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 "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /*
  91          * Note that we're grabbing a snapc ref here without holding
  92          * any locks!
  93          */
  94         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         if (ci->i_head_snapc == NULL)
  99                 ci->i_head_snapc = ceph_get_snap_context(snapc);
 100         ++ci->i_wrbuffer_ref_head;
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         ret = __set_page_dirty_nobuffers(page);
 121         WARN_ON(!PageLocked(page));
 122         WARN_ON(!page->mapping);
 123
 124         return ret;
 125 }
 126
 127 /*
 128  * If we are truncating the full page (i.e. offset == 0), adjust the
 129  * dirty page counters appropriately.  Only called if there is private
 130  * data on the page.
 131  */
 132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 133                                 unsigned int length)
 134 {
 135         struct inode *inode;
 136         struct ceph_inode_info *ci;
 137         struct ceph_snap_context *snapc = page_snap_context(page);
 138
 139         inode = page->mapping->host;
 140         ci = ceph_inode(inode);
 141
 142         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 143                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 144                      inode, page, page->index, offset, length);
 145                 return;
 146         }
 147
 148         ceph_invalidate_fscache_page(inode, page);
 149
 150         if (!PagePrivate(page))
 151                 return;
 152
 153         /*
 154          * We can get non-dirty pages here due to races between
 155          * set_page_dirty and truncate_complete_page; just spit out a
 156          * warning, in case we end up with accounting problems later.
 157          */
 158         if (!PageDirty(page))
 159                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 160
 161         ClearPageChecked(page);
 162
 163         dout("%p invalidatepage %p idx %lu full dirty page\n",
 164              inode, page, page->index);
 165
 166         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 167         ceph_put_snap_context(snapc);
 168         page->private = 0;
 169         ClearPagePrivate(page);
 170 }
 171
 172 static int ceph_releasepage(struct page *page, gfp_t g)
 173 {
 174         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 175         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 176         WARN_ON(PageDirty(page));
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * read a single page, without unlocking it.
 187  */
 188 static int readpage_nounlock(struct file *filp, struct page *page)
 189 {
 190         struct inode *inode = file_inode(filp);
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_osd_client *osdc =
 193                 &ceph_inode_to_client(inode)->client->osdc;
 194         int err = 0;
 195         u64 len = PAGE_CACHE_SIZE;
 196
 197         err = ceph_readpage_from_fscache(inode, page);
 198
 199         if (err == 0)
 200                 goto out;
 201
 202         dout("readpage inode %p file %p page %p index %lu\n",
 203              inode, filp, page, page->index);
 204         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 205                                   (u64) page_offset(page), &len,
 206                                   ci->i_truncate_seq, ci->i_truncate_size,
 207                                   &page, 1, 0);
 208         if (err == -ENOENT)
 209                 err = 0;
 210         if (err < 0) {
 211                 SetPageError(page);
 212                 goto out;
 213         } else if (err < PAGE_CACHE_SIZE) {
 214                 /* zero fill remainder of page */
 215                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 216         }
 217         SetPageUptodate(page);
 218
 219         if (err >= 0)
 220                 ceph_readpage_to_fscache(inode, 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         struct ceph_osd_data *osd_data;
 240         int rc = req->r_result;
 241         int bytes = le32_to_cpu(msg->hdr.data_len);
 242         int num_pages;
 243         int i;
 244
 245         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 246
 247         /* unlock all pages, zeroing any data we didn't read */
 248         osd_data = osd_req_op_extent_osd_data(req, 0);
 249         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 250         num_pages = calc_pages_for((u64)osd_data->alignment,
 251                                         (u64)osd_data->length);
 252         for (i = 0; i < num_pages; i++) {
 253                 struct page *page = osd_data->pages[i];
 254
 255                 if (bytes < (int)PAGE_CACHE_SIZE) {
 256                         /* zero (remainder of) page */
 257                         int s = bytes < 0 ? 0 : bytes;
 258                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 259                 }
 260                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 261                      page->index);
 262                 flush_dcache_page(page);
 263                 SetPageUptodate(page);
 264                 ceph_readpage_to_fscache(inode, page);
 265                 unlock_page(page);
 266                 page_cache_release(page);
 267                 bytes -= PAGE_CACHE_SIZE;
 268         }
 269         kfree(osd_data->pages);
 270 }
 271
 272 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 273 {
 274         int i;
 275
 276         for (i = 0; i < num_pages; i++)
 277                 unlock_page(pages[i]);
 278 }
 279
 280 /*
 281  * start an async read(ahead) operation.  return nr_pages we submitted
 282  * a read for on success, or negative error code.
