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 = &req->r_data_in;
 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         struct ceph_osd_req_op op;
 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, &op, 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                         page_cache_release(page);
 340                         dout("start_read %p add_to_page_cache failed %p\n",
 341                              inode, page);
 342                         nr_pages = i;
 343                         goto out_pages;
 344                 }
 345                 pages[i] = page;
 346         }
 347         ceph_osd_data_pages_init(&req->r_data_in, pages, len, 0,
 348                                         false, false);
 349         req->r_callback = finish_read;
 350         req->r_inode = inode;
 351
 352         ceph_osdc_build_request(req, off, 1, &op, 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         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 383                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 384                         >> PAGE_SHIFT;
 385
 386         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 387                 file, nr_pages,
 388              max);
 389         while (!list_empty(page_list)) {
 390                 rc = start_read(inode, page_list, max);
 391                 if (rc < 0)
 392                         goto out;
 393                 BUG_ON(rc == 0);
 394         }
 395 out:
 396         dout("readpages %p file %p ret %d\n", inode, file, rc);
 397         return rc;
 398 }
 399
 400 /*
 401  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 402  * only snap context we are allowed to write back.
 403  */
 404 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 405                                                     u64 *snap_size)
 406 {
 407         struct ceph_inode_info *ci = ceph_inode(inode);
 408         struct ceph_snap_context *snapc = NULL;
 409         struct ceph_cap_snap *capsnap = NULL;
 410
 411         spin_lock(&ci->i_ceph_lock);
 412         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 413                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 414                      capsnap->context, capsnap->dirty_pages);
 415                 if (capsnap->dirty_pages) {
 416                         snapc = ceph_get_snap_context(capsnap->context);
 417                         if (snap_size)
 418                                 *snap_size = capsnap->size;
 419                         break;
 420                 }
 421         }
 422         if (!snapc && ci->i_wrbuffer_ref_head) {
 423                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 424                 dout(" head snapc %p has %d dirty pages\n",
 425                      snapc, ci->i_wrbuffer_ref_head);
 426         }
 427         spin_unlock(&ci->i_ceph_lock);
 428         return snapc;
 429 }
 430
 431 /*
 432  * Write a single page, but leave the page locked.
 433  *
 434  * If we get a write error, set the page error bit, but still adjust the
 435  * dirty page accounting (i.e., page is no longer dirty).
 436  */
 437 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 438 {
 439         struct inode *inode;
 440         struct ceph_inode_info *ci;
 441         struct ceph_fs_client *fsc;
 442         struct ceph_osd_client *osdc;
 443         loff_t page_off = page_offset(page);
 444         int len = PAGE_CACHE_SIZE;
 445         loff_t i_size;
 446         int err = 0;
 447         struct ceph_snap_context *snapc, *oldest;
 448         u64 snap_size = 0;
 449         long writeback_stat;
 450
 451         dout("writepage %p idx %lu\n", page, page->index);
 452
 453         if (!page->mapping || !page->mapping->host) {
 454                 dout("writepage %p - no mapping\n", page);
 455                 return -EFAULT;
 456         }
 457         inode = page->mapping->host;
 458         ci = ceph_inode(inode);
 459         fsc = ceph_inode_to_client(inode);
 460         osdc = &fsc->client->osdc;
 461
 462         /* verify this is a writeable snap context */
 463         snapc = page_snap_context(page);
 464         if (snapc == NULL) {
 465                 dout("writepage %p page %p not dirty?\n", inode, page);
 466                 goto out;
 467         }
 468         oldest = get_oldest_context(inode, &snap_size);
 469         if (snapc->seq > oldest->seq) {
 470                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 471                      inode, page, snapc);
 472                 /* we should only noop if called by kswapd */
 473                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 474                 ceph_put_snap_context(oldest);
 475                 goto out;
 476         }
 477         ceph_put_snap_context(oldest);
 478
 479         /* is this a partial page at end of file? */
 480         if (snap_size)
 481                 i_size = snap_size;
 482         else
 483                 i_size = i_size_read(inode);
 484         if (i_size < page_off + len)
 485                 len = i_size - page_off;
 486
 487         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 488              inode, page, page->index, page_off, len, snapc);
 489
 490         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 491         if (writeback_stat >
 492             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 493                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 494
 495         set_page_writeback(page);
 496         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 497                                    &ci->i_layout, snapc,
 498                                    page_off, len,
 499                                    ci->i_truncate_seq, ci->i_truncate_size,
 500                                    &inode->i_mtime, &page, 1);
 501         if (err < 0) {
 502                 dout("writepage setting page/mapping error %d %p\n", err, page);
 503                 SetPageError(page);
 504                 mapping_set_error(&inode->i_data, err);
 505                 if (wbc)
 506                         wbc->pages_skipped++;
 507         } else {
 508                 dout("writepage cleaned page %p\n", page);
 509                 err = 0;  /* vfs expects us to return 0 */
 510         }
 511         page->private = 0;
 512         ClearPagePrivate(page);
 513         end_page_writeback(page);
 514         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 515         ceph_put_snap_context(snapc);  /* page's reference */
 516 out:
 517         return err;
 518 }
 519
 520 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 521 {
 522         int err;
 523         struct inode *inode = page->mapping->host;
 524         BUG_ON(!inode);
 525         ihold(inode);
 526         err = writepage_nounlock(page, wbc);
 527         unlock_page(page);
 528         iput(inode);
 529         return err;
 530 }
 531
 532
 533 /*
 534  * lame release_pages helper.  release_pages() isn't exported to
 535  * modules.
