fs/ceph/addr.c

   1 #include <linux/ceph/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 "mds_client.h"
  14 #include <linux/ceph/osd_client.h>
  15
  16 /*
  17  * Ceph address space ops.
  18  *
  19  * There are a few funny things going on here.
  20  *
  21  * The page->private field is used to reference a struct
  22  * ceph_snap_context for _every_ dirty page.  This indicates which
  23  * snapshot the page was logically dirtied in, and thus which snap
  24  * context needs to be associated with the osd write during writeback.
  25  *
  26  * Similarly, struct ceph_inode_info maintains a set of counters to
  27  * count dirty pages on the inode.  In the absence of snapshots,
  28  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  29  *
  30  * When a snapshot is taken (that is, when the client receives
  31  * notification that a snapshot was taken), each inode with caps and
  32  * with dirty pages (dirty pages implies there is a cap) gets a new
  33  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  34  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  35  * moved to capsnap->dirty. (Unless a sync write is currently in
  36  * progress.  In that case, the capsnap is said to be "pending", new
  37  * writes cannot start, and the capsnap isn't "finalized" until the
  38  * write completes (or fails) and a final size/mtime for the inode for
  39  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  40  *
  41  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  42  * we look for the first capsnap in i_cap_snaps and write out pages in
  43  * that snap context _only_.  Then we move on to the next capsnap,
  44  * eventually reaching the "live" or "head" context (i.e., pages that
  45  * are not yet snapped) and are writing the most recently dirtied
  46  * pages.
  47  *
  48  * Invalidate and so forth must take care to ensure the dirty page
  49  * accounting is preserved.
  50  */
  51
  52 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  53 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  54         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  55          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  56
  57 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  58 {
  59         if (PagePrivate(page))
  60                 return (void *)page->private;
  61         return NULL;
  62 }
  63
  64 /*
  65  * Dirty a page.  Optimistically adjust accounting, on the assumption
  66  * that we won't race with invalidate.  If we do, readjust.
  67  */
  68 static int ceph_set_page_dirty(struct page *page)
  69 {
  70         struct address_space *mapping = page->mapping;
  71         struct inode *inode;
  72         struct ceph_inode_info *ci;
  73         int undo = 0;
  74         struct ceph_snap_context *snapc;
  75
  76         if (unlikely(!mapping))
  77                 return !TestSetPageDirty(page);
  78
  79         if (TestSetPageDirty(page)) {
  80                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  81                      mapping->host, page, page->index);
  82                 return 0;
  83         }
  84
  85         inode = mapping->host;
  86         ci = ceph_inode(inode);
  87
  88         /*
  89          * Note that we're grabbing a snapc ref here without holding
  90          * any locks!
  91          */
  92         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  93
  94         /* dirty the head */
  95         spin_lock(&ci->i_ceph_lock);
  96         if (ci->i_head_snapc == NULL)
  97                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  98         ++ci->i_wrbuffer_ref_head;
  99         if (ci->i_wrbuffer_ref == 0)
 100                 ihold(inode);
 101         ++ci->i_wrbuffer_ref;
 102         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 103              "snapc %p seq %lld (%d snaps)\n",
 104              mapping->host, page, page->index,
 105              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 106              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 107              snapc, snapc->seq, snapc->num_snaps);
 108         spin_unlock(&ci->i_ceph_lock);
 109
 110         /* now adjust page */
 111         spin_lock_irq(&mapping->tree_lock);
 112         if (page->mapping) {    /* Race with truncate? */
 113                 WARN_ON_ONCE(!PageUptodate(page));
 114                 account_page_dirtied(page, page->mapping);
 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 = page_snap_context(page);
 151
 152         BUG_ON(!PageLocked(page));
 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(PagePrivate(page));
 190         return 0;
 191 }
 192
 193 /*
 194  * read a single page, without unlocking it.
 195  */
 196 static int readpage_nounlock(struct file *filp, struct page *page)
 197 {
 198         struct inode *inode = file_inode(filp);
 199         struct ceph_inode_info *ci = ceph_inode(inode);
 200         struct ceph_osd_client *osdc =
 201                 &ceph_inode_to_client(inode)->client->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                                   (u64) page_offset(page), &len,
 209                                   ci->i_truncate_seq, ci->i_truncate_size,
 210                                   &page, 1, 0);
 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  * Finish an async read(ahead) op.
