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_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         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 179              page, page->index, PageDirty(page) ? "" : "not ");
 180
 181         /* Can we release the page from the cache? */
 182         if (!ceph_release_fscache_page(page, g))
 183                 return 0;
 184
 185         return !PagePrivate(page);
 186 }
 187
 188 /*
 189  * read a single page, without unlocking it.
 190  */
 191 static int readpage_nounlock(struct file *filp, struct page *page)
 192 {
 193         struct inode *inode = file_inode(filp);
 194         struct ceph_inode_info *ci = ceph_inode(inode);
 195         struct ceph_osd_client *osdc =
 196                 &ceph_inode_to_client(inode)->client->osdc;
 197         int err = 0;
 198         u64 off = page_offset(page);
 199         u64 len = PAGE_SIZE;
 200
 201         if (off >= i_size_read(inode)) {
 202                 zero_user_segment(page, 0, PAGE_SIZE);
 203                 SetPageUptodate(page);
 204                 return 0;
 205         }
 206
 207         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 208                 /*
 209                  * Uptodate inline data should have been added
 210                  * into page cache while getting Fcr caps.
 211                  */
 212                 if (off == 0)
 213                         return -EINVAL;
 214                 zero_user_segment(page, 0, PAGE_SIZE);
 215                 SetPageUptodate(page);
 216                 return 0;
 217         }
 218
 219         err = ceph_readpage_from_fscache(inode, page);
 220         if (err == 0)
 221                 goto out;
 222
 223         dout("readpage inode %p file %p page %p index %lu\n",
 224              inode, filp, page, page->index);
 225         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 226                                   off, &len,
 227                                   ci->i_truncate_seq, ci->i_truncate_size,
 228                                   &page, 1, 0);
 229         if (err == -ENOENT)
 230                 err = 0;
 231         if (err < 0) {
 232                 SetPageError(page);
 233                 ceph_fscache_readpage_cancel(inode, page);
 234                 goto out;
 235         }
 236         if (err < PAGE_SIZE)
 237                 /* zero fill remainder of page */
 238                 zero_user_segment(page, err, PAGE_SIZE);
 239         else
 240                 flush_dcache_page(page);
 241
 242         SetPageUptodate(page);
 243         ceph_readpage_to_fscache(inode, page);
 244
 245 out:
 246         return err < 0 ? err : 0;
 247 }
 248
 249 static int ceph_readpage(struct file *filp, struct page *page)
 250 {
 251         int r = readpage_nounlock(filp, page);
 252         unlock_page(page);
 253         return r;
 254 }
 255
 256 /*
 257  * Finish an async read(ahead) op.
 258  */
 259 static void finish_read(struct ceph_osd_request *req)
 260 {
 261         struct inode *inode = req->r_inode;
 262         struct ceph_osd_data *osd_data;
 263         int rc = req->r_result <= 0 ? req->r_result : 0;
 264         int bytes = req->r_result >= 0 ? req->r_result : 0;
 265         int num_pages;
 266         int i;
 267
 268         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 269
 270         /* unlock all pages, zeroing any data we didn't read */
 271         osd_data = osd_req_op_extent_osd_data(req, 0);
 272         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 273         num_pages = calc_pages_for((u64)osd_data->alignment,
 274                                         (u64)osd_data->length);
 275         for (i = 0; i < num_pages; i++) {
 276                 struct page *page = osd_data->pages[i];
 277
 278                 if (rc < 0 && rc != -ENOENT) {
 279                         ceph_fscache_readpage_cancel(inode, page);
 280                         goto unlock;
 281                 }
 282                 if (bytes < (int)PAGE_SIZE) {
 283                         /* zero (remainder of) page */
 284                         int s = bytes < 0 ? 0 : bytes;
 285                         zero_user_segment(page, s, PAGE_SIZE);
 286                 }
 287                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 288                      page->index);
 289                 flush_dcache_page(page);
 290                 SetPageUptodate(page);
 291                 ceph_readpage_to_fscache(inode, page);
 292 unlock:
 293                 unlock_page(page);
 294                 put_page(page);
 295                 bytes -= PAGE_SIZE;
 296         }
 297         kfree(osd_data->pages);
 298 }
 299
 300 /*
 301  * start an async read(ahead) operation.  return nr_pages we submitted
 302  * a read for on success, or negative error code.
 303  */
 304 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 305 {
 306         struct ceph_osd_client *osdc =
 307                 &ceph_inode_to_client(inode)->client->osdc;
 308         struct ceph_inode_info *ci = ceph_inode(inode);
 309         struct page *page = list_entry(page_list->prev, struct page, lru);
 310         struct ceph_vino vino;
 311         struct ceph_osd_request *req;
 312         u64 off;
 313         u64 len;
 314         int i;
 315         struct page **pages;
 316         pgoff_t next_index;
 317         int nr_pages = 0;
 318         int got = 0;
 319         int ret = 0;
 320
 321         if (!current->journal_info) {
 322                 /* caller of readpages does not hold buffer and read caps
 323                  * (fadvise, madvise and readahead cases) */
 324                 int want = CEPH_CAP_FILE_CACHE;
 325                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 326                 if (ret < 0) {
 327                         dout("start_read %p, error getting cap\n", inode);
 328                 } else if (!(got & want)) {
 329                         dout("start_read %p, no cache cap\n", inode);
 330                         ret = 0;
 331                 }
 332                 if (ret <= 0) {
 333                         if (got)
 334                                 ceph_put_cap_refs(ci, got);
 335                         while (!list_empty(page_list)) {
 336                                 page = list_entry(page_list->prev,
 337                                                   struct page, lru);
 338                                 list_del(&page->lru);
 339                                 put_page(page);
 340                         }
 341                         return ret;
 342                 }
 343         }
 344
 345         off = (u64) page_offset(page);
 346
 347         /* count pages */
 348         next_index = page->index;
 349         list_for_each_entry_reverse(page, page_list, lru) {
 350                 if (page->index != next_index)
 351                         break;
 352                 nr_pages++;
 353                 next_index++;
 354                 if (max && nr_pages == max)
 355                         break;
 356         }
 357         len = nr_pages << PAGE_SHIFT;
 358         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 359              off, len);
 360         vino = ceph_vino(inode);
 361         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 362                                     0, 1, CEPH_OSD_OP_READ,
 363                                     CEPH_OSD_FLAG_READ, NULL,
 364                                     ci->i_truncate_seq, ci->i_truncate_size,
 365                                     false);
 366         if (IS_ERR(req)) {
 367                 ret = PTR_ERR(req);
 368                 goto out;
 369         }
 370
 371         /* build page vector */
 372         nr_pages = calc_pages_for(0, len);
 373         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 374         if (!pages) {
 375                 ret = -ENOMEM;
 376                 goto out_put;
 377         }
 378         for (i = 0; i < nr_pages; ++i) {
 379                 page = list_entry(page_list->prev, struct page, lru);
 380                 BUG_ON(PageLocked(page));
 381                 list_del(&page->lru);
 382
 383                 dout("start_read %p adding %p idx %lu\n", inode, page,
 384                      page->index);
 385                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 386                                           GFP_KERNEL)) {
 387                         ceph_fscache_uncache_page(inode, page);
 388                         put_page(page);
 389                         dout("start_read %p add_to_page_cache failed %p\n",
 390                              inode, page);
 391                         nr_pages = i;
 392                         if (nr_pages > 0) {
 393                                 len = nr_pages << PAGE_SHIFT;
 394                                 break;
 395                         }
 396                         goto out_pages;
 397                 }
 398                 pages[i] = page;
 399         }
 400         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 401         req->r_callback = finish_read;
 402         req->r_inode = inode;
 403
 404         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 405         ret = ceph_osdc_start_request(osdc, req, false);
 406         if (ret < 0)
 407                 goto out_pages;
 408         ceph_osdc_put_request(req);
 409
 410         /* After adding locked pages to page cache, the inode holds cache cap.
