fs/ceph/addr.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "osd_client.h"
  14
  15 /*
  16  * Ceph address space ops.
  17  *
  18  * There are a few funny things going on here.
  19  *
  20  * The page->private field is used to reference a struct
  21  * ceph_snap_context for _every_ dirty page.  This indicates which
  22  * snapshot the page was logically dirtied in, and thus which snap
  23  * context needs to be associated with the osd write during writeback.
  24  *
  25  * Similarly, struct ceph_inode_info maintains a set of counters to
  26  * count dirty pages on the inode.  In the absense of snapshots,
  27  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  28  *
  29  * When a snapshot is taken (that is, when the client receives
  30  * notification that a snapshot was taken), each inode with caps and
  31  * with dirty pages (dirty pages implies there is a cap) gets a new
  32  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  33  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  34  * moved to capsnap->dirty. (Unless a sync write is currently in
  35  * progress.  In that case, the capsnap is said to be "pending", new
  36  * writes cannot start, and the capsnap isn't "finalized" until the
  37  * write completes (or fails) and a final size/mtime for the inode for
  38  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  39  *
  40  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  41  * we look for the first capsnap in i_cap_snaps and write out pages in
  42  * that snap context _only_.  Then we move on to the next capsnap,
  43  * eventually reaching the "live" or "head" context (i.e., pages that
  44  * are not yet snapped) and are writing the most recently dirtied
  45  * pages.
  46  *
  47  * Invalidate and so forth must take care to ensure the dirty page
  48  * accounting is preserved.
  49  */
  50
  51 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  52 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  53         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  54          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  55
  56
  57
  58 /*
  59  * Dirty a page.  Optimistically adjust accounting, on the assumption
  60  * that we won't race with invalidate.  If we do, readjust.
  61  */
  62 static int ceph_set_page_dirty(struct page *page)
  63 {
  64         struct address_space *mapping = page->mapping;
  65         struct inode *inode;
  66         struct ceph_inode_info *ci;
  67         int undo = 0;
  68         struct ceph_snap_context *snapc;
  69
  70         if (unlikely(!mapping))
  71                 return !TestSetPageDirty(page);
  72
  73         if (TestSetPageDirty(page)) {
  74                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  75                      mapping->host, page, page->index);
  76                 return 0;
  77         }
  78
  79         inode = mapping->host;
  80         ci = ceph_inode(inode);
  81
  82         /*
  83          * Note that we're grabbing a snapc ref here without holding
  84          * any locks!
  85          */
  86         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  87
  88         /* dirty the head */
  89         spin_lock(&inode->i_lock);
  90         if (ci->i_wrbuffer_ref_head == 0)
  91                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  92         ++ci->i_wrbuffer_ref_head;
  93         if (ci->i_wrbuffer_ref == 0)
  94                 igrab(inode);
  95         ++ci->i_wrbuffer_ref;
  96         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
  97              "snapc %p seq %lld (%d snaps)\n",
  98              mapping->host, page, page->index,
  99              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 100              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 101              snapc, snapc->seq, snapc->num_snaps);
 102         spin_unlock(&inode->i_lock);
 103
 104         /* now adjust page */
 105         spin_lock_irq(&mapping->tree_lock);
 106         if (page->mapping) {    /* Race with truncate? */
 107                 WARN_ON_ONCE(!PageUptodate(page));
 108
 109                 if (mapping_cap_account_dirty(mapping)) {
 110                         __inc_zone_page_state(page, NR_FILE_DIRTY);
 111                         __inc_bdi_stat(mapping->backing_dev_info,
 112                                         BDI_RECLAIMABLE);
 113                         task_io_account_write(PAGE_CACHE_SIZE);
 114                 }
 115                 radix_tree_tag_set(&mapping->page_tree,
 116                                 page_index(page), PAGECACHE_TAG_DIRTY);
 117
 118                 /*
 119                  * Reference snap context in page->private.  Also set
 120                  * PagePrivate so that we get invalidatepage callback.
 121                  */
 122                 page->private = (unsigned long)snapc;
 123                 SetPagePrivate(page);
 124         } else {
 125                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 126                 undo = 1;
 127         }
 128
 129         spin_unlock_irq(&mapping->tree_lock);
 130
 131         if (undo)
 132                 /* whoops, we failed to dirty the page */
 133                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 134
 135         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 136
 137         BUG_ON(!PageDirty(page));
 138         return 1;
 139 }
 140
 141 /*
 142  * If we are truncating the full page (i.e. offset == 0), adjust the
 143  * dirty page counters appropriately.  Only called if there is private
 144  * data on the page.