 235  */
 236 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 237 {
 238         struct inode *inode = req->r_inode;
 239         int rc = req->r_result;
 240         int bytes = le32_to_cpu(msg->hdr.data_len);
 241         int i;
 242
 243         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 244
 245         /* unlock all pages, zeroing any data we didn't read */
 246         BUG_ON(req->r_data.type != CEPH_OSD_DATA_TYPE_PAGES);
 247         for (i = 0; i < req->r_data.num_pages; i++, bytes -= PAGE_CACHE_SIZE) {
 248                 struct page *page = req->r_data.pages[i];
 249
 250                 if (bytes < (int)PAGE_CACHE_SIZE) {
 251                         /* zero (remainder of) page */
 252                         int s = bytes < 0 ? 0 : bytes;
 253                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 254                 }
 255                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 256                      page->index);
 257                 flush_dcache_page(page);
 258                 SetPageUptodate(page);
 259                 unlock_page(page);
 260                 page_cache_release(page);
 261         }
 262         kfree(req->r_data.pages);
 263 }
 264
 265 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 266 {
 267         int i;
 268
 269         for (i = 0; i < num_pages; i++)
 270                 unlock_page(pages[i]);
 271 }
 272
 273 /*
 274  * start an async read(ahead) operation.  return nr_pages we submitted
 275  * a read for on success, or negative error code.
 276  */
 277 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 278 {
 279         struct ceph_osd_client *osdc =
 280                 &ceph_inode_to_client(inode)->client->osdc;
 281         struct ceph_inode_info *ci = ceph_inode(inode);
 282         struct page *page = list_entry(page_list->prev, struct page, lru);
 283         struct ceph_osd_request *req;
 284         u64 off;
 285         u64 len;
 286         int i;
 287         struct page **pages;
 288         pgoff_t next_index;
 289         int nr_pages = 0;
 290         int ret;
 291
 292         off = (u64) page_offset(page);
 293
 294         /* count pages */
 295         next_index = page->index;
 296         list_for_each_entry_reverse(page, page_list, lru) {
 297                 if (page->index != next_index)
 298                         break;
 299                 nr_pages++;
 300                 next_index++;
 301                 if (max && nr_pages == max)
 302                         break;
 303         }
 304         len = nr_pages << PAGE_CACHE_SHIFT;
 305         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 306              off, len);
 307
 308         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
 309                                     off, &len,
 310                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 311                                     NULL, 0,
 312                                     ci->i_truncate_seq, ci->i_truncate_size,
 313                                     NULL, false);
 314         if (IS_ERR(req))
 315                 return PTR_ERR(req);
 316
 317         /* build page vector */
 318         nr_pages = calc_pages_for(0, len);
 319         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 320         ret = -ENOMEM;
 321         if (!pages)
 322                 goto out;
 323         for (i = 0; i < nr_pages; ++i) {
 324                 page = list_entry(page_list->prev, struct page, lru);
 325                 BUG_ON(PageLocked(page));
 326                 list_del(&page->lru);
 327
 328                 dout("start_read %p adding %p idx %lu\n", inode, page,
 329                      page->index);
 330                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 331                                           GFP_NOFS)) {
 332                         page_cache_release(page);
 333                         dout("start_read %p add_to_page_cache failed %p\n",
 334                              inode, page);
 335                         nr_pages = i;
 336                         goto out_pages;
 337                 }
 338                 pages[i] = page;
 339         }
 340         req->r_data.type = CEPH_OSD_DATA_TYPE_PAGES;
 341         req->r_data.pages = pages;
 342         req->r_data.num_pages = nr_pages;
 343         req->r_data.alignment = 0;
 344         req->r_callback = finish_read;
 345         req->r_inode = inode;
 346
 347         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 348         ret = ceph_osdc_start_request(osdc, req, false);
 349         if (ret < 0)
 350                 goto out_pages;
 351         ceph_osdc_put_request(req);
 352         return nr_pages;
 353
 354 out_pages:
 355         ceph_unlock_page_vector(pages, nr_pages);
 356         ceph_release_page_vector(pages, nr_pages);
 357 out:
 358         ceph_osdc_put_request(req);
 359         return ret;
 360 }
 361
 362
 363 /*
 364  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 365  * the caller (VM) cleans them up.
 366  */
 367 static int ceph_readpages(struct file *file, struct address_space *mapping,
 368                           struct list_head *page_list, unsigned nr_pages)
 369 {
 370         struct inode *inode = file_inode(file);
 371         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 372         int rc = 0;
 373         int max = 0;
 374
 375         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 376                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 377                         >> PAGE_SHIFT;
 378
 379         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 380                 file, nr_pages,
 381              max);
 382         while (!list_empty(page_list)) {
 383                 rc = start_read(inode, page_list, max);
 384                 if (rc < 0)
 385                         goto out;
 386                 BUG_ON(rc == 0);
 387         }
 388 out:
 389         dout("readpages %p file %p ret %d\n", inode, file, rc);
 390         return rc;
 391 }
 392
 393 /*
 394  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 395  * only snap context we are allowed to write back.