 411          * So we can drop our cap refs. */
 412         if (got)
 413                 ceph_put_cap_refs(ci, got);
 414
 415         return nr_pages;
 416
 417 out_pages:
 418         for (i = 0; i < nr_pages; ++i) {
 419                 ceph_fscache_readpage_cancel(inode, pages[i]);
 420                 unlock_page(pages[i]);
 421         }
 422         ceph_put_page_vector(pages, nr_pages, false);
 423 out_put:
 424         ceph_osdc_put_request(req);
 425 out:
 426         if (got)
 427                 ceph_put_cap_refs(ci, got);
 428         return ret;
 429 }
 430
 431
 432 /*
 433  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 434  * the caller (VM) cleans them up.
 435  */
 436 static int ceph_readpages(struct file *file, struct address_space *mapping,
 437                           struct list_head *page_list, unsigned nr_pages)
 438 {
 439         struct inode *inode = file_inode(file);
 440         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 441         int rc = 0;
 442         int max = 0;
 443
 444         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 445                 return -EINVAL;
 446
 447         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 448                                          &nr_pages);
 449
 450         if (rc == 0)
 451                 goto out;
 452
 453         if (fsc->mount_options->rsize >= PAGE_SIZE)
 454                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 455                         >> PAGE_SHIFT;
 456
 457         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 458                 file, nr_pages,
 459              max);
 460         while (!list_empty(page_list)) {
 461                 rc = start_read(inode, page_list, max);
 462                 if (rc < 0)
 463                         goto out;
 464         }
 465 out:
 466         ceph_fscache_readpages_cancel(inode, page_list);
 467
 468         dout("readpages %p file %p ret %d\n", inode, file, rc);
 469         return rc;
 470 }
 471
 472 /*
 473  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 474  * only snap context we are allowed to write back.
 475  */
 476 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 477                                                     loff_t *snap_size,
 478                                                     u64 *truncate_size,
 479                                                     u32 *truncate_seq)
 480 {
 481         struct ceph_inode_info *ci = ceph_inode(inode);
 482         struct ceph_snap_context *snapc = NULL;
 483         struct ceph_cap_snap *capsnap = NULL;
 484
 485         spin_lock(&ci->i_ceph_lock);
 486         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 487                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 488                      capsnap->context, capsnap->dirty_pages);
 489                 if (capsnap->dirty_pages) {
 490                         snapc = ceph_get_snap_context(capsnap->context);
 491                         if (snap_size)
 492                                 *snap_size = capsnap->size;
 493                         if (truncate_size)
 494                                 *truncate_size = capsnap->truncate_size;
 495                         if (truncate_seq)
 496                                 *truncate_seq = capsnap->truncate_seq;
 497                         break;
 498                 }
 499         }
 500         if (!snapc && ci->i_wrbuffer_ref_head) {
 501                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 502                 dout(" head snapc %p has %d dirty pages\n",
 503                      snapc, ci->i_wrbuffer_ref_head);
 504                 if (truncate_size)
 505                         *truncate_size = capsnap->truncate_size;
 506                 if (truncate_seq)
 507                         *truncate_seq = capsnap->truncate_seq;
 508         }
 509         spin_unlock(&ci->i_ceph_lock);
 510         return snapc;
 511 }
 512
 513 /*
 514  * Write a single page, but leave the page locked.
 515  *
 516  * If we get a write error, set the page error bit, but still adjust the
 517  * dirty page accounting (i.e., page is no longer dirty).
 518  */
 519 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 520 {
 521         struct inode *inode;
 522         struct ceph_inode_info *ci;
 523         struct ceph_fs_client *fsc;
 524         struct ceph_osd_client *osdc;
 525         struct ceph_snap_context *snapc, *oldest;
 526         loff_t page_off = page_offset(page);
 527         loff_t snap_size = -1;
 528         long writeback_stat;
 529         u64 truncate_size;
 530         u32 truncate_seq;
 531         int err = 0, len = PAGE_SIZE;
 532
 533         dout("writepage %p idx %lu\n", page, page->index);
 534
 535         if (!page->mapping || !page->mapping->host) {
 536                 dout("writepage %p - no mapping\n", page);
 537                 return -EFAULT;
 538         }
 539         inode = page->mapping->host;
 540         ci = ceph_inode(inode);
 541         fsc = ceph_inode_to_client(inode);
 542         osdc = &fsc->client->osdc;
 543
 544         /* verify this is a writeable snap context */
 545         snapc = page_snap_context(page);
 546         if (snapc == NULL) {
 547                 dout("writepage %p page %p not dirty?\n", inode, page);
 548                 goto out;
 549         }
 550         oldest = get_oldest_context(inode, &snap_size,
 551                                     &truncate_size, &truncate_seq);
 552         if (snapc->seq > oldest->seq) {
 553                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 554                      inode, page, snapc);
 555                 /* we should only noop if called by kswapd */
 556                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 557                 ceph_put_snap_context(oldest);
 558                 goto out;
 559         }
 560         ceph_put_snap_context(oldest);
 561
 562         if (snap_size == -1)
 563                 snap_size = i_size_read(inode);
 564
 565         /* is this a partial page at end of file? */
 566         if (page_off >= snap_size) {
 567                 dout("%p page eof %llu\n", page, snap_size);
 568                 goto out;
 569         }
 570         if (snap_size < page_off + len)
 571                 len = snap_size - page_off;
 572
 573         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 574              inode, page, page->index, page_off, len, snapc);
 575
 576         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 577         if (writeback_stat >
 578             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 579                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 580
 581         set_page_writeback(page);
 582         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 583                                    &ci->i_layout, snapc,
 584                                    page_off, len,
 585                                    truncate_seq, truncate_size,
 586                                    &inode->i_mtime, &page, 1);
 587         if (err < 0) {
 588                 struct writeback_control tmp_wbc;
 589                 if (!wbc)
 590                         wbc = &tmp_wbc;
 591                 if (err == -ERESTARTSYS) {
 592                         /* killed by SIGKILL */
 593                         dout("writepage interrupted page %p\n", page);
 594                         redirty_page_for_writepage(wbc, page);
 595                         end_page_writeback(page);
 596                         goto out;
 597                 }
 598                 dout("writepage setting page/mapping error %d %p\n",
 599                      err, page);
 600                 SetPageError(page);
 601                 mapping_set_error(&inode->i_data, err);
 602                 wbc->pages_skipped++;
 603         } else {
 604                 dout("writepage cleaned page %p\n", page);
 605                 err = 0;  /* vfs expects us to return 0 */
 606         }
 607         page->private = 0;
 608         ClearPagePrivate(page);
 609         end_page_writeback(page);
 610         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 611         ceph_put_snap_context(snapc);  /* page's reference */
 612 out:
 613         return err;
 614 }
 615
 616 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 617 {
 618         int err;
 619         struct inode *inode = page->mapping->host;
 620         BUG_ON(!inode);
 621         ihold(inode);
 622         err = writepage_nounlock(page, wbc);
 623         if (err == -ERESTARTSYS) {
 624                 /* direct memory reclaimer was killed by SIGKILL. return 0
 625                  * to prevent caller from setting mapping/page error */
 626                 err = 0;
 627         }
 628         unlock_page(page);
 629         iput(inode);
 630         return err;
 631 }
 632
 633 /*
 634  * lame release_pages helper.  release_pages() isn't exported to
 635  * modules.
