fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /*
  91          * Note that we're grabbing a snapc ref here without holding
  92          * any locks!
  93          */
  94         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         if (ci->i_head_snapc == NULL)
  99                 ci->i_head_snapc = ceph_get_snap_context(snapc);
 100         ++ci->i_wrbuffer_ref_head;
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         ret = __set_page_dirty_nobuffers(page);
 121         WARN_ON(!PageLocked(page));
 122         WARN_ON(!page->mapping);
 123
 124         return ret;
 125 }
 126
 127 /*
 128  * If we are truncating the full page (i.e. offset == 0), adjust the
 129  * dirty page counters appropriately.  Only called if there is private
 130  * data on the page.
 131  */
 132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 133                                 unsigned int length)
 134 {
 135         struct inode *inode;
 136         struct ceph_inode_info *ci;
 137         struct ceph_snap_context *snapc = page_snap_context(page);
 138
 139         inode = page->mapping->host;
 140         ci = ceph_inode(inode);
 141
 142         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 143                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 144                      inode, page, page->index, offset, length);
 145                 return;
 146         }
 147
 148         ceph_invalidate_fscache_page(inode, page);
 149
 150         if (!PagePrivate(page))
 151                 return;
 152
 153         /*
 154          * We can get non-dirty pages here due to races between
 155          * set_page_dirty and truncate_complete_page; just spit out a
 156          * warning, in case we end up with accounting problems later.
 157          */
 158         if (!PageDirty(page))
 159                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 160
 161         ClearPageChecked(page);
 162
 163         dout("%p invalidatepage %p idx %lu full dirty page\n",
 164              inode, page, page->index);
 165
 166         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 167         ceph_put_snap_context(snapc);
 168         page->private = 0;
 169         ClearPagePrivate(page);
 170 }
 171
 172 static int ceph_releasepage(struct page *page, gfp_t g)
 173 {
 174         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 175         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 176         WARN_ON(PageDirty(page));
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * read a single page, without unlocking it.
 187  */
 188 static int readpage_nounlock(struct file *filp, struct page *page)
 189 {
 190         struct inode *inode = file_inode(filp);
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_osd_client *osdc =
 193                 &ceph_inode_to_client(inode)->client->osdc;
 194         int err = 0;
 195         u64 off = page_offset(page);
 196         u64 len = PAGE_CACHE_SIZE;
 197
 198         if (off >= i_size_read(inode)) {
 199                 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 200                 SetPageUptodate(page);
 201                 return 0;
 202         }
 203
 204         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 205                 /*
 206                  * Uptodate inline data should have been added
 207                  * into page cache while getting Fcr caps.
 208                  */
 209                 if (off == 0)
 210                         return -EINVAL;
 211                 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 212                 SetPageUptodate(page);
 213                 return 0;
 214         }
 215
 216         err = ceph_readpage_from_fscache(inode, page);
 217         if (err == 0)
 218                 goto out;
 219
 220         dout("readpage inode %p file %p page %p index %lu\n",
 221              inode, filp, page, page->index);
 222         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 223                                   off, &len,
 224                                   ci->i_truncate_seq, ci->i_truncate_size,
 225                                   &page, 1, 0);
 226         if (err == -ENOENT)
 227                 err = 0;
 228         if (err < 0) {
 229                 SetPageError(page);
 230                 ceph_fscache_readpage_cancel(inode, page);
 231                 goto out;
 232         }
 233         if (err < PAGE_CACHE_SIZE)
 234                 /* zero fill remainder of page */
 235                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 236         else
 237                 flush_dcache_page(page);
 238
 239         SetPageUptodate(page);
 240         ceph_readpage_to_fscache(inode, page);
 241
 242 out:
 243         return err < 0 ? err : 0;
 244 }
 245
 246 static int ceph_readpage(struct file *filp, struct page *page)
 247 {
 248         int r = readpage_nounlock(filp, page);
 249         unlock_page(page);
 250         return r;
 251 }
 252
 253 /*
 254  * Finish an async read(ahead) op.
 255  */
 256 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 257 {
 258         struct inode *inode = req->r_inode;
 259         struct ceph_osd_data *osd_data;
 260         int rc = req->r_result;
 261         int bytes = le32_to_cpu(msg->hdr.data_len);
 262         int num_pages;
 263         int i;
 264
 265         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 266
 267         /* unlock all pages, zeroing any data we didn't read */
 268         osd_data = osd_req_op_extent_osd_data(req, 0);
 269         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 270         num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                         (u64)osd_data->length);
 272         for (i = 0; i < num_pages; i++) {
 273                 struct page *page = osd_data->pages[i];
 274
 275                 if (rc < 0)
 276                         goto unlock;
 277                 if (bytes < (int)PAGE_CACHE_SIZE) {
 278                         /* zero (remainder of) page */
 279                         int s = bytes < 0 ? 0 : bytes;
 280                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 281                 }
 282                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 283                      page->index);
 284                 flush_dcache_page(page);
 285                 SetPageUptodate(page);
 286                 ceph_readpage_to_fscache(inode, page);
 287 unlock:
 288                 unlock_page(page);
 289                 page_cache_release(page);
 290                 bytes -= PAGE_CACHE_SIZE;
 291         }
 292         kfree(osd_data->pages);
 293 }
 294
 295 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 296 {
 297         int i;
 298
 299         for (i = 0; i < num_pages; i++)
 300                 unlock_page(pages[i]);
 301 }
 302
 303 /*
 304  * start an async read(ahead) operation.  return nr_pages we submitted
 305  * a read for on success, or negative error code.