 145  */
 146 static void ceph_invalidatepage(struct page *page, unsigned long offset)
 147 {
 148         struct inode *inode;
 149         struct ceph_inode_info *ci;
 150         struct ceph_snap_context *snapc = (void *)page->private;
 151
 152         BUG_ON(!PageLocked(page));
 153         BUG_ON(!page->private);
 154         BUG_ON(!PagePrivate(page));
 155         BUG_ON(!page->mapping);
 156
 157         inode = page->mapping->host;
 158
 159         /*
 160          * We can get non-dirty pages here due to races between
 161          * set_page_dirty and truncate_complete_page; just spit out a
 162          * warning, in case we end up with accounting problems later.
 163          */
 164         if (!PageDirty(page))
 165                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 166
 167         if (offset == 0)
 168                 ClearPageChecked(page);
 169
 170         ci = ceph_inode(inode);
 171         if (offset == 0) {
 172                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 173                      inode, page, page->index, offset);
 174                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 175                 ceph_put_snap_context(snapc);
 176                 page->private = 0;
 177                 ClearPagePrivate(page);
 178         } else {
 179                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
 180                      inode, page, page->index);
 181         }
 182 }
 183
 184 /* just a sanity check */
 185 static int ceph_releasepage(struct page *page, gfp_t g)
 186 {
 187         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 188         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 189         WARN_ON(PageDirty(page));
 190         WARN_ON(page->private);
 191         WARN_ON(PagePrivate(page));
 192         return 0;
 193 }
 194
 195 /*
 196  * read a single page, without unlocking it.
 197  */
 198 static int readpage_nounlock(struct file *filp, struct page *page)
 199 {
 200         struct inode *inode = filp->f_dentry->d_inode;
 201         struct ceph_inode_info *ci = ceph_inode(inode);
 202         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 203         int err = 0;
 204         u64 len = PAGE_CACHE_SIZE;
 205
 206         dout("readpage inode %p file %p page %p index %lu\n",
 207              inode, filp, page, page->index);
 208         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 209                                   page->index << PAGE_CACHE_SHIFT, &len,
 210                                   ci->i_truncate_seq, ci->i_truncate_size,
 211                                   &page, 1);
 212         if (err == -ENOENT)
 213                 err = 0;
 214         if (err < 0) {
 215                 SetPageError(page);
 216                 goto out;
 217         } else if (err < PAGE_CACHE_SIZE) {
 218                 /* zero fill remainder of page */
 219                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 220         }
 221         SetPageUptodate(page);
 222
 223 out:
 224         return err < 0 ? err : 0;
 225 }
 226
 227 static int ceph_readpage(struct file *filp, struct page *page)
 228 {
 229         int r = readpage_nounlock(filp, page);
 230         unlock_page(page);
 231         return r;
 232 }
 233
 234 /*
 235  * Build a vector of contiguous pages from the provided page list.
 236  */
 237 static struct page **page_vector_from_list(struct list_head *page_list,
 238                                            unsigned *nr_pages)
 239 {
 240         struct page **pages;
 241         struct page *page;
 242         int next_index, contig_pages = 0;
 243
 244         /* build page vector */
 245         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 246         if (!pages)
 247                 return ERR_PTR(-ENOMEM);
 248
 249         BUG_ON(list_empty(page_list));
 250         next_index = list_entry(page_list->prev, struct page, lru)->index;
 251         list_for_each_entry_reverse(page, page_list, lru) {
 252                 if (page->index == next_index) {
 253                         dout("readpages page %d %p\n", contig_pages, page);
 254                         pages[contig_pages] = page;
 255                         contig_pages++;
 256                         next_index++;
 257                 } else {
 258                         break;
 259                 }
 260         }
 261         *nr_pages = contig_pages;
 262         return pages;
 263 }
 264
 265 /*
 266  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 267  * the caller (VM) cleans them up.