 283  */
 284 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 285 {
 286         struct ceph_osd_client *osdc =
 287                 &ceph_inode_to_client(inode)->client->osdc;
 288         struct ceph_inode_info *ci = ceph_inode(inode);
 289         struct page *page = list_entry(page_list->prev, struct page, lru);
 290         struct ceph_vino vino;
 291         struct ceph_osd_request *req;
 292         u64 off;
 293         u64 len;
 294         int i;
 295         struct page **pages;
 296         pgoff_t next_index;
 297         int nr_pages = 0;
 298         int ret;
 299
 300         off = (u64) page_offset(page);
 301
 302         /* count pages */
 303         next_index = page->index;
 304         list_for_each_entry_reverse(page, page_list, lru) {
 305                 if (page->index != next_index)
 306                         break;
 307                 nr_pages++;
 308                 next_index++;
 309                 if (max && nr_pages == max)
 310                         break;
 311         }
 312         len = nr_pages << PAGE_CACHE_SHIFT;
 313         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 314              off, len);
 315         vino = ceph_vino(inode);
 316         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 317                                     1, CEPH_OSD_OP_READ,
 318                                     CEPH_OSD_FLAG_READ, NULL,
 319                                     ci->i_truncate_seq, ci->i_truncate_size,
 320                                     false);
 321         if (IS_ERR(req))
 322                 return PTR_ERR(req);
 323
 324         /* build page vector */
 325         nr_pages = calc_pages_for(0, len);
 326         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 327         ret = -ENOMEM;
 328         if (!pages)
 329                 goto out;
 330         for (i = 0; i < nr_pages; ++i) {
 331                 page = list_entry(page_list->prev, struct page, lru);
 332                 BUG_ON(PageLocked(page));
 333                 list_del(&page->lru);
 334
 335                 dout("start_read %p adding %p idx %lu\n", inode, page,
 336                      page->index);
 337                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 338                                           GFP_NOFS)) {
 339                         ceph_fscache_uncache_page(inode, page);
 340                         page_cache_release(page);
 341                         dout("start_read %p add_to_page_cache failed %p\n",
 342                              inode, page);
 343                         nr_pages = i;
 344                         goto out_pages;
 345                 }
 346                 pages[i] = page;
 347         }
 348         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 349         req->r_callback = finish_read;
 350         req->r_inode = inode;
 351
 352         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 353
 354         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 355         ret = ceph_osdc_start_request(osdc, req, false);
 356         if (ret < 0)
 357                 goto out_pages;
 358         ceph_osdc_put_request(req);
 359         return nr_pages;
 360
 361 out_pages:
 362         ceph_unlock_page_vector(pages, nr_pages);
 363         ceph_release_page_vector(pages, nr_pages);
 364 out:
 365         ceph_osdc_put_request(req);
 366         return ret;
 367 }
 368
 369
 370 /*
 371  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 372  * the caller (VM) cleans them up.
 373  */
 374 static int ceph_readpages(struct file *file, struct address_space *mapping,
 375                           struct list_head *page_list, unsigned nr_pages)
 376 {
 377         struct inode *inode = file_inode(file);
 378         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 379         int rc = 0;
 380         int max = 0;
 381
 382         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 383                                          &nr_pages);
 384
 385         if (rc == 0)
 386                 goto out;
 387
 388         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 389                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 390                         >> PAGE_SHIFT;
 391
 392         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 393                 file, nr_pages,
 394              max);
 395         while (!list_empty(page_list)) {
 396                 rc = start_read(inode, page_list, max);
 397                 if (rc < 0)
 398                         goto out;
 399                 BUG_ON(rc == 0);
 400         }
 401 out:
 402         ceph_fscache_readpages_cancel(inode, page_list);
 403
 404         dout("readpages %p file %p ret %d\n", inode, file, rc);
 405         return rc;
 406 }
 407
 408 /*
 409  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 410  * only snap context we are allowed to write back.
 411  */
 412 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 413                                                     u64 *snap_size)
 414 {
 415         struct ceph_inode_info *ci = ceph_inode(inode);
 416         struct ceph_snap_context *snapc = NULL;
 417         struct ceph_cap_snap *capsnap = NULL;
 418
 419         spin_lock(&ci->i_ceph_lock);
 420         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 421                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 422                      capsnap->context, capsnap->dirty_pages);
 423                 if (capsnap->dirty_pages) {
 424                         snapc = ceph_get_snap_context(capsnap->context);
 425                         if (snap_size)
 426                                 *snap_size = capsnap->size;
 427                         break;
 428                 }
 429         }
 430         if (!snapc && ci->i_wrbuffer_ref_head) {
 431                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 432                 dout(" head snapc %p has %d dirty pages\n",
 433                      snapc, ci->i_wrbuffer_ref_head);
 434         }
 435         spin_unlock(&ci->i_ceph_lock);
 436         return snapc;
 437 }
 438
 439 /*
 440  * Write a single page, but leave the page locked.