 306  */
 307 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 308 {
 309         struct ceph_osd_client *osdc =
 310                 &ceph_inode_to_client(inode)->client->osdc;
 311         struct ceph_inode_info *ci = ceph_inode(inode);
 312         struct page *page = list_entry(page_list->prev, struct page, lru);
 313         struct ceph_vino vino;
 314         struct ceph_osd_request *req;
 315         u64 off;
 316         u64 len;
 317         int i;
 318         struct page **pages;
 319         pgoff_t next_index;
 320         int nr_pages = 0;
 321         int ret;
 322
 323         off = (u64) page_offset(page);
 324
 325         /* count pages */
 326         next_index = page->index;
 327         list_for_each_entry_reverse(page, page_list, lru) {
 328                 if (page->index != next_index)
 329                         break;
 330                 nr_pages++;
 331                 next_index++;
 332                 if (max && nr_pages == max)
 333                         break;
 334         }
 335         len = nr_pages << PAGE_CACHE_SHIFT;
 336         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 337              off, len);
 338         vino = ceph_vino(inode);
 339         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 340                                     0, 1, CEPH_OSD_OP_READ,
 341                                     CEPH_OSD_FLAG_READ, NULL,
 342                                     ci->i_truncate_seq, ci->i_truncate_size,
 343                                     false);
 344         if (IS_ERR(req))
 345                 return PTR_ERR(req);
 346
 347         /* build page vector */
 348         nr_pages = calc_pages_for(0, len);
 349         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 350         ret = -ENOMEM;
 351         if (!pages)
 352                 goto out;
 353         for (i = 0; i < nr_pages; ++i) {
 354                 page = list_entry(page_list->prev, struct page, lru);
 355                 BUG_ON(PageLocked(page));
 356                 list_del(&page->lru);
 357
 358                 dout("start_read %p adding %p idx %lu\n", inode, page,
 359                      page->index);
 360                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 361                                           GFP_NOFS)) {
 362                         ceph_fscache_uncache_page(inode, page);
 363                         page_cache_release(page);
 364                         dout("start_read %p add_to_page_cache failed %p\n",
 365                              inode, page);
 366                         nr_pages = i;
 367                         goto out_pages;
 368                 }
 369                 pages[i] = page;
 370         }
 371         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 372         req->r_callback = finish_read;
 373         req->r_inode = inode;
 374
 375         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 376
 377         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 378         ret = ceph_osdc_start_request(osdc, req, false);
 379         if (ret < 0)
 380                 goto out_pages;
 381         ceph_osdc_put_request(req);
 382         return nr_pages;
 383
 384 out_pages:
 385         ceph_unlock_page_vector(pages, nr_pages);
 386         ceph_release_page_vector(pages, nr_pages);
 387 out:
 388         ceph_osdc_put_request(req);
 389         return ret;
 390 }
 391
 392
 393 /*
 394  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 395  * the caller (VM) cleans them up.
 396  */
 397 static int ceph_readpages(struct file *file, struct address_space *mapping,
 398                           struct list_head *page_list, unsigned nr_pages)
 399 {
 400         struct inode *inode = file_inode(file);
 401         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 402         int rc = 0;
 403         int max = 0;
 404
 405         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 406                 return -EINVAL;
 407
 408         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 409                                          &nr_pages);
 410
 411         if (rc == 0)
 412                 goto out;
 413
 414         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 415                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 416                         >> PAGE_SHIFT;
 417
 418         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 419                 file, nr_pages,
 420              max);
 421         while (!list_empty(page_list)) {
 422                 rc = start_read(inode, page_list, max);
 423                 if (rc < 0)
 424                         goto out;
 425                 BUG_ON(rc == 0);
 426         }
 427 out:
 428         ceph_fscache_readpages_cancel(inode, page_list);
 429
 430         dout("readpages %p file %p ret %d\n", inode, file, rc);
 431         return rc;
 432 }
 433
 434 /*
 435  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 436  * only snap context we are allowed to write back.
 437  */
 438 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 439                                                     u64 *snap_size)
 440 {
 441         struct ceph_inode_info *ci = ceph_inode(inode);
 442         struct ceph_snap_context *snapc = NULL;
 443         struct ceph_cap_snap *capsnap = NULL;
 444
 445         spin_lock(&ci->i_ceph_lock);
 446         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 447                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 448                      capsnap->context, capsnap->dirty_pages);
 449                 if (capsnap->dirty_pages) {
 450                         snapc = ceph_get_snap_context(capsnap->context);
 451                         if (snap_size)
 452                                 *snap_size = capsnap->size;
 453                         break;
 454                 }
 455         }
 456         if (!snapc && ci->i_wrbuffer_ref_head) {
 457                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 458                 dout(" head snapc %p has %d dirty pages\n",
 459                      snapc, ci->i_wrbuffer_ref_head);
 460         }
 461         spin_unlock(&ci->i_ceph_lock);
 462         return snapc;
 463 }
 464
 465 /*
 466  * Write a single page, but leave the page locked.
 467  *
 468  * If we get a write error, set the page error bit, but still adjust the
 469  * dirty page accounting (i.e., page is no longer dirty).
 470  */
 471 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 472 {
 473         struct inode *inode;
 474         struct ceph_inode_info *ci;
 475         struct ceph_fs_client *fsc;
 476         struct ceph_osd_client *osdc;
 477         struct ceph_snap_context *snapc, *oldest;
 478         loff_t page_off = page_offset(page);
 479         long writeback_stat;
 480         u64 truncate_size, snap_size = 0;
 481         u32 truncate_seq;
 482         int err = 0, len = PAGE_CACHE_SIZE;
 483
 484         dout("writepage %p idx %lu\n", page, page->index);
 485
 486         if (!page->mapping || !page->mapping->host) {
 487                 dout("writepage %p - no mapping\n", page);
 488                 return -EFAULT;
 489         }
 490         inode = page->mapping->host;
 491         ci = ceph_inode(inode);
 492         fsc = ceph_inode_to_client(inode);
 493         osdc = &fsc->client->osdc;
 494
 495         /* verify this is a writeable snap context */
 496         snapc = page_snap_context(page);
 497         if (snapc == NULL) {
 498                 dout("writepage %p page %p not dirty?\n", inode, page);
 499                 goto out;
 500         }
 501         oldest = get_oldest_context(inode, &snap_size);
 502         if (snapc->seq > oldest->seq) {
 503                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 504                      inode, page, snapc);
 505                 /* we should only noop if called by kswapd */
 506                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 507                 ceph_put_snap_context(oldest);
 508                 goto out;
 509         }
 510         ceph_put_snap_context(oldest);
 511
 512         spin_lock(&ci->i_ceph_lock);
 513         truncate_seq = ci->i_truncate_seq;
 514         truncate_size = ci->i_truncate_size;
 515         if (!snap_size)
 516                 snap_size = i_size_read(inode);
 517         spin_unlock(&ci->i_ceph_lock);
 518
 519         /* is this a partial page at end of file? */
 520         if (page_off >= snap_size) {
 521                 dout("%p page eof %llu\n", page, snap_size);
 522                 goto out;
 523         }
 524         if (snap_size < page_off + len)
 525                 len = snap_size - page_off;
 526
 527         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 528              inode, page, page->index, page_off, len, snapc);
 529
 530         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 531         if (writeback_stat >
 532             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 533                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 534
 535         ceph_readpage_to_fscache(inode, page);
 536
 537         set_page_writeback(page);
 538         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 539                                    &ci->i_layout, snapc,
 540                                    page_off, len,
 541                                    truncate_seq, truncate_size,
 542                                    &inode->i_mtime, &page, 1);
 543         if (err < 0) {
 544                 dout("writepage setting page/mapping error %d %p\n", err, page);
 545                 SetPageError(page);
 546                 mapping_set_error(&inode->i_data, err);
 547                 if (wbc)
 548                         wbc->pages_skipped++;
 549         } else {
 550                 dout("writepage cleaned page %p\n", page);
 551                 err = 0;  /* vfs expects us to return 0 */
 552         }
 553         page->private = 0;
 554         ClearPagePrivate(page);
 555         end_page_writeback(page);
 556         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 557         ceph_put_snap_context(snapc);  /* page's reference */
 558 out:
 559         return err;
 560 }
 561
 562 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 563 {
 564         int err;
 565         struct inode *inode = page->mapping->host;
 566         BUG_ON(!inode);
 567         ihold(inode);
 568         err = writepage_nounlock(page, wbc);
 569         unlock_page(page);
 570         iput(inode);
 571         return err;
 572 }
 573
 574
 575 /*
 576  * lame release_pages helper.  release_pages() isn't exported to
 577  * modules.
