162          */
 163         if (!PageDirty(page))
 164                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 165
 166         if (offset == 0)
 167                 ClearPageChecked(page);
 168
 169         ci = ceph_inode(inode);
 170         if (offset == 0) {
 171                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 172                      inode, page, page->index, offset);
 173                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 174                 ceph_put_snap_context(snapc);
 175                 page->private = 0;
 176                 ClearPagePrivate(page);
 177         } else {
 178                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
 179                      inode, page, page->index);
 180         }
 181 }
 182
 183 /* just a sanity check */
 184 static int ceph_releasepage(struct page *page, gfp_t g)
 185 {
 186         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 187         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 188         WARN_ON(PageDirty(page));
 189         WARN_ON(page->private);
 190         WARN_ON(PagePrivate(page));
 191         return 0;
 192 }
 193
 194 /*
 195  * read a single page, without unlocking it.
 196  */
 197 static int readpage_nounlock(struct file *filp, struct page *page)
 198 {
 199         struct inode *inode = filp->f_dentry->d_inode;
 200         struct ceph_inode_info *ci = ceph_inode(inode);
 201         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 202         int err = 0;
 203         u64 len = PAGE_CACHE_SIZE;
 204
 205         dout("readpage inode %p file %p page %p index %lu\n",
 206              inode, filp, page, page->index);
 207         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 208                                   page->index << PAGE_CACHE_SHIFT, &len,
 209                                   ci->i_truncate_seq, ci->i_truncate_size,
 210                                   &page, 1);
 211         if (err == -ENOENT)
 212                 err = 0;
 213         if (err < 0) {
 214                 SetPageError(page);
 215                 goto out;
 216         } else if (err < PAGE_CACHE_SIZE) {
 217                 /* zero fill remainder of page */
 218                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 219         }
 220         SetPageUptodate(page);
 221
 222 out:
 223         return err < 0 ? err : 0;
 224 }
 225
 226 static int ceph_readpage(struct file *filp, struct page *page)
 227 {
 228         int r = readpage_nounlock(filp, page);
 229         unlock_page(page);
 230         return r;
 231 }
 232
 233 /*
 234  * Build a vector of contiguous pages from the provided page list.
 235  */
 236 static struct page **page_vector_from_list(struct list_head *page_list,
 237                                            unsigned *nr_pages)
 238 {
 239         struct page **pages;
 240         struct page *page;
 241         int next_index, contig_pages = 0;
 242
 243         /* build page vector */
 244         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 245         if (!pages)
 246                 return ERR_PTR(-ENOMEM);
 247
 248         BUG_ON(list_empty(page_list));
 249         next_index = list_entry(page_list->prev, struct page, lru)->index;
 250         list_for_each_entry_reverse(page, page_list, lru) {
 251                 if (page->index == next_index) {
 252                         dout("readpages page %d %p\n", contig_pages, page);
 253                         pages[contig_pages] = page;
 254                         contig_pages++;
 255                         next_index++;
 256                 } else {
 257                         break;
 258                 }
 259         }
 260         *nr_pages = contig_pages;
 261         return pages;
 262 }
 263
 264 /*
 265  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 266  * the caller (VM) cleans them up.
 267  */
 268 static int ceph_readpages(struct file *file, struct address_space *mapping,
 269                           struct list_head *page_list, unsigned nr_pages)
 270 {
 271         struct inode *inode = file->f_dentry->d_inode;
 272         struct ceph_inode_info *ci = ceph_inode(inode);
 273         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 274         int rc = 0;
 275         struct page **pages;
 276         struct pagevec pvec;
 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         /* set uptodate and add to lru in pagevec-sized chunks */
 300         pagevec_init(&pvec, 0);
 301         for (; !list_empty(page_list) && len > 0;
 302              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
 303                 struct page *page =
 304                         list_entry(page_list->prev, struct page, lru);
 305
 306                 list_del(&page->lru);
 307
 308                 if (rc < (int)PAGE_CACHE_SIZE) {
 309                         /* zero (remainder of) page */
 310                         int s = rc < 0 ? 0 : rc;
 311                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 312                 }
 313
 314                 if (add_to_page_cache(page, mapping, page->index, GFP_NOFS)) {
 315                         page_cache_release(page);
 316                         dout("readpages %p add_to_page_cache failed %p\n",
 317                              inode, page);
 318                         continue;
 319                 }
 320                 dout("readpages %p adding %p idx %lu\n", inode, page,
 321                      page->index);
 322                 flush_dcache_page(page);
 323                 SetPageUptodate(page);
 324                 unlock_page(page);
 325                 if (pagevec_add(&pvec, page) == 0)
 326                         pagevec_lru_add_file(&pvec);   /* add to lru */
 327         }
 328         pagevec_lru_add_file(&pvec);
 329         rc = 0;
 330
 331 out:
 332         kfree(pages);
 333         return rc;
 334 }
 335
 336 /*
 337  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 338  * only snap context we are allowed to write back.