 441  *
 442  * If we get a write error, set the page error bit, but still adjust the
 443  * dirty page accounting (i.e., page is no longer dirty).
 444  */
 445 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 446 {
 447         struct inode *inode;
 448         struct ceph_inode_info *ci;
 449         struct ceph_fs_client *fsc;
 450         struct ceph_osd_client *osdc;
 451         struct ceph_snap_context *snapc, *oldest;
 452         loff_t page_off = page_offset(page);
 453         long writeback_stat;
 454         u64 truncate_size, snap_size = 0;
 455         u32 truncate_seq;
 456         int err = 0, len = PAGE_CACHE_SIZE;
 457
 458         dout("writepage %p idx %lu\n", page, page->index);
 459
 460         if (!page->mapping || !page->mapping->host) {
 461                 dout("writepage %p - no mapping\n", page);
 462                 return -EFAULT;
 463         }
 464         inode = page->mapping->host;
 465         ci = ceph_inode(inode);
 466         fsc = ceph_inode_to_client(inode);
 467         osdc = &fsc->client->osdc;
 468
 469         /* verify this is a writeable snap context */
 470         snapc = page_snap_context(page);
 471         if (snapc == NULL) {
 472                 dout("writepage %p page %p not dirty?\n", inode, page);
 473                 goto out;
 474         }
 475         oldest = get_oldest_context(inode, &snap_size);
 476         if (snapc->seq > oldest->seq) {
 477                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 478                      inode, page, snapc);
 479                 /* we should only noop if called by kswapd */
 480                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 481                 ceph_put_snap_context(oldest);
 482                 goto out;
 483         }
 484         ceph_put_snap_context(oldest);
 485
 486         spin_lock(&ci->i_ceph_lock);
 487         truncate_seq = ci->i_truncate_seq;
 488         truncate_size = ci->i_truncate_size;
 489         if (!snap_size)
 490                 snap_size = i_size_read(inode);
 491         spin_unlock(&ci->i_ceph_lock);
 492
 493         /* is this a partial page at end of file? */
 494         if (page_off >= snap_size) {
 495                 dout("%p page eof %llu\n", page, snap_size);
 496                 goto out;
 497         }
 498         if (snap_size < page_off + len)
 499                 len = snap_size - page_off;
 500
 501         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 502              inode, page, page->index, page_off, len, snapc);
 503
 504         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 505         if (writeback_stat >
 506             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 507                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 508
 509         ceph_readpage_to_fscache(inode, page);
 510
 511         set_page_writeback(page);
 512         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 513                                    &ci->i_layout, snapc,
 514                                    page_off, len,
 515                                    truncate_seq, truncate_size,
 516                                    &inode->i_mtime, &page, 1);
 517         if (err < 0) {
 518                 dout("writepage setting page/mapping error %d %p\n", err, page);
 519                 SetPageError(page);
 520                 mapping_set_error(&inode->i_data, err);
 521                 if (wbc)
 522                         wbc->pages_skipped++;
 523         } else {
 524                 dout("writepage cleaned page %p\n", page);
 525                 err = 0;  /* vfs expects us to return 0 */
 526         }
 527         page->private = 0;
 528         ClearPagePrivate(page);
 529         end_page_writeback(page);
 530         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 531         ceph_put_snap_context(snapc);  /* page's reference */
 532 out:
 533         return err;
 534 }
 535
 536 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 537 {
 538         int err;
 539         struct inode *inode = page->mapping->host;
 540         BUG_ON(!inode);
 541         ihold(inode);
 542         err = writepage_nounlock(page, wbc);
 543         unlock_page(page);
 544         iput(inode);
 545         return err;
 546 }
 547
 548
 549 /*
 550  * lame release_pages helper.  release_pages() isn't exported to
 551  * modules.
 552  */
 553 static void ceph_release_pages(struct page **pages, int num)
 554 {
 555         struct pagevec pvec;
 556         int i;
 557
 558         pagevec_init(&pvec, 0);
 559         for (i = 0; i < num; i++) {
 560                 if (pagevec_add(&pvec, pages[i]) == 0)
 561                         pagevec_release(&pvec);
 562         }
 563         pagevec_release(&pvec);
 564 }
 565
 566 /*
 567  * async writeback completion handler.
 568  *
 569  * If we get an error, set the mapping error bit, but not the individual
 570  * page error bits.