 578  */
 579 static void ceph_release_pages(struct page **pages, int num)
 580 {
 581         struct pagevec pvec;
 582         int i;
 583
 584         pagevec_init(&pvec, 0);
 585         for (i = 0; i < num; i++) {
 586                 if (pagevec_add(&pvec, pages[i]) == 0)
 587                         pagevec_release(&pvec);
 588         }
 589         pagevec_release(&pvec);
 590 }
 591
 592 /*
 593  * async writeback completion handler.
 594  *
 595  * If we get an error, set the mapping error bit, but not the individual
 596  * page error bits.
 597  */
 598 static void writepages_finish(struct ceph_osd_request *req,
 599                               struct ceph_msg *msg)
 600 {
 601         struct inode *inode = req->r_inode;
 602         struct ceph_inode_info *ci = ceph_inode(inode);
 603         struct ceph_osd_data *osd_data;
 604         unsigned wrote;
 605         struct page *page;
 606         int num_pages;
 607         int i;
 608         struct ceph_snap_context *snapc = req->r_snapc;
 609         struct address_space *mapping = inode->i_mapping;
 610         int rc = req->r_result;
 611         u64 bytes = req->r_ops[0].extent.length;
 612         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 613         long writeback_stat;
 614         unsigned issued = ceph_caps_issued(ci);
 615
 616         osd_data = osd_req_op_extent_osd_data(req, 0);
 617         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 618         num_pages = calc_pages_for((u64)osd_data->alignment,
 619                                         (u64)osd_data->length);
 620         if (rc >= 0) {
 621                 /*
 622                  * Assume we wrote the pages we originally sent.  The
 623                  * osd might reply with fewer pages if our writeback
 624                  * raced with a truncation and was adjusted at the osd,
 625                  * so don't believe the reply.
 626                  */
 627                 wrote = num_pages;
 628         } else {
 629                 wrote = 0;
 630                 mapping_set_error(mapping, rc);
 631         }
 632         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 633              inode, rc, bytes, wrote);
 634
 635         /* clean all pages */
 636         for (i = 0; i < num_pages; i++) {
 637                 page = osd_data->pages[i];
 638                 BUG_ON(!page);
 639                 WARN_ON(!PageUptodate(page));
 640
 641                 writeback_stat =
 642                         atomic_long_dec_return(&fsc->writeback_count);
 643                 if (writeback_stat <
 644                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 645                         clear_bdi_congested(&fsc->backing_dev_info,
 646                                             BLK_RW_ASYNC);
 647
 648                 ceph_put_snap_context(page_snap_context(page));
 649                 page->private = 0;
 650                 ClearPagePrivate(page);
 651                 dout("unlocking %d %p\n", i, page);
 652                 end_page_writeback(page);
 653
 654                 /*
 655                  * We lost the cache cap, need to truncate the page before
 656                  * it is unlocked, otherwise we'd truncate it later in the
 657                  * page truncation thread, possibly losing some data that
 658                  * raced its way in
 659                  */
 660                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 661                         generic_error_remove_page(inode->i_mapping, page);
 662
 663                 unlock_page(page);
 664         }
 665         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 666         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 667
 668         ceph_release_pages(osd_data->pages, num_pages);
 669         if (osd_data->pages_from_pool)
 670                 mempool_free(osd_data->pages,
 671                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 672         else
 673                 kfree(osd_data->pages);
 674         ceph_osdc_put_request(req);
 675 }
 676
 677 /*
 678  * initiate async writeback
 679  */
 680 static int ceph_writepages_start(struct address_space *mapping,
 681                                  struct writeback_control *wbc)
 682 {
 683         struct inode *inode = mapping->host;
 684         struct ceph_inode_info *ci = ceph_inode(inode);
 685         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 686         struct ceph_vino vino = ceph_vino(inode);
 687         pgoff_t index, start, end;
 688         int range_whole = 0;
 689         int should_loop = 1;
 690         pgoff_t max_pages = 0, max_pages_ever = 0;
 691         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 692         struct pagevec pvec;
 693         int done = 0;
 694         int rc = 0;
 695         unsigned wsize = 1 << inode->i_blkbits;
 696         struct ceph_osd_request *req = NULL;
 697         int do_sync = 0;
 698         u64 truncate_size, snap_size;
 699         u32 truncate_seq;
 700
 701         /*
 702          * Include a 'sync' in the OSD request if this is a data
 703          * integrity write (e.g., O_SYNC write or fsync()), or if our
 704          * cap is being revoked.