 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         spin_lock(&inode->i_lock);
 348         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 349                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 350                      capsnap->context, capsnap->dirty_pages);
 351                 if (capsnap->dirty_pages) {
 352                         snapc = ceph_get_snap_context(capsnap->context);
 353                         if (snap_size)
 354                                 *snap_size = capsnap->size;
 355                         break;
 356                 }
 357         }
 358         if (!snapc && ci->i_head_snapc) {
 359                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 360                 dout(" head snapc %p has %d dirty pages\n",
 361                      snapc, ci->i_wrbuffer_ref_head);
 362         }
 363         spin_unlock(&inode->i_lock);
 364         return snapc;
 365 }
 366
 367 /*
 368  * Write a single page, but leave the page locked.
 369  *
 370  * If we get a write error, set the page error bit, but still adjust the
 371  * dirty page accounting (i.e., page is no longer dirty).
 372  */
 373 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 374 {
 375         struct inode *inode;
 376         struct ceph_inode_info *ci;
 377         struct ceph_client *client;
 378         struct ceph_osd_client *osdc;
 379         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 380         int len = PAGE_CACHE_SIZE;
 381         loff_t i_size;
 382         int err = 0;
 383         struct ceph_snap_context *snapc, *oldest;
 384         u64 snap_size = 0;
 385         long writeback_stat;
 386
 387         dout("writepage %p idx %lu\n", page, page->index);
 388
 389         if (!page->mapping || !page->mapping->host) {
 390                 dout("writepage %p - no mapping\n", page);
 391                 return -EFAULT;
 392         }
 393         inode = page->mapping->host;
 394         ci = ceph_inode(inode);
 395         client = ceph_inode_to_client(inode);
 396         osdc = &client->osdc;
 397
 398         /* verify this is a writeable snap context */
 399         snapc = (void *)page->private;
 400         if (snapc == NULL) {
 401                 dout("writepage %p page %p not dirty?\n", inode, page);
 402                 goto out;
 403         }
 404         oldest = get_oldest_context(inode, &snap_size);
 405         if (snapc->seq > oldest->seq) {
 406                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 407                      inode, page, (void *)page->private);
 408                 /* we should only noop if called by kswapd */
 409                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 410                 ceph_put_snap_context(oldest);
 411                 goto out;
 412         }
 413         ceph_put_snap_context(oldest);
 414
 415         /* is this a partial page at end of file? */
 416         if (snap_size)
 417                 i_size = snap_size;
 418         else
 419                 i_size = i_size_read(inode);
 420         if (i_size < page_off + len)
 421                 len = i_size - page_off;
 422
 423         dout("writepage %p page %p index %lu on %llu~%u\n",
 424              inode, page, page->index, page_off, len);
 425
 426         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 427         if (writeback_stat >
 428             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
 429                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 430
 431         set_page_writeback(page);
 432         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 433                                    &ci->i_layout, snapc,
 434                                    page_off, len,
 435                                    ci->i_truncate_seq, ci->i_truncate_size,
 436                                    &inode->i_mtime,
 437                                    &page, 1, 0, 0, true);
 438         if (err < 0) {
 439                 dout("writepage setting page/mapping error %d %p\n", err, page);
 440                 SetPageError(page);
 441                 mapping_set_error(&inode->i_data, err);
 442                 if (wbc)
 443                         wbc->pages_skipped++;
 444         } else {
 445                 dout("writepage cleaned page %p\n", page);
 446                 err = 0;  /* vfs expects us to return 0 */
 447         }
 448         page->private = 0;
 449         ClearPagePrivate(page);
 450         end_page_writeback(page);
 451         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 452         ceph_put_snap_context(snapc);  /* page's reference */
 453 out:
 454         return err;
 455 }
 456
 457 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 458 {
 459         int err;
 460         struct inode *inode = page->mapping->host;
 461         BUG_ON(!inode);
 462         igrab(inode);
 463         err = writepage_nounlock(page, wbc);
 464         unlock_page(page);
 465         iput(inode);
 466         return err;
 467 }
 468
 469
 470 /*
 471  * lame release_pages helper.  release_pages() isn't exported to
 472  * modules.