 571  */
 572 static void writepages_finish(struct ceph_osd_request *req,
 573                               struct ceph_msg *msg)
 574 {
 575         struct inode *inode = req->r_inode;
 576         struct ceph_inode_info *ci = ceph_inode(inode);
 577         struct ceph_osd_data *osd_data;
 578         unsigned wrote;
 579         struct page *page;
 580         int num_pages;
 581         int i;
 582         struct ceph_snap_context *snapc = req->r_snapc;
 583         struct address_space *mapping = inode->i_mapping;
 584         int rc = req->r_result;
 585         u64 bytes = req->r_ops[0].extent.length;
 586         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 587         long writeback_stat;
 588         unsigned issued = ceph_caps_issued(ci);
 589
 590         osd_data = osd_req_op_extent_osd_data(req, 0);
 591         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 592         num_pages = calc_pages_for((u64)osd_data->alignment,
 593                                         (u64)osd_data->length);
 594         if (rc >= 0) {
 595                 /*
 596                  * Assume we wrote the pages we originally sent.  The
 597                  * osd might reply with fewer pages if our writeback
 598                  * raced with a truncation and was adjusted at the osd,
 599                  * so don't believe the reply.
 600                  */
 601                 wrote = num_pages;
 602         } else {
 603                 wrote = 0;
 604                 mapping_set_error(mapping, rc);
 605         }
 606         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 607              inode, rc, bytes, wrote);
 608
 609         /* clean all pages */
 610         for (i = 0; i < num_pages; i++) {
 611                 page = osd_data->pages[i];
 612                 BUG_ON(!page);
 613                 WARN_ON(!PageUptodate(page));
 614
 615                 writeback_stat =
 616                         atomic_long_dec_return(&fsc->writeback_count);
 617                 if (writeback_stat <
 618                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 619                         clear_bdi_congested(&fsc->backing_dev_info,
 620                                             BLK_RW_ASYNC);
 621
 622                 ceph_put_snap_context(page_snap_context(page));
 623                 page->private = 0;
 624                 ClearPagePrivate(page);
 625                 dout("unlocking %d %p\n", i, page);
 626                 end_page_writeback(page);
 627
 628                 /*
 629                  * We lost the cache cap, need to truncate the page before
 630                  * it is unlocked, otherwise we'd truncate it later in the
 631                  * page truncation thread, possibly losing some data that
 632                  * raced its way in
 633                  */
 634                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 635                         generic_error_remove_page(inode->i_mapping, page);
 636
 637                 unlock_page(page);
 638         }
 639         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 640         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 641
 642         ceph_release_pages(osd_data->pages, num_pages);
 643         if (osd_data->pages_from_pool)
 644                 mempool_free(osd_data->pages,
 645                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 646         else
 647                 kfree(osd_data->pages);
 648         ceph_osdc_put_request(req);
 649 }
 650
 651 /*
 652  * initiate async writeback
 653  */
 654 static int ceph_writepages_start(struct address_space *mapping,
 655                                  struct writeback_control *wbc)
 656 {
 657         struct inode *inode = mapping->host;
 658         struct ceph_inode_info *ci = ceph_inode(inode);
 659         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 660         struct ceph_vino vino = ceph_vino(inode);
 661         pgoff_t index, start, end;
 662         int range_whole = 0;
 663         int should_loop = 1;
 664         pgoff_t max_pages = 0, max_pages_ever = 0;
 665         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 666         struct pagevec pvec;
 667         int done = 0;
 668         int rc = 0;
 669         unsigned wsize = 1 << inode->i_blkbits;
 670         struct ceph_osd_request *req = NULL;
 671         int do_sync;
 672         u64 truncate_size, snap_size;
 673         u32 truncate_seq;
 674
 675         /*
 676          * Include a 'sync' in the OSD request if this is a data
 677          * integrity write (e.g., O_SYNC write or fsync()), or if our
 678          * cap is being revoked.