 268  */
 269 static int ceph_readpages(struct file *file, struct address_space *mapping,
 270                           struct list_head *page_list, unsigned nr_pages)
 271 {
 272         struct inode *inode = file->f_dentry->d_inode;
 273         struct ceph_inode_info *ci = ceph_inode(inode);
 274         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 275         int rc = 0;
 276         struct page **pages;
 277         struct pagevec pvec;
 278         loff_t offset;
 279         u64 len;
 280
 281         dout("readpages %p file %p nr_pages %d\n",
 282              inode, file, nr_pages);
 283
 284         pages = page_vector_from_list(page_list, &nr_pages);
 285         if (IS_ERR(pages))
 286                 return PTR_ERR(pages);
 287
 288         /* guess read extent */
 289         offset = pages[0]->index << PAGE_CACHE_SHIFT;
 290         len = nr_pages << PAGE_CACHE_SHIFT;
 291         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 292                                  offset, &len,
 293                                  ci->i_truncate_seq, ci->i_truncate_size,
 294                                  pages, nr_pages);
 295         if (rc == -ENOENT)
 296                 rc = 0;
 297         if (rc < 0)
 298                 goto out;
 299
 300         /* set uptodate and add to lru in pagevec-sized chunks */
 301         pagevec_init(&pvec, 0);
 302         for (; !list_empty(page_list) && len > 0;
 303              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
 304                 struct page *page =
 305                         list_entry(page_list->prev, struct page, lru);
 306
 307                 list_del(&page->lru);
 308
 309                 if (rc < (int)PAGE_CACHE_SIZE) {
 310                         /* zero (remainder of) page */
 311                         int s = rc < 0 ? 0 : rc;
 312                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 313                 }
 314
 315                 if (add_to_page_cache(page, mapping, page->index, GFP_NOFS)) {
 316                         page_cache_release(page);
 317                         dout("readpages %p add_to_page_cache failed %p\n",
 318                              inode, page);
 319                         continue;
 320                 }
 321                 dout("readpages %p adding %p idx %lu\n", inode, page,
 322                      page->index);
 323                 flush_dcache_page(page);
 324                 SetPageUptodate(page);
 325                 unlock_page(page);
 326                 if (pagevec_add(&pvec, page) == 0)
 327                         pagevec_lru_add_file(&pvec);   /* add to lru */
 328         }
 329         pagevec_lru_add_file(&pvec);
 330         rc = 0;
 331
 332 out:
 333         kfree(pages);
 334         return rc;
 335 }
 336
 337 /*
 338  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 339  * only snap context we are allowed to write back.
 340  */
 341 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 342                                                     u64 *snap_size)
 343 {
 344         struct ceph_inode_info *ci = ceph_inode(inode);
 345         struct ceph_snap_context *snapc = NULL;
 346         struct ceph_cap_snap *capsnap = NULL;
 347
 348         spin_lock(&inode->i_lock);
 349         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 350                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 351                      capsnap->context, capsnap->dirty_pages);
 352                 if (capsnap->dirty_pages) {
 353                         snapc = ceph_get_snap_context(capsnap->context);
 354                         if (snap_size)
 355                                 *snap_size = capsnap->size;
 356                         break;
 357                 }
 358         }
 359         if (!snapc && ci->i_head_snapc) {
 360                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 361                 dout(" head snapc %p has %d dirty pages\n",
 362                      snapc, ci->i_wrbuffer_ref_head);
 363         }
 364         spin_unlock(&inode->i_lock);
 365         return snapc;
 366 }
 367
 368 /*
 369  * Write a single page, but leave the page locked.
 370  *
 371  * If we get a write error, set the page error bit, but still adjust the
 372  * dirty page accounting (i.e., page is no longer dirty).
 373  */
 374 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 375 {
 376         struct inode *inode;
 377         struct ceph_inode_info *ci;
 378         struct ceph_client *client;
 379         struct ceph_osd_client *osdc;
 380         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 381         int len = PAGE_CACHE_SIZE;
 382         loff_t i_size;
 383         int err = 0;
 384         struct ceph_snap_context *snapc, *oldest;
 385         u64 snap_size = 0;
 386         long writeback_stat;
 387
 388         dout("writepage %p idx %lu\n", page, page->index);
 389
 390         if (!page->mapping || !page->mapping->host) {
 391                 dout("writepage %p - no mapping\n", page);
 392                 return -EFAULT;
 393         }
 394         inode = page->mapping->host;
 395         ci = ceph_inode(inode);
 396         client = ceph_inode_to_client(inode);
 397         osdc = &client->osdc;
 398
 399         /* verify this is a writeable snap context */
 400         snapc = (void *)page->private;
 401         if (snapc == NULL) {
 402                 dout("writepage %p page %p not dirty?\n", inode, page);
 403                 goto out;
 404         }
 405         oldest = get_oldest_context(inode, &snap_size);
 406         if (snapc->seq > oldest->seq) {
 407                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 408                      inode, page, (void *)page->private);
 409                 /* we should only noop if called by kswapd */
 410                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 411                 ceph_put_snap_context(oldest);
 412                 goto out;
 413         }
 414         ceph_put_snap_context(oldest);
 415
 416         /* is this a partial page at end of file? */
 417         if (snap_size)
 418                 i_size = snap_size;
 419         else
 420                 i_size = i_size_read(inode);
 421         if (i_size < page_off + len)
 422                 len = i_size - page_off;
 423
 424         dout("writepage %p page %p index %lu on %llu~%u\n",
 425              inode, page, page->index, page_off, len);
 426
 427         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 428         if (writeback_stat >
 429             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
 430                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 431
 432         set_page_writeback(page);
 433         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 434                                    &ci->i_layout, snapc,
 435                                    page_off, len,
 436                                    ci->i_truncate_seq, ci->i_truncate_size,
 437                                    &inode->i_mtime,
 438                                    &page, 1, 0, 0, true);
 439         if (err < 0) {
 440                 dout("writepage setting page/mapping error %d %p\n", err, page);
 441                 SetPageError(page);
 442                 mapping_set_error(&inode->i_data, err);
 443                 if (wbc)
 444                         wbc->pages_skipped++;
 445         } else {
 446                 dout("writepage cleaned page %p\n", page);
 447                 err = 0;  /* vfs expects us to return 0 */
 448         }
 449         page->private = 0;
 450         ClearPagePrivate(page);
 451         end_page_writeback(page);
 452         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 453         ceph_put_snap_context(snapc);  /* page's reference */
 454 out:
 455         return err;
 456 }
 457
 458 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 459 {
 460         int err;
 461         struct inode *inode = page->mapping->host;
 462         BUG_ON(!inode);
 463         igrab(inode);
 464         err = writepage_nounlock(page, wbc);
 465         unlock_page(page);
 466         iput(inode);
 467         return err;
 468 }
 469
 470
 471 /*
 472  * lame release_pages helper.  release_pages() isn't exported to
 473  * modules.