 396  */
 397 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 398                                                     u64 *snap_size)
 399 {
 400         struct ceph_inode_info *ci = ceph_inode(inode);
 401         struct ceph_snap_context *snapc = NULL;
 402         struct ceph_cap_snap *capsnap = NULL;
 403
 404         spin_lock(&ci->i_ceph_lock);
 405         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 406                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 407                      capsnap->context, capsnap->dirty_pages);
 408                 if (capsnap->dirty_pages) {
 409                         snapc = ceph_get_snap_context(capsnap->context);
 410                         if (snap_size)
 411                                 *snap_size = capsnap->size;
 412                         break;
 413                 }
 414         }
 415         if (!snapc && ci->i_wrbuffer_ref_head) {
 416                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 417                 dout(" head snapc %p has %d dirty pages\n",
 418                      snapc, ci->i_wrbuffer_ref_head);
 419         }
 420         spin_unlock(&ci->i_ceph_lock);
 421         return snapc;
 422 }
 423
 424 /*
 425  * Write a single page, but leave the page locked.
 426  *
 427  * If we get a write error, set the page error bit, but still adjust the
 428  * dirty page accounting (i.e., page is no longer dirty).
 429  */
 430 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 431 {
 432         struct inode *inode;
 433         struct ceph_inode_info *ci;
 434         struct ceph_fs_client *fsc;
 435         struct ceph_osd_client *osdc;
 436         loff_t page_off = page_offset(page);
 437         int len = PAGE_CACHE_SIZE;
 438         loff_t i_size;
 439         int err = 0;
 440         struct ceph_snap_context *snapc, *oldest;
 441         u64 snap_size = 0;
 442         long writeback_stat;
 443
 444         dout("writepage %p idx %lu\n", page, page->index);
 445
 446         if (!page->mapping || !page->mapping->host) {
 447                 dout("writepage %p - no mapping\n", page);
 448                 return -EFAULT;
 449         }
 450         inode = page->mapping->host;
 451         ci = ceph_inode(inode);
 452         fsc = ceph_inode_to_client(inode);
 453         osdc = &fsc->client->osdc;
 454
 455         /* verify this is a writeable snap context */
 456         snapc = page_snap_context(page);
 457         if (snapc == NULL) {
 458                 dout("writepage %p page %p not dirty?\n", inode, page);
 459                 goto out;
 460         }
 461         oldest = get_oldest_context(inode, &snap_size);
 462         if (snapc->seq > oldest->seq) {
 463                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 464                      inode, page, snapc);
 465                 /* we should only noop if called by kswapd */
 466                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 467                 ceph_put_snap_context(oldest);
 468                 goto out;
 469         }
 470         ceph_put_snap_context(oldest);
 471
 472         /* is this a partial page at end of file? */
 473         if (snap_size)
 474                 i_size = snap_size;
 475         else
 476                 i_size = i_size_read(inode);
 477         if (i_size < page_off + len)
 478                 len = i_size - page_off;
 479
 480         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 481              inode, page, page->index, page_off, len, snapc);
 482
 483         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 484         if (writeback_stat >
 485             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 486                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 487
 488         set_page_writeback(page);
 489         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 490                                    &ci->i_layout, snapc,
 491                                    page_off, len,
 492                                    ci->i_truncate_seq, ci->i_truncate_size,
 493                                    &inode->i_mtime, &page, 1);
 494         if (err < 0) {
 495                 dout("writepage setting page/mapping error %d %p\n", err, page);
 496                 SetPageError(page);
 497                 mapping_set_error(&inode->i_data, err);
 498                 if (wbc)
 499                         wbc->pages_skipped++;
 500         } else {
 501                 dout("writepage cleaned page %p\n", page);
 502                 err = 0;  /* vfs expects us to return 0 */
 503         }
 504         page->private = 0;
 505         ClearPagePrivate(page);
 506         end_page_writeback(page);
 507         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 508         ceph_put_snap_context(snapc);  /* page's reference */
 509 out:
 510         return err;
 511 }
 512
 513 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 514 {
 515         int err;
 516         struct inode *inode = page->mapping->host;
 517         BUG_ON(!inode);
 518         ihold(inode);
 519         err = writepage_nounlock(page, wbc);
 520         unlock_page(page);
 521         iput(inode);
 522         return err;
 523 }
 524
 525
 526 /*
 527  * lame release_pages helper.  release_pages() isn't exported to
 528  * modules.
 529  */
 530 static void ceph_release_pages(struct page **pages, int num)
 531 {
 532         struct pagevec pvec;
 533         int i;
 534
 535         pagevec_init(&pvec, 0);
 536         for (i = 0; i < num; i++) {
 537                 if (pagevec_add(&pvec, pages[i]) == 0)
 538                         pagevec_release(&pvec);
 539         }
 540         pagevec_release(&pvec);
 541 }
 542
 543
 544 /*
 545  * async writeback completion handler.
