Pileus Git - ~andy/linux/blob - 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         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                 unlock_page(page);
 265                 page_cache_release(page);
 266                 bytes -= PAGE_CACHE_SIZE;
 267         }
 268         kfree(osd_data->pages);
 269 }
 270
 271 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 272 {
 273         int i;
 274
 275         for (i = 0; i < num_pages; i++)
 276                 unlock_page(pages[i]);
 277 }
 278
 279 /*
 280  * start an async read(ahead) operation.  return nr_pages we submitted
 281  * a read for on success, or negative error code.
 282  */
 283 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 284 {
 285         struct ceph_osd_client *osdc =
 286                 &ceph_inode_to_client(inode)->client->osdc;
 287         struct ceph_inode_info *ci = ceph_inode(inode);
 288         struct page *page = list_entry(page_list->prev, struct page, lru);
 289         struct ceph_vino vino;
 290         struct ceph_osd_request *req;
 291         u64 off;
 292         u64 len;
 293         int i;
 294         struct page **pages;
 295         pgoff_t next_index;
 296         int nr_pages = 0;
 297         int ret;
 298
 299         off = (u64) page_offset(page);
 300
 301         /* count pages */
 302         next_index = page->index;
 303         list_for_each_entry_reverse(page, page_list, lru) {
 304                 if (page->index != next_index)
 305                         break;
 306                 nr_pages++;
 307                 next_index++;
 308                 if (max && nr_pages == max)
 309                         break;
 310         }
 311         len = nr_pages << PAGE_CACHE_SHIFT;
 312         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 313              off, len);
 314         vino = ceph_vino(inode);
 315         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 316                                     1, CEPH_OSD_OP_READ,
 317                                     CEPH_OSD_FLAG_READ, NULL,
 318                                     ci->i_truncate_seq, ci->i_truncate_size,
 319                                     false);
 320         if (IS_ERR(req))
 321                 return PTR_ERR(req);
 322
 323         /* build page vector */
 324         nr_pages = calc_pages_for(0, len);
 325         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 326         ret = -ENOMEM;
 327         if (!pages)
 328                 goto out;
 329         for (i = 0; i < nr_pages; ++i) {
 330                 page = list_entry(page_list->prev, struct page, lru);
 331                 BUG_ON(PageLocked(page));
 332                 list_del(&page->lru);
 333
 334                 dout("start_read %p adding %p idx %lu\n", inode, page,
 335                      page->index);
 336                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 337                                           GFP_NOFS)) {
 338                         page_cache_release(page);
 339                         dout("start_read %p add_to_page_cache failed %p\n",
 340                              inode, page);
 341                         nr_pages = i;
 342                         goto out_pages;
 343                 }
 344                 pages[i] = page;
 345         }
 346         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 347         req->r_callback = finish_read;
 348         req->r_inode = inode;
 349
 350         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 351
 352         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 353         ret = ceph_osdc_start_request(osdc, req, false);
 354         if (ret < 0)
 355                 goto out_pages;
 356         ceph_osdc_put_request(req);
 357         return nr_pages;
 358
 359 out_pages:
 360         ceph_unlock_page_vector(pages, nr_pages);
 361         ceph_release_page_vector(pages, nr_pages);
 362 out:
 363         ceph_osdc_put_request(req);
 364         return ret;
 365 }
 366
 367
 368 /*
 369  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 370  * the caller (VM) cleans them up.
 371  */
 372 static int ceph_readpages(struct file *file, struct address_space *mapping,
 373                           struct list_head *page_list, unsigned nr_pages)
 374 {
 375         struct inode *inode = file_inode(file);
 376         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 377         int rc = 0;
 378         int max = 0;
 379
 380         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 381                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 382                         >> PAGE_SHIFT;
 383
 384         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 385                 file, nr_pages,
 386              max);
 387         while (!list_empty(page_list)) {
 388                 rc = start_read(inode, page_list, max);
 389                 if (rc < 0)
 390                         goto out;
 391                 BUG_ON(rc == 0);
 392         }
 393 out:
 394         dout("readpages %p file %p ret %d\n", inode, file, rc);
 395         return rc;
 396 }
 397
 398 /*
 399  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 400  * only snap context we are allowed to write back.