 474  */
 475 static void ceph_release_pages(struct page **pages, int num)
 476 {
 477         struct pagevec pvec;
 478         int i;
 479
 480         pagevec_init(&pvec, 0);
 481         for (i = 0; i < num; i++) {
 482                 if (pagevec_add(&pvec, pages[i]) == 0)
 483                         pagevec_release(&pvec);
 484         }
 485         pagevec_release(&pvec);
 486 }
 487
 488
 489 /*
 490  * async writeback completion handler.
 491  *
 492  * If we get an error, set the mapping error bit, but not the individual
 493  * page error bits.
 494  */
 495 static void writepages_finish(struct ceph_osd_request *req,
 496                               struct ceph_msg *msg)
 497 {
 498         struct inode *inode = req->r_inode;
 499         struct ceph_osd_reply_head *replyhead;
 500         struct ceph_osd_op *op;
 501         struct ceph_inode_info *ci = ceph_inode(inode);
 502         unsigned wrote;
 503         struct page *page;
 504         int i;
 505         struct ceph_snap_context *snapc = req->r_snapc;
 506         struct address_space *mapping = inode->i_mapping;
 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);
 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                 ceph_put_snap_context((void *)page->private);
 549                 page->private = 0;
 550                 ClearPagePrivate(page);
 551                 dout("unlocking %d %p\n", i, page);
 552                 end_page_writeback(page);
 553
 554                 /*
 555                  * We lost the cache cap, need to truncate the page before
 556                  * it is unlocked, otherwise we'd truncate it later in the
 557                  * page truncation thread, possibly losing some data that
 558                  * raced its way in
 559                  */
 560                 if ((issued & CEPH_CAP_FILE_CACHE) == 0)
 561                         generic_error_remove_page(inode->i_mapping, page);
 562
 563                 unlock_page(page);
 564         }
 565         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 566         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 567
 568         ceph_release_pages(req->r_pages, req->r_num_pages);
 569         if (req->r_pages_from_pool)
 570                 mempool_free(req->r_pages,
 571                              ceph_client(inode->i_sb)->wb_pagevec_pool);
 572         else
 573                 kfree(req->r_pages);
 574         ceph_osdc_put_request(req);
 575 }
 576
 577 /*
 578  * allocate a page vec, either directly, or if necessary, via a the
 579  * mempool.  we avoid the mempool if we can because req->r_num_pages
 580  * may be less than the maximum write size.
 581  */
 582 static void alloc_page_vec(struct ceph_client *client,
 583                            struct ceph_osd_request *req)
 584 {
 585         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 586                                GFP_NOFS);
 587         if (!req->r_pages) {
 588                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
 589                 req->r_pages_from_pool = 1;
 590                 WARN_ON(!req->r_pages);
 591         }
 592 }
 593
 594 /*
 595  * initiate async writeback
 596  */
 597 static int ceph_writepages_start(struct address_space *mapping,
 598                                  struct writeback_control *wbc)
 599 {
 600         struct inode *inode = mapping->host;
 601         struct backing_dev_info *bdi = mapping->backing_dev_info;
 602         struct ceph_inode_info *ci = ceph_inode(inode);
 603         struct ceph_client *client;
 604         pgoff_t index, start, end;
 605         int range_whole = 0;
 606         int should_loop = 1;
 607         pgoff_t max_pages = 0, max_pages_ever = 0;
 608         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 609         struct pagevec pvec;
 610         int done = 0;
 611         int rc = 0;
 612         unsigned wsize = 1 << inode->i_blkbits;
 613         struct ceph_osd_request *req = NULL;
 614         int do_sync;
 615         u64 snap_size = 0;
 616
 617         /*
 618          * Include a 'sync' in the OSD request if this is a data
 619          * integrity write (e.g., O_SYNC write or fsync()), or if our
 620          * cap is being revoked.
