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 && rc != ENOENT)
 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_KERNEL);
 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_KERNEL)) {
 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                                                     loff_t *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         loff_t snap_size = -1;
 484         long writeback_stat;
 485         u64 truncate_size;
 486         u32 truncate_seq;
 487         int err = 0, len = PAGE_CACHE_SIZE;
 488
 489         dout("writepage %p idx %lu\n", page, page->index);
 490
 491         if (!page->mapping || !page->mapping->host) {
 492                 dout("writepage %p - no mapping\n", page);
 493                 return -EFAULT;
 494         }
 495         inode = page->mapping->host;
 496         ci = ceph_inode(inode);
 497         fsc = ceph_inode_to_client(inode);
 498         osdc = &fsc->client->osdc;
 499
 500         /* verify this is a writeable snap context */
 501         snapc = page_snap_context(page);
 502         if (snapc == NULL) {
 503                 dout("writepage %p page %p not dirty?\n", inode, page);
 504                 goto out;
 505         }
 506         oldest = get_oldest_context(inode, &snap_size);
 507         if (snapc->seq > oldest->seq) {
 508                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 509                      inode, page, snapc);
 510                 /* we should only noop if called by kswapd */
 511                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 512                 ceph_put_snap_context(oldest);
 513                 goto out;
 514         }
 515         ceph_put_snap_context(oldest);
 516
 517         spin_lock(&ci->i_ceph_lock);
 518         truncate_seq = ci->i_truncate_seq;
 519         truncate_size = ci->i_truncate_size;
 520         if (snap_size == -1)
 521                 snap_size = i_size_read(inode);
 522         spin_unlock(&ci->i_ceph_lock);
 523
 524         /* is this a partial page at end of file? */
 525         if (page_off >= snap_size) {
 526                 dout("%p page eof %llu\n", page, snap_size);
 527                 goto out;
 528         }
 529         if (snap_size < page_off + len)
 530                 len = snap_size - page_off;
 531
 532         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 533              inode, page, page->index, page_off, len, snapc);
 534
 535         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 536         if (writeback_stat >
 537             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 538                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 539
 540         ceph_readpage_to_fscache(inode, page);
 541
 542         set_page_writeback(page);
 543         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 544                                    &ci->i_layout, snapc,
 545                                    page_off, len,
 546                                    truncate_seq, truncate_size,
 547                                    &inode->i_mtime, &page, 1);
 548         if (err < 0) {
 549                 dout("writepage setting page/mapping error %d %p\n", err, page);
 550                 SetPageError(page);
 551                 mapping_set_error(&inode->i_data, err);
 552                 if (wbc)
 553                         wbc->pages_skipped++;
 554         } else {
 555                 dout("writepage cleaned page %p\n", page);
 556                 err = 0;  /* vfs expects us to return 0 */
 557         }
 558         page->private = 0;
 559         ClearPagePrivate(page);
 560         end_page_writeback(page);
 561         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 562         ceph_put_snap_context(snapc);  /* page's reference */
 563 out:
 564         return err;
 565 }
 566
 567 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 568 {
 569         int err;
 570         struct inode *inode = page->mapping->host;
 571         BUG_ON(!inode);
 572         ihold(inode);
 573         err = writepage_nounlock(page, wbc);
 574         unlock_page(page);
 575         iput(inode);
 576         return err;
 577 }
 578
 579
 580 /*
 581  * lame release_pages helper.  release_pages() isn't exported to
 582  * modules.
 583  */
 584 static void ceph_release_pages(struct page **pages, int num)
 585 {
 586         struct pagevec pvec;
 587         int i;
 588
 589         pagevec_init(&pvec, 0);
 590         for (i = 0; i < num; i++) {
 591                 if (pagevec_add(&pvec, pages[i]) == 0)
 592                         pagevec_release(&pvec);
 593         }
 594         pagevec_release(&pvec);
 595 }
 596
 597 /*
 598  * async writeback completion handler.
 599  *
 600  * If we get an error, set the mapping error bit, but not the individual
 601  * page error bits.
 602  */
 603 static void writepages_finish(struct ceph_osd_request *req,
 604                               struct ceph_msg *msg)
 605 {
 606         struct inode *inode = req->r_inode;
 607         struct ceph_inode_info *ci = ceph_inode(inode);
 608         struct ceph_osd_data *osd_data;
 609         unsigned wrote;
 610         struct page *page;
 611         int num_pages;
 612         int i;
 613         struct ceph_snap_context *snapc = req->r_snapc;
 614         struct address_space *mapping = inode->i_mapping;
 615         int rc = req->r_result;
 616         u64 bytes = req->r_ops[0].extent.length;
 617         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 618         long writeback_stat;
 619         unsigned issued = ceph_caps_issued(ci);
 620
 621         osd_data = osd_req_op_extent_osd_data(req, 0);
 622         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 623         num_pages = calc_pages_for((u64)osd_data->alignment,
 624                                         (u64)osd_data->length);
 625         if (rc >= 0) {
 626                 /*
 627                  * Assume we wrote the pages we originally sent.  The
 628                  * osd might reply with fewer pages if our writeback
 629                  * raced with a truncation and was adjusted at the osd,
 630                  * so don't believe the reply.