 401  */
 402 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 403                                                     u64 *snap_size)
 404 {
 405         struct ceph_inode_info *ci = ceph_inode(inode);
 406         struct ceph_snap_context *snapc = NULL;
 407         struct ceph_cap_snap *capsnap = NULL;
 408
 409         spin_lock(&ci->i_ceph_lock);
 410         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 411                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 412                      capsnap->context, capsnap->dirty_pages);
 413                 if (capsnap->dirty_pages) {
 414                         snapc = ceph_get_snap_context(capsnap->context);
 415                         if (snap_size)
 416                                 *snap_size = capsnap->size;
 417                         break;
 418                 }
 419         }
 420         if (!snapc && ci->i_wrbuffer_ref_head) {
 421                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 422                 dout(" head snapc %p has %d dirty pages\n",
 423                      snapc, ci->i_wrbuffer_ref_head);
 424         }
 425         spin_unlock(&ci->i_ceph_lock);
 426         return snapc;
 427 }
 428
 429 /*
 430  * Write a single page, but leave the page locked.
 431  *
 432  * If we get a write error, set the page error bit, but still adjust the
 433  * dirty page accounting (i.e., page is no longer dirty).
 434  */
 435 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 436 {
 437         struct inode *inode;
 438         struct ceph_inode_info *ci;
 439         struct ceph_fs_client *fsc;
 440         struct ceph_osd_client *osdc;
 441         loff_t page_off = page_offset(page);
 442         int len = PAGE_CACHE_SIZE;
 443         loff_t i_size;
 444         int err = 0;
 445         struct ceph_snap_context *snapc, *oldest;
 446         u64 snap_size = 0;
 447         long writeback_stat;
 448
 449         dout("writepage %p idx %lu\n", page, page->index);
 450
 451         if (!page->mapping || !page->mapping->host) {
 452                 dout("writepage %p - no mapping\n", page);
 453                 return -EFAULT;
 454         }
 455         inode = page->mapping->host;
 456         ci = ceph_inode(inode);
 457         fsc = ceph_inode_to_client(inode);
 458         osdc = &fsc->client->osdc;
 459
 460         /* verify this is a writeable snap context */
 461         snapc = page_snap_context(page);
 462         if (snapc == NULL) {
 463                 dout("writepage %p page %p not dirty?\n", inode, page);
 464                 goto out;
 465         }
 466         oldest = get_oldest_context(inode, &snap_size);
 467         if (snapc->seq > oldest->seq) {
 468                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 469                      inode, page, snapc);
 470                 /* we should only noop if called by kswapd */
 471                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 472                 ceph_put_snap_context(oldest);
 473                 goto out;
 474         }
 475         ceph_put_snap_context(oldest);
 476
 477         /* is this a partial page at end of file? */
 478         if (snap_size)
 479                 i_size = snap_size;
 480         else
 481                 i_size = i_size_read(inode);
 482         if (i_size < page_off + len)
 483                 len = i_size - page_off;
 484
 485         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 486              inode, page, page->index, page_off, len, snapc);
 487
 488         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 489         if (writeback_stat >
 490             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 491                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 492
 493         set_page_writeback(page);
 494         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 495                                    &ci->i_layout, snapc,
 496                                    page_off, len,
 497                                    ci->i_truncate_seq, ci->i_truncate_size,
 498                                    &inode->i_mtime, &page, 1);
 499         if (err < 0) {
 500                 dout("writepage setting page/mapping error %d %p\n", err, page);
 501                 SetPageError(page);
 502                 mapping_set_error(&inode->i_data, err);
 503                 if (wbc)
 504                         wbc->pages_skipped++;
 505         } else {
 506                 dout("writepage cleaned page %p\n", page);
 507                 err = 0;  /* vfs expects us to return 0 */
 508         }
 509         page->private = 0;
 510         ClearPagePrivate(page);
 511         end_page_writeback(page);
 512         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 513         ceph_put_snap_context(snapc);  /* page's reference */
 514 out:
 515         return err;
 516 }
 517
 518 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 519 {
 520         int err;
 521         struct inode *inode = page->mapping->host;
 522         BUG_ON(!inode);
 523         ihold(inode);
 524         err = writepage_nounlock(page, wbc);
 525         unlock_page(page);
 526         iput(inode);
 527         return err;
 528 }
 529
 530
 531 /*
 532  * lame release_pages helper.  release_pages() isn't exported to
 533  * modules.