 636  */
 637 static void ceph_release_pages(struct page **pages, int num)
 638 {
 639         struct pagevec pvec;
 640         int i;
 641
 642         pagevec_init(&pvec, 0);
 643         for (i = 0; i < num; i++) {
 644                 if (pagevec_add(&pvec, pages[i]) == 0)
 645                         pagevec_release(&pvec);
 646         }
 647         pagevec_release(&pvec);
 648 }
 649
 650 /*
 651  * async writeback completion handler.
 652  *
 653  * If we get an error, set the mapping error bit, but not the individual
 654  * page error bits.
 655  */
 656 static void writepages_finish(struct ceph_osd_request *req)
 657 {
 658         struct inode *inode = req->r_inode;
 659         struct ceph_inode_info *ci = ceph_inode(inode);
 660         struct ceph_osd_data *osd_data;
 661         struct page *page;
 662         int num_pages, total_pages = 0;
 663         int i, j;
 664         int rc = req->r_result;
 665         struct ceph_snap_context *snapc = req->r_snapc;
 666         struct address_space *mapping = inode->i_mapping;
 667         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 668         bool remove_page;
 669
 670         dout("writepages_finish %p rc %d\n", inode, rc);
 671         if (rc < 0)
 672                 mapping_set_error(mapping, rc);
 673
 674         /*
 675          * We lost the cache cap, need to truncate the page before
 676          * it is unlocked, otherwise we'd truncate it later in the
 677          * page truncation thread, possibly losing some data that
 678          * raced its way in
 679          */
 680         remove_page = !(ceph_caps_issued(ci) &
 681                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 682
 683         /* clean all pages */
 684         for (i = 0; i < req->r_num_ops; i++) {
 685                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 686                         break;
 687
 688                 osd_data = osd_req_op_extent_osd_data(req, i);
 689                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 690                 num_pages = calc_pages_for((u64)osd_data->alignment,
 691                                            (u64)osd_data->length);
 692                 total_pages += num_pages;
 693                 for (j = 0; j < num_pages; j++) {
 694                         page = osd_data->pages[j];
 695                         BUG_ON(!page);
 696                         WARN_ON(!PageUptodate(page));
 697
 698                         if (atomic_long_dec_return(&fsc->writeback_count) <
 699                              CONGESTION_OFF_THRESH(
 700                                         fsc->mount_options->congestion_kb))
 701                                 clear_bdi_congested(&fsc->backing_dev_info,
 702                                                     BLK_RW_ASYNC);
 703
 704                         if (rc < 0)
 705                                 SetPageError(page);
 706
 707                         ceph_put_snap_context(page_snap_context(page));
 708                         page->private = 0;
 709                         ClearPagePrivate(page);
 710                         dout("unlocking %p\n", page);
 711                         end_page_writeback(page);
 712
 713                         if (remove_page)
 714                                 generic_error_remove_page(inode->i_mapping,
 715                                                           page);
 716
 717                         unlock_page(page);
 718                 }
 719                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 720                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 721
 722                 ceph_release_pages(osd_data->pages, num_pages);
 723         }
 724
 725         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 726
 727         osd_data = osd_req_op_extent_osd_data(req, 0);
 728         if (osd_data->pages_from_pool)
 729                 mempool_free(osd_data->pages,
 730                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 731         else
 732                 kfree(osd_data->pages);
 733         ceph_osdc_put_request(req);
 734 }
 735
 736 /*
 737  * initiate async writeback
 738  */
 739 static int ceph_writepages_start(struct address_space *mapping,
 740                                  struct writeback_control *wbc)
 741 {
 742         struct inode *inode = mapping->host;
 743         struct ceph_inode_info *ci = ceph_inode(inode);
 744         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 745         struct ceph_vino vino = ceph_vino(inode);
 746         pgoff_t index, start, end;
 747         int range_whole = 0;
 748         int should_loop = 1;
 749         pgoff_t max_pages = 0, max_pages_ever = 0;
 750         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 751         struct pagevec pvec;
 752         int done = 0;
 753         int rc = 0;
 754         unsigned wsize = 1 << inode->i_blkbits;
 755         struct ceph_osd_request *req = NULL;
 756         int do_sync = 0;
 757         loff_t snap_size, i_size;
 758         u64 truncate_size;
 759         u32 truncate_seq;
 760
 761         /*
 762          * Include a 'sync' in the OSD request if this is a data
 763          * integrity write (e.g., O_SYNC write or fsync()), or if our
 764          * cap is being revoked.