 546  *
 547  * If we get an error, set the mapping error bit, but not the individual
 548  * page error bits.
 549  */
 550 static void writepages_finish(struct ceph_osd_request *req,
 551                               struct ceph_msg *msg)
 552 {
 553         struct inode *inode = req->r_inode;
 554         struct ceph_inode_info *ci = ceph_inode(inode);
 555         unsigned wrote;
 556         struct page *page;
 557         int i;
 558         struct ceph_snap_context *snapc = req->r_snapc;
 559         struct address_space *mapping = inode->i_mapping;
 560         int rc = req->r_result;
 561         u64 bytes = le64_to_cpu(req->r_request_ops[0].extent.length);
 562         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 563         long writeback_stat;
 564         unsigned issued = ceph_caps_issued(ci);
 565
 566         BUG_ON(req->r_data.type != CEPH_OSD_DATA_TYPE_PAGES);
 567         if (rc >= 0) {
 568                 /*
 569                  * Assume we wrote the pages we originally sent.  The
 570                  * osd might reply with fewer pages if our writeback
 571                  * raced with a truncation and was adjusted at the osd,
 572                  * so don't believe the reply.
 573                  */
 574                 wrote = req->r_data.num_pages;
 575         } else {
 576                 wrote = 0;
 577                 mapping_set_error(mapping, rc);
 578         }
 579         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 580              inode, rc, bytes, wrote);
 581
 582         /* clean all pages */
 583         for (i = 0; i < req->r_data.num_pages; i++) {
 584                 page = req->r_data.pages[i];
 585                 BUG_ON(!page);
 586                 WARN_ON(!PageUptodate(page));
 587
 588                 writeback_stat =
 589                         atomic_long_dec_return(&fsc->writeback_count);
 590                 if (writeback_stat <
 591                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 592                         clear_bdi_congested(&fsc->backing_dev_info,
 593                                             BLK_RW_ASYNC);
 594
 595                 ceph_put_snap_context(page_snap_context(page));
 596                 page->private = 0;
 597                 ClearPagePrivate(page);
 598                 dout("unlocking %d %p\n", i, page);
 599                 end_page_writeback(page);
 600
 601                 /*
 602                  * We lost the cache cap, need to truncate the page before
 603                  * it is unlocked, otherwise we'd truncate it later in the
 604                  * page truncation thread, possibly losing some data that
 605                  * raced its way in
 606                  */
 607                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 608                         generic_error_remove_page(inode->i_mapping, page);
 609
 610                 unlock_page(page);
 611         }
 612         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 613         ceph_put_wrbuffer_cap_refs(ci, req->r_data.num_pages, snapc);
 614
 615         ceph_release_pages(req->r_data.pages, req->r_data.num_pages);
 616         if (req->r_data.pages_from_pool)
 617                 mempool_free(req->r_data.pages,
 618                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 619         else
 620                 kfree(req->r_data.pages);
 621         ceph_osdc_put_request(req);
 622 }
 623
 624 /*
 625  * allocate a page vec, either directly, or if necessary, via a the
 626  * mempool.  we avoid the mempool if we can because req->r_data.num_pages
 627  * may be less than the maximum write size.
 628  */
 629 static void alloc_page_vec(struct ceph_fs_client *fsc,
 630                            struct ceph_osd_request *req)
 631 {
 632         req->r_data.pages = kmalloc(sizeof(struct page *) * req->r_data.num_pages,
 633                                GFP_NOFS);
 634         if (!req->r_data.pages) {
 635                 req->r_data.pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
 636                 req->r_data.pages_from_pool = 1;
 637                 WARN_ON(!req->r_data.pages);
 638         }
 639 }
 640
 641 /*
 642  * initiate async writeback
 643  */
 644 static int ceph_writepages_start(struct address_space *mapping,
 645                                  struct writeback_control *wbc)
 646 {
 647         struct inode *inode = mapping->host;
 648         struct ceph_inode_info *ci = ceph_inode(inode);
 649         struct ceph_fs_client *fsc;
 650         pgoff_t index, start, end;
 651         int range_whole = 0;
 652         int should_loop = 1;
 653         pgoff_t max_pages = 0, max_pages_ever = 0;
 654         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 655         struct pagevec pvec;
 656         int done = 0;
 657         int rc = 0;
 658         unsigned wsize = 1 << inode->i_blkbits;
 659         struct ceph_osd_request *req = NULL;
 660         int do_sync;
 661         u64 snap_size = 0;
 662
 663         /*
 664          * Include a 'sync' in the OSD request if this is a data
 665          * integrity write (e.g., O_SYNC write or fsync()), or if our
 666          * cap is being revoked.