 536  */
 537 static void ceph_release_pages(struct page **pages, int num)
 538 {
 539         struct pagevec pvec;
 540         int i;
 541
 542         pagevec_init(&pvec, 0);
 543         for (i = 0; i < num; i++) {
 544                 if (pagevec_add(&pvec, pages[i]) == 0)
 545                         pagevec_release(&pvec);
 546         }
 547         pagevec_release(&pvec);
 548 }
 549
 550
 551 /*
 552  * async writeback completion handler.
 553  *
 554  * If we get an error, set the mapping error bit, but not the individual
 555  * page error bits.
 556  */
 557 static void writepages_finish(struct ceph_osd_request *req,
 558                               struct ceph_msg *msg)
 559 {
 560         struct inode *inode = req->r_inode;
 561         struct ceph_inode_info *ci = ceph_inode(inode);
 562         struct ceph_osd_data *osd_data;
 563         unsigned wrote;
 564         struct page *page;
 565         int num_pages;
 566         int i;
 567         struct ceph_snap_context *snapc = req->r_snapc;
 568         struct address_space *mapping = inode->i_mapping;
 569         int rc = req->r_result;
 570         u64 bytes = le64_to_cpu(req->r_request_ops[0].extent.length);
 571         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 572         long writeback_stat;
 573         unsigned issued = ceph_caps_issued(ci);
 574
 575         osd_data = &req->r_data_out;
 576         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 577         num_pages = calc_pages_for((u64)osd_data->alignment,
 578                                         (u64)osd_data->length);
 579         if (rc >= 0) {
 580                 /*
 581                  * Assume we wrote the pages we originally sent.  The
 582                  * osd might reply with fewer pages if our writeback
 583                  * raced with a truncation and was adjusted at the osd,
 584                  * so don't believe the reply.
 585                  */
 586                 wrote = num_pages;
 587         } else {
 588                 wrote = 0;
 589                 mapping_set_error(mapping, rc);
 590         }
 591         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 592              inode, rc, bytes, wrote);
 593
 594         /* clean all pages */
 595         for (i = 0; i < num_pages; i++) {
 596                 page = osd_data->pages[i];
 597                 BUG_ON(!page);
 598                 WARN_ON(!PageUptodate(page));
 599
 600                 writeback_stat =
 601                         atomic_long_dec_return(&fsc->writeback_count);
 602                 if (writeback_stat <
 603                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 604                         clear_bdi_congested(&fsc->backing_dev_info,
 605                                             BLK_RW_ASYNC);
 606
 607                 ceph_put_snap_context(page_snap_context(page));
 608                 page->private = 0;
 609                 ClearPagePrivate(page);
 610                 dout("unlocking %d %p\n", i, page);
 611                 end_page_writeback(page);
 612
 613                 /*
 614                  * We lost the cache cap, need to truncate the page before
 615                  * it is unlocked, otherwise we'd truncate it later in the
 616                  * page truncation thread, possibly losing some data that
 617                  * raced its way in
 618                  */
 619                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 620                         generic_error_remove_page(inode->i_mapping, page);
 621
 622                 unlock_page(page);
 623         }
 624         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 625         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 626
 627         ceph_release_pages(osd_data->pages, num_pages);
 628         if (osd_data->pages_from_pool)
 629                 mempool_free(osd_data->pages,
 630                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 631         else
 632                 kfree(osd_data->pages);
 633         ceph_osdc_put_request(req);
 634 }
 635
 636 static struct ceph_osd_request *
 637 ceph_writepages_osd_request(struct inode *inode, u64 offset, u64 *len,
 638                                 struct ceph_snap_context *snapc,
 639                                 int num_ops, struct ceph_osd_req_op *ops)
 640 {
 641         struct ceph_fs_client *fsc;
 642         struct ceph_inode_info *ci;
 643         struct ceph_vino vino;
 644
 645         fsc = ceph_inode_to_client(inode);
 646         ci = ceph_inode(inode);
 647         vino = ceph_vino(inode);
 648         /* BUG_ON(vino.snap != CEPH_NOSNAP); */
 649
 650         return ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 651                         vino, offset, len, num_ops, ops, CEPH_OSD_OP_WRITE,
 652                         CEPH_OSD_FLAG_WRITE|CEPH_OSD_FLAG_ONDISK,
 653                         snapc, ci->i_truncate_seq, ci->i_truncate_size, true);
 654 }
 655
 656 /*
 657  * initiate async writeback