 679          */
 680         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 681                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 682                 do_sync = 1;
 683         dout("writepages_start %p dosync=%d (mode=%s)\n",
 684              inode, do_sync,
 685              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 686              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 687
 688         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 689                 pr_warning("writepage_start %p on forced umount\n", inode);
 690                 return -EIO; /* we're in a forced umount, don't write! */
 691         }
 692         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 693                 wsize = fsc->mount_options->wsize;
 694         if (wsize < PAGE_CACHE_SIZE)
 695                 wsize = PAGE_CACHE_SIZE;
 696         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 697
 698         pagevec_init(&pvec, 0);
 699
 700         /* where to start/end? */
 701         if (wbc->range_cyclic) {
 702                 start = mapping->writeback_index; /* Start from prev offset */
 703                 end = -1;
 704                 dout(" cyclic, start at %lu\n", start);
 705         } else {
 706                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 707                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 708                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 709                         range_whole = 1;
 710                 should_loop = 0;
 711                 dout(" not cyclic, %lu to %lu\n", start, end);
 712         }
 713         index = start;
 714
 715 retry:
 716         /* find oldest snap context with dirty data */
 717         ceph_put_snap_context(snapc);
 718         snap_size = 0;
 719         snapc = get_oldest_context(inode, &snap_size);
 720         if (!snapc) {
 721                 /* hmm, why does writepages get called when there
 722                    is no dirty data? */
 723                 dout(" no snap context with dirty data?\n");
 724                 goto out;
 725         }
 726         if (snap_size == 0)
 727                 snap_size = i_size_read(inode);
 728         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 729              snapc, snapc->seq, snapc->num_snaps);
 730
 731         spin_lock(&ci->i_ceph_lock);
 732         truncate_seq = ci->i_truncate_seq;
 733         truncate_size = ci->i_truncate_size;
 734         if (!snap_size)
 735                 snap_size = i_size_read(inode);
 736         spin_unlock(&ci->i_ceph_lock);
 737
 738         if (last_snapc && snapc != last_snapc) {
 739                 /* if we switched to a newer snapc, restart our scan at the
 740                  * start of the original file range. */
 741                 dout("  snapc differs from last pass, restarting at %lu\n",
 742                      index);
 743                 index = start;
 744         }
 745         last_snapc = snapc;
 746
 747         while (!done && index <= end) {
 748                 int num_ops = do_sync ? 2 : 1;
 749                 unsigned i;
 750                 int first;
 751                 pgoff_t next;
 752                 int pvec_pages, locked_pages;
 753                 struct page **pages = NULL;
 754                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 755                 struct page *page;
 756                 int want;
 757                 u64 offset, len;
 758                 long writeback_stat;
 759
 760                 next = 0;
 761                 locked_pages = 0;
 762                 max_pages = max_pages_ever;
 763
 764 get_more_pages:
 765                 first = -1;
 766                 want = min(end - index,
 767                            min((pgoff_t)PAGEVEC_SIZE,
 768                                max_pages - (pgoff_t)locked_pages) - 1)
 769                         + 1;
 770                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 771                                                 PAGECACHE_TAG_DIRTY,
 772                                                 want);
 773                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 774                 if (!pvec_pages && !locked_pages)
 775                         break;
 776                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 777                         page = pvec.pages[i];
 778                         dout("? %p idx %lu\n", page, page->index);
 779                         if (locked_pages == 0)
 780                                 lock_page(page);  /* first page */
 781                         else if (!trylock_page(page))
 782                                 break;
 783
 784                         /* only dirty pages, or our accounting breaks */
 785                         if (unlikely(!PageDirty(page)) ||
 786                             unlikely(page->mapping != mapping)) {
 787                                 dout("!dirty or !mapping %p\n", page);
 788                                 unlock_page(page);
 789                                 break;
 790                         }
 791                         if (!wbc->range_cyclic && page->index > end) {
 792                                 dout("end of range %p\n", page);
 793                                 done = 1;
 794                                 unlock_page(page);
 795                                 break;
 796                         }
 797                         if (next && (page->index != next)) {
 798                                 dout("not consecutive %p\n", page);
 799                                 unlock_page(page);
 800                                 break;
 801                         }
 802                         if (wbc->sync_mode != WB_SYNC_NONE) {
 803                                 dout("waiting on writeback %p\n", page);
 804                                 wait_on_page_writeback(page);
 805                         }
 806                         if (page_offset(page) >= snap_size) {
 807                                 dout("%p page eof %llu\n", page, snap_size);
 808                                 done = 1;
 809                                 unlock_page(page);
 810                                 break;
 811                         }
 812                         if (PageWriteback(page)) {
 813                                 dout("%p under writeback\n", page);
 814                                 unlock_page(page);
 815                                 break;
 816                         }
 817
 818                         /* only if matching snap context */
 819                         pgsnapc = page_snap_context(page);
 820                         if (pgsnapc->seq > snapc->seq) {
 821                                 dout("page snapc %p %lld > oldest %p %lld\n",
 822                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 823                                 unlock_page(page);
 824                                 if (!locked_pages)
 825                                         continue; /* keep looking for snap */
 826                                 break;
 827                         }
 828
 829                         if (!clear_page_dirty_for_io(page)) {
 830                                 dout("%p !clear_page_dirty_for_io\n", page);
 831                                 unlock_page(page);
 832                                 break;
 833                         }
 834
 835                         /*
 836                          * We have something to write.  If this is
 837                          * the first locked page this time through,
 838                          * allocate an osd request and a page array
 839                          * that it will use.