 473  */
 474 static void ceph_release_pages(struct page **pages, int num)
 475 {
 476         struct pagevec pvec;
 477         int i;
 478
 479         pagevec_init(&pvec, 0);
 480         for (i = 0; i < num; i++) {
 481                 if (pagevec_add(&pvec, pages[i]) == 0)
 482                         pagevec_release(&pvec);
 483         }
 484         pagevec_release(&pvec);
 485 }
 486
 487
 488 /*
 489  * async writeback completion handler.
 490  *
 491  * If we get an error, set the mapping error bit, but not the individual
 492  * page error bits.
 493  */
 494 static void writepages_finish(struct ceph_osd_request *req,
 495                               struct ceph_msg *msg)
 496 {
 497         struct inode *inode = req->r_inode;
 498         struct ceph_osd_reply_head *replyhead;
 499         struct ceph_osd_op *op;
 500         struct ceph_inode_info *ci = ceph_inode(inode);
 501         unsigned wrote;
 502         struct page *page;
 503         int i;
 504         struct ceph_snap_context *snapc = req->r_snapc;
 505         struct address_space *mapping = inode->i_mapping;
 506         struct writeback_control *wbc = req->r_wbc;
 507         __s32 rc = -EIO;
 508         u64 bytes = 0;
 509         struct ceph_client *client = ceph_inode_to_client(inode);
 510         long writeback_stat;
 511         unsigned issued = __ceph_caps_issued(ci, NULL);
 512
 513         /* parse reply */
 514         replyhead = msg->front.iov_base;
 515         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 516         op = (void *)(replyhead + 1);
 517         rc = le32_to_cpu(replyhead->result);
 518         bytes = le64_to_cpu(op->extent.length);
 519
 520         if (rc >= 0) {
 521                 /*
 522                  * Assume we wrote the pages we originally sent.  The
 523                  * osd might reply with fewer pages if our writeback
 524                  * raced with a truncation and was adjusted at the osd,
 525                  * so don't believe the reply.
 526                  */
 527                 wrote = req->r_num_pages;
 528         } else {
 529                 wrote = 0;
 530                 mapping_set_error(mapping, rc);
 531         }
 532         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 533              inode, rc, bytes, wrote);
 534
 535         /* clean all pages */
 536         for (i = 0; i < req->r_num_pages; i++) {
 537                 page = req->r_pages[i];
 538                 BUG_ON(!page);
 539                 WARN_ON(!PageUptodate(page));
 540
 541                 writeback_stat =
 542                         atomic_long_dec_return(&client->writeback_count);
 543                 if (writeback_stat <
 544                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
 545                         clear_bdi_congested(&client->backing_dev_info,
 546                                             BLK_RW_ASYNC);
 547
 548                 if (i >= wrote) {
 549                         dout("inode %p skipping page %p\n", inode, page);
 550                         wbc->pages_skipped++;
 551                 }
 552                 ceph_put_snap_context((void *)page->private);
 553                 page->private = 0;
 554                 ClearPagePrivate(page);
 555                 dout("unlocking %d %p\n", i, page);
 556                 end_page_writeback(page);
 557
 558                 /*
 559                  * We lost the cache cap, need to truncate the page before
 560                  * it is unlocked, otherwise we'd truncate it later in the
 561                  * page truncation thread, possibly losing some data that
 562                  * raced its way in
 563                  */
 564                 if ((issued & CEPH_CAP_FILE_CACHE) == 0)
 565                         generic_error_remove_page(inode->i_mapping, page);
 566
 567                 unlock_page(page);
 568         }
 569         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 570         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 571
 572         ceph_release_pages(req->r_pages, req->r_num_pages);
 573         if (req->r_pages_from_pool)
 574                 mempool_free(req->r_pages,
 575                              ceph_client(inode->i_sb)->wb_pagevec_pool);
 576         else
 577                 kfree(req->r_pages);
 578         ceph_osdc_put_request(req);
 579 }
 580
 581 /*
 582  * allocate a page vec, either directly, or if necessary, via a the
 583  * mempool.  we avoid the mempool if we can because req->r_num_pages
 584  * may be less than the maximum write size.