 765          */
 766         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 767                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 768                 do_sync = 1;
 769         dout("writepages_start %p dosync=%d (mode=%s)\n",
 770              inode, do_sync,
 771              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 772              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 773
 774         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 775                 if (ci->i_wrbuffer_ref > 0) {
 776                         pr_warn_ratelimited(
 777                                 "writepage_start %p %lld forced umount\n",
 778                                 inode, ceph_ino(inode));
 779                 }
 780                 mapping_set_error(mapping, -EIO);
 781                 return -EIO; /* we're in a forced umount, don't write! */
 782         }
 783         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 784                 wsize = fsc->mount_options->wsize;
 785         if (wsize < PAGE_SIZE)
 786                 wsize = PAGE_SIZE;
 787         max_pages_ever = wsize >> PAGE_SHIFT;
 788
 789         pagevec_init(&pvec, 0);
 790
 791         /* where to start/end? */
 792         if (wbc->range_cyclic) {
 793                 start = mapping->writeback_index; /* Start from prev offset */
 794                 end = -1;
 795                 dout(" cyclic, start at %lu\n", start);
 796         } else {
 797                 start = wbc->range_start >> PAGE_SHIFT;
 798                 end = wbc->range_end >> PAGE_SHIFT;
 799                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 800                         range_whole = 1;
 801                 should_loop = 0;
 802                 dout(" not cyclic, %lu to %lu\n", start, end);
 803         }
 804         index = start;
 805
 806 retry:
 807         /* find oldest snap context with dirty data */
 808         ceph_put_snap_context(snapc);
 809         snap_size = -1;
 810         snapc = get_oldest_context(inode, &snap_size,
 811                                    &truncate_size, &truncate_seq);
 812         if (!snapc) {
 813                 /* hmm, why does writepages get called when there
 814                    is no dirty data? */
 815                 dout(" no snap context with dirty data?\n");
 816                 goto out;
 817         }
 818         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 819              snapc, snapc->seq, snapc->num_snaps);
 820
 821         i_size = i_size_read(inode);
 822
 823         if (last_snapc && snapc != last_snapc) {
 824                 /* if we switched to a newer snapc, restart our scan at the
 825                  * start of the original file range. */
 826                 dout("  snapc differs from last pass, restarting at %lu\n",
 827                      index);
 828                 index = start;
 829         }
 830         last_snapc = snapc;
 831
 832         while (!done && index <= end) {
 833                 unsigned i;
 834                 int first;
 835                 pgoff_t strip_unit_end = 0;
 836                 int num_ops = 0, op_idx;
 837                 int pvec_pages, locked_pages = 0;
 838                 struct page **pages = NULL, **data_pages;
 839                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 840                 struct page *page;
 841                 int want;
 842                 u64 offset = 0, len = 0;
 843
 844                 max_pages = max_pages_ever;
 845
 846 get_more_pages:
 847                 first = -1;
 848                 want = min(end - index,
 849                            min((pgoff_t)PAGEVEC_SIZE,
 850                                max_pages - (pgoff_t)locked_pages) - 1)
 851                         + 1;
 852                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 853                                                 PAGECACHE_TAG_DIRTY,
 854                                                 want);
 855                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 856                 if (!pvec_pages && !locked_pages)
 857                         break;
 858                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 859                         page = pvec.pages[i];
 860                         dout("? %p idx %lu\n", page, page->index);
 861                         if (locked_pages == 0)
 862                                 lock_page(page);  /* first page */
 863                         else if (!trylock_page(page))
 864                                 break;
 865
 866                         /* only dirty pages, or our accounting breaks */
 867                         if (unlikely(!PageDirty(page)) ||
 868                             unlikely(page->mapping != mapping)) {
 869                                 dout("!dirty or !mapping %p\n", page);
 870                                 unlock_page(page);
 871                                 break;
 872                         }
 873                         if (!wbc->range_cyclic && page->index > end) {
 874                                 dout("end of range %p\n", page);
 875                                 done = 1;
 876                                 unlock_page(page);
 877                                 break;
 878                         }
 879                         if (strip_unit_end && (page->index > strip_unit_end)) {
 880                                 dout("end of strip unit %p\n", page);
 881                                 unlock_page(page);
 882                                 break;
 883                         }
 884                         if (wbc->sync_mode != WB_SYNC_NONE) {
 885                                 dout("waiting on writeback %p\n", page);
 886                                 wait_on_page_writeback(page);
 887                         }
 888                         if (page_offset(page) >=
 889                             (snap_size == -1 ? i_size : snap_size)) {
 890                                 dout("%p page eof %llu\n", page,
 891                                      (snap_size == -1 ? i_size : snap_size));
 892                                 done = 1;
 893                                 unlock_page(page);
 894                                 break;
 895                         }
 896                         if (PageWriteback(page)) {
 897                                 dout("%p under writeback\n", page);
 898                                 unlock_page(page);
 899                                 break;
 900                         }
 901
 902                         /* only if matching snap context */
 903                         pgsnapc = page_snap_context(page);
 904                         if (pgsnapc->seq > snapc->seq) {
 905                                 dout("page snapc %p %lld > oldest %p %lld\n",
 906                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 907                                 unlock_page(page);
 908                                 if (!locked_pages)
 909                                         continue; /* keep looking for snap */
 910                                 break;
 911                         }
 912
 913                         if (!clear_page_dirty_for_io(page)) {
 914                                 dout("%p !clear_page_dirty_for_io\n", page);
 915                                 unlock_page(page);
 916                                 break;
 917                         }
 918
 919                         /*
 920                          * We have something to write.  If this is
 921                          * the first locked page this time through,
 922                          * calculate max possinle write size and
 923                          * allocate a page array
 924                          */
 925                         if (locked_pages == 0) {
 926                                 u64 objnum;
 927                                 u64 objoff;
 928
 929                                 /* prepare async write request */
 930                                 offset = (u64)page_offset(page);
 931                                 len = wsize;
 932
 933                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 934                                                                 offset, len,
 935                                                                 &objnum, &objoff,
 936                                                                 &len);
 937                                 if (rc < 0) {
 938                                         unlock_page(page);
 939                                         break;
 940                                 }
 941
 942                                 num_ops = 1 + do_sync;
 943                                 strip_unit_end = page->index +
 944                                         ((len - 1) >> PAGE_SHIFT);
 945
 946                                 BUG_ON(pages);
 947                                 max_pages = calc_pages_for(0, (u64)len);
 948                                 pages = kmalloc(max_pages * sizeof (*pages),
 949                                                 GFP_NOFS);
 950                                 if (!pages) {
 951                                         pool = fsc->wb_pagevec_pool;
 952                                         pages = mempool_alloc(pool, GFP_NOFS);
 953                                         BUG_ON(!pages);
 954                                 }
 955
 956                                 len = 0;
 957                         } else if (page->index !=
 958                                    (offset + len) >> PAGE_SHIFT) {