 658  */
 659 static int ceph_writepages_start(struct address_space *mapping,
 660                                  struct writeback_control *wbc)
 661 {
 662         struct inode *inode = mapping->host;
 663         struct ceph_inode_info *ci = ceph_inode(inode);
 664         struct ceph_fs_client *fsc;
 665         pgoff_t index, start, end;
 666         int range_whole = 0;
 667         int should_loop = 1;
 668         pgoff_t max_pages = 0, max_pages_ever = 0;
 669         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 670         struct pagevec pvec;
 671         int done = 0;
 672         int rc = 0;
 673         unsigned wsize = 1 << inode->i_blkbits;
 674         struct ceph_osd_request *req = NULL;
 675         int do_sync;
 676         u64 snap_size = 0;
 677
 678         /*
 679          * Include a 'sync' in the OSD request if this is a data
 680          * integrity write (e.g., O_SYNC write or fsync()), or if our
 681          * cap is being revoked.
 682          */
 683         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 684         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 685                 do_sync = 1;
 686         dout("writepages_start %p dosync=%d (mode=%s)\n",
 687              inode, do_sync,
 688              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 689              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 690
 691         fsc = ceph_inode_to_client(inode);
 692         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 693                 pr_warning("writepage_start %p on forced umount\n", inode);
 694                 return -EIO; /* we're in a forced umount, don't write! */
 695         }
 696         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 697                 wsize = fsc->mount_options->wsize;
 698         if (wsize < PAGE_CACHE_SIZE)
 699                 wsize = PAGE_CACHE_SIZE;
 700         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 701
 702         pagevec_init(&pvec, 0);
 703
 704         /* where to start/end? */
 705         if (wbc->range_cyclic) {
 706                 start = mapping->writeback_index; /* Start from prev offset */
 707                 end = -1;
 708                 dout(" cyclic, start at %lu\n", start);
 709         } else {
 710                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 711                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 712                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 713                         range_whole = 1;
 714                 should_loop = 0;
 715                 dout(" not cyclic, %lu to %lu\n", start, end);
 716         }
 717         index = start;
 718
 719 retry:
 720         /* find oldest snap context with dirty data */
 721         ceph_put_snap_context(snapc);
 722         snapc = get_oldest_context(inode, &snap_size);
 723         if (!snapc) {
 724                 /* hmm, why does writepages get called when there
 725                    is no dirty data? */
 726                 dout(" no snap context with dirty data?\n");
 727                 goto out;
 728         }
 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                 struct ceph_osd_req_op ops[2];
 742                 int num_ops = do_sync ? 2 : 1;
 743                 struct ceph_vino vino;
 744                 unsigned i;
 745                 int first;
 746                 pgoff_t next;
 747                 int pvec_pages, locked_pages;
 748                 struct page **pages = NULL;
 749                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 750                 struct page *page;
 751                 int want;
 752                 u64 offset, len;
 753                 long writeback_stat;
 754
 755                 next = 0;
 756                 locked_pages = 0;
 757                 max_pages = max_pages_ever;
 758
 759 get_more_pages:
 760                 first = -1;
 761                 want = min(end - index,
 762                            min((pgoff_t)PAGEVEC_SIZE,
 763                                max_pages - (pgoff_t)locked_pages) - 1)
 764                         + 1;
 765                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 766                                                 PAGECACHE_TAG_DIRTY,
 767                                                 want);
 768                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 769                 if (!pvec_pages && !locked_pages)
 770                         break;
 771                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 772                         page = pvec.pages[i];
 773                         dout("? %p idx %lu\n", page, page->index);
 774                         if (locked_pages == 0)
 775                                 lock_page(page);  /* first page */
 776                         else if (!trylock_page(page))
 777                                 break;
 778
 779                         /* only dirty pages, or our accounting breaks */