 534  */
 535 static void ceph_release_pages(struct page **pages, int num)
 536 {
 537         struct pagevec pvec;
 538         int i;
 539
 540         pagevec_init(&pvec, 0);
 541         for (i = 0; i < num; i++) {
 542                 if (pagevec_add(&pvec, pages[i]) == 0)
 543                         pagevec_release(&pvec);
 544         }
 545         pagevec_release(&pvec);
 546 }
 547
 548
 549 /*
 550  * async writeback completion handler.
 551  *
 552  * If we get an error, set the mapping error bit, but not the individual
 553  * page error bits.
 554  */
 555 static void writepages_finish(struct ceph_osd_request *req,
 556                               struct ceph_msg *msg)
 557 {
 558         struct inode *inode = req->r_inode;
 559         struct ceph_inode_info *ci = ceph_inode(inode);
 560         struct ceph_osd_data *osd_data;
 561         unsigned wrote;
 562         struct page *page;
 563         int num_pages;
 564         int i;
 565         struct ceph_snap_context *snapc = req->r_snapc;
 566         struct address_space *mapping = inode->i_mapping;
 567         int rc = req->r_result;
 568         u64 bytes = req->r_ops[0].extent.length;
 569         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 570         long writeback_stat;
 571         unsigned issued = ceph_caps_issued(ci);
 572
 573         osd_data = osd_req_op_extent_osd_data(req, 0);
 574         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 575         num_pages = calc_pages_for((u64)osd_data->alignment,
 576                                         (u64)osd_data->length);
 577         if (rc >= 0) {
 578                 /*
 579                  * Assume we wrote the pages we originally sent.  The
 580                  * osd might reply with fewer pages if our writeback
 581                  * raced with a truncation and was adjusted at the osd,
 582                  * so don't believe the reply.
 583                  */
 584                 wrote = num_pages;
 585         } else {
 586                 wrote = 0;
 587                 mapping_set_error(mapping, rc);
 588         }
 589         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 590              inode, rc, bytes, wrote);
 591
 592         /* clean all pages */
 593         for (i = 0; i < num_pages; i++) {
 594                 page = osd_data->pages[i];
 595                 BUG_ON(!page);
 596                 WARN_ON(!PageUptodate(page));
 597
 598                 writeback_stat =
 599                         atomic_long_dec_return(&fsc->writeback_count);
 600                 if (writeback_stat <
 601                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 602                         clear_bdi_congested(&fsc->backing_dev_info,
 603                                             BLK_RW_ASYNC);
 604
 605                 ceph_put_snap_context(page_snap_context(page));
 606                 page->private = 0;
 607                 ClearPagePrivate(page);
 608                 dout("unlocking %d %p\n", i, page);
 609                 end_page_writeback(page);
 610
 611                 /*
 612                  * We lost the cache cap, need to truncate the page before
 613                  * it is unlocked, otherwise we'd truncate it later in the
 614                  * page truncation thread, possibly losing some data that
 615                  * raced its way in
 616                  */
 617                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 618                         generic_error_remove_page(inode->i_mapping, page);
 619
 620                 unlock_page(page);
 621         }
 622         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 623         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 624
 625         ceph_release_pages(osd_data->pages, num_pages);
 626         if (osd_data->pages_from_pool)
 627                 mempool_free(osd_data->pages,
 628                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 629         else
 630                 kfree(osd_data->pages);
 631         ceph_osdc_put_request(req);
 632 }
 633
 634 static struct ceph_osd_request *
 635 ceph_writepages_osd_request(struct inode *inode, u64 offset, u64 *len,
 636                                 struct ceph_snap_context *snapc, int num_ops)
 637 {
 638         struct ceph_fs_client *fsc;
 639         struct ceph_inode_info *ci;
 640         struct ceph_vino vino;
 641
 642         fsc = ceph_inode_to_client(inode);
 643         ci = ceph_inode(inode);
 644         vino = ceph_vino(inode);
 645         /* BUG_ON(vino.snap != CEPH_NOSNAP); */
 646
 647         return ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 648                         vino, offset, len, num_ops, CEPH_OSD_OP_WRITE,
 649                         CEPH_OSD_FLAG_WRITE|CEPH_OSD_FLAG_ONDISK,
 650                         snapc, ci->i_truncate_seq, ci->i_truncate_size, true);
 651 }
 652
 653 /*
 654  * initiate async writeback
 655  */
 656 static int ceph_writepages_start(struct address_space *mapping,
 657                                  struct writeback_control *wbc)
 658 {
 659         struct inode *inode = mapping->host;
 660         struct ceph_inode_info *ci = ceph_inode(inode);
 661         struct ceph_fs_client *fsc;