 667          */
 668         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 669         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 670                 do_sync = 1;
 671         dout("writepages_start %p dosync=%d (mode=%s)\n",
 672              inode, do_sync,
 673              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 674              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 675
 676         fsc = ceph_inode_to_client(inode);
 677         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 678                 pr_warning("writepage_start %p on forced umount\n", inode);
 679                 return -EIO; /* we're in a forced umount, don't write! */
 680         }
 681         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 682                 wsize = fsc->mount_options->wsize;
 683         if (wsize < PAGE_CACHE_SIZE)
 684                 wsize = PAGE_CACHE_SIZE;
 685         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 686
 687         pagevec_init(&pvec, 0);
 688
 689         /* where to start/end? */
 690         if (wbc->range_cyclic) {
 691                 start = mapping->writeback_index; /* Start from prev offset */
 692                 end = -1;
 693                 dout(" cyclic, start at %lu\n", start);
 694         } else {
 695                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 696                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 697                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 698                         range_whole = 1;
 699                 should_loop = 0;
 700                 dout(" not cyclic, %lu to %lu\n", start, end);
 701         }
 702         index = start;
 703
 704 retry:
 705         /* find oldest snap context with dirty data */
 706         ceph_put_snap_context(snapc);
 707         snapc = get_oldest_context(inode, &snap_size);
 708         if (!snapc) {
 709                 /* hmm, why does writepages get called when there
 710                    is no dirty data? */
 711                 dout(" no snap context with dirty data?\n");
 712                 goto out;
 713         }
 714         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 715              snapc, snapc->seq, snapc->num_snaps);
 716         if (last_snapc && snapc != last_snapc) {
 717                 /* if we switched to a newer snapc, restart our scan at the
 718                  * start of the original file range. */
 719                 dout("  snapc differs from last pass, restarting at %lu\n",
 720                      index);
 721                 index = start;
 722         }
 723         last_snapc = snapc;
 724
 725         while (!done && index <= end) {
 726                 unsigned i;
 727                 int first;
 728                 pgoff_t next;
 729                 int pvec_pages, locked_pages;
 730                 struct page *page;
 731                 int want;
 732                 u64 offset, len;
 733                 long writeback_stat;
 734
 735                 next = 0;
 736                 locked_pages = 0;
 737                 max_pages = max_pages_ever;
 738
 739 get_more_pages:
 740                 first = -1;
 741                 want = min(end - index,
 742                            min((pgoff_t)PAGEVEC_SIZE,
 743                                max_pages - (pgoff_t)locked_pages) - 1)
 744                         + 1;
 745                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 746                                                 PAGECACHE_TAG_DIRTY,
 747                                                 want);
 748                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 749                 if (!pvec_pages && !locked_pages)
 750                         break;
 751                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 752                         page = pvec.pages[i];
 753                         dout("? %p idx %lu\n", page, page->index);
 754                         if (locked_pages == 0)
 755                                 lock_page(page);  /* first page */
 756                         else if (!trylock_page(page))
 757                                 break;
 758
 759                         /* only dirty pages, or our accounting breaks */
 760                         if (unlikely(!PageDirty(page)) ||
 761                             unlikely(page->mapping != mapping)) {
 762                                 dout("!dirty or !mapping %p\n", page);
 763                                 unlock_page(page);
 764                                 break;
 765                         }
 766                         if (!wbc->range_cyclic && page->index > end) {
 767                                 dout("end of range %p\n", page);
 768                                 done = 1;
 769                                 unlock_page(page);
 770                                 break;
 771                         }
 772                         if (next && (page->index != next)) {
 773                                 dout("not consecutive %p\n", page);
 774                                 unlock_page(page);
 775                                 break;
 776                         }
 777                         if (wbc->sync_mode != WB_SYNC_NONE) {
 778                                 dout("waiting on writeback %p\n", page);
 779                                 wait_on_page_writeback(page);
 780                         }
 781                         if ((snap_size && page_offset(page) > snap_size) ||
 782                             (!snap_size &&
 783                              page_offset(page) > i_size_read(inode))) {
 784                                 dout("%p page eof %llu\n", page, snap_size ?