 780                         if (unlikely(!PageDirty(page)) ||
 781                             unlikely(page->mapping != mapping)) {
 782                                 dout("!dirty or !mapping %p\n", page);
 783                                 unlock_page(page);
 784                                 break;
 785                         }
 786                         if (!wbc->range_cyclic && page->index > end) {
 787                                 dout("end of range %p\n", page);
 788                                 done = 1;
 789                                 unlock_page(page);
 790                                 break;
 791                         }
 792                         if (next && (page->index != next)) {
 793                                 dout("not consecutive %p\n", page);
 794                                 unlock_page(page);
 795                                 break;
 796                         }
 797                         if (wbc->sync_mode != WB_SYNC_NONE) {
 798                                 dout("waiting on writeback %p\n", page);
 799                                 wait_on_page_writeback(page);
 800                         }
 801                         if ((snap_size && page_offset(page) > snap_size) ||
 802                             (!snap_size &&
 803                              page_offset(page) > i_size_read(inode))) {
 804                                 dout("%p page eof %llu\n", page, snap_size ?
 805                                      snap_size : i_size_read(inode));
 806                                 done = 1;
 807                                 unlock_page(page);
 808                                 break;
 809                         }
 810                         if (PageWriteback(page)) {
 811                                 dout("%p under writeback\n", page);
 812                                 unlock_page(page);
 813                                 break;
 814                         }
 815
 816                         /* only if matching snap context */
 817                         pgsnapc = page_snap_context(page);
 818                         if (pgsnapc->seq > snapc->seq) {
 819                                 dout("page snapc %p %lld > oldest %p %lld\n",
 820                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 821                                 unlock_page(page);
 822                                 if (!locked_pages)
 823                                         continue; /* keep looking for snap */
 824                                 break;
 825                         }
 826
 827                         if (!clear_page_dirty_for_io(page)) {
 828                                 dout("%p !clear_page_dirty_for_io\n", page);
 829                                 unlock_page(page);
 830                                 break;
 831                         }
 832
 833                         /*
 834                          * We have something to write.  If this is
 835                          * the first locked page this time through,
 836                          * allocate an osd request and a page array
 837                          * that it will use.
 838                          */
 839                         if (locked_pages == 0) {
 840                                 size_t size;
 841
 842                                 BUG_ON(pages);
 843
 844                                 /* prepare async write request */
 845                                 offset = (u64)page_offset(page);
 846                                 len = wsize;
 847                                 req = ceph_writepages_osd_request(inode,
 848                                                         offset, &len, snapc,
 849                                                         num_ops, ops);
 850
 851                                 if (IS_ERR(req)) {
 852                                         rc = PTR_ERR(req);
 853                                         unlock_page(page);
 854                                         break;
 855                                 }
 856
 857                                 req->r_callback = writepages_finish;
 858                                 req->r_inode = inode;
 859
 860                                 max_pages = calc_pages_for(0, (u64)len);
 861                                 size = max_pages * sizeof (*pages);
 862                                 pages = kmalloc(size, GFP_NOFS);
 863                                 if (!pages) {
 864                                         pool = fsc->wb_pagevec_pool;
 865                                         pages = mempool_alloc(pool, GFP_NOFS);
 866                                         BUG_ON(!pages);
 867                                 }
 868                         }
 869
 870                         /* note position of first page in pvec */
 871                         if (first < 0)
 872                                 first = i;
 873                         dout("%p will write page %p idx %lu\n",