 840                          */
 841                         if (locked_pages == 0) {
 842                                 BUG_ON(pages);
 843                                 /* prepare async write request */
 844                                 offset = (u64)page_offset(page);
 845                                 len = wsize;
 846                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 847                                                         &ci->i_layout, vino,
 848                                                         offset, &len, num_ops,
 849                                                         CEPH_OSD_OP_WRITE,
 850                                                         CEPH_OSD_FLAG_WRITE |
 851                                                         CEPH_OSD_FLAG_ONDISK,
 852                                                         snapc, truncate_seq,
 853                                                         truncate_size, true);
 854                                 if (IS_ERR(req)) {
 855                                         rc = PTR_ERR(req);
 856                                         unlock_page(page);
 857                                         break;
 858                                 }
 859
 860                                 req->r_callback = writepages_finish;
 861                                 req->r_inode = inode;
 862
 863                                 max_pages = calc_pages_for(0, (u64)len);
 864                                 pages = kmalloc(max_pages * sizeof (*pages),
 865                                                 GFP_NOFS);
 866                                 if (!pages) {
 867                                         pool = fsc->wb_pagevec_pool;
 868                                         pages = mempool_alloc(pool, GFP_NOFS);
 869                                         BUG_ON(!pages);
 870                                 }
 871                         }
 872
 873                         /* note position of first page in pvec */
 874                         if (first < 0)
 875                                 first = i;
 876                         dout("%p will write page %p idx %lu\n",
 877                              inode, page, page->index);
 878
 879                         writeback_stat =
 880                                atomic_long_inc_return(&fsc->writeback_count);
 881                         if (writeback_stat > CONGESTION_ON_THRESH(
 882                                     fsc->mount_options->congestion_kb)) {
 883                                 set_bdi_congested(&fsc->backing_dev_info,
 884                                                   BLK_RW_ASYNC);
 885                         }
 886
 887                         set_page_writeback(page);
 888                         pages[locked_pages] = page;
 889                         locked_pages++;
 890                         next = page->index + 1;
 891                 }
 892
 893                 /* did we get anything? */
 894                 if (!locked_pages)
 895                         goto release_pvec_pages;
 896                 if (i) {
 897                         int j;
 898                         BUG_ON(!locked_pages || first < 0);
 899
 900                         if (pvec_pages && i == pvec_pages &&
 901                             locked_pages < max_pages) {
 902                                 dout("reached end pvec, trying for more\n");
 903                                 pagevec_reinit(&pvec);
 904                                 goto get_more_pages;
 905                         }
 906
 907                         /* shift unused pages over in the pvec...  we
 908                          * will need to release them below. */
 909                         for (j = i; j < pvec_pages; j++) {
 910                                 dout(" pvec leftover page %p\n",
 911                                      pvec.pages[j]);
 912                                 pvec.pages[j-i+first] = pvec.pages[j];
 913                         }
 914                         pvec.nr -= i-first;
 915                 }
 916
 917                 /* Format the osd request message and submit the write */
 918
 919                 offset = page_offset(pages[0]);
 920                 len = min(snap_size - offset,
 921                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 922                 dout("writepages got %d pages at %llu~%llu\n",
 923                      locked_pages, offset, len);
 924
 925                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 926                                                         !!pool, false);
 927
 928                 pages = NULL;   /* request message now owns the pages array */
 929                 pool = NULL;
 930
 931                 /* Update the write op length in case we changed it */
 932
 933                 osd_req_op_extent_update(req, 0, len);
 934
 935                 vino = ceph_vino(inode);
 936                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 937                                         &inode->i_mtime);
 938
 939                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 940                 BUG_ON(rc);
 941                 req = NULL;
 942
 943                 /* continue? */
 944                 index = next;
 945                 wbc->nr_to_write -= locked_pages;
 946                 if (wbc->nr_to_write <= 0)
 947                         done = 1;
 948
 949 release_pvec_pages:
 950                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 951                      pvec.nr ? pvec.pages[0] : NULL);
 952                 pagevec_release(&pvec);
 953
 954                 if (locked_pages && !done)
 955                         goto retry;
 956         }
 957
 958         if (should_loop && !done) {
 959                 /* more to do; loop back to beginning of file */
 960                 dout("writepages looping back to beginning of file\n");
 961                 should_loop = 0;
 962                 index = 0;
 963                 goto retry;
 964         }
 965
 966         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 967                 mapping->writeback_index = index;
 968
 969 out:
 970         if (req)
 971                 ceph_osdc_put_request(req);
 972         ceph_put_snap_context(snapc);
 973         dout("writepages done, rc = %d\n", rc);
 974         return rc;
 975 }
 976
 977
 978
 979 /*
 980  * See if a given @snapc is either writeable, or already written.