 585  */
 586 static void alloc_page_vec(struct ceph_client *client,
 587                            struct ceph_osd_request *req)
 588 {
 589         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 590                                GFP_NOFS);
 591         if (!req->r_pages) {
 592                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
 593                 req->r_pages_from_pool = 1;
 594                 WARN_ON(!req->r_pages);
 595         }
 596 }
 597
 598 /*
 599  * initiate async writeback
 600  */
 601 static int ceph_writepages_start(struct address_space *mapping,
 602                                  struct writeback_control *wbc)
 603 {
 604         struct inode *inode = mapping->host;
 605         struct backing_dev_info *bdi = mapping->backing_dev_info;
 606         struct ceph_inode_info *ci = ceph_inode(inode);
 607         struct ceph_client *client;
 608         pgoff_t index, start, end;
 609         int range_whole = 0;
 610         int should_loop = 1;
 611         pgoff_t max_pages = 0, max_pages_ever = 0;
 612         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 613         struct pagevec pvec;
 614         int done = 0;
 615         int rc = 0;
 616         unsigned wsize = 1 << inode->i_blkbits;
 617         struct ceph_osd_request *req = NULL;
 618         int do_sync;
 619         u64 snap_size = 0;
 620
 621         /*
 622          * Include a 'sync' in the OSD request if this is a data
 623          * integrity write (e.g., O_SYNC write or fsync()), or if our
 624          * cap is being revoked.
 625          */
 626         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 627         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 628                 do_sync = 1;
 629         dout("writepages_start %p dosync=%d (mode=%s)\n",
 630              inode, do_sync,
 631              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 632              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 633
 634         client = ceph_inode_to_client(inode);
 635         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
 636                 pr_warning("writepage_start %p on forced umount\n", inode);
 637                 return -EIO; /* we're in a forced umount, don't write! */
 638         }
 639         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
 640                 wsize = client->mount_args->wsize;
 641         if (wsize < PAGE_CACHE_SIZE)
 642                 wsize = PAGE_CACHE_SIZE;
 643         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 644
 645         pagevec_init(&pvec, 0);
 646
 647         /* ?? */
 648         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 649                 dout(" writepages congested\n");
 650                 wbc->encountered_congestion = 1;
 651                 goto out_final;
 652         }
 653
 654         /* where to start/end? */
 655         if (wbc->range_cyclic) {
 656                 start = mapping->writeback_index; /* Start from prev offset */
 657                 end = -1;
 658                 dout(" cyclic, start at %lu\n", start);
 659         } else {
 660                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 661                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 662                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 663                         range_whole = 1;
 664                 should_loop = 0;
 665                 dout(" not cyclic, %lu to %lu\n", start, end);
 666         }
 667         index = start;
 668
 669 retry:
 670         /* find oldest snap context with dirty data */
 671         ceph_put_snap_context(snapc);
 672         snapc = get_oldest_context(inode, &snap_size);
 673         if (!snapc) {
 674                 /* hmm, why does writepages get called when there
 675                    is no dirty data? */
 676                 dout(" no snap context with dirty data?\n");
 677                 goto out;
 678         }
 679         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 680              snapc, snapc->seq, snapc->num_snaps);
 681         if (last_snapc && snapc != last_snapc) {
 682                 /* if we switched to a newer snapc, restart our scan at the
 683                  * start of the original file range. */
 684                 dout("  snapc differs from last pass, restarting at %lu\n",
 685                      index);
 686                 index = start;
 687         }
 688         last_snapc = snapc;
 689
 690         while (!done && index <= end) {
 691                 unsigned i;
 692                 int first;
 693                 pgoff_t next;
 694                 int pvec_pages, locked_pages;
 695                 struct page *page;
 696                 int want;
 697                 u64 offset, len;
 698                 struct ceph_osd_request_head *reqhead;
 699                 struct ceph_osd_op *op;
 700                 long writeback_stat;
 701
 702                 next = 0;
 703                 locked_pages = 0;
 704                 max_pages = max_pages_ever;
 705
 706 get_more_pages:
 707                 first = -1;
 708                 want = min(end - index,
 709                            min((pgoff_t)PAGEVEC_SIZE,
 710                                max_pages - (pgoff_t)locked_pages) - 1)
 711                         + 1;
 712                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 713                                                 PAGECACHE_TAG_DIRTY,
 714                                                 want);