 705          */
 706         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 707                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 708                 do_sync = 1;
 709         dout("writepages_start %p dosync=%d (mode=%s)\n",
 710              inode, do_sync,
 711              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 712              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 713
 714         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 715                 pr_warn("writepage_start %p on forced umount\n", inode);
 716                 return -EIO; /* we're in a forced umount, don't write! */
 717         }
 718         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 719                 wsize = fsc->mount_options->wsize;
 720         if (wsize < PAGE_CACHE_SIZE)
 721                 wsize = PAGE_CACHE_SIZE;
 722         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 723
 724         pagevec_init(&pvec, 0);
 725
 726         /* where to start/end? */
 727         if (wbc->range_cyclic) {
 728                 start = mapping->writeback_index; /* Start from prev offset */
 729                 end = -1;
 730                 dout(" cyclic, start at %lu\n", start);
 731         } else {
 732                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 733                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 734                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 735                         range_whole = 1;
 736                 should_loop = 0;
 737                 dout(" not cyclic, %lu to %lu\n", start, end);
 738         }
 739         index = start;
 740
 741 retry:
 742         /* find oldest snap context with dirty data */
 743         ceph_put_snap_context(snapc);
 744         snap_size = 0;
 745         snapc = get_oldest_context(inode, &snap_size);
 746         if (!snapc) {
 747                 /* hmm, why does writepages get called when there
 748                    is no dirty data? */
 749                 dout(" no snap context with dirty data?\n");
 750                 goto out;
 751         }
 752         if (snap_size == 0)
 753                 snap_size = i_size_read(inode);
 754         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 755              snapc, snapc->seq, snapc->num_snaps);
 756
 757         spin_lock(&ci->i_ceph_lock);
 758         truncate_seq = ci->i_truncate_seq;
 759         truncate_size = ci->i_truncate_size;
 760         if (!snap_size)
 761                 snap_size = i_size_read(inode);
 762         spin_unlock(&ci->i_ceph_lock);
 763
 764         if (last_snapc && snapc != last_snapc) {
 765                 /* if we switched to a newer snapc, restart our scan at the
 766                  * start of the original file range. */
 767                 dout("  snapc differs from last pass, restarting at %lu\n",
 768                      index);
 769                 index = start;
 770         }
 771         last_snapc = snapc;
 772
 773         while (!done && index <= end) {
 774                 unsigned i;
 775                 int first;
 776                 pgoff_t next;
 777                 int pvec_pages, locked_pages;
 778                 struct page **pages = NULL;
 779                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 780                 struct page *page;
 781                 int want;
 782                 u64 offset, len;
 783                 long writeback_stat;
 784
 785                 next = 0;
 786                 locked_pages = 0;
 787                 max_pages = max_pages_ever;
 788
 789 get_more_pages:
 790                 first = -1;
 791                 want = min(end - index,
 792                            min((pgoff_t)PAGEVEC_SIZE,
 793                                max_pages - (pgoff_t)locked_pages) - 1)
 794                         + 1;
 795                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 796                                                 PAGECACHE_TAG_DIRTY,
 797                                                 want);
 798                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 799                 if (!pvec_pages && !locked_pages)
 800                         break;
 801                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 802                         page = pvec.pages[i];
 803                         dout("? %p idx %lu\n", page, page->index);
 804                         if (locked_pages == 0)
 805                                 lock_page(page);  /* first page */
 806                         else if (!trylock_page(page))
 807                                 break;
 808
 809                         /* only dirty pages, or our accounting breaks */
 810                         if (unlikely(!PageDirty(page)) ||
 811                             unlikely(page->mapping != mapping)) {
 812                                 dout("!dirty or !mapping %p\n", page);
 813                                 unlock_page(page);
 814                                 break;
 815                         }
 816                         if (!wbc->range_cyclic && page->index > end) {
 817                                 dout("end of range %p\n", page);
 818                                 done = 1;
 819                                 unlock_page(page);
 820                                 break;
 821                         }
 822                         if (next && (page->index != next)) {
 823                                 dout("not consecutive %p\n", page);
 824                                 unlock_page(page);
 825                                 break;
 826                         }
 827                         if (wbc->sync_mode != WB_SYNC_NONE) {
 828                                 dout("waiting on writeback %p\n", page);
 829                                 wait_on_page_writeback(page);
 830                         }
 831                         if (page_offset(page) >= snap_size) {
 832                                 dout("%p page eof %llu\n", page, snap_size);
 833                                 done = 1;
 834                                 unlock_page(page);
 835                                 break;
 836                         }
 837                         if (PageWriteback(page)) {
 838                                 dout("%p under writeback\n", page);
 839                                 unlock_page(page);
 840                                 break;
 841                         }
 842
 843                         /* only if matching snap context */
 844                         pgsnapc = page_snap_context(page);
 845                         if (pgsnapc->seq > snapc->seq) {
 846                                 dout("page snapc %p %lld > oldest %p %lld\n",
 847                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 848                                 unlock_page(page);
 849                                 if (!locked_pages)
 850                                         continue; /* keep looking for snap */
 851                                 break;
 852                         }
 853
 854                         if (!clear_page_dirty_for_io(page)) {
 855                                 dout("%p !clear_page_dirty_for_io\n", page);
 856                                 unlock_page(page);
 857                                 break;
 858                         }
 859
 860                         /*
 861                          * We have something to write.  If this is
 862                          * the first locked page this time through,
 863                          * allocate an osd request and a page array
 864                          * that it will use.
 865                          */
 866                         if (locked_pages == 0) {
 867                                 BUG_ON(pages);
 868                                 /* prepare async write request */
 869                                 offset = (u64)page_offset(page);
 870                                 len = wsize;
 871                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 872                                                         &ci->i_layout, vino,
 873                                                         offset, &len, 0,
 874                                                         do_sync ? 2 : 1,
 875                                                         CEPH_OSD_OP_WRITE,
 876                                                         CEPH_OSD_FLAG_WRITE |
 877                                                         CEPH_OSD_FLAG_ONDISK,
 878                                                         snapc, truncate_seq,
 879                                                         truncate_size, true);
 880                                 if (IS_ERR(req)) {
 881                                         rc = PTR_ERR(req);
 882                                         unlock_page(page);
 883                                         break;
 884                                 }
 885
 886                                 if (do_sync)
 887                                         osd_req_op_init(req, 1,
 888                                                         CEPH_OSD_OP_STARTSYNC, 0);
 889
 890                                 req->r_callback = writepages_finish;
 891                                 req->r_inode = inode;
 892
 893                                 max_pages = calc_pages_for(0, (u64)len);
 894                                 pages = kmalloc(max_pages * sizeof (*pages),
 895                                                 GFP_NOFS);
 896                                 if (!pages) {
 897                                         pool = fsc->wb_pagevec_pool;
 898                                         pages = mempool_alloc(pool, GFP_NOFS);
 899                                         BUG_ON(!pages);
 900                                 }
 901                         }
 902
 903                         /* note position of first page in pvec */
 904                         if (first < 0)
 905                                 first = i;
 906                         dout("%p will write page %p idx %lu\n",
 907                              inode, page, page->index);
 908
 909                         writeback_stat =
 910                                atomic_long_inc_return(&fsc->writeback_count);
 911                         if (writeback_stat > CONGESTION_ON_THRESH(
 912                                     fsc->mount_options->congestion_kb)) {
 913                                 set_bdi_congested(&fsc->backing_dev_info,
 914                                                   BLK_RW_ASYNC);
 915                         }
 916
 917                         set_page_writeback(page);
 918                         pages[locked_pages] = page;
 919                         locked_pages++;
 920                         next = page->index + 1;
 921                 }
 922
 923                 /* did we get anything? */
 924                 if (!locked_pages)
 925                         goto release_pvec_pages;
 926                 if (i) {
 927                         int j;
 928                         BUG_ON(!locked_pages || first < 0);
 929
 930                         if (pvec_pages && i == pvec_pages &&
 931                             locked_pages < max_pages) {
 932                                 dout("reached end pvec, trying for more\n");
 933                                 pagevec_reinit(&pvec);
 934                                 goto get_more_pages;
 935                         }
 936
 937                         /* shift unused pages over in the pvec...  we
 938                          * will need to release them below. */
 939                         for (j = i; j < pvec_pages; j++) {
 940                                 dout(" pvec leftover page %p\n",
 941                                      pvec.pages[j]);
 942                                 pvec.pages[j-i+first] = pvec.pages[j];
 943                         }
 944                         pvec.nr -= i-first;
 945                 }
 946
 947                 /* Format the osd request message and submit the write */
 948
 949                 offset = page_offset(pages[0]);
 950                 len = min(snap_size - offset,
 951                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 952                 dout("writepages got %d pages at %llu~%llu\n",
 953                      locked_pages, offset, len);
 954
 955                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 956                                                         !!pool, false);
 957
 958                 pages = NULL;   /* request message now owns the pages array */
 959                 pool = NULL;
 960
 961                 /* Update the write op length in case we changed it */
 962
 963                 osd_req_op_extent_update(req, 0, len);
 964
 965                 vino = ceph_vino(inode);
 966                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 967                                         &inode->i_mtime);
 968
 969                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 970                 BUG_ON(rc);
 971                 req = NULL;
 972
 973                 /* continue? */
 974                 index = next;
 975                 wbc->nr_to_write -= locked_pages;
 976                 if (wbc->nr_to_write <= 0)
 977                         done = 1;
 978
 979 release_pvec_pages:
 980                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 981                      pvec.nr ? pvec.pages[0] : NULL);
 982                 pagevec_release(&pvec);
 983
 984                 if (locked_pages && !done)
 985                         goto retry;
 986         }
 987
 988         if (should_loop && !done) {
 989                 /* more to do; loop back to beginning of file */
 990                 dout("writepages looping back to beginning of file\n");
 991                 should_loop = 0;
 992                 index = 0;
 993                 goto retry;
 994         }
 995
 996         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 997                 mapping->writeback_index = index;
 998
 999 out:
1000         if (req)
1001                 ceph_osdc_put_request(req);
1002         ceph_put_snap_context(snapc);
1003         dout("writepages done, rc = %d\n", rc);
1004         return rc;
1005 }
1006
1007
1008
1009 /*
1010  * See if a given @snapc is either writeable, or already written.