 631                  */
 632                 wrote = num_pages;
 633         } else {
 634                 wrote = 0;
 635                 mapping_set_error(mapping, rc);
 636         }
 637         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 638              inode, rc, bytes, wrote);
 639
 640         /* clean all pages */
 641         for (i = 0; i < num_pages; i++) {
 642                 page = osd_data->pages[i];
 643                 BUG_ON(!page);
 644                 WARN_ON(!PageUptodate(page));
 645
 646                 writeback_stat =
 647                         atomic_long_dec_return(&fsc->writeback_count);
 648                 if (writeback_stat <
 649                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 650                         clear_bdi_congested(&fsc->backing_dev_info,
 651                                             BLK_RW_ASYNC);
 652
 653                 ceph_put_snap_context(page_snap_context(page));
 654                 page->private = 0;
 655                 ClearPagePrivate(page);
 656                 dout("unlocking %d %p\n", i, page);
 657                 end_page_writeback(page);
 658
 659                 /*
 660                  * We lost the cache cap, need to truncate the page before
 661                  * it is unlocked, otherwise we'd truncate it later in the
 662                  * page truncation thread, possibly losing some data that
 663                  * raced its way in
 664                  */
 665                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 666                         generic_error_remove_page(inode->i_mapping, page);
 667
 668                 unlock_page(page);
 669         }
 670         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 671         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 672
 673         ceph_release_pages(osd_data->pages, num_pages);
 674         if (osd_data->pages_from_pool)
 675                 mempool_free(osd_data->pages,
 676                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 677         else
 678                 kfree(osd_data->pages);
 679         ceph_osdc_put_request(req);
 680 }
 681
 682 /*
 683  * initiate async writeback
 684  */
 685 static int ceph_writepages_start(struct address_space *mapping,
 686                                  struct writeback_control *wbc)
 687 {
 688         struct inode *inode = mapping->host;
 689         struct ceph_inode_info *ci = ceph_inode(inode);
 690         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 691         struct ceph_vino vino = ceph_vino(inode);
 692         pgoff_t index, start, end;
 693         int range_whole = 0;
 694         int should_loop = 1;
 695         pgoff_t max_pages = 0, max_pages_ever = 0;
 696         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 697         struct pagevec pvec;
 698         int done = 0;
 699         int rc = 0;
 700         unsigned wsize = 1 << inode->i_blkbits;
 701         struct ceph_osd_request *req = NULL;
 702         int do_sync = 0;
 703         loff_t snap_size, i_size;
 704         u64 truncate_size;
 705         u32 truncate_seq;
 706
 707         /*
 708          * Include a 'sync' in the OSD request if this is a data
 709          * integrity write (e.g., O_SYNC write or fsync()), or if our
 710          * cap is being revoked.
 711          */
 712         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 713                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 714                 do_sync = 1;
 715         dout("writepages_start %p dosync=%d (mode=%s)\n",
 716              inode, do_sync,
 717              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 718              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 719
 720         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 721                 pr_warn("writepage_start %p on forced umount\n", inode);
 722                 truncate_pagecache(inode, 0);
 723                 mapping_set_error(mapping, -EIO);
 724                 return -EIO; /* we're in a forced umount, don't write! */
 725         }
 726         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 727                 wsize = fsc->mount_options->wsize;
 728         if (wsize < PAGE_CACHE_SIZE)
 729                 wsize = PAGE_CACHE_SIZE;
 730         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 731
 732         pagevec_init(&pvec, 0);
 733
 734         /* where to start/end? */
 735         if (wbc->range_cyclic) {
 736                 start = mapping->writeback_index; /* Start from prev offset */
 737                 end = -1;
 738                 dout(" cyclic, start at %lu\n", start);
 739         } else {
 740                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 741                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 742                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 743                         range_whole = 1;
 744                 should_loop = 0;
 745                 dout(" not cyclic, %lu to %lu\n", start, end);
 746         }
 747         index = start;
 748
 749 retry:
 750         /* find oldest snap context with dirty data */
 751         ceph_put_snap_context(snapc);
 752         snap_size = -1;
 753         snapc = get_oldest_context(inode, &snap_size);
 754         if (!snapc) {
 755                 /* hmm, why does writepages get called when there
 756                    is no dirty data? */
 757                 dout(" no snap context with dirty data?\n");
 758                 goto out;
 759         }
 760         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 761              snapc, snapc->seq, snapc->num_snaps);
 762
 763         spin_lock(&ci->i_ceph_lock);
 764         truncate_seq = ci->i_truncate_seq;
 765         truncate_size = ci->i_truncate_size;
 766         i_size = i_size_read(inode);
 767         spin_unlock(&ci->i_ceph_lock);
 768
 769         if (last_snapc && snapc != last_snapc) {
 770                 /* if we switched to a newer snapc, restart our scan at the
 771                  * start of the original file range. */
 772                 dout("  snapc differs from last pass, restarting at %lu\n",
 773                      index);
 774                 index = start;
 775         }
 776         last_snapc = snapc;
 777
 778         while (!done && index <= end) {
 779                 unsigned i;
 780                 int first;
 781                 pgoff_t next;
 782                 int pvec_pages, locked_pages;
 783                 struct page **pages = NULL;
 784                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 785                 struct page *page;
 786                 int want;
 787                 u64 offset, len;
 788                 long writeback_stat;
 789
 790                 next = 0;
 791                 locked_pages = 0;
 792                 max_pages = max_pages_ever;
 793
 794 get_more_pages:
 795                 first = -1;
 796                 want = min(end - index,
 797                            min((pgoff_t)PAGEVEC_SIZE,
 798                                max_pages - (pgoff_t)locked_pages) - 1)
 799                         + 1;
 800                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 801                                                 PAGECACHE_TAG_DIRTY,
 802                                                 want);
 803                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 804                 if (!pvec_pages && !locked_pages)
 805                         break;
 806                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 807                         page = pvec.pages[i];