 959                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 960                                                         CEPH_OSD_MAX_OPS)) {
 961                                         redirty_page_for_writepage(wbc, page);
 962                                         unlock_page(page);
 963                                         break;
 964                                 }
 965
 966                                 num_ops++;
 967                                 offset = (u64)page_offset(page);
 968                                 len = 0;
 969                         }
 970
 971                         /* note position of first page in pvec */
 972                         if (first < 0)
 973                                 first = i;
 974                         dout("%p will write page %p idx %lu\n",
 975                              inode, page, page->index);
 976
 977                         if (atomic_long_inc_return(&fsc->writeback_count) >
 978                             CONGESTION_ON_THRESH(
 979                                     fsc->mount_options->congestion_kb)) {
 980                                 set_bdi_congested(&fsc->backing_dev_info,
 981                                                   BLK_RW_ASYNC);
 982                         }
 983
 984                         pages[locked_pages] = page;
 985                         locked_pages++;
 986                         len += PAGE_SIZE;
 987                 }
 988
 989                 /* did we get anything? */
 990                 if (!locked_pages)
 991                         goto release_pvec_pages;
 992                 if (i) {
 993                         int j;
 994                         BUG_ON(!locked_pages || first < 0);
 995
 996                         if (pvec_pages && i == pvec_pages &&
 997                             locked_pages < max_pages) {
 998                                 dout("reached end pvec, trying for more\n");
 999                                 pagevec_reinit(&pvec);
1000                                 goto get_more_pages;
1001                         }
1002
1003                         /* shift unused pages over in the pvec...  we
1004                          * will need to release them below. */
1005                         for (j = i; j < pvec_pages; j++) {
1006                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
1007                                 pvec.pages[j-i+first] = pvec.pages[j];
1008                         }
1009                         pvec.nr -= i-first;
1010                 }
1011
1012 new_request:
1013                 offset = page_offset(pages[0]);
1014                 len = wsize;
1015
1016                 req = ceph_osdc_new_request(&fsc->client->osdc,
1017                                         &ci->i_layout, vino,
1018                                         offset, &len, 0, num_ops,
1019                                         CEPH_OSD_OP_WRITE,
1020                                         CEPH_OSD_FLAG_WRITE |
1021                                         CEPH_OSD_FLAG_ONDISK,
1022                                         snapc, truncate_seq,
1023                                         truncate_size, false);
1024                 if (IS_ERR(req)) {
1025                         req = ceph_osdc_new_request(&fsc->client->osdc,
1026                                                 &ci->i_layout, vino,
1027                                                 offset, &len, 0,
1028                                                 min(num_ops,
1029                                                     CEPH_OSD_SLAB_OPS),
1030                                                 CEPH_OSD_OP_WRITE,
1031                                                 CEPH_OSD_FLAG_WRITE |
1032                                                 CEPH_OSD_FLAG_ONDISK,
1033                                                 snapc, truncate_seq,
1034                                                 truncate_size, true);
1035                         BUG_ON(IS_ERR(req));
1036                 }
1037                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1038                              PAGE_SIZE - offset);
1039
1040                 req->r_callback = writepages_finish;
1041                 req->r_inode = inode;
1042
1043                 /* Format the osd request message and submit the write */
1044                 len = 0;
1045                 data_pages = pages;
1046                 op_idx = 0;
1047                 for (i = 0; i < locked_pages; i++) {
1048                         u64 cur_offset = page_offset(pages[i]);
1049                         if (offset + len != cur_offset) {
1050                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1051                                         break;
1052                                 osd_req_op_extent_dup_last(req, op_idx,
1053                                                            cur_offset - offset);
1054                                 dout("writepages got pages at %llu~%llu\n",
1055                                      offset, len);
1056                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1057                                                         data_pages, len, 0,
1058                                                         !!pool, false);
1059                                 osd_req_op_extent_update(req, op_idx, len);
1060
1061                                 len = 0;
1062                                 offset = cur_offset;
1063                                 data_pages = pages + i;
1064                                 op_idx++;
1065                         }
1066
1067                         set_page_writeback(pages[i]);
1068                         len += PAGE_SIZE;
1069                 }
1070
1071                 if (snap_size != -1) {
1072                         len = min(len, snap_size - offset);
1073                 } else if (i == locked_pages) {
1074                         /* writepages_finish() clears writeback pages
1075                          * according to the data length, so make sure
1076                          * data length covers all locked pages */
1077                         u64 min_len = len + 1 - PAGE_SIZE;
1078                         len = min(len, (u64)i_size_read(inode) - offset);
1079                         len = max(len, min_len);
1080                 }
1081                 dout("writepages got pages at %llu~%llu\n", offset, len);
1082
1083                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1084                                                  0, !!pool, false);
1085                 osd_req_op_extent_update(req, op_idx, len);
1086
1087                 if (do_sync) {
1088                         op_idx++;
1089                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1090                 }
1091                 BUG_ON(op_idx + 1 != req->r_num_ops);
1092
1093                 pool = NULL;
1094                 if (i < locked_pages) {
1095                         BUG_ON(num_ops <= req->r_num_ops);
1096                         num_ops -= req->r_num_ops;
1097                         num_ops += do_sync;
1098                         locked_pages -= i;
1099
1100                         /* allocate new pages array for next request */
1101                         data_pages = pages;
1102                         pages = kmalloc(locked_pages * sizeof (*pages),
1103                                         GFP_NOFS);
1104                         if (!pages) {
1105                                 pool = fsc->wb_pagevec_pool;
1106                                 pages = mempool_alloc(pool, GFP_NOFS);
1107                                 BUG_ON(!pages);
1108                         }
1109                         memcpy(pages, data_pages + i,
1110                                locked_pages * sizeof(*pages));
1111                         memset(data_pages + i, 0,
1112                                locked_pages * sizeof(*pages));
1113                 } else {
1114                         BUG_ON(num_ops != req->r_num_ops);
1115                         index = pages[i - 1]->index + 1;
1116                         /* request message now owns the pages array */
1117                         pages = NULL;
1118                 }
1119
1120                 req->r_mtime = inode->i_mtime;
1121                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1122                 BUG_ON(rc);
1123                 req = NULL;
1124
1125                 wbc->nr_to_write -= i;
1126                 if (pages)
1127                         goto new_request;
1128
1129                 if (wbc->nr_to_write <= 0)
1130                         done = 1;
1131
1132 release_pvec_pages:
1133                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1134                      pvec.nr ? pvec.pages[0] : NULL);
1135                 pagevec_release(&pvec);
1136
1137                 if (locked_pages && !done)
1138                         goto retry;
1139         }
1140
1141         if (should_loop && !done) {
1142                 /* more to do; loop back to beginning of file */
1143                 dout("writepages looping back to beginning of file\n");
1144                 should_loop = 0;
1145                 index = 0;
1146                 goto retry;
1147         }
1148
1149         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1150                 mapping->writeback_index = index;
1151
1152 out:
1153         ceph_osdc_put_request(req);
1154         ceph_put_snap_context(snapc);
1155         dout("writepages done, rc = %d\n", rc);
1156         return rc;
1157 }
1158
1159
1160
1161 /*
1162  * See if a given @snapc is either writeable, or already written.