1011  */
1012 static int context_is_writeable_or_written(struct inode *inode,
1013                                            struct ceph_snap_context *snapc)
1014 {
1015         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1016         int ret = !oldest || snapc->seq <= oldest->seq;
1017
1018         ceph_put_snap_context(oldest);
1019         return ret;
1020 }
1021
1022 /*
1023  * We are only allowed to write into/dirty the page if the page is
1024  * clean, or already dirty within the same snap context.
1025  *
1026  * called with page locked.
1027  * return success with page locked,
1028  * or any failure (incl -EAGAIN) with page unlocked.
1029  */
1030 static int ceph_update_writeable_page(struct file *file,
1031                             loff_t pos, unsigned len,
1032                             struct page *page)
1033 {
1034         struct inode *inode = file_inode(file);
1035         struct ceph_inode_info *ci = ceph_inode(inode);
1036         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1037         loff_t page_off = pos & PAGE_CACHE_MASK;
1038         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1039         int end_in_page = pos_in_page + len;
1040         loff_t i_size;
1041         int r;
1042         struct ceph_snap_context *snapc, *oldest;
1043
1044 retry_locked:
1045         /* writepages currently holds page lock, but if we change that later, */
1046         wait_on_page_writeback(page);
1047
1048         /* check snap context */
1049         BUG_ON(!ci->i_snap_realm);
1050         down_read(&mdsc->snap_rwsem);
1051         BUG_ON(!ci->i_snap_realm->cached_context);
1052         snapc = page_snap_context(page);
1053         if (snapc && snapc != ci->i_head_snapc) {
1054                 /*
1055                  * this page is already dirty in another (older) snap
1056                  * context!  is it writeable now?
1057                  */
1058                 oldest = get_oldest_context(inode, NULL);
1059                 up_read(&mdsc->snap_rwsem);
1060
1061                 if (snapc->seq > oldest->seq) {
1062                         ceph_put_snap_context(oldest);
1063                         dout(" page %p snapc %p not current or oldest\n",
1064                              page, snapc);
1065                         /*
1066                          * queue for writeback, and wait for snapc to
1067                          * be writeable or written
1068                          */
1069                         snapc = ceph_get_snap_context(snapc);
1070                         unlock_page(page);
1071                         ceph_queue_writeback(inode);
1072                         r = wait_event_interruptible(ci->i_cap_wq,
1073                                context_is_writeable_or_written(inode, snapc));
1074                         ceph_put_snap_context(snapc);
1075                         if (r == -ERESTARTSYS)
1076                                 return r;
1077                         return -EAGAIN;
1078                 }
1079                 ceph_put_snap_context(oldest);
1080
1081                 /* yay, writeable, do it now (without dropping page lock) */
1082                 dout(" page %p snapc %p not current, but oldest\n",
1083                      page, snapc);
1084                 if (!clear_page_dirty_for_io(page))
1085                         goto retry_locked;
1086                 r = writepage_nounlock(page, NULL);
1087                 if (r < 0)
1088                         goto fail_nosnap;
1089                 goto retry_locked;
1090         }
1091
1092         if (PageUptodate(page)) {
1093                 dout(" page %p already uptodate\n", page);
1094                 return 0;
1095         }
1096
1097         /* full page? */
1098         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1099                 return 0;
1100
1101         /* past end of file? */
1102         i_size = inode->i_size;   /* caller holds i_mutex */
1103
1104         if (page_off >= i_size ||
1105             (pos_in_page == 0 && (pos+len) >= i_size &&
1106              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1107                 dout(" zeroing %p 0 - %d and %d - %d\n",
1108                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1109                 zero_user_segments(page,
1110                                    0, pos_in_page,
1111                                    end_in_page, PAGE_CACHE_SIZE);
1112                 return 0;
1113         }
1114
1115         /* we need to read it. */
1116         up_read(&mdsc->snap_rwsem);
1117         r = readpage_nounlock(file, page);
1118         if (r < 0)
1119                 goto fail_nosnap;
1120         goto retry_locked;
1121 fail_nosnap:
1122         unlock_page(page);
1123         return r;
1124 }
1125
1126 /*
1127  * We are only allowed to write into/dirty the page if the page is
1128  * clean, or already dirty within the same snap context.
1129  */
1130 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1131                             loff_t pos, unsigned len, unsigned flags,
1132                             struct page **pagep, void **fsdata)
1133 {
1134         struct inode *inode = file_inode(file);
1135         struct page *page;
1136         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1137         int r;
1138
1139         do {
1140                 /* get a page */
1141                 page = grab_cache_page_write_begin(mapping, index, 0);
1142                 if (!page)
1143                         return -ENOMEM;
1144                 *pagep = page;
1145
1146                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1147                      inode, page, (int)pos, (int)len);
1148
1149                 r = ceph_update_writeable_page(file, pos, len, page);
1150                 if (r < 0)
1151                         page_cache_release(page);
1152                 else
1153                         *pagep = page;
1154         } while (r == -EAGAIN);
1155
1156         return r;
1157 }
1158
1159 /*
1160  * we don't do anything in here that simple_write_end doesn't do
1161  * except adjust dirty page accounting and drop read lock on
1162  * mdsc->snap_rwsem.