 808                         dout("? %p idx %lu\n", page, page->index);
 809                         if (locked_pages == 0)
 810                                 lock_page(page);  /* first page */
 811                         else if (!trylock_page(page))
 812                                 break;
 813
 814                         /* only dirty pages, or our accounting breaks */
 815                         if (unlikely(!PageDirty(page)) ||
 816                             unlikely(page->mapping != mapping)) {
 817                                 dout("!dirty or !mapping %p\n", page);
 818                                 unlock_page(page);
 819                                 break;
 820                         }
 821                         if (!wbc->range_cyclic && page->index > end) {
 822                                 dout("end of range %p\n", page);
 823                                 done = 1;
 824                                 unlock_page(page);
 825                                 break;
 826                         }
 827                         if (next && (page->index != next)) {
 828                                 dout("not consecutive %p\n", page);
 829                                 unlock_page(page);
 830                                 break;
 831                         }
 832                         if (wbc->sync_mode != WB_SYNC_NONE) {
 833                                 dout("waiting on writeback %p\n", page);
 834                                 wait_on_page_writeback(page);
 835                         }
 836                         if (page_offset(page) >=
 837                             (snap_size == -1 ? i_size : snap_size)) {
 838                                 dout("%p page eof %llu\n", page,
 839                                      (snap_size == -1 ? i_size : snap_size));
 840                                 done = 1;
 841                                 unlock_page(page);
 842                                 break;
 843                         }
 844                         if (PageWriteback(page)) {
 845                                 dout("%p under writeback\n", page);
 846                                 unlock_page(page);
 847                                 break;
 848                         }
 849
 850                         /* only if matching snap context */
 851                         pgsnapc = page_snap_context(page);
 852                         if (pgsnapc->seq > snapc->seq) {
 853                                 dout("page snapc %p %lld > oldest %p %lld\n",
 854                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 855                                 unlock_page(page);
 856                                 if (!locked_pages)
 857                                         continue; /* keep looking for snap */
 858                                 break;
 859                         }
 860
 861                         if (!clear_page_dirty_for_io(page)) {
 862                                 dout("%p !clear_page_dirty_for_io\n", page);
 863                                 unlock_page(page);
 864                                 break;
 865                         }
 866
 867                         /*
 868                          * We have something to write.  If this is
 869                          * the first locked page this time through,
 870                          * allocate an osd request and a page array
 871                          * that it will use.
 872                          */
 873                         if (locked_pages == 0) {
 874                                 BUG_ON(pages);
 875                                 /* prepare async write request */
 876                                 offset = (u64)page_offset(page);
 877                                 len = wsize;
 878                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 879                                                         &ci->i_layout, vino,
 880                                                         offset, &len, 0,
 881                                                         do_sync ? 2 : 1,
 882                                                         CEPH_OSD_OP_WRITE,
 883                                                         CEPH_OSD_FLAG_WRITE |
 884                                                         CEPH_OSD_FLAG_ONDISK,
 885                                                         snapc, truncate_seq,
 886                                                         truncate_size, true);
 887                                 if (IS_ERR(req)) {
 888                                         rc = PTR_ERR(req);
 889                                         unlock_page(page);
 890                                         break;
 891                                 }
 892
 893                                 if (do_sync)
 894                                         osd_req_op_init(req, 1,
 895                                                         CEPH_OSD_OP_STARTSYNC, 0);
 896
 897                                 req->r_callback = writepages_finish;
 898                                 req->r_inode = inode;
 899
 900                                 max_pages = calc_pages_for(0, (u64)len);
 901                                 pages = kmalloc(max_pages * sizeof (*pages),
 902                                                 GFP_NOFS);
 903                                 if (!pages) {
 904                                         pool = fsc->wb_pagevec_pool;
 905                                         pages = mempool_alloc(pool, GFP_NOFS);
 906                                         BUG_ON(!pages);
 907                                 }
 908                         }
 909
 910                         /* note position of first page in pvec */
 911                         if (first < 0)
 912                                 first = i;
 913                         dout("%p will write page %p idx %lu\n",
 914                              inode, page, page->index);
 915
 916                         writeback_stat =
 917                                atomic_long_inc_return(&fsc->writeback_count);
 918                         if (writeback_stat > CONGESTION_ON_THRESH(
 919                                     fsc->mount_options->congestion_kb)) {
 920                                 set_bdi_congested(&fsc->backing_dev_info,
 921                                                   BLK_RW_ASYNC);
 922                         }
 923
 924                         set_page_writeback(page);
 925                         pages[locked_pages] = page;
 926                         locked_pages++;
 927                         next = page->index + 1;
 928                 }
 929
 930                 /* did we get anything? */
 931                 if (!locked_pages)
 932                         goto release_pvec_pages;
 933                 if (i) {
 934                         int j;
 935                         BUG_ON(!locked_pages || first < 0);
 936
 937                         if (pvec_pages && i == pvec_pages &&
 938                             locked_pages < max_pages) {
 939                                 dout("reached end pvec, trying for more\n");
 940                                 pagevec_reinit(&pvec);
 941                                 goto get_more_pages;
 942                         }
 943
 944                         /* shift unused pages over in the pvec...  we
 945                          * will need to release them below. */
 946                         for (j = i; j < pvec_pages; j++) {
 947                                 dout(" pvec leftover page %p\n",
 948                                      pvec.pages[j]);