 785                                      snap_size : i_size_read(inode));
 786                                 done = 1;
 787                                 unlock_page(page);
 788                                 break;
 789                         }
 790                         if (PageWriteback(page)) {
 791                                 dout("%p under writeback\n", page);
 792                                 unlock_page(page);
 793                                 break;
 794                         }
 795
 796                         /* only if matching snap context */
 797                         pgsnapc = page_snap_context(page);
 798                         if (pgsnapc->seq > snapc->seq) {
 799                                 dout("page snapc %p %lld > oldest %p %lld\n",
 800                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 801                                 unlock_page(page);
 802                                 if (!locked_pages)
 803                                         continue; /* keep looking for snap */
 804                                 break;
 805                         }
 806
 807                         if (!clear_page_dirty_for_io(page)) {
 808                                 dout("%p !clear_page_dirty_for_io\n", page);
 809                                 unlock_page(page);
 810                                 break;
 811                         }
 812
 813                         /* ok */
 814                         if (locked_pages == 0) {
 815                                 /* prepare async write request */
 816                                 offset = (u64) page_offset(page);
 817                                 len = wsize;
 818                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 819                                             &ci->i_layout,
 820                                             ceph_vino(inode),
 821                                             offset, &len,
 822                                             CEPH_OSD_OP_WRITE,
 823                                             CEPH_OSD_FLAG_WRITE |
 824                                                     CEPH_OSD_FLAG_ONDISK,
 825                                             snapc, do_sync,
 826                                             ci->i_truncate_seq,
 827                                             ci->i_truncate_size,
 828                                             &inode->i_mtime, true);
 829
 830                                 if (IS_ERR(req)) {
 831                                         rc = PTR_ERR(req);
 832                                         unlock_page(page);
 833                                         break;
 834                                 }
 835
 836                                 req->r_data.type = CEPH_OSD_DATA_TYPE_PAGES;
 837                                 req->r_data.num_pages = calc_pages_for(0, len);
 838                                 req->r_data.alignment = 0;
 839                                 max_pages = req->r_data.num_pages;
 840
 841                                 alloc_page_vec(fsc, req);
 842                                 req->r_callback = writepages_finish;
 843                                 req->r_inode = inode;
 844                         }
 845
 846                         /* note position of first page in pvec */
 847                         if (first < 0)
 848                                 first = i;
 849                         dout("%p will write page %p idx %lu\n",
 850                              inode, page, page->index);
 851
 852                         writeback_stat =
 853                                atomic_long_inc_return(&fsc->writeback_count);
 854                         if (writeback_stat > CONGESTION_ON_THRESH(
 855                                     fsc->mount_options->congestion_kb)) {
 856                                 set_bdi_congested(&fsc->backing_dev_info,
 857                                                   BLK_RW_ASYNC);
 858                         }
 859
 860                         set_page_writeback(page);
 861                         req->r_data.pages[locked_pages] = page;
 862                         locked_pages++;
 863                         next = page->index + 1;
 864                 }
 865
 866                 /* did we get anything? */
 867                 if (!locked_pages)
 868                         goto release_pvec_pages;
 869                 if (i) {
 870                         int j;
 871                         BUG_ON(!locked_pages || first < 0);
 872
 873                         if (pvec_pages && i == pvec_pages &&
 874                             locked_pages < max_pages) {
 875                                 dout("reached end pvec, trying for more\n");
 876                                 pagevec_reinit(&pvec);
 877                                 goto get_more_pages;
 878                         }
 879
 880                         /* shift unused pages over in the pvec...  we
 881                          * will need to release them below. */
 882                         for (j = i; j < pvec_pages; j++) {
 883                                 dout(" pvec leftover page %p\n",
 884                                      pvec.pages[j]);
 885                                 pvec.pages[j-i+first] = pvec.pages[j];
 886                         }
 887                         pvec.nr -= i-first;
 888                 }
 889
 890                 /* submit the write */
 891                 offset = req->r_data.pages[0]->index << PAGE_CACHE_SHIFT;
 892                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 893                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 894                 dout("writepages got %d pages at %llu~%llu\n",
 895                      locked_pages, offset, len);
 896
 897                 /* revise final length, page count */
 898                 req->r_data.num_pages = locked_pages;
 899                 req->r_request_ops[0].extent.length = cpu_to_le64(len);
 900                 req->r_request_ops[0].payload_len = cpu_to_le32(len);
 901                 req->r_request->hdr.data_len = cpu_to_le32(len);
 902
 903                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 904                 BUG_ON(rc);
 905                 req = NULL;
 906
 907                 /* continue? */
 908                 index = next;
 909                 wbc->nr_to_write -= locked_pages;
 910                 if (wbc->nr_to_write <= 0)
 911                         done = 1;
 912
 913 release_pvec_pages:
 914                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 915                      pvec.nr ? pvec.pages[0] : NULL);
 916                 pagevec_release(&pvec);
 917
 918                 if (locked_pages && !done)
 919                         goto retry;
 920         }
 921
 922         if (should_loop && !done) {
 923                 /* more to do; loop back to beginning of file */
 924                 dout("writepages looping back to beginning of file\n");
 925                 should_loop = 0;
 926                 index = 0;
 927                 goto retry;
 928         }
 929
 930         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 931                 mapping->writeback_index = index;
 932
 933 out:
 934         if (req)
 935                 ceph_osdc_put_request(req);
 936         ceph_put_snap_context(snapc);
 937         dout("writepages done, rc = %d\n", rc);
 938         return rc;
 939 }
 940
 941
 942
 943 /*
 944  * See if a given @snapc is either writeable, or already written.