 662         pgoff_t index, start, end;
 663         int range_whole = 0;
 664         int should_loop = 1;
 665         pgoff_t max_pages = 0, max_pages_ever = 0;
 666         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 667         struct pagevec pvec;
 668         int done = 0;
 669         int rc = 0;
 670         unsigned wsize = 1 << inode->i_blkbits;
 671         struct ceph_osd_request *req = NULL;
 672         int do_sync;
 673         u64 snap_size;
 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         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 681         if (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         fsc = ceph_inode_to_client(inode);
 689         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 690                 pr_warning("writepage_start %p on forced umount\n", inode);
 691                 return -EIO; /* we're in a forced umount, don't write! */
 692         }
 693         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 694                 wsize = fsc->mount_options->wsize;
 695         if (wsize < PAGE_CACHE_SIZE)
 696                 wsize = PAGE_CACHE_SIZE;
 697         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 698
 699         pagevec_init(&pvec, 0);
 700
 701         /* where to start/end? */
 702         if (wbc->range_cyclic) {
 703                 start = mapping->writeback_index; /* Start from prev offset */
 704                 end = -1;
 705                 dout(" cyclic, start at %lu\n", start);
 706         } else {
 707                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 708                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 709                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 710                         range_whole = 1;
 711                 should_loop = 0;
 712                 dout(" not cyclic, %lu to %lu\n", start, end);
 713         }
 714         index = start;
 715
 716 retry:
 717         /* find oldest snap context with dirty data */
 718         ceph_put_snap_context(snapc);
 719         snap_size = 0;
 720         snapc = get_oldest_context(inode, &snap_size);
 721         if (!snapc) {
 722                 /* hmm, why does writepages get called when there
 723                    is no dirty data? */
 724                 dout(" no snap context with dirty data?\n");
 725                 goto out;
 726         }
 727         if (snap_size == 0)
 728                 snap_size = i_size_read(inode);
 729         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 730              snapc, snapc->seq, snapc->num_snaps);
 731         if (last_snapc && snapc != last_snapc) {
 732                 /* if we switched to a newer snapc, restart our scan at the
 733                  * start of the original file range. */
 734                 dout("  snapc differs from last pass, restarting at %lu\n",
 735                      index);
 736                 index = start;
 737         }
 738         last_snapc = snapc;
 739
 740         while (!done && index <= end) {
 741                 int num_ops = do_sync ? 2 : 1;
 742                 struct ceph_vino vino;
 743                 unsigned i;
 744                 int first;
 745                 pgoff_t next;
 746                 int pvec_pages, locked_pages;
 747                 struct page **pages = NULL;
 748                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 749                 struct page *page;
 750                 int want;
 751                 u64 offset, len;
 752                 long writeback_stat;
 753
 754                 next = 0;
 755                 locked_pages = 0;
 756                 max_pages = max_pages_ever;
 757
 758 get_more_pages:
 759                 first = -1;
 760                 want = min(end - index,
 761                            min((pgoff_t)PAGEVEC_SIZE,
 762                                max_pages - (pgoff_t)locked_pages) - 1)
 763                         + 1;
 764                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 765                                                 PAGECACHE_TAG_DIRTY,
 766                                                 want);
 767                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 768                 if (!pvec_pages && !locked_pages)
 769                         break;
 770                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 771                         page = pvec.pages[i];
 772                         dout("? %p idx %lu\n", page, page->index);
 773                         if (locked_pages == 0)
 774                                 lock_page(page);  /* first page */
 775                         else if (!trylock_page(page))
 776                                 break;
 777
 778                         /* only dirty pages, or our accounting breaks */
 779                         if (unlikely(!PageDirty(page)) ||
 780                             unlikely(page->mapping != mapping)) {
 781                                 dout("!dirty or !mapping %p\n", page);
 782                                 unlock_page(page);
 783                                 break;
 784                         }
 785                         if (!wbc->range_cyclic && page->index > end) {
 786                                 dout("end of range %p\n", page);