 949                                 pvec.pages[j-i+first] = pvec.pages[j];
 950                         }
 951                         pvec.nr -= i-first;
 952                 }
 953
 954                 /* Format the osd request message and submit the write */
 955                 offset = page_offset(pages[0]);
 956                 len = (u64)locked_pages << PAGE_CACHE_SHIFT;
 957                 if (snap_size == -1) {
 958                         len = min(len, (u64)i_size_read(inode) - offset);
 959                          /* writepages_finish() clears writeback pages
 960                           * according to the data length, so make sure
 961                           * data length covers all locked pages */
 962                         len = max(len, 1 +
 963                                 ((u64)(locked_pages - 1) << PAGE_CACHE_SHIFT));
 964                 } else {
 965                         len = min(len, snap_size - offset);
 966                 }
 967                 dout("writepages got %d pages at %llu~%llu\n",
 968                      locked_pages, offset, len);
 969
 970                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 971                                                         !!pool, false);
 972
 973                 pages = NULL;   /* request message now owns the pages array */
 974                 pool = NULL;
 975
 976                 /* Update the write op length in case we changed it */
 977
 978                 osd_req_op_extent_update(req, 0, len);
 979
 980                 vino = ceph_vino(inode);
 981                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 982                                         &inode->i_mtime);
 983
 984                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 985                 BUG_ON(rc);
 986                 req = NULL;
 987
 988                 /* continue? */
 989                 index = next;
 990                 wbc->nr_to_write -= locked_pages;
 991                 if (wbc->nr_to_write <= 0)
 992                         done = 1;
 993
 994 release_pvec_pages:
 995                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 996                      pvec.nr ? pvec.pages[0] : NULL);
 997                 pagevec_release(&pvec);
 998
 999                 if (locked_pages && !done)
1000                         goto retry;
1001         }
1002
1003         if (should_loop && !done) {
1004                 /* more to do; loop back to beginning of file */
1005                 dout("writepages looping back to beginning of file\n");
1006                 should_loop = 0;
1007                 index = 0;
1008                 goto retry;
1009         }
1010
1011         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1012                 mapping->writeback_index = index;
1013
1014 out:
1015         if (req)
1016                 ceph_osdc_put_request(req);
1017         ceph_put_snap_context(snapc);
1018         dout("writepages done, rc = %d\n", rc);
1019         return rc;
1020 }
1021
1022
1023
1024 /*
1025  * See if a given @snapc is either writeable, or already written.
1026  */
1027 static int context_is_writeable_or_written(struct inode *inode,
1028                                            struct ceph_snap_context *snapc)
1029 {
1030         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1031         int ret = !oldest || snapc->seq <= oldest->seq;
1032
1033         ceph_put_snap_context(oldest);
1034         return ret;
1035 }
1036
1037 /*
1038  * We are only allowed to write into/dirty the page if the page is
1039  * clean, or already dirty within the same snap context.
1040  *
1041  * called with page locked.
1042  * return success with page locked,
1043  * or any failure (incl -EAGAIN) with page unlocked.
1044  */
1045 static int ceph_update_writeable_page(struct file *file,
1046                             loff_t pos, unsigned len,
1047                             struct page *page)
1048 {
1049         struct inode *inode = file_inode(file);
1050         struct ceph_inode_info *ci = ceph_inode(inode);
1051         loff_t page_off = pos & PAGE_CACHE_MASK;
1052         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1053         int end_in_page = pos_in_page + len;
1054         loff_t i_size;
1055         int r;
1056         struct ceph_snap_context *snapc, *oldest;
1057
1058 retry_locked:
1059         /* writepages currently holds page lock, but if we change that later, */
1060         wait_on_page_writeback(page);
1061
1062         snapc = page_snap_context(page);
1063         if (snapc && snapc != ci->i_head_snapc) {
1064                 /*
1065                  * this page is already dirty in another (older) snap
1066                  * context!  is it writeable now?
1067                  */
1068                 oldest = get_oldest_context(inode, NULL);
1069
1070                 if (snapc->seq > oldest->seq) {
1071                         ceph_put_snap_context(oldest);
1072                         dout(" page %p snapc %p not current or oldest\n",
1073                              page, snapc);
1074                         /*
1075                          * queue for writeback, and wait for snapc to
1076                          * be writeable or written
1077                          */
1078                         snapc = ceph_get_snap_context(snapc);
1079                         unlock_page(page);
1080                         ceph_queue_writeback(inode);
1081                         r = wait_event_interruptible(ci->i_cap_wq,
1082                                context_is_writeable_or_written(inode, snapc));
1083                         ceph_put_snap_context(snapc);
1084                         if (r == -ERESTARTSYS)
1085                                 return r;
1086                         return -EAGAIN;
1087                 }
1088                 ceph_put_snap_context(oldest);
1089
1090                 /* yay, writeable, do it now (without dropping page lock) */
1091                 dout(" page %p snapc %p not current, but oldest\n",
1092                      page, snapc);
1093                 if (!clear_page_dirty_for_io(page))
1094                         goto retry_locked;
1095                 r = writepage_nounlock(page, NULL);
1096                 if (r < 0)
1097                         goto fail_nosnap;
1098                 goto retry_locked;
1099         }
1100
1101         if (PageUptodate(page)) {
1102                 dout(" page %p already uptodate\n", page);
1103                 return 0;
1104         }
1105
1106         /* full page? */
1107         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1108                 return 0;
1109
1110         /* past end of file? */
1111         i_size = i_size_read(inode);
1112
1113         if (page_off >= i_size ||
1114             (pos_in_page == 0 && (pos+len) >= i_size &&
1115              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1116                 dout(" zeroing %p 0 - %d and %d - %d\n",
1117                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1118                 zero_user_segments(page,
1119                                    0, pos_in_page,
1120                                    end_in_page, PAGE_CACHE_SIZE);
1121                 return 0;
1122         }
1123
1124         /* we need to read it. */
1125         r = readpage_nounlock(file, page);
1126         if (r < 0)
1127                 goto fail_nosnap;
1128         goto retry_locked;
1129 fail_nosnap:
1130         unlock_page(page);
1131         return r;
1132 }
1133
1134 /*
1135  * We are only allowed to write into/dirty the page if the page is
1136  * clean, or already dirty within the same snap context.
