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