 981  */
 982 static int context_is_writeable_or_written(struct inode *inode,
 983                                            struct ceph_snap_context *snapc)
 984 {
 985         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 986         int ret = !oldest || snapc->seq <= oldest->seq;
 987
 988         ceph_put_snap_context(oldest);
 989         return ret;
 990 }
 991
 992 /*
 993  * We are only allowed to write into/dirty the page if the page is
 994  * clean, or already dirty within the same snap context.
 995  *
 996  * called with page locked.
 997  * return success with page locked,
 998  * or any failure (incl -EAGAIN) with page unlocked.
 999  */
1000 static int ceph_update_writeable_page(struct file *file,
1001                             loff_t pos, unsigned len,
1002                             struct page *page)
1003 {
1004         struct inode *inode = file_inode(file);
1005         struct ceph_inode_info *ci = ceph_inode(inode);
1006         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1007         loff_t page_off = pos & PAGE_CACHE_MASK;
1008         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1009         int end_in_page = pos_in_page + len;
1010         loff_t i_size;
1011         int r;
1012         struct ceph_snap_context *snapc, *oldest;
1013
1014 retry_locked:
1015         /* writepages currently holds page lock, but if we change that later, */
1016         wait_on_page_writeback(page);
1017
1018         /* check snap context */
1019         BUG_ON(!ci->i_snap_realm);
1020         down_read(&mdsc->snap_rwsem);
1021         BUG_ON(!ci->i_snap_realm->cached_context);
1022         snapc = page_snap_context(page);
1023         if (snapc && snapc != ci->i_head_snapc) {
1024                 /*
1025                  * this page is already dirty in another (older) snap
1026                  * context!  is it writeable now?
1027                  */
1028                 oldest = get_oldest_context(inode, NULL);
1029                 up_read(&mdsc->snap_rwsem);
1030
1031                 if (snapc->seq > oldest->seq) {
1032                         ceph_put_snap_context(oldest);
1033                         dout(" page %p snapc %p not current or oldest\n",
1034                              page, snapc);
1035                         /*
1036                          * queue for writeback, and wait for snapc to
1037                          * be writeable or written
1038                          */
1039                         snapc = ceph_get_snap_context(snapc);
1040                         unlock_page(page);
1041                         ceph_queue_writeback(inode);
1042                         r = wait_event_interruptible(ci->i_cap_wq,
1043                                context_is_writeable_or_written(inode, snapc));
1044                         ceph_put_snap_context(snapc);
1045                         if (r == -ERESTARTSYS)
1046                                 return r;
1047                         return -EAGAIN;
1048                 }
1049                 ceph_put_snap_context(oldest);
1050
1051                 /* yay, writeable, do it now (without dropping page lock) */
1052                 dout(" page %p snapc %p not current, but oldest\n",
1053                      page, snapc);
1054                 if (!clear_page_dirty_for_io(page))
1055                         goto retry_locked;
1056                 r = writepage_nounlock(page, NULL);
1057                 if (r < 0)
1058                         goto fail_nosnap;
1059                 goto retry_locked;
1060         }
1061
1062         if (PageUptodate(page)) {
1063                 dout(" page %p already uptodate\n", page);
1064                 return 0;
1065         }
1066
1067         /* full page? */
1068         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1069                 return 0;
1070
1071         /* past end of file? */
1072         i_size = inode->i_size;   /* caller holds i_mutex */
1073
1074         if (i_size + len > inode->i_sb->s_maxbytes) {
1075                 /* file is too big */
1076                 r = -EINVAL;
1077                 goto fail;
1078         }
1079
1080         if (page_off >= i_size ||
1081             (pos_in_page == 0 && (pos+len) >= i_size &&
1082              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1083                 dout(" zeroing %p 0 - %d and %d - %d\n",
1084                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1085                 zero_user_segments(page,
1086                                    0, pos_in_page,
1087                                    end_in_page, PAGE_CACHE_SIZE);
1088                 return 0;
1089         }
1090
1091         /* we need to read it. */
1092         up_read(&mdsc->snap_rwsem);
1093         r = readpage_nounlock(file, page);
1094         if (r < 0)
1095                 goto fail_nosnap;
1096         goto retry_locked;
1097
1098 fail:
1099         up_read(&mdsc->snap_rwsem);
1100 fail_nosnap:
1101         unlock_page(page);
1102         return r;
1103 }
1104
1105 /*
1106  * We are only allowed to write into/dirty the page if the page is
1107  * clean, or already dirty within the same snap context.