 874                              inode, page, page->index);
 875
 876                         writeback_stat =
 877                                atomic_long_inc_return(&fsc->writeback_count);
 878                         if (writeback_stat > CONGESTION_ON_THRESH(
 879                                     fsc->mount_options->congestion_kb)) {
 880                                 set_bdi_congested(&fsc->backing_dev_info,
 881                                                   BLK_RW_ASYNC);
 882                         }
 883
 884                         set_page_writeback(page);
 885                         pages[locked_pages] = page;
 886                         locked_pages++;
 887                         next = page->index + 1;
 888                 }
 889
 890                 /* did we get anything? */
 891                 if (!locked_pages)
 892                         goto release_pvec_pages;
 893                 if (i) {
 894                         int j;
 895                         BUG_ON(!locked_pages || first < 0);
 896
 897                         if (pvec_pages && i == pvec_pages &&
 898                             locked_pages < max_pages) {
 899                                 dout("reached end pvec, trying for more\n");
 900                                 pagevec_reinit(&pvec);
 901                                 goto get_more_pages;
 902                         }
 903
 904                         /* shift unused pages over in the pvec...  we
 905                          * will need to release them below. */
 906                         for (j = i; j < pvec_pages; j++) {
 907                                 dout(" pvec leftover page %p\n",
 908                                      pvec.pages[j]);
 909                                 pvec.pages[j-i+first] = pvec.pages[j];
 910                         }
 911                         pvec.nr -= i-first;
 912                 }
 913
 914                 /* Format the osd request message and submit the write */
 915
 916                 offset = page_offset(pages[0]);
 917                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 918                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 919                 dout("writepages got %d pages at %llu~%llu\n",
 920                      locked_pages, offset, len);
 921
 922                 ceph_osd_data_pages_init(&req->r_data_out, pages, len, 0,
 923                                                 !!pool, false);
 924
 925                 pages = NULL;   /* request message now owns the pages array */
 926                 pool = NULL;
 927
 928                 /* Update the write op length in case we changed it */
 929
 930                 osd_req_op_extent_update(&ops[0], len);
 931
 932                 vino = ceph_vino(inode);
 933                 ceph_osdc_build_request(req, offset, num_ops, ops,
 934                                         snapc, vino.snap, &inode->i_mtime);
 935
 936                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 937                 BUG_ON(rc);
 938                 req = NULL;
 939
 940                 /* continue? */
 941                 index = next;
 942                 wbc->nr_to_write -= locked_pages;
 943                 if (wbc->nr_to_write <= 0)
 944                         done = 1;
 945
 946 release_pvec_pages:
 947                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 948                      pvec.nr ? pvec.pages[0] : NULL);
 949                 pagevec_release(&pvec);
 950
 951                 if (locked_pages && !done)
 952                         goto retry;
 953         }
 954
 955         if (should_loop && !done) {
 956                 /* more to do; loop back to beginning of file */
 957                 dout("writepages looping back to beginning of file\n");
 958                 should_loop = 0;
 959                 index = 0;
 960                 goto retry;
 961         }
 962
 963         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 964                 mapping->writeback_index = index;
 965
 966 out:
 967         if (req)
 968                 ceph_osdc_put_request(req);
 969         ceph_put_snap_context(snapc);
 970         dout("writepages done, rc = %d\n", rc);
 971         return rc;
 972 }
 973
 974
 975
 976 /*
 977  * See if a given @snapc is either writeable, or already written.
 978  */
 979 static int context_is_writeable_or_written(struct inode *inode,
 980                                            struct ceph_snap_context *snapc)
 981 {
 982         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 983         int ret = !oldest || snapc->seq <= oldest->seq;
 984
 985         ceph_put_snap_context(oldest);
 986         return ret;
 987 }
 988
 989 /*
 990  * We are only allowed to write into/dirty the page if the page is
 991  * clean, or already dirty within the same snap context.
 992  *
 993  * called with page locked.
 994  * return success with page locked,
 995  * or any failure (incl -EAGAIN) with page unlocked.