 621          */
 622         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 623         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 624                 do_sync = 1;
 625         dout("writepages_start %p dosync=%d (mode=%s)\n",
 626              inode, do_sync,
 627              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 628              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 629
 630         client = ceph_inode_to_client(inode);
 631         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
 632                 pr_warning("writepage_start %p on forced umount\n", inode);
 633                 return -EIO; /* we're in a forced umount, don't write! */
 634         }
 635         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
 636                 wsize = client->mount_args->wsize;
 637         if (wsize < PAGE_CACHE_SIZE)
 638                 wsize = PAGE_CACHE_SIZE;
 639         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 640
 641         pagevec_init(&pvec, 0);
 642
 643         /* ?? */
 644         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 645                 dout(" writepages congested\n");
 646                 wbc->encountered_congestion = 1;
 647                 goto out_final;
 648         }
 649
 650         /* where to start/end? */
 651         if (wbc->range_cyclic) {
 652                 start = mapping->writeback_index; /* Start from prev offset */
 653                 end = -1;
 654                 dout(" cyclic, start at %lu\n", start);
 655         } else {
 656                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 657                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 658                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 659                         range_whole = 1;
 660                 should_loop = 0;
 661                 dout(" not cyclic, %lu to %lu\n", start, end);
 662         }
 663         index = start;
 664
 665 retry:
 666         /* find oldest snap context with dirty data */
 667         ceph_put_snap_context(snapc);
 668         snapc = get_oldest_context(inode, &snap_size);
 669         if (!snapc) {
 670                 /* hmm, why does writepages get called when there
 671                    is no dirty data? */
 672                 dout(" no snap context with dirty data?\n");
 673                 goto out;
 674         }
 675         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 676              snapc, snapc->seq, snapc->num_snaps);
 677         if (last_snapc && snapc != last_snapc) {
 678                 /* if we switched to a newer snapc, restart our scan at the
 679                  * start of the original file range. */
 680                 dout("  snapc differs from last pass, restarting at %lu\n",
 681                      index);
 682                 index = start;
 683         }
 684         last_snapc = snapc;
 685
 686         while (!done && index <= end) {
 687                 unsigned i;
 688                 int first;
 689                 pgoff_t next;
 690                 int pvec_pages, locked_pages;
 691                 struct page *page;
 692                 int want;
 693                 u64 offset, len;
 694                 struct ceph_osd_request_head *reqhead;
 695                 struct ceph_osd_op *op;
 696                 long writeback_stat;
 697
 698                 next = 0;
 699                 locked_pages = 0;
 700                 max_pages = max_pages_ever;
 701
 702 get_more_pages:
 703                 first = -1;
 704                 want = min(end - index,
 705                            min((pgoff_t)PAGEVEC_SIZE,
 706                                max_pages - (pgoff_t)locked_pages) - 1)
 707                         + 1;
 708                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 709                                                 PAGECACHE_TAG_DIRTY,
 710                                                 want);
 711                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 712                 if (!pvec_pages && !locked_pages)
 713                         break;
 714                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 715                         page = pvec.pages[i];
 716                         dout("? %p idx %lu\n", page, page->index);
 717                         if (locked_pages == 0)
 718                                 lock_page(page);  /* first page */
 719                         else if (!trylock_page(page))
 720                                 break;
 721
 722                         /* only dirty pages, or our accounting breaks */
 723                         if (unlikely(!PageDirty(page)) ||
 724                             unlikely(page->mapping != mapping)) {
 725                                 dout("!dirty or !mapping %p\n", page);
 726                                 unlock_page(page);
 727                                 break;
 728                         }
 729                         if (!wbc->range_cyclic && page->index > end) {
 730                                 dout("end of range %p\n", page);
 731                                 done = 1;
 732                                 unlock_page(page);
 733                                 break;
 734                         }
 735                         if (next && (page->index != next)) {
 736                                 dout("not consecutive %p\n", page);
 737                                 unlock_page(page);
 738                                 break;
 739                         }
 740                         if (wbc->sync_mode != WB_SYNC_NONE) {
 741                                 dout("waiting on writeback %p\n", page);
 742                                 wait_on_page_writeback(page);
 743                         }
 744                         if ((snap_size && page_offset(page) > snap_size) ||
 745                             (!snap_size &&
 746                              page_offset(page) > i_size_read(inode))) {
 747                                 dout("%p page eof %llu\n", page, snap_size ?