1108  */
1109 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1110                             loff_t pos, unsigned len, unsigned flags,
1111                             struct page **pagep, void **fsdata)
1112 {
1113         struct inode *inode = file_inode(file);
1114         struct page *page;
1115         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1116         int r;
1117
1118         do {
1119                 /* get a page */
1120                 page = grab_cache_page_write_begin(mapping, index, 0);
1121                 if (!page)
1122                         return -ENOMEM;
1123                 *pagep = page;
1124
1125                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1126                      inode, page, (int)pos, (int)len);
1127
1128                 r = ceph_update_writeable_page(file, pos, len, page);
1129         } while (r == -EAGAIN);
1130
1131         return r;
1132 }
1133
1134 /*
1135  * we don't do anything in here that simple_write_end doesn't do
1136  * except adjust dirty page accounting and drop read lock on
1137  * mdsc->snap_rwsem.
1138  */
1139 static int ceph_write_end(struct file *file, struct address_space *mapping,
1140                           loff_t pos, unsigned len, unsigned copied,
1141                           struct page *page, void *fsdata)
1142 {
1143         struct inode *inode = file_inode(file);
1144         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1145         struct ceph_mds_client *mdsc = fsc->mdsc;
1146         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1147         int check_cap = 0;
1148
1149         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1150              inode, page, (int)pos, (int)copied, (int)len);
1151
1152         /* zero the stale part of the page if we did a short copy */
1153         if (copied < len)
1154                 zero_user_segment(page, from+copied, len);
1155
1156         /* did file size increase? */
1157         /* (no need for i_size_read(); we caller holds i_mutex */
1158         if (pos+copied > inode->i_size)
1159                 check_cap = ceph_inode_set_size(inode, pos+copied);
1160
1161         if (!PageUptodate(page))
1162                 SetPageUptodate(page);
1163
1164         set_page_dirty(page);
1165
1166         unlock_page(page);
1167         up_read(&mdsc->snap_rwsem);
1168         page_cache_release(page);
1169
1170         if (check_cap)
1171                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1172
1173         return copied;
1174 }
1175
1176 /*
1177  * we set .direct_IO to indicate direct io is supported, but since we
1178  * intercept O_DIRECT reads and writes early, this function should
1179  * never get called.
1180  */
1181 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1182                               const struct iovec *iov,
1183                               loff_t pos, unsigned long nr_segs)
1184 {
1185         WARN_ON(1);
1186         return -EINVAL;
1187 }
1188
1189 const struct address_space_operations ceph_aops = {
1190         .readpage = ceph_readpage,
1191         .readpages = ceph_readpages,
1192         .writepage = ceph_writepage,
1193         .writepages = ceph_writepages_start,
1194         .write_begin = ceph_write_begin,
1195         .write_end = ceph_write_end,
1196         .set_page_dirty = ceph_set_page_dirty,
1197         .invalidatepage = ceph_invalidatepage,
1198         .releasepage = ceph_releasepage,
1199         .direct_IO = ceph_direct_io,
1200 };
1201
1202
1203 /*
1204  * vm ops
1205  */
1206
1207 /*
1208  * Reuse write_begin here for simplicity.
1209  */
1210 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1211 {
1212         struct inode *inode = file_inode(vma->vm_file);
1213         struct page *page = vmf->page;
1214         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1215         loff_t off = page_offset(page);
1216         loff_t size, len;
1217         int ret;
1218
1219         /* Update time before taking page lock */
1220         file_update_time(vma->vm_file);
1221
1222         size = i_size_read(inode);
1223         if (off + PAGE_CACHE_SIZE <= size)
1224                 len = PAGE_CACHE_SIZE;
1225         else
1226                 len = size & ~PAGE_CACHE_MASK;
1227
1228         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1229              off, len, page, page->index);
1230
1231         lock_page(page);
1232
1233         ret = VM_FAULT_NOPAGE;
1234         if ((off > size) ||
1235             (page->mapping != inode->i_mapping))
1236                 goto out;
1237
1238         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1239         if (ret == 0) {
1240                 /* success.  we'll keep the page locked. */
1241                 set_page_dirty(page);
1242                 up_read(&mdsc->snap_rwsem);
1243                 ret = VM_FAULT_LOCKED;
1244         } else {
1245                 if (ret == -ENOMEM)
1246                         ret = VM_FAULT_OOM;
1247                 else
1248                         ret = VM_FAULT_SIGBUS;
1249         }
1250 out:
1251         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1252         if (ret != VM_FAULT_LOCKED)
1253                 unlock_page(page);
1254         return ret;
1255 }
1256
1257 static struct vm_operations_struct ceph_vmops = {
1258         .fault          = filemap_fault,
1259         .page_mkwrite   = ceph_page_mkwrite,
1260         .remap_pages    = generic_file_remap_pages,
1261 };
1262
1263 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1264 {
1265         struct address_space *mapping = file->f_mapping;
1266
1267         if (!mapping->a_ops->readpage)
1268                 return -ENOEXEC;
1269         file_accessed(file);
1270         vma->vm_ops = &ceph_vmops;
1271         return 0;
1272 }