 996  */
 997 static int ceph_update_writeable_page(struct file *file,
 998                             loff_t pos, unsigned len,
 999                             struct page *page)
1000 {
1001         struct inode *inode = file_inode(file);
1002         struct ceph_inode_info *ci = ceph_inode(inode);
1003         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1004         loff_t page_off = pos & PAGE_CACHE_MASK;
1005         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1006         int end_in_page = pos_in_page + len;
1007         loff_t i_size;
1008         int r;
1009         struct ceph_snap_context *snapc, *oldest;
1010
1011 retry_locked:
1012         /* writepages currently holds page lock, but if we change that later, */
1013         wait_on_page_writeback(page);
1014
1015         /* check snap context */
1016         BUG_ON(!ci->i_snap_realm);
1017         down_read(&mdsc->snap_rwsem);
1018         BUG_ON(!ci->i_snap_realm->cached_context);
1019         snapc = page_snap_context(page);
1020         if (snapc && snapc != ci->i_head_snapc) {
1021                 /*
1022                  * this page is already dirty in another (older) snap
1023                  * context!  is it writeable now?
1024                  */
1025                 oldest = get_oldest_context(inode, NULL);
1026                 up_read(&mdsc->snap_rwsem);
1027
1028                 if (snapc->seq > oldest->seq) {
1029                         ceph_put_snap_context(oldest);
1030                         dout(" page %p snapc %p not current or oldest\n",
1031                              page, snapc);
1032                         /*
1033                          * queue for writeback, and wait for snapc to
1034                          * be writeable or written
1035                          */
1036                         snapc = ceph_get_snap_context(snapc);
1037                         unlock_page(page);
1038                         ceph_queue_writeback(inode);
1039                         r = wait_event_interruptible(ci->i_cap_wq,
1040                                context_is_writeable_or_written(inode, snapc));
1041                         ceph_put_snap_context(snapc);
1042                         if (r == -ERESTARTSYS)
1043                                 return r;
1044                         return -EAGAIN;
1045                 }
1046                 ceph_put_snap_context(oldest);
1047
1048                 /* yay, writeable, do it now (without dropping page lock) */
1049                 dout(" page %p snapc %p not current, but oldest\n",
1050                      page, snapc);
1051                 if (!clear_page_dirty_for_io(page))
1052                         goto retry_locked;
1053                 r = writepage_nounlock(page, NULL);
1054                 if (r < 0)
1055                         goto fail_nosnap;
1056                 goto retry_locked;
1057         }
1058
1059         if (PageUptodate(page)) {
1060                 dout(" page %p already uptodate\n", page);
1061                 return 0;
1062         }
1063
1064         /* full page? */
1065         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1066                 return 0;
1067
1068         /* past end of file? */
1069         i_size = inode->i_size;   /* caller holds i_mutex */
1070
1071         if (i_size + len > inode->i_sb->s_maxbytes) {
1072                 /* file is too big */
1073                 r = -EINVAL;
1074                 goto fail;
1075         }
1076
1077         if (page_off >= i_size ||
1078             (pos_in_page == 0 && (pos+len) >= i_size &&
1079              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1080                 dout(" zeroing %p 0 - %d and %d - %d\n",
1081                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1082                 zero_user_segments(page,
1083                                    0, pos_in_page,
1084                                    end_in_page, PAGE_CACHE_SIZE);
1085                 return 0;
1086         }
1087
1088         /* we need to read it. */
1089         up_read(&mdsc->snap_rwsem);
1090         r = readpage_nounlock(file, page);
1091         if (r < 0)
1092                 goto fail_nosnap;
1093         goto retry_locked;
1094
1095 fail:
1096         up_read(&mdsc->snap_rwsem);
1097 fail_nosnap:
1098         unlock_page(page);
1099         return r;
1100 }
1101
1102 /*
1103  * We are only allowed to write into/dirty the page if the page is
1104  * clean, or already dirty within the same snap context.
1105  */
1106 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1107                             loff_t pos, unsigned len, unsigned flags,
1108                             struct page **pagep, void **fsdata)
1109 {
1110         struct inode *inode = file_inode(file);
1111         struct page *page;
1112         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1113         int r;
1114
1115         do {
1116                 /* get a page */
1117                 page = grab_cache_page_write_begin(mapping, index, 0);
1118                 if (!page)
1119                         return -ENOMEM;
1120                 *pagep = page;
1121
1122                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1123                      inode, page, (int)pos, (int)len);
1124
1125                 r = ceph_update_writeable_page(file, pos, len, page);
1126         } while (r == -EAGAIN);
1127
1128         return r;
1129 }
1130
1131 /*
1132  * we don't do anything in here that simple_write_end doesn't do
1133  * except adjust dirty page accounting and drop read lock on
1134  * mdsc->snap_rwsem.