 748                                      snap_size : i_size_read(inode));
 749                                 done = 1;
 750                                 unlock_page(page);
 751                                 break;
 752                         }
 753                         if (PageWriteback(page)) {
 754                                 dout("%p under writeback\n", page);
 755                                 unlock_page(page);
 756                                 break;
 757                         }
 758
 759                         /* only if matching snap context */
 760                         pgsnapc = (void *)page->private;
 761                         if (pgsnapc->seq > snapc->seq) {
 762                                 dout("page snapc %p %lld > oldest %p %lld\n",
 763                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 764                                 unlock_page(page);
 765                                 if (!locked_pages)
 766                                         continue; /* keep looking for snap */
 767                                 break;
 768                         }
 769
 770                         if (!clear_page_dirty_for_io(page)) {
 771                                 dout("%p !clear_page_dirty_for_io\n", page);
 772                                 unlock_page(page);
 773                                 break;
 774                         }
 775
 776                         /* ok */
 777                         if (locked_pages == 0) {
 778                                 /* prepare async write request */
 779                                 offset = page->index << PAGE_CACHE_SHIFT;
 780                                 len = wsize;
 781                                 req = ceph_osdc_new_request(&client->osdc,
 782                                             &ci->i_layout,
 783                                             ceph_vino(inode),
 784                                             offset, &len,
 785                                             CEPH_OSD_OP_WRITE,
 786                                             CEPH_OSD_FLAG_WRITE |
 787                                                     CEPH_OSD_FLAG_ONDISK,
 788                                             snapc, do_sync,
 789                                             ci->i_truncate_seq,
 790                                             ci->i_truncate_size,
 791                                             &inode->i_mtime, true, 1);
 792                                 max_pages = req->r_num_pages;
 793
 794                                 alloc_page_vec(client, req);
 795                                 req->r_callback = writepages_finish;
 796                                 req->r_inode = inode;
 797                         }
 798
 799                         /* note position of first page in pvec */
 800                         if (first < 0)
 801                                 first = i;
 802                         dout("%p will write page %p idx %lu\n",
 803                              inode, page, page->index);
 804
 805                         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 806                         if (writeback_stat > CONGESTION_ON_THRESH(client->mount_args->congestion_kb)) {
 807                                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 808                         }
 809
 810                         set_page_writeback(page);
 811                         req->r_pages[locked_pages] = page;
 812                         locked_pages++;
 813                         next = page->index + 1;
 814                 }
 815
 816                 /* did we get anything? */
 817                 if (!locked_pages)
 818                         goto release_pvec_pages;
 819                 if (i) {
 820                         int j;
 821                         BUG_ON(!locked_pages || first < 0);
 822
 823                         if (pvec_pages && i == pvec_pages &&
 824                             locked_pages < max_pages) {
 825                                 dout("reached end pvec, trying for more\n");
 826                                 pagevec_reinit(&pvec);
 827                                 goto get_more_pages;
 828                         }
 829
 830                         /* shift unused pages over in the pvec...  we
 831                          * will need to release them below. */
 832                         for (j = i; j < pvec_pages; j++) {
 833                                 dout(" pvec leftover page %p\n",
 834                                      pvec.pages[j]);
 835                                 pvec.pages[j-i+first] = pvec.pages[j];
 836                         }
 837                         pvec.nr -= i-first;
 838                 }
 839
 840                 /* submit the write */
 841                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 842                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 843                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 844                 dout("writepages got %d pages at %llu~%llu\n",
 845                      locked_pages, offset, len);
 846
 847                 /* revise final length, page count */
 848                 req->r_num_pages = locked_pages;
 849                 reqhead = req->r_request->front.iov_base;
 850                 op = (void *)(reqhead + 1);
 851                 op->extent.length = cpu_to_le64(len);
 852                 op->payload_len = cpu_to_le32(len);
 853                 req->r_request->hdr.data_len = cpu_to_le32(len);
 854
 855                 ceph_osdc_start_request(&client->osdc, req, true);
 856                 req = NULL;
 857
 858                 /* continue? */
 859                 index = next;
 860                 wbc->nr_to_write -= locked_pages;
 861                 if (wbc->nr_to_write <= 0)
 862                         done = 1;
 863
 864 release_pvec_pages:
 865                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 866                      pvec.nr ? pvec.pages[0] : NULL);
 867                 pagevec_release(&pvec);
 868
 869                 if (locked_pages && !done)
 870                         goto retry;
 871         }
 872
 873         if (should_loop && !done) {
 874                 /* more to do; loop back to beginning of file */
 875                 dout("writepages looping back to beginning of file\n");
 876                 should_loop = 0;
 877                 index = 0;
 878                 goto retry;
 879         }
 880
 881         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 882                 mapping->writeback_index = index;
 883
 884 out:
 885         if (req)
 886                 ceph_osdc_put_request(req);
 887         if (rc > 0)
 888                 rc = 0;  /* vfs expects us to return 0 */
 889         ceph_put_snap_context(snapc);
 890         dout("writepages done, rc = %d\n", rc);
 891 out_final:
 892         return rc;
 893 }
 894
 895
 896
 897 /*
 898  * See if a given @snapc is either writeable, or already written.