 715                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 716                 if (!pvec_pages && !locked_pages)
 717                         break;
 718                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 719                         page = pvec.pages[i];
 720                         dout("? %p idx %lu\n", page, page->index);
 721                         if (locked_pages == 0)
 722                                 lock_page(page);  /* first page */
 723                         else if (!trylock_page(page))
 724                                 break;
 725
 726                         /* only dirty pages, or our accounting breaks */
 727                         if (unlikely(!PageDirty(page)) ||
 728                             unlikely(page->mapping != mapping)) {
 729                                 dout("!dirty or !mapping %p\n", page);
 730                                 unlock_page(page);
 731                                 break;
 732                         }
 733                         if (!wbc->range_cyclic && page->index > end) {
 734                                 dout("end of range %p\n", page);
 735                                 done = 1;
 736                                 unlock_page(page);
 737                                 break;
 738                         }
 739                         if (next && (page->index != next)) {
 740                                 dout("not consecutive %p\n", page);
 741                                 unlock_page(page);
 742                                 break;
 743                         }
 744                         if (wbc->sync_mode != WB_SYNC_NONE) {
 745                                 dout("waiting on writeback %p\n", page);
 746                                 wait_on_page_writeback(page);
 747                         }
 748                         if ((snap_size && page_offset(page) > snap_size) ||
 749                             (!snap_size &&
 750                              page_offset(page) > i_size_read(inode))) {
 751                                 dout("%p page eof %llu\n", page, snap_size ?
 752                                      snap_size : i_size_read(inode));
 753                                 done = 1;
 754                                 unlock_page(page);
 755                                 break;
 756                         }
 757                         if (PageWriteback(page)) {
 758                                 dout("%p under writeback\n", page);
 759                                 unlock_page(page);
 760                                 break;
 761                         }
 762
 763                         /* only if matching snap context */
 764                         pgsnapc = (void *)page->private;
 765                         if (pgsnapc->seq > snapc->seq) {
 766                                 dout("page snapc %p %lld > oldest %p %lld\n",
 767                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 768                                 unlock_page(page);
 769                                 if (!locked_pages)
 770                                         continue; /* keep looking for snap */
 771                                 break;
 772                         }
 773
 774                         if (!clear_page_dirty_for_io(page)) {
 775                                 dout("%p !clear_page_dirty_for_io\n", page);
 776                                 unlock_page(page);
 777                                 break;
 778                         }
 779
 780                         /* ok */
 781                         if (locked_pages == 0) {
 782                                 /* prepare async write request */
 783                                 offset = page->index << PAGE_CACHE_SHIFT;
 784                                 len = wsize;
 785                                 req = ceph_osdc_new_request(&client->osdc,
 786                                             &ci->i_layout,
 787                                             ceph_vino(inode),
 788                                             offset, &len,
 789                                             CEPH_OSD_OP_WRITE,
 790                                             CEPH_OSD_FLAG_WRITE |
 791                                                     CEPH_OSD_FLAG_ONDISK,
 792                                             snapc, do_sync,
 793                                             ci->i_truncate_seq,
 794                                             ci->i_truncate_size,
 795                                             &inode->i_mtime, true, 1);
 796                                 max_pages = req->r_num_pages;
 797
 798                                 alloc_page_vec(client, req);
 799                                 req->r_callback = writepages_finish;
 800                                 req->r_inode = inode;
 801                                 req->r_wbc = wbc;
 802                         }
 803
 804                         /* note position of first page in pvec */
 805                         if (first < 0)
 806                                 first = i;
 807                         dout("%p will write page %p idx %lu\n",
 808                              inode, page, page->index);
 809
 810                         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 811                         if (writeback_stat > CONGESTION_ON_THRESH(client->mount_args->congestion_kb)) {
 812                                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 813                         }
 814