1163  */
1164 static int ceph_write_end(struct file *file, struct address_space *mapping,
1165                           loff_t pos, unsigned len, unsigned copied,
1166                           struct page *page, void *fsdata)
1167 {
1168         struct inode *inode = file_inode(file);
1169         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1170         struct ceph_mds_client *mdsc = fsc->mdsc;
1171         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1172         int check_cap = 0;
1173
1174         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1175              inode, page, (int)pos, (int)copied, (int)len);
1176
1177         /* zero the stale part of the page if we did a short copy */
1178         if (copied < len)
1179                 zero_user_segment(page, from+copied, len);
1180
1181         /* did file size increase? */
1182         /* (no need for i_size_read(); we caller holds i_mutex */
1183         if (pos+copied > inode->i_size)
1184                 check_cap = ceph_inode_set_size(inode, pos+copied);
1185
1186         if (!PageUptodate(page))
1187                 SetPageUptodate(page);
1188
1189         set_page_dirty(page);
1190
1191         unlock_page(page);
1192         up_read(&mdsc->snap_rwsem);
1193         page_cache_release(page);
1194
1195         if (check_cap)
1196                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1197
1198         return copied;
1199 }
1200
1201 /*
1202  * we set .direct_IO to indicate direct io is supported, but since we
1203  * intercept O_DIRECT reads and writes early, this function should
1204  * never get called.
1205  */
1206 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter,
1207                               loff_t pos)
1208 {
1209         WARN_ON(1);
1210         return -EINVAL;
1211 }
1212
1213 const struct address_space_operations ceph_aops = {
1214         .readpage = ceph_readpage,
1215         .readpages = ceph_readpages,
1216         .writepage = ceph_writepage,
1217         .writepages = ceph_writepages_start,
1218         .write_begin = ceph_write_begin,
1219         .write_end = ceph_write_end,
1220         .set_page_dirty = ceph_set_page_dirty,
1221         .invalidatepage = ceph_invalidatepage,
1222         .releasepage = ceph_releasepage,
1223         .direct_IO = ceph_direct_io,
1224 };
1225
1226
1227 /*
1228  * vm ops
1229  */
1230 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1231 {
1232         struct inode *inode = file_inode(vma->vm_file);
1233         struct ceph_inode_info *ci = ceph_inode(inode);
1234         struct ceph_file_info *fi = vma->vm_file->private_data;
1235         struct page *pinned_page = NULL;
1236         loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1237         int want, got, ret;
1238
1239         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1240              inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1241         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1242                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1243         else
1244                 want = CEPH_CAP_FILE_CACHE;
1245         while (1) {
1246                 got = 0;
1247                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1248                                     -1, &got, &pinned_page);
1249                 if (ret == 0)
1250                         break;
1251                 if (ret != -ERESTARTSYS) {
1252                         WARN_ON(1);
1253                         return VM_FAULT_SIGBUS;
1254                 }
1255         }
1256         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1257              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1258
1259         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1260             ci->i_inline_version == CEPH_INLINE_NONE)
1261                 ret = filemap_fault(vma, vmf);
1262         else
1263                 ret = -EAGAIN;
1264
1265         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1266              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1267         if (pinned_page)
1268                 page_cache_release(pinned_page);
1269         ceph_put_cap_refs(ci, got);
1270
1271         if (ret != -EAGAIN)
1272                 return ret;
1273
1274         /* read inline data */
1275         if (off >= PAGE_CACHE_SIZE) {
1276                 /* does not support inline data > PAGE_SIZE */
1277                 ret = VM_FAULT_SIGBUS;
1278         } else {
1279                 int ret1;
1280                 struct address_space *mapping = inode->i_mapping;
1281                 struct page *page = find_or_create_page(mapping, 0,
1282                                                 mapping_gfp_mask(mapping) &
1283                                                 ~__GFP_FS);
1284                 if (!page) {
1285                         ret = VM_FAULT_OOM;
1286                         goto out;
1287                 }
1288                 ret1 = __ceph_do_getattr(inode, page,
1289                                          CEPH_STAT_CAP_INLINE_DATA, true);
1290                 if (ret1 < 0 || off >= i_size_read(inode)) {
1291                         unlock_page(page);
1292                         page_cache_release(page);
1293                         ret = VM_FAULT_SIGBUS;
1294                         goto out;
1295                 }
1296                 if (ret1 < PAGE_CACHE_SIZE)
1297                         zero_user_segment(page, ret1, PAGE_CACHE_SIZE);
1298                 else
1299                         flush_dcache_page(page);
1300                 SetPageUptodate(page);
1301                 vmf->page = page;
1302                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1303         }
1304 out:
1305         dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1306              inode, off, (size_t)PAGE_CACHE_SIZE, ret);
1307         return ret;
1308 }
1309
1310 /*
1311  * Reuse write_begin here for simplicity.