 787                                 done = 1;
 788                                 unlock_page(page);
 789                                 break;
 790                         }
 791                         if (next && (page->index != next)) {
 792                                 dout("not consecutive %p\n", page);
 793                                 unlock_page(page);
 794                                 break;
 795                         }
 796                         if (wbc->sync_mode != WB_SYNC_NONE) {
 797                                 dout("waiting on writeback %p\n", page);
 798                                 wait_on_page_writeback(page);
 799                         }
 800                         if (page_offset(page) >= snap_size) {
 801                                 dout("%p page eof %llu\n", page, snap_size);
 802                                 done = 1;
 803                                 unlock_page(page);
 804                                 break;
 805                         }
 806                         if (PageWriteback(page)) {
 807                                 dout("%p under writeback\n", page);
 808                                 unlock_page(page);
 809                                 break;
 810                         }
 811
 812                         /* only if matching snap context */
 813                         pgsnapc = page_snap_context(page);
 814                         if (pgsnapc->seq > snapc->seq) {
 815                                 dout("page snapc %p %lld > oldest %p %lld\n",
 816                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 817                                 unlock_page(page);
 818                                 if (!locked_pages)
 819                                         continue; /* keep looking for snap */
 820                                 break;
 821                         }
 822
 823                         if (!clear_page_dirty_for_io(page)) {
 824                                 dout("%p !clear_page_dirty_for_io\n", page);
 825                                 unlock_page(page);
 826                                 break;
 827                         }
 828
 829                         /*
 830                          * We have something to write.  If this is
 831                          * the first locked page this time through,
 832                          * allocate an osd request and a page array
 833                          * that it will use.
 834                          */
 835                         if (locked_pages == 0) {
 836                                 size_t size;
 837
 838                                 BUG_ON(pages);
 839
 840                                 /* prepare async write request */
 841                                 offset = (u64)page_offset(page);
 842                                 len = wsize;
 843                                 req = ceph_writepages_osd_request(inode,
 844                                                         offset, &len, snapc,
 845                                                         num_ops);
 846
 847                                 if (IS_ERR(req)) {
 848                                         rc = PTR_ERR(req);
 849                                         unlock_page(page);
 850                                         break;
 851                                 }
 852
 853                                 req->r_callback = writepages_finish;
 854                                 req->r_inode = inode;
 855
 856                                 max_pages = calc_pages_for(0, (u64)len);
 857                                 size = max_pages * sizeof (*pages);
 858                                 pages = kmalloc(size, GFP_NOFS);
 859                                 if (!pages) {
 860                                         pool = fsc->wb_pagevec_pool;
 861                                         pages = mempool_alloc(pool, GFP_NOFS);
 862                                         BUG_ON(!pages);
 863                                 }
 864                         }
 865
 866                         /* note position of first page in pvec */
 867                         if (first < 0)
 868                                 first = i;
 869                         dout("%p will write page %p idx %lu\n",
 870                              inode, page, page->index);
 871
 872                         writeback_stat =
 873                                atomic_long_inc_return(&fsc->writeback_count);
 874                         if (writeback_stat > CONGESTION_ON_THRESH(
 875                                     fsc->mount_options->congestion_kb)) {
 876                                 set_bdi_congested(&fsc->backing_dev_info,
 877                                                   BLK_RW_ASYNC);
 878                         }
 879
 880                         set_page_writeback(page);
 881                         pages[locked_pages] = page;
 882                         locked_pages++;
 883                         next = page->index + 1;
 884                 }
 885
 886                 /* did we get anything? */
 887                 if (!locked_pages)
 888                         goto release_pvec_pages;
 889                 if (i) {