1137  */
1138 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1139                             loff_t pos, unsigned len, unsigned flags,
1140                             struct page **pagep, void **fsdata)
1141 {
1142         struct inode *inode = file_inode(file);
1143         struct page *page;
1144         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1145         int r;
1146
1147         do {
1148                 /* get a page */
1149                 page = grab_cache_page_write_begin(mapping, index, 0);
1150                 if (!page)
1151                         return -ENOMEM;
1152
1153                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1154                      inode, page, (int)pos, (int)len);
1155
1156                 r = ceph_update_writeable_page(file, pos, len, page);
1157                 if (r < 0)
1158                         page_cache_release(page);
1159                 else
1160                         *pagep = page;
1161         } while (r == -EAGAIN);
1162
1163         return r;
1164 }
1165
1166 /*
1167  * we don't do anything in here that simple_write_end doesn't do
1168  * except adjust dirty page accounting
1169  */
1170 static int ceph_write_end(struct file *file, struct address_space *mapping,
1171                           loff_t pos, unsigned len, unsigned copied,
1172                           struct page *page, void *fsdata)
1173 {
1174         struct inode *inode = file_inode(file);
1175         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1176         int check_cap = 0;
1177
1178         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1179              inode, page, (int)pos, (int)copied, (int)len);
1180
1181         /* zero the stale part of the page if we did a short copy */
1182         if (copied < len)
1183                 zero_user_segment(page, from+copied, len);
1184
1185         /* did file size increase? */
1186         if (pos+copied > i_size_read(inode))
1187                 check_cap = ceph_inode_set_size(inode, pos+copied);
1188
1189         if (!PageUptodate(page))
1190                 SetPageUptodate(page);
1191
1192         set_page_dirty(page);
1193
1194         unlock_page(page);
1195         page_cache_release(page);
1196
1197         if (check_cap)
1198                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1199
1200         return copied;
1201 }
1202
1203 /*
1204  * we set .direct_IO to indicate direct io is supported, but since we
1205  * intercept O_DIRECT reads and writes early, this function should
1206  * never get called.
1207  */
1208 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter,
1209                               loff_t pos)
1210 {
1211         WARN_ON(1);
1212         return -EINVAL;
1213 }
1214
1215 const struct address_space_operations ceph_aops = {
1216         .readpage = ceph_readpage,
1217         .readpages = ceph_readpages,
1218         .writepage = ceph_writepage,
1219         .writepages = ceph_writepages_start,
1220         .write_begin = ceph_write_begin,
1221         .write_end = ceph_write_end,
1222         .set_page_dirty = ceph_set_page_dirty,
1223         .invalidatepage = ceph_invalidatepage,
1224         .releasepage = ceph_releasepage,
1225         .direct_IO = ceph_direct_io,
1226 };
1227
1228
1229 /*
1230  * vm ops
1231  */
1232 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1233 {
1234         struct inode *inode = file_inode(vma->vm_file);
1235         struct ceph_inode_info *ci = ceph_inode(inode);
1236         struct ceph_file_info *fi = vma->vm_file->private_data;
1237         struct page *pinned_page = NULL;
1238         loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1239         int want, got, ret;
1240
1241         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1242              inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1243         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1244                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1245         else
1246                 want = CEPH_CAP_FILE_CACHE;
1247         while (1) {
1248                 got = 0;
1249                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1250                                     -1, &got, &pinned_page);
1251                 if (ret == 0)
1252                         break;
1253                 if (ret != -ERESTARTSYS) {
1254                         WARN_ON(1);
1255                         return VM_FAULT_SIGBUS;
1256                 }
1257         }
1258         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1259              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1260
1261         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1262             ci->i_inline_version == CEPH_INLINE_NONE)
1263                 ret = filemap_fault(vma, vmf);
1264         else
1265                 ret = -EAGAIN;
1266
1267         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1268              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1269         if (pinned_page)
1270                 page_cache_release(pinned_page);
1271         ceph_put_cap_refs(ci, got);
1272
1273         if (ret != -EAGAIN)
1274                 return ret;
1275
1276         /* read inline data */
1277         if (off >= PAGE_CACHE_SIZE) {
1278                 /* does not support inline data > PAGE_SIZE */
1279                 ret = VM_FAULT_SIGBUS;
1280         } else {
1281                 int ret1;
1282                 struct address_space *mapping = inode->i_mapping;
1283                 struct page *page = find_or_create_page(mapping, 0,
1284                                                 mapping_gfp_constraint(mapping,
1285                                                 ~__GFP_FS));
1286                 if (!page) {
1287                         ret = VM_FAULT_OOM;
1288                         goto out;
1289                 }
1290                 ret1 = __ceph_do_getattr(inode, page,
1291                                          CEPH_STAT_CAP_INLINE_DATA, true);
1292                 if (ret1 < 0 || off >= i_size_read(inode)) {
1293                         unlock_page(page);
1294                         page_cache_release(page);
1295                         ret = VM_FAULT_SIGBUS;
1296                         goto out;
1297                 }
1298                 if (ret1 < PAGE_CACHE_SIZE)
1299                         zero_user_segment(page, ret1, PAGE_CACHE_SIZE);
1300                 else
1301                         flush_dcache_page(page);
1302                 SetPageUptodate(page);
1303                 vmf->page = page;
1304                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1305         }
1306 out:
1307         dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1308              inode, off, (size_t)PAGE_CACHE_SIZE, ret);
1309         return ret;
1310 }
1311
1312 /*
1313  * Reuse write_begin here for simplicity.