 899  */
 900 static int context_is_writeable_or_written(struct inode *inode,
 901                                            struct ceph_snap_context *snapc)
 902 {
 903         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 904         int ret = !oldest || snapc->seq <= oldest->seq;
 905
 906         ceph_put_snap_context(oldest);
 907         return ret;
 908 }
 909
 910 /*
 911  * We are only allowed to write into/dirty the page if the page is
 912  * clean, or already dirty within the same snap context.
 913  *
 914  * called with page locked.
 915  * return success with page locked,
 916  * or any failure (incl -EAGAIN) with page unlocked.
 917  */
 918 static int ceph_update_writeable_page(struct file *file,
 919                             loff_t pos, unsigned len,
 920                             struct page *page)
 921 {
 922         struct inode *inode = file->f_dentry->d_inode;
 923         struct ceph_inode_info *ci = ceph_inode(inode);
 924         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 925         loff_t page_off = pos & PAGE_CACHE_MASK;
 926         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 927         int end_in_page = pos_in_page + len;
 928         loff_t i_size;
 929         int r;
 930         struct ceph_snap_context *snapc, *oldest;
 931
 932 retry_locked:
 933         /* writepages currently holds page lock, but if we change that later, */
 934         wait_on_page_writeback(page);
 935
 936         /* check snap context */
 937         BUG_ON(!ci->i_snap_realm);
 938         down_read(&mdsc->snap_rwsem);
 939         BUG_ON(!ci->i_snap_realm->cached_context);
 940         snapc = (void *)page->private;
 941         if (snapc && snapc != ci->i_head_snapc) {
 942                 /*
 943                  * this page is already dirty in another (older) snap
 944                  * context!  is it writeable now?
 945                  */
 946                 oldest = get_oldest_context(inode, NULL);
 947                 up_read(&mdsc->snap_rwsem);
 948
 949                 if (snapc->seq > oldest->seq) {
 950                         ceph_put_snap_context(oldest);
 951                         dout(" page %p snapc %p not current or oldest\n",
 952                              page, snapc);
 953                         /*
 954                          * queue for writeback, and wait for snapc to
 955                          * be writeable or written
 956                          */
 957                         snapc = ceph_get_snap_context(snapc);
 958                         unlock_page(page);
 959                         ceph_queue_writeback(inode);
 960                         r = wait_event_interruptible(ci->i_cap_wq,
 961                                context_is_writeable_or_written(inode, snapc));
 962                         ceph_put_snap_context(snapc);
 963                         if (r == -ERESTARTSYS)
 964                                 return r;
 965                         return -EAGAIN;
 966                 }
 967                 ceph_put_snap_context(oldest);
 968
 969                 /* yay, writeable, do it now (without dropping page lock) */
 970                 dout(" page %p snapc %p not current, but oldest\n",
 971                      page, snapc);
 972                 if (!clear_page_dirty_for_io(page))
 973                         goto retry_locked;
 974                 r = writepage_nounlock(page, NULL);
 975                 if (r < 0)
 976                         goto fail_nosnap;
 977                 goto retry_locked;
 978         }
 979
 980         if (PageUptodate(page)) {
 981                 dout(" page %p already uptodate\n", page);
 982                 return 0;
 983         }
 984
 985         /* full page? */
 986         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 987                 return 0;
 988
 989         /* past end of file? */
 990         i_size = inode->i_size;   /* caller holds i_mutex */
 991
 992         if (i_size + len > inode->i_sb->s_maxbytes) {
 993                 /* file is too big */
 994                 r = -EINVAL;
 995                 goto fail;
 996         }
 997
 998         if (page_off >= i_size ||
 999             (pos_in_page == 0 && (pos+len) >= i_size &&
1000              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1001                 dout(" zeroing %p 0 - %d and %d - %d\n",
1002                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1003                 zero_user_segments(page,
1004                                    0, pos_in_page,
1005                                    end_in_page, PAGE_CACHE_SIZE);
1006                 return 0;
1007         }
1008
1009         /* we need to read it. */
1010         up_read(&mdsc->snap_rwsem);
1011         r = readpage_nounlock(file, page);
1012         if (r < 0)
1013                 goto fail_nosnap;
1014         goto retry_locked;
1015
1016 fail:
1017         up_read(&mdsc->snap_rwsem);
1018 fail_nosnap:
1019         unlock_page(page);
1020         return r;
1021 }
1022
1023 /*
1024  * We are only allowed to write into/dirty the page if the page is
1025  * clean, or already dirty within the same snap context.