 815                         set_page_writeback(page);
 816                         req->r_pages[locked_pages] = page;
 817                         locked_pages++;
 818                         next = page->index + 1;
 819                 }
 820
 821                 /* did we get anything? */
 822                 if (!locked_pages)
 823                         goto release_pvec_pages;
 824                 if (i) {
 825                         int j;
 826                         BUG_ON(!locked_pages || first < 0);
 827
 828                         if (pvec_pages && i == pvec_pages &&
 829                             locked_pages < max_pages) {
 830                                 dout("reached end pvec, trying for more\n");
 831                                 pagevec_reinit(&pvec);
 832                                 goto get_more_pages;
 833                         }
 834
 835                         /* shift unused pages over in the pvec...  we
 836                          * will need to release them below. */
 837                         for (j = i; j < pvec_pages; j++) {
 838                                 dout(" pvec leftover page %p\n",
 839                                      pvec.pages[j]);
 840                                 pvec.pages[j-i+first] = pvec.pages[j];
 841                         }
 842                         pvec.nr -= i-first;
 843                 }
 844
 845                 /* submit the write */
 846                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 847                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 848                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 849                 dout("writepages got %d pages at %llu~%llu\n",
 850                      locked_pages, offset, len);
 851
 852                 /* revise final length, page count */
 853                 req->r_num_pages = locked_pages;
 854                 reqhead = req->r_request->front.iov_base;
 855                 op = (void *)(reqhead + 1);
 856                 op->extent.length = cpu_to_le64(len);
 857                 op->payload_len = cpu_to_le32(len);
 858                 req->r_request->hdr.data_len = cpu_to_le32(len);
 859
 860                 ceph_osdc_start_request(&client->osdc, req, true);
 861                 req = NULL;
 862
 863                 /* continue? */
 864                 index = next;
 865                 wbc->nr_to_write -= locked_pages;
 866                 if (wbc->nr_to_write <= 0)
 867                         done = 1;
 868
 869 release_pvec_pages:
 870                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 871                      pvec.nr ? pvec.pages[0] : NULL);
 872                 pagevec_release(&pvec);
 873
 874                 if (locked_pages && !done)
 875                         goto retry;
 876         }
 877
 878         if (should_loop && !done) {
 879                 /* more to do; loop back to beginning of file */
 880                 dout("writepages looping back to beginning of file\n");
 881                 should_loop = 0;
 882                 index = 0;
 883                 goto retry;
 884         }
 885
 886         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 887                 mapping->writeback_index = index;
 888
 889 out:
 890         if (req)
 891                 ceph_osdc_put_request(req);
 892         if (rc > 0)
 893                 rc = 0;  /* vfs expects us to return 0 */
 894         ceph_put_snap_context(snapc);
 895         dout("writepages done, rc = %d\n", rc);
 896 out_final:
 897         return rc;
 898 }
 899
 900
 901
 902 /*
 903  * See if a given @snapc is either writeable, or already written.
 904  */
 905 static int context_is_writeable_or_written(struct inode *inode,
 906                                            struct ceph_snap_context *snapc)
 907 {
 908         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 909         int ret = !oldest || snapc->seq <= oldest->seq;
 910
 911         ceph_put_snap_context(oldest);
 912         return ret;
 913 }
 914
 915 /*
 916  * We are only allowed to write into/dirty the page if the page is
 917  * clean, or already dirty within the same snap context.
 918  *
 919  * called with page locked.
 920  * return success with page locked,
 921  * or any failure (incl -EAGAIN) with page unlocked.
 922  */
 923 static int ceph_update_writeable_page(struct file *file,
 924                             loff_t pos, unsigned len,
 925                             struct page *page)
 926 {
 927         struct inode *inode = file->f_dentry->d_inode;
 928         struct ceph_inode_info *ci = ceph_inode(inode);
 929         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 930         loff_t page_off = pos & PAGE_CACHE_MASK;
 931         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 932         int end_in_page = pos_in_page + len;
 933         loff_t i_size;
 934         int r;
 935         struct ceph_snap_context *snapc, *oldest;
 936
 937 retry_locked:
 938         /* writepages currently holds page lock, but if we change that later, */
 939         wait_on_page_writeback(page);
 940
 941         /* check snap context */
 942         BUG_ON(!ci->i_snap_realm);
 943         down_read(&mdsc->snap_rwsem);
 944         BUG_ON(!ci->i_snap_realm->cached_context);
 945         snapc = (void *)page->private;
 946         if (snapc && snapc != ci->i_head_snapc) {
 947                 /*
 948                  * this page is already dirty in another (older) snap
 949                  * context!  is it writeable now?