1314  */
1315 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1316 {
1317         struct inode *inode = file_inode(vma->vm_file);
1318         struct ceph_inode_info *ci = ceph_inode(inode);
1319         struct ceph_file_info *fi = vma->vm_file->private_data;
1320         struct ceph_cap_flush *prealloc_cf;
1321         struct page *page = vmf->page;
1322         loff_t off = page_offset(page);
1323         loff_t size = i_size_read(inode);
1324         size_t len;
1325         int want, got, ret;
1326
1327         prealloc_cf = ceph_alloc_cap_flush();
1328         if (!prealloc_cf)
1329                 return VM_FAULT_SIGBUS;
1330
1331         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1332                 struct page *locked_page = NULL;
1333                 if (off == 0) {
1334                         lock_page(page);
1335                         locked_page = page;
1336                 }
1337                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1338                 if (locked_page)
1339                         unlock_page(locked_page);
1340                 if (ret < 0) {
1341                         ret = VM_FAULT_SIGBUS;
1342                         goto out_free;
1343                 }
1344         }
1345
1346         if (off + PAGE_CACHE_SIZE <= size)
1347                 len = PAGE_CACHE_SIZE;
1348         else
1349                 len = size & ~PAGE_CACHE_MASK;
1350
1351         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1352              inode, ceph_vinop(inode), off, len, size);
1353         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1354                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1355         else
1356                 want = CEPH_CAP_FILE_BUFFER;
1357         while (1) {
1358                 got = 0;
1359                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1360                                     &got, NULL);
1361                 if (ret == 0)
1362                         break;
1363                 if (ret != -ERESTARTSYS) {
1364                         WARN_ON(1);
1365                         ret = VM_FAULT_SIGBUS;
1366                         goto out_free;
1367                 }
1368         }
1369         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1370              inode, off, len, ceph_cap_string(got));
1371
1372         /* Update time before taking page lock */
1373         file_update_time(vma->vm_file);
1374
1375         lock_page(page);
1376
1377         ret = VM_FAULT_NOPAGE;
1378         if ((off > size) ||
1379             (page->mapping != inode->i_mapping)) {
1380                 unlock_page(page);
1381                 goto out;
1382         }
1383
1384         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1385         if (ret >= 0) {
1386                 /* success.  we'll keep the page locked. */
1387                 set_page_dirty(page);
1388                 ret = VM_FAULT_LOCKED;
1389         } else {
1390                 if (ret == -ENOMEM)
1391                         ret = VM_FAULT_OOM;
1392                 else
1393                         ret = VM_FAULT_SIGBUS;
1394         }
1395 out:
1396         if (ret == VM_FAULT_LOCKED ||
1397             ci->i_inline_version != CEPH_INLINE_NONE) {
1398                 int dirty;
1399                 spin_lock(&ci->i_ceph_lock);
1400                 ci->i_inline_version = CEPH_INLINE_NONE;
1401                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1402                                                &prealloc_cf);
1403                 spin_unlock(&ci->i_ceph_lock);
1404                 if (dirty)
1405                         __mark_inode_dirty(inode, dirty);
1406         }
1407
1408         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1409              inode, off, len, ceph_cap_string(got), ret);
1410         ceph_put_cap_refs(ci, got);
1411 out_free:
1412         ceph_free_cap_flush(prealloc_cf);
1413
1414         return ret;
1415 }
1416
1417 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1418                            char *data, size_t len)
1419 {
1420         struct address_space *mapping = inode->i_mapping;
1421         struct page *page;
1422
1423         if (locked_page) {
1424                 page = locked_page;
1425         } else {
1426                 if (i_size_read(inode) == 0)
1427                         return;
1428                 page = find_or_create_page(mapping, 0,
1429                                            mapping_gfp_constraint(mapping,
1430                                            ~__GFP_FS));
1431                 if (!page)
1432                         return;
1433                 if (PageUptodate(page)) {
1434                         unlock_page(page);
1435                         page_cache_release(page);
1436                         return;
1437                 }
1438         }
1439
1440         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1441              inode, ceph_vinop(inode), len, locked_page);
1442
1443         if (len > 0) {
1444                 void *kaddr = kmap_atomic(page);
1445                 memcpy(kaddr, data, len);
1446                 kunmap_atomic(kaddr);
1447         }
1448
1449         if (page != locked_page) {
1450                 if (len < PAGE_CACHE_SIZE)
1451                         zero_user_segment(page, len, PAGE_CACHE_SIZE);
1452                 else
1453                         flush_dcache_page(page);
1454
1455                 SetPageUptodate(page);
1456                 unlock_page(page);
1457                 page_cache_release(page);
1458         }
1459 }
1460
1461 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1462 {
1463         struct inode *inode = file_inode(filp);
1464         struct ceph_inode_info *ci = ceph_inode(inode);
1465         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1466         struct ceph_osd_request *req;
1467         struct page *page = NULL;
1468         u64 len, inline_version;
1469         int err = 0;
1470         bool from_pagecache = false;
1471
1472         spin_lock(&ci->i_ceph_lock);
1473         inline_version = ci->i_inline_version;
1474         spin_unlock(&ci->i_ceph_lock);
1475
1476         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1477              inode, ceph_vinop(inode), inline_version);
1478
1479         if (inline_version == 1 || /* initial version, no data */
1480             inline_version == CEPH_INLINE_NONE)
1481                 goto out;
1482