1163  */
1164 static int context_is_writeable_or_written(struct inode *inode,
1165                                            struct ceph_snap_context *snapc)
1166 {
1167         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL,
1168                                                               NULL, NULL);
1169         int ret = !oldest || snapc->seq <= oldest->seq;
1170
1171         ceph_put_snap_context(oldest);
1172         return ret;
1173 }
1174
1175 /*
1176  * We are only allowed to write into/dirty the page if the page is
1177  * clean, or already dirty within the same snap context.
1178  *
1179  * called with page locked.
1180  * return success with page locked,
1181  * or any failure (incl -EAGAIN) with page unlocked.
1182  */
1183 static int ceph_update_writeable_page(struct file *file,
1184                             loff_t pos, unsigned len,
1185                             struct page *page)
1186 {
1187         struct inode *inode = file_inode(file);
1188         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1189         struct ceph_inode_info *ci = ceph_inode(inode);
1190         loff_t page_off = pos & PAGE_MASK;
1191         int pos_in_page = pos & ~PAGE_MASK;
1192         int end_in_page = pos_in_page + len;
1193         loff_t i_size;
1194         int r;
1195         struct ceph_snap_context *snapc, *oldest;
1196
1197         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1198                 dout(" page %p forced umount\n", page);
1199                 unlock_page(page);
1200                 return -EIO;
1201         }
1202
1203 retry_locked:
1204         /* writepages currently holds page lock, but if we change that later, */
1205         wait_on_page_writeback(page);
1206
1207         snapc = page_snap_context(page);
1208         if (snapc && snapc != ci->i_head_snapc) {
1209                 /*
1210                  * this page is already dirty in another (older) snap
1211                  * context!  is it writeable now?
1212                  */
1213                 oldest = get_oldest_context(inode, NULL, NULL, NULL);
1214
1215                 if (snapc->seq > oldest->seq) {
1216                         ceph_put_snap_context(oldest);
1217                         dout(" page %p snapc %p not current or oldest\n",
1218                              page, snapc);
1219                         /*
1220                          * queue for writeback, and wait for snapc to
1221                          * be writeable or written
1222                          */
1223                         snapc = ceph_get_snap_context(snapc);
1224                         unlock_page(page);
1225                         ceph_queue_writeback(inode);
1226                         r = wait_event_killable(ci->i_cap_wq,
1227                                context_is_writeable_or_written(inode, snapc));
1228                         ceph_put_snap_context(snapc);
1229                         if (r == -ERESTARTSYS)
1230                                 return r;
1231                         return -EAGAIN;
1232                 }
1233                 ceph_put_snap_context(oldest);
1234
1235                 /* yay, writeable, do it now (without dropping page lock) */
1236                 dout(" page %p snapc %p not current, but oldest\n",
1237                      page, snapc);
1238                 if (!clear_page_dirty_for_io(page))
1239                         goto retry_locked;
1240                 r = writepage_nounlock(page, NULL);
1241                 if (r < 0)
1242                         goto fail_nosnap;
1243                 goto retry_locked;
1244         }
1245
1246         if (PageUptodate(page)) {
1247                 dout(" page %p already uptodate\n", page);
1248                 return 0;
1249         }
1250
1251         /* full page? */
1252         if (pos_in_page == 0 && len == PAGE_SIZE)
1253                 return 0;
1254
1255         /* past end of file? */
1256         i_size = i_size_read(inode);
1257
1258         if (page_off >= i_size ||
1259             (pos_in_page == 0 && (pos+len) >= i_size &&
1260              end_in_page - pos_in_page != PAGE_SIZE)) {
1261                 dout(" zeroing %p 0 - %d and %d - %d\n",
1262                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1263                 zero_user_segments(page,
1264                                    0, pos_in_page,
1265                                    end_in_page, PAGE_SIZE);
1266                 return 0;
1267         }
1268
1269         /* we need to read it. */
1270         r = readpage_nounlock(file, page);
1271         if (r < 0)
1272                 goto fail_nosnap;
1273         goto retry_locked;
1274 fail_nosnap:
1275         unlock_page(page);
1276         return r;
1277 }
1278
1279 /*
1280  * We are only allowed to write into/dirty the page if the page is
1281  * clean, or already dirty within the same snap context.
1282  */
1283 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1284                             loff_t pos, unsigned len, unsigned flags,
1285                             struct page **pagep, void **fsdata)
1286 {
1287         struct inode *inode = file_inode(file);
1288         struct page *page;
1289         pgoff_t index = pos >> PAGE_SHIFT;
1290         int r;
1291
1292         do {
1293                 /* get a page */
1294                 page = grab_cache_page_write_begin(mapping, index, 0);
1295                 if (!page)
1296                         return -ENOMEM;
1297
1298                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1299                      inode, page, (int)pos, (int)len);
1300
1301                 r = ceph_update_writeable_page(file, pos, len, page);
1302                 if (r < 0)
1303                         put_page(page);
1304                 else
1305                         *pagep = page;
1306         } while (r == -EAGAIN);
1307
1308         return r;
1309 }
1310
1311 /*
1312  * we don't do anything in here that simple_write_end doesn't do
1313  * except adjust dirty page accounting
1314  */
1315 static int ceph_write_end(struct file *file, struct address_space *mapping,
1316                           loff_t pos, unsigned len, unsigned copied,
1317                           struct page *page, void *fsdata)
1318 {
1319         struct inode *inode = file_inode(file);
1320         int check_cap = 0;
1321
1322         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1323              inode, page, (int)pos, (int)copied, (int)len);
1324
1325         /* zero the stale part of the page if we did a short copy */
1326         if (!PageUptodate(page)) {
1327                 if (copied < len) {
1328                         copied = 0;
1329                         goto out;
1330                 }
1331                 SetPageUptodate(page);
1332         }
1333
1334         /* did file size increase? */
1335         if (pos+copied > i_size_read(inode))
1336                 check_cap = ceph_inode_set_size(inode, pos+copied);
1337
1338         set_page_dirty(page);
1339
1340 out:
1341         unlock_page(page);
1342         put_page(page);
1343
1344         if (check_cap)
1345                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1346
1347         return copied;
1348 }
1349
1350 /*
1351  * we set .direct_IO to indicate direct io is supported, but since we
1352  * intercept O_DIRECT reads and writes early, this function should
1353  * never get called.