 945  */
 946 static int context_is_writeable_or_written(struct inode *inode,
 947                                            struct ceph_snap_context *snapc)
 948 {
 949         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 950         int ret = !oldest || snapc->seq <= oldest->seq;
 951
 952         ceph_put_snap_context(oldest);
 953         return ret;
 954 }
 955
 956 /*
 957  * We are only allowed to write into/dirty the page if the page is
 958  * clean, or already dirty within the same snap context.
 959  *
 960  * called with page locked.
 961  * return success with page locked,
 962  * or any failure (incl -EAGAIN) with page unlocked.
 963  */
 964 static int ceph_update_writeable_page(struct file *file,
 965                             loff_t pos, unsigned len,
 966                             struct page *page)
 967 {
 968         struct inode *inode = file_inode(file);
 969         struct ceph_inode_info *ci = ceph_inode(inode);
 970         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 971         loff_t page_off = pos & PAGE_CACHE_MASK;
 972         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 973         int end_in_page = pos_in_page + len;
 974         loff_t i_size;
 975         int r;
 976         struct ceph_snap_context *snapc, *oldest;
 977
 978 retry_locked:
 979         /* writepages currently holds page lock, but if we change that later, */
 980         wait_on_page_writeback(page);
 981
 982         /* check snap context */
 983         BUG_ON(!ci->i_snap_realm);
 984         down_read(&mdsc->snap_rwsem);
 985         BUG_ON(!ci->i_snap_realm->cached_context);
 986         snapc = page_snap_context(page);
 987         if (snapc && snapc != ci->i_head_snapc) {
 988                 /*
 989                  * this page is already dirty in another (older) snap
 990                  * context!  is it writeable now?
 991                  */
 992                 oldest = get_oldest_context(inode, NULL);
 993                 up_read(&mdsc->snap_rwsem);
 994
 995                 if (snapc->seq > oldest->seq) {
 996                         ceph_put_snap_context(oldest);
 997                         dout(" page %p snapc %p not current or oldest\n",
 998                              page, snapc);
 999                         /*
1000                          * queue for writeback, and wait for snapc to
1001                          * be writeable or written
1002                          */
1003                         snapc = ceph_get_snap_context(snapc);
1004                         unlock_page(page);
1005                         ceph_queue_writeback(inode);
1006                         r = wait_event_interruptible(ci->i_cap_wq,
1007                                context_is_writeable_or_written(inode, snapc));
1008                         ceph_put_snap_context(snapc);
1009                         if (r == -ERESTARTSYS)
1010                                 return r;
1011                         return -EAGAIN;
1012                 }
1013                 ceph_put_snap_context(oldest);
1014
1015                 /* yay, writeable, do it now (without dropping page lock) */
1016                 dout(" page %p snapc %p not current, but oldest\n",
1017                      page, snapc);
1018                 if (!clear_page_dirty_for_io(page))
1019                         goto retry_locked;
1020                 r = writepage_nounlock(page, NULL);
1021                 if (r < 0)
1022                         goto fail_nosnap;
1023                 goto retry_locked;
1024         }
1025
1026         if (PageUptodate(page)) {
1027                 dout(" page %p already uptodate\n", page);
1028                 return 0;
1029         }
1030
1031         /* full page? */
1032         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1033                 return 0;
1034
1035         /* past end of file? */
1036         i_size = inode->i_size;   /* caller holds i_mutex */
1037
1038         if (i_size + len > inode->i_sb->s_maxbytes) {
1039                 /* file is too big */
1040                 r = -EINVAL;
1041                 goto fail;
1042         }
1043
1044         if (page_off >= i_size ||
1045             (pos_in_page == 0 && (pos+len) >= i_size &&
1046              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1047                 dout(" zeroing %p 0 - %d and %d - %d\n",
1048                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1049                 zero_user_segments(page,
1050                                    0, pos_in_page,
1051                                    end_in_page, PAGE_CACHE_SIZE);
1052                 return 0;
1053         }
1054
1055         /* we need to read it. */
1056         up_read(&mdsc->snap_rwsem);
1057         r = readpage_nounlock(file, page);
1058         if (r < 0)
1059                 goto fail_nosnap;
1060         goto retry_locked;
1061
1062 fail:
1063         up_read(&mdsc->snap_rwsem);
1064 fail_nosnap:
1065         unlock_page(page);
1066         return r;
1067 }
1068
1069 /*
1070  * We are only allowed to write into/dirty the page if the page is
1071  * clean, or already dirty within the same snap context.