 950                  */
 951                 oldest = get_oldest_context(inode, NULL);
 952                 up_read(&mdsc->snap_rwsem);
 953
 954                 if (snapc->seq > oldest->seq) {
 955                         ceph_put_snap_context(oldest);
 956                         dout(" page %p snapc %p not current or oldest\n",
 957                              page, snapc);
 958                         /*
 959                          * queue for writeback, and wait for snapc to
 960                          * be writeable or written
 961                          */
 962                         snapc = ceph_get_snap_context(snapc);
 963                         unlock_page(page);
 964                         ceph_queue_writeback(inode);
 965                         r = wait_event_interruptible(ci->i_cap_wq,
 966                                context_is_writeable_or_written(inode, snapc));
 967                         ceph_put_snap_context(snapc);
 968                         if (r == -ERESTARTSYS)
 969                                 return r;
 970                         return -EAGAIN;
 971                 }
 972                 ceph_put_snap_context(oldest);
 973
 974                 /* yay, writeable, do it now (without dropping page lock) */
 975                 dout(" page %p snapc %p not current, but oldest\n",
 976                      page, snapc);
 977                 if (!clear_page_dirty_for_io(page))
 978                         goto retry_locked;
 979                 r = writepage_nounlock(page, NULL);
 980                 if (r < 0)
 981                         goto fail_nosnap;
 982                 goto retry_locked;
 983         }
 984
 985         if (PageUptodate(page)) {
 986                 dout(" page %p already uptodate\n", page);
 987                 return 0;
 988         }
 989
 990         /* full page? */
 991         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 992                 return 0;
 993
 994         /* past end of file? */
 995         i_size = inode->i_size;   /* caller holds i_mutex */
 996
 997         if (i_size + len > inode->i_sb->s_maxbytes) {
 998                 /* file is too big */
 999                 r = -EINVAL;
1000                 goto fail;
1001         }
1002
1003         if (page_off >= i_size ||
1004             (pos_in_page == 0 && (pos+len) >= i_size &&
1005              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1006                 dout(" zeroing %p 0 - %d and %d - %d\n",
1007                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1008                 zero_user_segments(page,
1009                                    0, pos_in_page,
1010                                    end_in_page, PAGE_CACHE_SIZE);
1011                 return 0;
1012         }
1013
1014         /* we need to read it. */
1015         up_read(&mdsc->snap_rwsem);
1016         r = readpage_nounlock(file, page);
1017         if (r < 0)
1018                 goto fail_nosnap;
1019         goto retry_locked;
1020
1021 fail:
1022         up_read(&mdsc->snap_rwsem);
1023 fail_nosnap:
1024         unlock_page(page);
1025         return r;
1026 }
1027
1028 /*
1029  * We are only allowed to write into/dirty the page if the page is
1030  * clean, or already dirty within the same snap context.
1031  */
1032 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1033                             loff_t pos, unsigned len, unsigned flags,
1034                             struct page **pagep, void **fsdata)
1035 {
1036         struct inode *inode = file->f_dentry->d_inode;
1037         struct page *page;
1038         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1039         int r;
1040
1041         do {
1042                 /* get a page */
1043                 page = grab_cache_page_write_begin(mapping, index, 0);
1044                 if (!page)
1045                         return -ENOMEM;
1046                 *pagep = page;
1047
1048                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1049                 inode, page, (int)pos, (int)len);
1050
1051                 r = ceph_update_writeable_page(file, pos, len, page);
1052         } while (r == -EAGAIN);
1053
1054         return r;
1055 }
1056
1057 /*
1058  * we don't do anything in here that simple_write_end doesn't do
1059  * except adjust dirty page accounting and drop read lock on
1060  * mdsc->snap_rwsem.