1483         if (locked_page) {
1484                 page = locked_page;
1485                 WARN_ON(!PageUptodate(page));
1486         } else if (ceph_caps_issued(ci) &
1487                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1488                 page = find_get_page(inode->i_mapping, 0);
1489                 if (page) {
1490                         if (PageUptodate(page)) {
1491                                 from_pagecache = true;
1492                                 lock_page(page);
1493                         } else {
1494                                 page_cache_release(page);
1495                                 page = NULL;
1496                         }
1497                 }
1498         }
1499
1500         if (page) {
1501                 len = i_size_read(inode);
1502                 if (len > PAGE_CACHE_SIZE)
1503                         len = PAGE_CACHE_SIZE;
1504         } else {
1505                 page = __page_cache_alloc(GFP_NOFS);
1506                 if (!page) {
1507                         err = -ENOMEM;
1508                         goto out;
1509                 }
1510                 err = __ceph_do_getattr(inode, page,
1511                                         CEPH_STAT_CAP_INLINE_DATA, true);
1512                 if (err < 0) {
1513                         /* no inline data */
1514                         if (err == -ENODATA)
1515                                 err = 0;
1516                         goto out;
1517                 }
1518                 len = err;
1519         }
1520
1521         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1522                                     ceph_vino(inode), 0, &len, 0, 1,
1523                                     CEPH_OSD_OP_CREATE,
1524                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1525                                     ceph_empty_snapc, 0, 0, false);
1526         if (IS_ERR(req)) {
1527                 err = PTR_ERR(req);
1528                 goto out;
1529         }
1530
1531         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1532         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1533         if (!err)
1534                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1535         ceph_osdc_put_request(req);
1536         if (err < 0)
1537                 goto out;
1538
1539         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1540                                     ceph_vino(inode), 0, &len, 1, 3,
1541                                     CEPH_OSD_OP_WRITE,
1542                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1543                                     ceph_empty_snapc,
1544                                     ci->i_truncate_seq, ci->i_truncate_size,
1545                                     false);
1546         if (IS_ERR(req)) {
1547                 err = PTR_ERR(req);
1548                 goto out;
1549         }
1550
1551         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1552
1553         {
1554                 __le64 xattr_buf = cpu_to_le64(inline_version);
1555                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1556                                             "inline_version", &xattr_buf,
1557                                             sizeof(xattr_buf),
1558                                             CEPH_OSD_CMPXATTR_OP_GT,
1559                                             CEPH_OSD_CMPXATTR_MODE_U64);
1560                 if (err)
1561                         goto out_put;
1562         }
1563
1564         {
1565                 char xattr_buf[32];
1566                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1567                                          "%llu", inline_version);
1568                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1569                                             "inline_version",
1570                                             xattr_buf, xattr_len, 0, 0);
1571                 if (err)
1572                         goto out_put;
1573         }
1574
1575         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1576         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1577         if (!err)
1578                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1579 out_put:
1580         ceph_osdc_put_request(req);
1581         if (err == -ECANCELED)
1582                 err = 0;
1583 out:
1584         if (page && page != locked_page) {
1585                 if (from_pagecache) {
1586                         unlock_page(page);
1587                         page_cache_release(page);
1588                 } else
1589                         __free_pages(page, 0);
1590         }
1591
1592         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1593              inode, ceph_vinop(inode), inline_version, err);
1594         return err;
1595 }
1596
1597 static const struct vm_operations_struct ceph_vmops = {
1598         .fault          = ceph_filemap_fault,
1599         .page_mkwrite   = ceph_page_mkwrite,
1600 };
1601
1602 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1603 {
1604         struct address_space *mapping = file->f_mapping;
1605
1606         if (!mapping->a_ops->readpage)
1607                 return -ENOEXEC;
1608         file_accessed(file);
1609         vma->vm_ops = &ceph_vmops;
1610         return 0;
1611 }
1612
1613 enum {
1614         POOL_READ       = 1,
1615         POOL_WRITE      = 2,
1616 };
1617
1618 static int __ceph_pool_perm_get(struct ceph_inode_info *ci, u32 pool)
1619 {
1620         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1621         struct ceph_mds_client *mdsc = fsc->mdsc;
1622         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1623         struct rb_node **p, *parent;
1624         struct ceph_pool_perm *perm;
1625         struct page **pages;
1626         int err = 0, err2 = 0, have = 0;
1627
1628         down_read(&mdsc->pool_perm_rwsem);
1629         p = &mdsc->pool_perm_tree.rb_node;
1630         while (*p) {
1631                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1632                 if (pool < perm->pool)
1633                         p = &(*p)->rb_left;
1634                 else if (pool > perm->pool)
1635                         p = &(*p)->rb_right;
1636                 else {
1637                         have = perm->perm;
1638                         break;
1639                 }
1640         }
1641         up_read(&mdsc->pool_perm_rwsem);
1642         if (*p)