1312  */
1313 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1314 {
1315         struct inode *inode = file_inode(vma->vm_file);
1316         struct ceph_inode_info *ci = ceph_inode(inode);
1317         struct ceph_file_info *fi = vma->vm_file->private_data;
1318         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1319         struct page *page = vmf->page;
1320         loff_t off = page_offset(page);
1321         loff_t size = i_size_read(inode);
1322         size_t len;
1323         int want, got, ret;
1324
1325         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1326                 struct page *locked_page = NULL;
1327                 if (off == 0) {
1328                         lock_page(page);
1329                         locked_page = page;
1330                 }
1331                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1332                 if (locked_page)
1333                         unlock_page(locked_page);
1334                 if (ret < 0)
1335                         return VM_FAULT_SIGBUS;
1336         }
1337
1338         if (off + PAGE_CACHE_SIZE <= size)
1339                 len = PAGE_CACHE_SIZE;
1340         else
1341                 len = size & ~PAGE_CACHE_MASK;
1342
1343         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1344              inode, ceph_vinop(inode), off, len, size);
1345         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1346                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1347         else
1348                 want = CEPH_CAP_FILE_BUFFER;
1349         while (1) {
1350                 got = 0;
1351                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1352                                     &got, NULL);
1353                 if (ret == 0)
1354                         break;
1355                 if (ret != -ERESTARTSYS) {
1356                         WARN_ON(1);
1357                         return VM_FAULT_SIGBUS;
1358                 }
1359         }
1360         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1361              inode, off, len, ceph_cap_string(got));
1362
1363         /* Update time before taking page lock */
1364         file_update_time(vma->vm_file);
1365
1366         lock_page(page);
1367
1368         ret = VM_FAULT_NOPAGE;
1369         if ((off > size) ||
1370             (page->mapping != inode->i_mapping))
1371                 goto out;
1372
1373         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1374         if (ret == 0) {
1375                 /* success.  we'll keep the page locked. */
1376                 set_page_dirty(page);
1377                 up_read(&mdsc->snap_rwsem);
1378                 ret = VM_FAULT_LOCKED;
1379         } else {
1380                 if (ret == -ENOMEM)
1381                         ret = VM_FAULT_OOM;
1382                 else
1383                         ret = VM_FAULT_SIGBUS;
1384         }
1385 out:
1386         if (ret != VM_FAULT_LOCKED)
1387                 unlock_page(page);
1388         if (ret == VM_FAULT_LOCKED ||
1389             ci->i_inline_version != CEPH_INLINE_NONE) {
1390                 int dirty;
1391                 spin_lock(&ci->i_ceph_lock);
1392                 ci->i_inline_version = CEPH_INLINE_NONE;
1393                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1394                 spin_unlock(&ci->i_ceph_lock);
1395                 if (dirty)
1396                         __mark_inode_dirty(inode, dirty);
1397         }
1398
1399         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1400              inode, off, len, ceph_cap_string(got), ret);
1401         ceph_put_cap_refs(ci, got);
1402
1403         return ret;
1404 }
1405
1406 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1407                            char *data, size_t len)
1408 {
1409         struct address_space *mapping = inode->i_mapping;
1410         struct page *page;
1411
1412         if (locked_page) {
1413                 page = locked_page;
1414         } else {
1415                 if (i_size_read(inode) == 0)
1416                         return;
1417                 page = find_or_create_page(mapping, 0,
1418                                            mapping_gfp_mask(mapping) & ~__GFP_FS);
1419                 if (!page)
1420                         return;
1421                 if (PageUptodate(page)) {
1422                         unlock_page(page);
1423                         page_cache_release(page);
1424                         return;
1425                 }
1426         }
1427
1428         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1429              inode, ceph_vinop(inode), len, locked_page);
1430
1431         if (len > 0) {
1432                 void *kaddr = kmap_atomic(page);
1433                 memcpy(kaddr, data, len);
1434                 kunmap_atomic(kaddr);
1435         }
1436
1437         if (page != locked_page) {
1438                 if (len < PAGE_CACHE_SIZE)
1439                         zero_user_segment(page, len, PAGE_CACHE_SIZE);
1440                 else
1441                         flush_dcache_page(page);
1442
1443                 SetPageUptodate(page);
1444                 unlock_page(page);
1445                 page_cache_release(page);
1446         }
1447 }
1448
1449 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1450 {
1451         struct inode *inode = file_inode(filp);
1452         struct ceph_inode_info *ci = ceph_inode(inode);
1453         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1454         struct ceph_osd_request *req;
1455         struct page *page = NULL;
1456         u64 len, inline_version;
1457         int err = 0;
1458         bool from_pagecache = false;
1459
1460         spin_lock(&ci->i_ceph_lock);
1461         inline_version = ci->i_inline_version;
1462         spin_unlock(&ci->i_ceph_lock);
1463
1464         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1465              inode, ceph_vinop(inode), inline_version);
1466
1467         if (inline_version == 1 || /* initial version, no data */
1468             inline_version == CEPH_INLINE_NONE)
1469                 goto out;
1470
1471         if (locked_page) {
1472                 page = locked_page;
1473                 WARN_ON(!PageUptodate(page));
1474         } else if (ceph_caps_issued(ci) &
1475                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1476                 page = find_get_page(inode->i_mapping, 0);
1477                 if (page) {
1478                         if (PageUptodate(page)) {
1479                                 from_pagecache = true;
1480                                 lock_page(page);
1481                         } else {
1482                                 page_cache_release(page);
1483                                 page = NULL;
1484                         }
1485                 }
1486         }
1487
1488         if (page) {
1489                 len = i_size_read(inode);
1490                 if (len > PAGE_CACHE_SIZE)
1491                         len = PAGE_CACHE_SIZE;
1492         } else {
1493                 page = __page_cache_alloc(GFP_NOFS);
1494                 if (!page) {
1495                         err = -ENOMEM;
1496                         goto out;
1497                 }
1498                 err = __ceph_do_getattr(inode, page,
1499                                         CEPH_STAT_CAP_INLINE_DATA, true);
1500                 if (err < 0) {
1501                         /* no inline data */
1502                         if (err == -ENODATA)
1503                                 err = 0;
1504                         goto out;
1505                 }
1506                 len = err;
1507         }
1508
1509         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1510                                     ceph_vino(inode), 0, &len, 0, 1,
1511                                     CEPH_OSD_OP_CREATE,
1512                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1513                                     ceph_empty_snapc, 0, 0, false);
1514         if (IS_ERR(req)) {
1515                 err = PTR_ERR(req);
1516                 goto out;
1517         }
1518
1519         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1520         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1521         if (!err)
1522                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1523         ceph_osdc_put_request(req);
1524         if (err < 0)
1525                 goto out;
1526
1527         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1528                                     ceph_vino(inode), 0, &len, 1, 3,
1529                                     CEPH_OSD_OP_WRITE,
1530                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1531                                     ceph_empty_snapc,
1532                                     ci->i_truncate_seq, ci->i_truncate_size,
1533                                     false);
1534         if (IS_ERR(req)) {
1535                 err = PTR_ERR(req);
1536                 goto out;
1537         }
1538
1539         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1540
1541         {
1542                 __le64 xattr_buf = cpu_to_le64(inline_version);
1543                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1544                                             "inline_version", &xattr_buf,
1545                                             sizeof(xattr_buf),
1546                                             CEPH_OSD_CMPXATTR_OP_GT,
1547                                             CEPH_OSD_CMPXATTR_MODE_U64);