 890                         int j;
 891                         BUG_ON(!locked_pages || first < 0);
 892
 893                         if (pvec_pages && i == pvec_pages &&
 894                             locked_pages < max_pages) {
 895                                 dout("reached end pvec, trying for more\n");
 896                                 pagevec_reinit(&pvec);
 897                                 goto get_more_pages;
 898                         }
 899
 900                         /* shift unused pages over in the pvec...  we
 901                          * will need to release them below. */
 902                         for (j = i; j < pvec_pages; j++) {
 903                                 dout(" pvec leftover page %p\n",
 904                                      pvec.pages[j]);
 905                                 pvec.pages[j-i+first] = pvec.pages[j];
 906                         }
 907                         pvec.nr -= i-first;
 908                 }
 909
 910                 /* Format the osd request message and submit the write */
 911
 912                 offset = page_offset(pages[0]);
 913                 len = min(snap_size - offset,
 914                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 915                 dout("writepages got %d pages at %llu~%llu\n",
 916                      locked_pages, offset, len);
 917
 918                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 919                                                         !!pool, false);
 920
 921                 pages = NULL;   /* request message now owns the pages array */
 922                 pool = NULL;
 923
 924                 /* Update the write op length in case we changed it */
 925
 926                 osd_req_op_extent_update(req, 0, len);
 927
 928                 vino = ceph_vino(inode);
 929                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 930                                         &inode->i_mtime);
 931
 932                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 933                 BUG_ON(rc);
 934                 req = NULL;
 935
 936                 /* continue? */
 937                 index = next;
 938                 wbc->nr_to_write -= locked_pages;
 939                 if (wbc->nr_to_write <= 0)
 940                         done = 1;
 941
 942 release_pvec_pages:
 943                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 944                      pvec.nr ? pvec.pages[0] : NULL);
 945                 pagevec_release(&pvec);
 946
 947                 if (locked_pages && !done)
 948                         goto retry;
 949         }
 950
 951         if (should_loop && !done) {
 952                 /* more to do; loop back to beginning of file */
 953                 dout("writepages looping back to beginning of file\n");
 954                 should_loop = 0;
 955                 index = 0;
 956                 goto retry;
 957         }
 958
 959         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 960                 mapping->writeback_index = index;
 961
 962 out:
 963         if (req)
 964                 ceph_osdc_put_request(req);
 965         ceph_put_snap_context(snapc);
 966         dout("writepages done, rc = %d\n", rc);
 967         return rc;
 968 }
 969
 970
 971
 972 /*
 973  * See if a given @snapc is either writeable, or already written.
 974  */
 975 static int context_is_writeable_or_written(struct inode *inode,
 976                                            struct ceph_snap_context *snapc)
 977 {
 978         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 979         int ret = !oldest || snapc->seq <= oldest->seq;
 980
 981         ceph_put_snap_context(oldest);
 982         return ret;
 983 }
 984
 985 /*
 986  * We are only allowed to write into/dirty the page if the page is
 987  * clean, or already dirty within the same snap context.
 988  *
 989  * called with page locked.
 990  * return success with page locked,
 991  * or any failure (incl -EAGAIN) with page unlocked.
 992  */
 993 static int ceph_update_writeable_page(struct file *file,
 994                             loff_t pos, unsigned len,
 995                             struct page *page)
 996 {
 997         struct inode *inode = file_inode(file);
 998         struct ceph_inode_info *ci = ceph_inode(inode);
 999         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1000         loff_t page_off = pos & PAGE_CACHE_MASK;
1001         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1002         int end_in_page = pos_in_page + len;
1003         loff_t i_size;
1004         int r;
1005         struct ceph_snap_context *snapc, *oldest;
1006
1007 retry_locked:
1008         /* writepages currently holds page lock, but if we change that later, */
1009         wait_on_page_writeback(page);
1010
1011         /* check snap context */
1012         BUG_ON(!ci->i_snap_realm);
1013         down_read(&mdsc->snap_rwsem);
1014         BUG_ON(!ci->i_snap_realm->cached_context);
1015         snapc = page_snap_context(page);
1016         if (snapc && snapc != ci->i_head_snapc) {
1017                 /*
1018                  * this page is already dirty in another (older) snap
1019                  * context!  is it writeable now?