1354  */
1355 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1356 {
1357         WARN_ON(1);
1358         return -EINVAL;
1359 }
1360
1361 const struct address_space_operations ceph_aops = {
1362         .readpage = ceph_readpage,
1363         .readpages = ceph_readpages,
1364         .writepage = ceph_writepage,
1365         .writepages = ceph_writepages_start,
1366         .write_begin = ceph_write_begin,
1367         .write_end = ceph_write_end,
1368         .set_page_dirty = ceph_set_page_dirty,
1369         .invalidatepage = ceph_invalidatepage,
1370         .releasepage = ceph_releasepage,
1371         .direct_IO = ceph_direct_io,
1372 };
1373
1374 static void ceph_block_sigs(sigset_t *oldset)
1375 {
1376         sigset_t mask;
1377         siginitsetinv(&mask, sigmask(SIGKILL));
1378         sigprocmask(SIG_BLOCK, &mask, oldset);
1379 }
1380
1381 static void ceph_restore_sigs(sigset_t *oldset)
1382 {
1383         sigprocmask(SIG_SETMASK, oldset, NULL);
1384 }
1385
1386 /*
1387  * vm ops
1388  */
1389 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1390 {
1391         struct inode *inode = file_inode(vma->vm_file);
1392         struct ceph_inode_info *ci = ceph_inode(inode);
1393         struct ceph_file_info *fi = vma->vm_file->private_data;
1394         struct page *pinned_page = NULL;
1395         loff_t off = vmf->pgoff << PAGE_SHIFT;
1396         int want, got, ret;
1397         sigset_t oldset;
1398
1399         ceph_block_sigs(&oldset);
1400
1401         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1402              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1403         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1404                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1405         else
1406                 want = CEPH_CAP_FILE_CACHE;
1407
1408         got = 0;
1409         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1410         if (ret < 0)
1411                 goto out_restore;
1412
1413         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1414              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1415
1416         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1417             ci->i_inline_version == CEPH_INLINE_NONE) {
1418                 current->journal_info = vma->vm_file;
1419                 ret = filemap_fault(vma, vmf);
1420                 current->journal_info = NULL;
1421         } else
1422                 ret = -EAGAIN;
1423
1424         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1425              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1426         if (pinned_page)
1427                 put_page(pinned_page);
1428         ceph_put_cap_refs(ci, got);
1429
1430         if (ret != -EAGAIN)
1431                 goto out_restore;
1432
1433         /* read inline data */
1434         if (off >= PAGE_SIZE) {
1435                 /* does not support inline data > PAGE_SIZE */
1436                 ret = VM_FAULT_SIGBUS;
1437         } else {
1438                 int ret1;
1439                 struct address_space *mapping = inode->i_mapping;
1440                 struct page *page = find_or_create_page(mapping, 0,
1441                                                 mapping_gfp_constraint(mapping,
1442                                                 ~__GFP_FS));
1443                 if (!page) {
1444                         ret = VM_FAULT_OOM;
1445                         goto out_inline;
1446                 }
1447                 ret1 = __ceph_do_getattr(inode, page,
1448                                          CEPH_STAT_CAP_INLINE_DATA, true);
1449                 if (ret1 < 0 || off >= i_size_read(inode)) {
1450                         unlock_page(page);
1451                         put_page(page);
1452                         if (ret1 < 0)
1453                                 ret = ret1;
1454                         else
1455                                 ret = VM_FAULT_SIGBUS;
1456                         goto out_inline;
1457                 }
1458                 if (ret1 < PAGE_SIZE)
1459                         zero_user_segment(page, ret1, PAGE_SIZE);
1460                 else
1461                         flush_dcache_page(page);
1462                 SetPageUptodate(page);
1463                 vmf->page = page;
1464                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1465 out_inline:
1466                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1467                      inode, off, (size_t)PAGE_SIZE, ret);
1468         }
1469 out_restore:
1470         ceph_restore_sigs(&oldset);
1471         if (ret < 0)
1472                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1473
1474         return ret;
1475 }
1476
1477 /*
1478  * Reuse write_begin here for simplicity.
1479  */
1480 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1481 {
1482         struct inode *inode = file_inode(vma->vm_file);
1483         struct ceph_inode_info *ci = ceph_inode(inode);
1484         struct ceph_file_info *fi = vma->vm_file->private_data;
1485         struct ceph_cap_flush *prealloc_cf;
1486         struct page *page = vmf->page;
1487         loff_t off = page_offset(page);
1488         loff_t size = i_size_read(inode);
1489         size_t len;
1490         int want, got, ret;
1491         sigset_t oldset;
1492
1493         prealloc_cf = ceph_alloc_cap_flush();
1494         if (!prealloc_cf)
1495                 return VM_FAULT_OOM;
1496
1497         ceph_block_sigs(&oldset);
1498
1499         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1500                 struct page *locked_page = NULL;
1501                 if (off == 0) {
1502                         lock_page(page);
1503                         locked_page = page;
1504                 }
1505                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1506                 if (locked_page)
1507                         unlock_page(locked_page);
1508                 if (ret < 0)
1509                         goto out_free;
1510         }
1511
1512         if (off + PAGE_SIZE <= size)
1513                 len = PAGE_SIZE;
1514         else
1515                 len = size & ~PAGE_MASK;
1516
1517         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1518              inode, ceph_vinop(inode), off, len, size);
1519         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1520                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1521         else
1522                 want = CEPH_CAP_FILE_BUFFER;
1523
1524         got = 0;
1525         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1526                             &got, NULL);
1527         if (ret < 0)
1528                 goto out_free;
1529
1530         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1531              inode, off, len, ceph_cap_string(got));
1532
1533         /* Update time before taking page lock */
1534         file_update_time(vma->vm_file);
1535
1536         do {
1537                 lock_page(page);
1538
1539                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1540                         unlock_page(page);
1541                         ret = VM_FAULT_NOPAGE;
1542                         break;
1543                 }
1544
1545                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1546                 if (ret >= 0) {
1547                         /* success.  we'll keep the page locked. */
1548                         set_page_dirty(page);
1549                         ret = VM_FAULT_LOCKED;
1550                 }
1551         } while (ret == -EAGAIN);
1552
1553         if (ret == VM_FAULT_LOCKED ||
1554             ci->i_inline_version != CEPH_INLINE_NONE) {
1555                 int dirty;
1556                 spin_lock(&ci->i_ceph_lock);
1557                 ci->i_inline_version = CEPH_INLINE_NONE;
1558                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1559                                                &prealloc_cf);
1560                 spin_unlock(&ci->i_ceph_lock);
1561                 if (dirty)
1562                         __mark_inode_dirty(inode, dirty);
1563         }
1564
1565         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1566              inode, off, len, ceph_cap_string(got), ret);
1567         ceph_put_cap_refs(ci, got);
1568 out_free:
1569         ceph_restore_sigs(&oldset);
1570         ceph_free_cap_flush(prealloc_cf);
1571         if (ret < 0)
1572                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1573         return ret;
1574 }
1575
1576 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1577                            char *data, size_t len)
1578 {