1135  */
1136 static int ceph_write_end(struct file *file, struct address_space *mapping,
1137                           loff_t pos, unsigned len, unsigned copied,
1138                           struct page *page, void *fsdata)
1139 {
1140         struct inode *inode = file_inode(file);
1141         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1142         struct ceph_mds_client *mdsc = fsc->mdsc;
1143         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1144         int check_cap = 0;
1145
1146         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1147              inode, page, (int)pos, (int)copied, (int)len);
1148
1149         /* zero the stale part of the page if we did a short copy */
1150         if (copied < len)
1151                 zero_user_segment(page, from+copied, len);
1152
1153         /* did file size increase? */
1154         /* (no need for i_size_read(); we caller holds i_mutex */
1155         if (pos+copied > inode->i_size)
1156                 check_cap = ceph_inode_set_size(inode, pos+copied);
1157
1158         if (!PageUptodate(page))
1159                 SetPageUptodate(page);
1160
1161         set_page_dirty(page);
1162
1163         unlock_page(page);
1164         up_read(&mdsc->snap_rwsem);
1165         page_cache_release(page);
1166
1167         if (check_cap)
1168                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1169
1170         return copied;
1171 }
1172
1173 /*
1174  * we set .direct_IO to indicate direct io is supported, but since we
1175  * intercept O_DIRECT reads and writes early, this function should
1176  * never get called.
1177  */
1178 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1179                               const struct iovec *iov,
1180                               loff_t pos, unsigned long nr_segs)
1181 {
1182         WARN_ON(1);
1183         return -EINVAL;
1184 }
1185
1186 const struct address_space_operations ceph_aops = {
1187         .readpage = ceph_readpage,
1188         .readpages = ceph_readpages,
1189         .writepage = ceph_writepage,
1190         .writepages = ceph_writepages_start,
1191         .write_begin = ceph_write_begin,
1192         .write_end = ceph_write_end,
1193         .set_page_dirty = ceph_set_page_dirty,
1194         .invalidatepage = ceph_invalidatepage,
1195         .releasepage = ceph_releasepage,
1196         .direct_IO = ceph_direct_io,
1197 };
1198
1199
1200 /*
1201  * vm ops
1202  */
1203
1204 /*
1205  * Reuse write_begin here for simplicity.
1206  */
1207 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1208 {
1209         struct inode *inode = file_inode(vma->vm_file);
1210         struct page *page = vmf->page;
1211         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1212         loff_t off = page_offset(page);
1213         loff_t size, len;
1214         int ret;
1215
1216         /* Update time before taking page lock */
1217         file_update_time(vma->vm_file);
1218
1219         size = i_size_read(inode);
1220         if (off + PAGE_CACHE_SIZE <= size)
1221                 len = PAGE_CACHE_SIZE;
1222         else
1223                 len = size & ~PAGE_CACHE_MASK;
1224
1225         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1226              off, len, page, page->index);
1227
1228         lock_page(page);
1229
1230         ret = VM_FAULT_NOPAGE;
1231         if ((off > size) ||
1232             (page->mapping != inode->i_mapping))
1233                 goto out;
1234
1235         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1236         if (ret == 0) {
1237                 /* success.  we'll keep the page locked. */
1238                 set_page_dirty(page);
1239                 up_read(&mdsc->snap_rwsem);
1240                 ret = VM_FAULT_LOCKED;
1241         } else {
1242                 if (ret == -ENOMEM)
1243                         ret = VM_FAULT_OOM;
1244                 else
1245                         ret = VM_FAULT_SIGBUS;
1246         }
1247 out:
1248         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1249         if (ret != VM_FAULT_LOCKED)
1250                 unlock_page(page);
1251         return ret;
1252 }
1253
1254 static struct vm_operations_struct ceph_vmops = {
1255         .fault          = filemap_fault,
1256         .page_mkwrite   = ceph_page_mkwrite,
1257         .remap_pages    = generic_file_remap_pages,
1258 };
1259
1260 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1261 {
1262         struct address_space *mapping = file->f_mapping;
1263
1264         if (!mapping->a_ops->readpage)
1265                 return -ENOEXEC;
1266         file_accessed(file);
1267         vma->vm_ops = &ceph_vmops;
1268         return 0;
1269 }