1548                 if (err)
1549                         goto out_put;
1550         }
1551
1552         {
1553                 char xattr_buf[32];
1554                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1555                                          "%llu", inline_version);
1556                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1557                                             "inline_version",
1558                                             xattr_buf, xattr_len, 0, 0);
1559                 if (err)
1560                         goto out_put;
1561         }
1562
1563         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1564         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1565         if (!err)
1566                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1567 out_put:
1568         ceph_osdc_put_request(req);
1569         if (err == -ECANCELED)
1570                 err = 0;
1571 out:
1572         if (page && page != locked_page) {
1573                 if (from_pagecache) {
1574                         unlock_page(page);
1575                         page_cache_release(page);
1576                 } else
1577                         __free_pages(page, 0);
1578         }
1579
1580         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1581              inode, ceph_vinop(inode), inline_version, err);
1582         return err;
1583 }
1584
1585 static struct vm_operations_struct ceph_vmops = {
1586         .fault          = ceph_filemap_fault,
1587         .page_mkwrite   = ceph_page_mkwrite,
1588 };
1589
1590 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1591 {
1592         struct address_space *mapping = file->f_mapping;
1593
1594         if (!mapping->a_ops->readpage)
1595                 return -ENOEXEC;
1596         file_accessed(file);
1597         vma->vm_ops = &ceph_vmops;
1598         return 0;
1599 }
1600
1601 enum {
1602         POOL_READ       = 1,
1603         POOL_WRITE      = 2,
1604 };
1605
1606 static int __ceph_pool_perm_get(struct ceph_inode_info *ci, u32 pool)
1607 {
1608         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1609         struct ceph_mds_client *mdsc = fsc->mdsc;
1610         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1611         struct rb_node **p, *parent;
1612         struct ceph_pool_perm *perm;
1613         struct page **pages;
1614         int err = 0, err2 = 0, have = 0;
1615
1616         down_read(&mdsc->pool_perm_rwsem);
1617         p = &mdsc->pool_perm_tree.rb_node;
1618         while (*p) {
1619                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1620                 if (pool < perm->pool)
1621                         p = &(*p)->rb_left;
1622                 else if (pool > perm->pool)
1623                         p = &(*p)->rb_right;
1624                 else {
1625                         have = perm->perm;
1626                         break;
1627                 }
1628         }
1629         up_read(&mdsc->pool_perm_rwsem);
1630         if (*p)
1631                 goto out;
1632
1633         dout("__ceph_pool_perm_get pool %u no perm cached\n", pool);
1634
1635         down_write(&mdsc->pool_perm_rwsem);
1636         parent = NULL;
1637         while (*p) {
1638                 parent = *p;
1639                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1640                 if (pool < perm->pool)
1641                         p = &(*p)->rb_left;
1642                 else if (pool > perm->pool)
1643                         p = &(*p)->rb_right;
1644                 else {
1645                         have = perm->perm;
1646                         break;
1647                 }
1648         }
1649         if (*p) {
1650                 up_write(&mdsc->pool_perm_rwsem);
1651                 goto out;
1652         }
1653
1654         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1655                                          ceph_empty_snapc,
1656                                          1, false, GFP_NOFS);
1657         if (!rd_req) {
1658                 err = -ENOMEM;
1659                 goto out_unlock;
1660         }
1661
1662         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1663         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1664         rd_req->r_base_oloc.pool = pool;
1665         snprintf(rd_req->r_base_oid.name, sizeof(rd_req->r_base_oid.name),
1666                  "%llx.00000000", ci->i_vino.ino);
1667         rd_req->r_base_oid.name_len = strlen(rd_req->r_base_oid.name);
1668
1669         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1670                                          ceph_empty_snapc,
1671                                          1, false, GFP_NOFS);
1672         if (!wr_req) {
1673                 err = -ENOMEM;
1674                 goto out_unlock;
1675         }
1676
1677         wr_req->r_flags = CEPH_OSD_FLAG_WRITE |
1678                           CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK;
1679         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1680         wr_req->r_base_oloc.pool = pool;
1681         wr_req->r_base_oid = rd_req->r_base_oid;
1682
1683         /* one page should be large enough for STAT data */
1684         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1685         if (IS_ERR(pages)) {
1686                 err = PTR_ERR(pages);
1687                 goto out_unlock;
1688         }
1689
1690         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1691                                      0, false, true);
1692         ceph_osdc_build_request(rd_req, 0, NULL, CEPH_NOSNAP,
1693                                 &ci->vfs_inode.i_mtime);
1694         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1695
1696         ceph_osdc_build_request(wr_req, 0, NULL, CEPH_NOSNAP,
1697                                 &ci->vfs_inode.i_mtime);
1698         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1699
1700         if (!err)
1701                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1702         if (!err2)
1703                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1704
1705         if (err >= 0 || err == -ENOENT)
1706                 have |= POOL_READ;
1707         else if (err != -EPERM)
1708                 goto out_unlock;
1709
1710         if (err2 == 0 || err2 == -EEXIST)
1711                 have |= POOL_WRITE;
1712         else if (err2 != -EPERM) {
1713                 err = err2;
1714                 goto out_unlock;
1715         }
1716
1717         perm = kmalloc(sizeof(*perm), GFP_NOFS);
1718         if (!perm) {
1719                 err = -ENOMEM;
1720                 goto out_unlock;
1721         }
1722
1723         perm->pool = pool;
1724         perm->perm = have;
1725         rb_link_node(&perm->node, parent, p);
1726         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1727         err = 0;
1728 out_unlock:
1729         up_write(&mdsc->pool_perm_rwsem);
1730
1731         if (rd_req)
1732                 ceph_osdc_put_request(rd_req);
1733         if (wr_req)
1734                 ceph_osdc_put_request(wr_req);
1735 out:
1736         if (!err)
1737                 err = have;
1738         dout("__ceph_pool_perm_get pool %u result = %d\n", pool, err);
1739         return err;
1740 }
1741
1742 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1743 {
1744         u32 pool;
1745         int ret, flags;
1746
1747         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1748                                 NOPOOLPERM))
1749                 return 0;
1750
1751         spin_lock(&ci->i_ceph_lock);
1752         flags = ci->i_ceph_flags;
1753         pool = ceph_file_layout_pg_pool(ci->i_layout);
1754         spin_unlock(&ci->i_ceph_lock);
1755 check:
1756         if (flags & CEPH_I_POOL_PERM) {
1757                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1758                         dout("ceph_pool_perm_check pool %u no read perm\n",
1759                              pool);
1760                         return -EPERM;
1761                 }
1762                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1763                         dout("ceph_pool_perm_check pool %u no write perm\n",
1764                              pool);
1765                         return -EPERM;
1766                 }
1767                 return 0;
1768         }
1769
1770         ret = __ceph_pool_perm_get(ci, pool);
1771         if (ret < 0)
1772                 return ret;
1773
1774         flags = CEPH_I_POOL_PERM;
1775         if (ret & POOL_READ)
1776                 flags |= CEPH_I_POOL_RD;
1777         if (ret & POOL_WRITE)
1778                 flags |= CEPH_I_POOL_WR;
1779
1780         spin_lock(&ci->i_ceph_lock);
1781         if (pool == ceph_file_layout_pg_pool(ci->i_layout)) {
1782                 ci->i_ceph_flags = flags;
1783         } else {
1784                 pool = ceph_file_layout_pg_pool(ci->i_layout);
1785                 flags = ci->i_ceph_flags;
1786         }
1787         spin_unlock(&ci->i_ceph_lock);
1788         goto check;
1789 }
1790
1791 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1792 {
1793         struct ceph_pool_perm *perm;
1794         struct rb_node *n;
1795
1796         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1797                 n = rb_first(&mdsc->pool_perm_tree);
1798                 perm = rb_entry(n, struct ceph_pool_perm, node);
1799                 rb_erase(n, &mdsc->pool_perm_tree);
1800                 kfree(perm);
1801         }
1802 }