1026  */
1027 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1028                             loff_t pos, unsigned len, unsigned flags,
1029                             struct page **pagep, void **fsdata)
1030 {
1031         struct inode *inode = file->f_dentry->d_inode;
1032         struct page *page;
1033         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1034         int r;
1035
1036         do {
1037                 /* get a page */
1038                 page = grab_cache_page_write_begin(mapping, index, 0);
1039                 if (!page)
1040                         return -ENOMEM;
1041                 *pagep = page;
1042
1043                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1044                 inode, page, (int)pos, (int)len);
1045
1046                 r = ceph_update_writeable_page(file, pos, len, page);
1047         } while (r == -EAGAIN);
1048
1049         return r;
1050 }
1051
1052 /*
1053  * we don't do anything in here that simple_write_end doesn't do
1054  * except adjust dirty page accounting and drop read lock on
1055  * mdsc->snap_rwsem.
1056  */
1057 static int ceph_write_end(struct file *file, struct address_space *mapping,
1058                           loff_t pos, unsigned len, unsigned copied,
1059                           struct page *page, void *fsdata)
1060 {
1061         struct inode *inode = file->f_dentry->d_inode;
1062         struct ceph_client *client = ceph_inode_to_client(inode);
1063         struct ceph_mds_client *mdsc = &client->mdsc;
1064         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1065         int check_cap = 0;
1066
1067         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1068              inode, page, (int)pos, (int)copied, (int)len);
1069
1070         /* zero the stale part of the page if we did a short copy */
1071         if (copied < len)
1072                 zero_user_segment(page, from+copied, len);
1073
1074         /* did file size increase? */
1075         /* (no need for i_size_read(); we caller holds i_mutex */
1076         if (pos+copied > inode->i_size)
1077                 check_cap = ceph_inode_set_size(inode, pos+copied);
1078
1079         if (!PageUptodate(page))
1080                 SetPageUptodate(page);
1081
1082         set_page_dirty(page);
1083
1084         unlock_page(page);
1085         up_read(&mdsc->snap_rwsem);
1086         page_cache_release(page);
1087
1088         if (check_cap)
1089                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1090
1091         return copied;
1092 }
1093
1094 /*
1095  * we set .direct_IO to indicate direct io is supported, but since we
1096  * intercept O_DIRECT reads and writes early, this function should
1097  * never get called.
1098  */
1099 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1100                               const struct iovec *iov,
1101                               loff_t pos, unsigned long nr_segs)
1102 {
1103         WARN_ON(1);
1104         return -EINVAL;
1105 }
1106
1107 const struct address_space_operations ceph_aops = {
1108         .readpage = ceph_readpage,
1109         .readpages = ceph_readpages,
1110         .writepage = ceph_writepage,
1111         .writepages = ceph_writepages_start,
1112         .write_begin = ceph_write_begin,
1113         .write_end = ceph_write_end,
1114         .set_page_dirty = ceph_set_page_dirty,
1115         .invalidatepage = ceph_invalidatepage,
1116         .releasepage = ceph_releasepage,
1117         .direct_IO = ceph_direct_io,
1118 };
1119
1120
1121 /*
1122  * vm ops
1123  */
1124
1125 /*
1126  * Reuse write_begin here for simplicity.
1127  */
1128 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1129 {
1130         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1131         struct page *page = vmf->page;
1132         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1133         loff_t off = page->index << PAGE_CACHE_SHIFT;
1134         loff_t size, len;
1135         int ret;
1136
1137         size = i_size_read(inode);
1138         if (off + PAGE_CACHE_SIZE <= size)
1139                 len = PAGE_CACHE_SIZE;
1140         else
1141                 len = size & ~PAGE_CACHE_MASK;
1142
1143         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1144              off, len, page, page->index);
1145
1146         lock_page(page);
1147
1148         ret = VM_FAULT_NOPAGE;
1149         if ((off > size) ||
1150             (page->mapping != inode->i_mapping))
1151                 goto out;
1152
1153         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1154         if (ret == 0) {
1155                 /* success.  we'll keep the page locked. */
1156                 set_page_dirty(page);
1157                 up_read(&mdsc->snap_rwsem);
1158                 ret = VM_FAULT_LOCKED;
1159         } else {
1160                 if (ret == -ENOMEM)
1161                         ret = VM_FAULT_OOM;
1162                 else
1163                         ret = VM_FAULT_SIGBUS;
1164         }
1165 out:
1166         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1167         if (ret != VM_FAULT_LOCKED)
1168                 unlock_page(page);
1169         return ret;
1170 }
1171
1172 static struct vm_operations_struct ceph_vmops = {
1173         .fault          = filemap_fault,
1174         .page_mkwrite   = ceph_page_mkwrite,
1175 };
1176
1177 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1178 {
1179         struct address_space *mapping = file->f_mapping;
1180
1181         if (!mapping->a_ops->readpage)
1182                 return -ENOEXEC;
1183         file_accessed(file);
1184         vma->vm_ops = &ceph_vmops;
1185         vma->vm_flags |= VM_CAN_NONLINEAR;
1186         return 0;
1187 }