1643                 goto out;
1644
1645         dout("__ceph_pool_perm_get pool %u no perm cached\n", pool);
1646
1647         down_write(&mdsc->pool_perm_rwsem);
1648         parent = NULL;
1649         while (*p) {
1650                 parent = *p;
1651                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1652                 if (pool < perm->pool)
1653                         p = &(*p)->rb_left;
1654                 else if (pool > perm->pool)
1655                         p = &(*p)->rb_right;
1656                 else {
1657                         have = perm->perm;
1658                         break;
1659                 }
1660         }
1661         if (*p) {
1662                 up_write(&mdsc->pool_perm_rwsem);
1663                 goto out;
1664         }
1665
1666         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1667                                          ceph_empty_snapc,
1668                                          1, false, GFP_NOFS);
1669         if (!rd_req) {
1670                 err = -ENOMEM;
1671                 goto out_unlock;
1672         }
1673
1674         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1675         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1676         rd_req->r_base_oloc.pool = pool;
1677         snprintf(rd_req->r_base_oid.name, sizeof(rd_req->r_base_oid.name),
1678                  "%llx.00000000", ci->i_vino.ino);
1679         rd_req->r_base_oid.name_len = strlen(rd_req->r_base_oid.name);
1680
1681         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1682                                          ceph_empty_snapc,
1683                                          1, false, GFP_NOFS);
1684         if (!wr_req) {
1685                 err = -ENOMEM;
1686                 goto out_unlock;
1687         }
1688
1689         wr_req->r_flags = CEPH_OSD_FLAG_WRITE |
1690                           CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK;
1691         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1692         wr_req->r_base_oloc.pool = pool;
1693         wr_req->r_base_oid = rd_req->r_base_oid;
1694
1695         /* one page should be large enough for STAT data */
1696         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1697         if (IS_ERR(pages)) {
1698                 err = PTR_ERR(pages);
1699                 goto out_unlock;
1700         }
1701
1702         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1703                                      0, false, true);
1704         ceph_osdc_build_request(rd_req, 0, NULL, CEPH_NOSNAP,
1705                                 &ci->vfs_inode.i_mtime);
1706         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1707
1708         ceph_osdc_build_request(wr_req, 0, NULL, CEPH_NOSNAP,
1709                                 &ci->vfs_inode.i_mtime);
1710         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1711
1712         if (!err)
1713                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1714         if (!err2)
1715                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1716
1717         if (err >= 0 || err == -ENOENT)
1718                 have |= POOL_READ;
1719         else if (err != -EPERM)
1720                 goto out_unlock;
1721
1722         if (err2 == 0 || err2 == -EEXIST)
1723                 have |= POOL_WRITE;
1724         else if (err2 != -EPERM) {
1725                 err = err2;
1726                 goto out_unlock;
1727         }
1728
1729         perm = kmalloc(sizeof(*perm), GFP_NOFS);
1730         if (!perm) {
1731                 err = -ENOMEM;
1732                 goto out_unlock;
1733         }
1734
1735         perm->pool = pool;
1736         perm->perm = have;
1737         rb_link_node(&perm->node, parent, p);
1738         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1739         err = 0;
1740 out_unlock:
1741         up_write(&mdsc->pool_perm_rwsem);
1742
1743         if (rd_req)
1744                 ceph_osdc_put_request(rd_req);
1745         if (wr_req)
1746                 ceph_osdc_put_request(wr_req);
1747 out:
1748         if (!err)
1749                 err = have;
1750         dout("__ceph_pool_perm_get pool %u result = %d\n", pool, err);
1751         return err;
1752 }
1753
1754 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1755 {
1756         u32 pool;
1757         int ret, flags;
1758
1759         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1760                                 NOPOOLPERM))
1761                 return 0;
1762
1763         spin_lock(&ci->i_ceph_lock);
1764         flags = ci->i_ceph_flags;
1765         pool = ceph_file_layout_pg_pool(ci->i_layout);
1766         spin_unlock(&ci->i_ceph_lock);
1767 check:
1768         if (flags & CEPH_I_POOL_PERM) {
1769                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1770                         dout("ceph_pool_perm_check pool %u no read perm\n",
1771                              pool);
1772                         return -EPERM;
1773                 }
1774                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1775                         dout("ceph_pool_perm_check pool %u no write perm\n",
1776                              pool);
1777                         return -EPERM;
1778                 }
1779                 return 0;
1780         }
1781
1782         ret = __ceph_pool_perm_get(ci, pool);
1783         if (ret < 0)
1784                 return ret;
1785
1786         flags = CEPH_I_POOL_PERM;
1787         if (ret & POOL_READ)
1788                 flags |= CEPH_I_POOL_RD;
1789         if (ret & POOL_WRITE)
1790                 flags |= CEPH_I_POOL_WR;
1791
1792         spin_lock(&ci->i_ceph_lock);
1793         if (pool == ceph_file_layout_pg_pool(ci->i_layout)) {
1794                 ci->i_ceph_flags = flags;
1795         } else {
1796                 pool = ceph_file_layout_pg_pool(ci->i_layout);
1797                 flags = ci->i_ceph_flags;
1798         }
1799         spin_unlock(&ci->i_ceph_lock);
1800         goto check;
1801 }
1802
1803 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1804 {
1805         struct ceph_pool_perm *perm;
1806         struct rb_node *n;
1807
1808         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1809                 n = rb_first(&mdsc->pool_perm_tree);
1810                 perm = rb_entry(n, struct ceph_pool_perm, node);
1811                 rb_erase(n, &mdsc->pool_perm_tree);
1812                 kfree(perm);
1813         }
1814 }