1020                  */
1021                 oldest = get_oldest_context(inode, NULL);
1022                 up_read(&mdsc->snap_rwsem);
1023
1024                 if (snapc->seq > oldest->seq) {
1025                         ceph_put_snap_context(oldest);
1026                         dout(" page %p snapc %p not current or oldest\n",
1027                              page, snapc);
1028                         /*
1029                          * queue for writeback, and wait for snapc to
1030                          * be writeable or written
1031                          */
1032                         snapc = ceph_get_snap_context(snapc);
1033                         unlock_page(page);
1034                         ceph_queue_writeback(inode);
1035                         r = wait_event_interruptible(ci->i_cap_wq,
1036                                context_is_writeable_or_written(inode, snapc));
1037                         ceph_put_snap_context(snapc);
1038                         if (r == -ERESTARTSYS)
1039                                 return r;
1040                         return -EAGAIN;
1041                 }
1042                 ceph_put_snap_context(oldest);
1043
1044                 /* yay, writeable, do it now (without dropping page lock) */
1045                 dout(" page %p snapc %p not current, but oldest\n",
1046                      page, snapc);
1047                 if (!clear_page_dirty_for_io(page))
1048                         goto retry_locked;
1049                 r = writepage_nounlock(page, NULL);
1050                 if (r < 0)
1051                         goto fail_nosnap;
1052                 goto retry_locked;
1053         }
1054
1055         if (PageUptodate(page)) {
1056                 dout(" page %p already uptodate\n", page);
1057                 return 0;
1058         }
1059
1060         /* full page? */
1061         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1062                 return 0;
1063
1064         /* past end of file? */
1065         i_size = inode->i_size;   /* caller holds i_mutex */
1066
1067         if (i_size + len > inode->i_sb->s_maxbytes) {
1068                 /* file is too big */
1069                 r = -EINVAL;
1070                 goto fail;
1071         }
1072
1073         if (page_off >= i_size ||
1074             (pos_in_page == 0 && (pos+len) >= i_size &&
1075              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1076                 dout(" zeroing %p 0 - %d and %d - %d\n",
1077                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1078                 zero_user_segments(page,
1079                                    0, pos_in_page,
1080                                    end_in_page, PAGE_CACHE_SIZE);
1081                 return 0;
1082         }
1083
1084         /* we need to read it. */
1085         up_read(&mdsc->snap_rwsem);
1086         r = readpage_nounlock(file, page);
1087         if (r < 0)
1088                 goto fail_nosnap;
1089         goto retry_locked;
1090
1091 fail:
1092         up_read(&mdsc->snap_rwsem);
1093 fail_nosnap:
1094         unlock_page(page);
1095         return r;
1096 }
1097
1098 /*
1099  * We are only allowed to write into/dirty the page if the page is
1100  * clean, or already dirty within the same snap context.
1101  */
1102 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1103                             loff_t pos, unsigned len, unsigned flags,
1104                             struct page **pagep, void **fsdata)
1105 {
1106         struct inode *inode = file_inode(file);
1107         struct page *page;
1108         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1109         int r;
1110
1111         do {
1112                 /* get a page */
1113                 page = grab_cache_page_write_begin(mapping, index, 0);
1114                 if (!page)
1115                         return -ENOMEM;
1116                 *pagep = page;
1117
1118                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1119                      inode, page, (int)pos, (int)len);
1120
1121                 r = ceph_update_writeable_page(file, pos, len, page);
1122         } while (r == -EAGAIN);
1123
1124         return r;
1125 }
1126
1127 /*
1128  * we don't do anything in here that simple_write_end doesn't do
1129  * except adjust dirty page accounting and drop read lock on
1130  * mdsc->snap_rwsem.