1061  */
1062 static int ceph_write_end(struct file *file, struct address_space *mapping,
1063                           loff_t pos, unsigned len, unsigned copied,
1064                           struct page *page, void *fsdata)
1065 {
1066         struct inode *inode = file->f_dentry->d_inode;
1067         struct ceph_client *client = ceph_inode_to_client(inode);
1068         struct ceph_mds_client *mdsc = &client->mdsc;
1069         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1070         int check_cap = 0;
1071
1072         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1073              inode, page, (int)pos, (int)copied, (int)len);
1074
1075         /* zero the stale part of the page if we did a short copy */
1076         if (copied < len)
1077                 zero_user_segment(page, from+copied, len);
1078
1079         /* did file size increase? */
1080         /* (no need for i_size_read(); we caller holds i_mutex */
1081         if (pos+copied > inode->i_size)
1082                 check_cap = ceph_inode_set_size(inode, pos+copied);
1083
1084         if (!PageUptodate(page))
1085                 SetPageUptodate(page);
1086
1087         set_page_dirty(page);
1088
1089         unlock_page(page);
1090         up_read(&mdsc->snap_rwsem);
1091         page_cache_release(page);
1092
1093         if (check_cap)
1094                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1095
1096         return copied;
1097 }
1098
1099 /*
1100  * we set .direct_IO to indicate direct io is supported, but since we
1101  * intercept O_DIRECT reads and writes early, this function should
1102  * never get called.
1103  */
1104 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1105                               const struct iovec *iov,
1106                               loff_t pos, unsigned long nr_segs)
1107 {
1108         WARN_ON(1);
1109         return -EINVAL;
1110 }
1111
1112 const struct address_space_operations ceph_aops = {
1113         .readpage = ceph_readpage,
1114         .readpages = ceph_readpages,
1115         .writepage = ceph_writepage,
1116         .writepages = ceph_writepages_start,
1117         .write_begin = ceph_write_begin,
1118         .write_end = ceph_write_end,
1119         .set_page_dirty = ceph_set_page_dirty,
1120         .invalidatepage = ceph_invalidatepage,
1121         .releasepage = ceph_releasepage,
1122         .direct_IO = ceph_direct_io,
1123 };
1124
1125
1126 /*
1127  * vm ops
1128  */
1129
1130 /*
1131  * Reuse write_begin here for simplicity.
1132  */
1133 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1134 {
1135         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1136         struct page *page = vmf->page;
1137         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1138         loff_t off = page->index << PAGE_CACHE_SHIFT;
1139         loff_t size, len;
1140         int ret;
1141
1142         size = i_size_read(inode);
1143         if (off + PAGE_CACHE_SIZE <= size)
1144                 len = PAGE_CACHE_SIZE;
1145         else
1146                 len = size & ~PAGE_CACHE_MASK;
1147
1148         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1149              off, len, page, page->index);
1150
1151         lock_page(page);
1152
1153         ret = VM_FAULT_NOPAGE;
1154         if ((off > size) ||
1155             (page->mapping != inode->i_mapping))
1156                 goto out;
1157
1158         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1159         if (ret == 0) {
1160                 /* success.  we'll keep the page locked. */
1161                 set_page_dirty(page);
1162                 up_read(&mdsc->snap_rwsem);
1163                 ret = VM_FAULT_LOCKED;
1164         } else {
1165                 if (ret == -ENOMEM)
1166                         ret = VM_FAULT_OOM;
1167                 else
1168                         ret = VM_FAULT_SIGBUS;
1169         }
1170 out:
1171         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1172         if (ret != VM_FAULT_LOCKED)
1173                 unlock_page(page);
1174         return ret;
1175 }
1176
1177 static struct vm_operations_struct ceph_vmops = {
1178         .fault          = filemap_fault,
1179         .page_mkwrite   = ceph_page_mkwrite,
1180 };
1181
1182 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1183 {
1184         struct address_space *mapping = file->f_mapping;
1185
1186         if (!mapping->a_ops->readpage)
1187                 return -ENOEXEC;
1188         file_accessed(file);
1189         vma->vm_ops = &ceph_vmops;
1190         vma->vm_flags |= VM_CAN_NONLINEAR;
1191         return 0;
1192 }