1072  */
1073 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1074                             loff_t pos, unsigned len, unsigned flags,
1075                             struct page **pagep, void **fsdata)
1076 {
1077         struct inode *inode = file_inode(file);
1078         struct page *page;
1079         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1080         int r;
1081
1082         do {
1083                 /* get a page */
1084                 page = grab_cache_page_write_begin(mapping, index, 0);
1085                 if (!page)
1086                         return -ENOMEM;
1087                 *pagep = page;
1088
1089                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1090                      inode, page, (int)pos, (int)len);
1091
1092                 r = ceph_update_writeable_page(file, pos, len, page);
1093         } while (r == -EAGAIN);
1094
1095         return r;
1096 }
1097
1098 /*
1099  * we don't do anything in here that simple_write_end doesn't do
1100  * except adjust dirty page accounting and drop read lock on
1101  * mdsc->snap_rwsem.
1102  */
1103 static int ceph_write_end(struct file *file, struct address_space *mapping,
1104                           loff_t pos, unsigned len, unsigned copied,
1105                           struct page *page, void *fsdata)
1106 {
1107         struct inode *inode = file_inode(file);
1108         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1109         struct ceph_mds_client *mdsc = fsc->mdsc;
1110         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1111         int check_cap = 0;
1112
1113         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1114              inode, page, (int)pos, (int)copied, (int)len);
1115
1116         /* zero the stale part of the page if we did a short copy */
1117         if (copied < len)
1118                 zero_user_segment(page, from+copied, len);
1119
1120         /* did file size increase? */
1121         /* (no need for i_size_read(); we caller holds i_mutex */
1122         if (pos+copied > inode->i_size)
1123                 check_cap = ceph_inode_set_size(inode, pos+copied);
1124
1125         if (!PageUptodate(page))
1126                 SetPageUptodate(page);
1127
1128         set_page_dirty(page);
1129
1130         unlock_page(page);
1131         up_read(&mdsc->snap_rwsem);
1132         page_cache_release(page);
1133
1134         if (check_cap)
1135                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1136
1137         return copied;
1138 }
1139
1140 /*
1141  * we set .direct_IO to indicate direct io is supported, but since we
1142  * intercept O_DIRECT reads and writes early, this function should
1143  * never get called.
1144  */
1145 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1146                               const struct iovec *iov,
1147                               loff_t pos, unsigned long nr_segs)
1148 {
1149         WARN_ON(1);
1150         return -EINVAL;
1151 }
1152
1153 const struct address_space_operations ceph_aops = {
1154         .readpage = ceph_readpage,
1155         .readpages = ceph_readpages,
1156         .writepage = ceph_writepage,
1157         .writepages = ceph_writepages_start,
1158         .write_begin = ceph_write_begin,
1159         .write_end = ceph_write_end,
1160         .set_page_dirty = ceph_set_page_dirty,
1161         .invalidatepage = ceph_invalidatepage,
1162         .releasepage = ceph_releasepage,
1163         .direct_IO = ceph_direct_io,
1164 };
1165
1166
1167 /*
1168  * vm ops
1169  */
1170
1171 /*
1172  * Reuse write_begin here for simplicity.
1173  */
1174 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1175 {
1176         struct inode *inode = file_inode(vma->vm_file);
1177         struct page *page = vmf->page;
1178         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1179         loff_t off = page_offset(page);
1180         loff_t size, len;
1181         int ret;
1182
1183         /* Update time before taking page lock */
1184         file_update_time(vma->vm_file);
1185
1186         size = i_size_read(inode);
1187         if (off + PAGE_CACHE_SIZE <= size)
1188                 len = PAGE_CACHE_SIZE;
1189         else
1190                 len = size & ~PAGE_CACHE_MASK;
1191
1192         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1193              off, len, page, page->index);
1194
1195         lock_page(page);
1196
1197         ret = VM_FAULT_NOPAGE;
1198         if ((off > size) ||
1199             (page->mapping != inode->i_mapping))
1200                 goto out;
1201
1202         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1203         if (ret == 0) {
1204                 /* success.  we'll keep the page locked. */
1205                 set_page_dirty(page);
1206                 up_read(&mdsc->snap_rwsem);
1207                 ret = VM_FAULT_LOCKED;
1208         } else {
1209                 if (ret == -ENOMEM)
1210                         ret = VM_FAULT_OOM;
1211                 else
1212                         ret = VM_FAULT_SIGBUS;
1213         }
1214 out:
1215         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1216         if (ret != VM_FAULT_LOCKED)
1217                 unlock_page(page);
1218         return ret;
1219 }
1220
1221 static struct vm_operations_struct ceph_vmops = {
1222         .fault          = filemap_fault,
1223         .page_mkwrite   = ceph_page_mkwrite,
1224         .remap_pages    = generic_file_remap_pages,
1225 };
1226
1227 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1228 {
1229         struct address_space *mapping = file->f_mapping;
1230
1231         if (!mapping->a_ops->readpage)
1232                 return -ENOEXEC;
1233         file_accessed(file);
1234         vma->vm_ops = &ceph_vmops;
1235         return 0;
1236 }