1131  */
1132 static int ceph_write_end(struct file *file, struct address_space *mapping,
1133                           loff_t pos, unsigned len, unsigned copied,
1134                           struct page *page, void *fsdata)
1135 {
1136         struct inode *inode = file_inode(file);
1137         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1138         struct ceph_mds_client *mdsc = fsc->mdsc;
1139         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1140         int check_cap = 0;
1141
1142         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1143              inode, page, (int)pos, (int)copied, (int)len);
1144
1145         /* zero the stale part of the page if we did a short copy */
1146         if (copied < len)
1147                 zero_user_segment(page, from+copied, len);
1148
1149         /* did file size increase? */
1150         /* (no need for i_size_read(); we caller holds i_mutex */
1151         if (pos+copied > inode->i_size)
1152                 check_cap = ceph_inode_set_size(inode, pos+copied);
1153
1154         if (!PageUptodate(page))
1155                 SetPageUptodate(page);
1156
1157         set_page_dirty(page);
1158
1159         unlock_page(page);
1160         up_read(&mdsc->snap_rwsem);
1161         page_cache_release(page);
1162
1163         if (check_cap)
1164                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1165
1166         return copied;
1167 }
1168
1169 /*
1170  * we set .direct_IO to indicate direct io is supported, but since we
1171  * intercept O_DIRECT reads and writes early, this function should
1172  * never get called.
1173  */
1174 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1175                               const struct iovec *iov,
1176                               loff_t pos, unsigned long nr_segs)
1177 {
1178         WARN_ON(1);
1179         return -EINVAL;
1180 }
1181
1182 const struct address_space_operations ceph_aops = {
1183         .readpage = ceph_readpage,
1184         .readpages = ceph_readpages,
1185         .writepage = ceph_writepage,
1186         .writepages = ceph_writepages_start,
1187         .write_begin = ceph_write_begin,
1188         .write_end = ceph_write_end,
1189         .set_page_dirty = ceph_set_page_dirty,
1190         .invalidatepage = ceph_invalidatepage,
1191         .releasepage = ceph_releasepage,
1192         .direct_IO = ceph_direct_io,
1193 };
1194
1195
1196 /*
1197  * vm ops
1198  */
1199
1200 /*
1201  * Reuse write_begin here for simplicity.
1202  */
1203 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1204 {
1205         struct inode *inode = file_inode(vma->vm_file);
1206         struct page *page = vmf->page;
1207         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1208         loff_t off = page_offset(page);
1209         loff_t size, len;
1210         int ret;
1211
1212         /* Update time before taking page lock */
1213         file_update_time(vma->vm_file);
1214
1215         size = i_size_read(inode);
1216         if (off + PAGE_CACHE_SIZE <= size)
1217                 len = PAGE_CACHE_SIZE;
1218         else
1219                 len = size & ~PAGE_CACHE_MASK;
1220
1221         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1222              off, len, page, page->index);
1223
1224         lock_page(page);
1225
1226         ret = VM_FAULT_NOPAGE;
1227         if ((off > size) ||
1228             (page->mapping != inode->i_mapping))
1229                 goto out;
1230
1231         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1232         if (ret == 0) {
1233                 /* success.  we'll keep the page locked. */
1234                 set_page_dirty(page);
1235                 up_read(&mdsc->snap_rwsem);
1236                 ret = VM_FAULT_LOCKED;
1237         } else {
1238                 if (ret == -ENOMEM)
1239                         ret = VM_FAULT_OOM;
1240                 else
1241                         ret = VM_FAULT_SIGBUS;
1242         }
1243 out:
1244         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1245         if (ret != VM_FAULT_LOCKED)
1246                 unlock_page(page);
1247         return ret;
1248 }
1249
1250 static struct vm_operations_struct ceph_vmops = {
1251         .fault          = filemap_fault,
1252         .page_mkwrite   = ceph_page_mkwrite,
1253         .remap_pages    = generic_file_remap_pages,
1254 };
1255
1256 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1257 {
1258         struct address_space *mapping = file->f_mapping;
1259
1260         if (!mapping->a_ops->readpage)
1261                 return -ENOEXEC;
1262         file_accessed(file);
1263         vma->vm_ops = &ceph_vmops;
1264         return 0;
1265 }