1579         struct address_space *mapping = inode->i_mapping;
1580         struct page *page;
1581
1582         if (locked_page) {
1583                 page = locked_page;
1584         } else {
1585                 if (i_size_read(inode) == 0)
1586                         return;
1587                 page = find_or_create_page(mapping, 0,
1588                                            mapping_gfp_constraint(mapping,
1589                                            ~__GFP_FS));
1590                 if (!page)
1591                         return;
1592                 if (PageUptodate(page)) {
1593                         unlock_page(page);
1594                         put_page(page);
1595                         return;
1596                 }
1597         }
1598
1599         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1600              inode, ceph_vinop(inode), len, locked_page);
1601
1602         if (len > 0) {
1603                 void *kaddr = kmap_atomic(page);
1604                 memcpy(kaddr, data, len);
1605                 kunmap_atomic(kaddr);
1606         }
1607
1608         if (page != locked_page) {
1609                 if (len < PAGE_SIZE)
1610                         zero_user_segment(page, len, PAGE_SIZE);
1611                 else
1612                         flush_dcache_page(page);
1613
1614                 SetPageUptodate(page);
1615                 unlock_page(page);
1616                 put_page(page);
1617         }
1618 }
1619
1620 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1621 {
1622         struct inode *inode = file_inode(filp);
1623         struct ceph_inode_info *ci = ceph_inode(inode);
1624         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1625         struct ceph_osd_request *req;
1626         struct page *page = NULL;
1627         u64 len, inline_version;
1628         int err = 0;
1629         bool from_pagecache = false;
1630
1631         spin_lock(&ci->i_ceph_lock);
1632         inline_version = ci->i_inline_version;
1633         spin_unlock(&ci->i_ceph_lock);
1634
1635         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1636              inode, ceph_vinop(inode), inline_version);
1637
1638         if (inline_version == 1 || /* initial version, no data */
1639             inline_version == CEPH_INLINE_NONE)
1640                 goto out;
1641
1642         if (locked_page) {
1643                 page = locked_page;
1644                 WARN_ON(!PageUptodate(page));
1645         } else if (ceph_caps_issued(ci) &
1646                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1647                 page = find_get_page(inode->i_mapping, 0);
1648                 if (page) {
1649                         if (PageUptodate(page)) {
1650                                 from_pagecache = true;
1651                                 lock_page(page);
1652                         } else {
1653                                 put_page(page);
1654                                 page = NULL;
1655                         }
1656                 }
1657         }
1658
1659         if (page) {
1660                 len = i_size_read(inode);
1661                 if (len > PAGE_SIZE)
1662                         len = PAGE_SIZE;
1663         } else {
1664                 page = __page_cache_alloc(GFP_NOFS);
1665                 if (!page) {
1666                         err = -ENOMEM;
1667                         goto out;
1668                 }
1669                 err = __ceph_do_getattr(inode, page,
1670                                         CEPH_STAT_CAP_INLINE_DATA, true);
1671                 if (err < 0) {
1672                         /* no inline data */
1673                         if (err == -ENODATA)
1674                                 err = 0;
1675                         goto out;
1676                 }
1677                 len = err;
1678         }
1679
1680         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1681                                     ceph_vino(inode), 0, &len, 0, 1,
1682                                     CEPH_OSD_OP_CREATE,
1683                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1684                                     NULL, 0, 0, false);
1685         if (IS_ERR(req)) {
1686                 err = PTR_ERR(req);
1687                 goto out;
1688         }
1689
1690         req->r_mtime = inode->i_mtime;
1691         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1692         if (!err)
1693                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1694         ceph_osdc_put_request(req);
1695         if (err < 0)
1696                 goto out;
1697
1698         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1699                                     ceph_vino(inode), 0, &len, 1, 3,
1700                                     CEPH_OSD_OP_WRITE,
1701                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1702                                     NULL, ci->i_truncate_seq,
1703                                     ci->i_truncate_size, false);
1704         if (IS_ERR(req)) {
1705                 err = PTR_ERR(req);
1706                 goto out;
1707         }
1708
1709         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1710
1711         {
1712                 __le64 xattr_buf = cpu_to_le64(inline_version);
1713                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1714                                             "inline_version", &xattr_buf,
1715                                             sizeof(xattr_buf),
1716                                             CEPH_OSD_CMPXATTR_OP_GT,
1717                                             CEPH_OSD_CMPXATTR_MODE_U64);
1718                 if (err)
1719                         goto out_put;
1720         }
1721
1722         {
1723                 char xattr_buf[32];
1724                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1725                                          "%llu", inline_version);
1726                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1727                                             "inline_version",
1728                                             xattr_buf, xattr_len, 0, 0);
1729                 if (err)
1730                         goto out_put;
1731         }
1732
1733         req->r_mtime = inode->i_mtime;
1734         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1735         if (!err)
1736                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1737 out_put:
1738         ceph_osdc_put_request(req);
1739         if (err == -ECANCELED)
1740                 err = 0;
1741 out:
1742         if (page && page != locked_page) {
1743                 if (from_pagecache) {
1744                         unlock_page(page);
1745                         put_page(page);
1746                 } else
1747                         __free_pages(page, 0);
1748         }
1749
1750         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1751              inode, ceph_vinop(inode), inline_version, err);
1752         return err;
1753 }
1754
1755 static const struct vm_operations_struct ceph_vmops = {
1756         .fault          = ceph_filemap_fault,
1757         .page_mkwrite   = ceph_page_mkwrite,
1758 };
1759
1760 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1761 {
1762         struct address_space *mapping = file->f_mapping;
1763
1764         if (!mapping->a_ops->readpage)
1765                 return -ENOEXEC;
1766         file_accessed(file);
1767         vma->vm_ops = &ceph_vmops;
1768         return 0;
1769 }
1770
1771 enum {
1772         POOL_READ       = 1,
1773         POOL_WRITE      = 2,
1774 };
1775
1776 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1777                                 s64 pool, struct ceph_string *pool_ns)
1778 {
1779         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1780         struct ceph_mds_client *mdsc = fsc->mdsc;
1781         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1782         struct rb_node **p, *parent;
1783         struct ceph_pool_perm *perm;
1784         struct page **pages;
1785         size_t pool_ns_len;
1786         int err = 0, err2 = 0, have = 0;
1787
1788         down_read(&mdsc->pool_perm_rwsem);
1789         p = &mdsc->pool_perm_tree.rb_node;
1790         while (*p) {
1791                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1792                 if (pool < perm->pool)
1793                         p = &(*p)->rb_left;
1794                 else if (pool > perm->pool)
1795                         p = &(*p)->rb_right;
1796                 else {
1797                         int ret = ceph_compare_string(pool_ns,
1798                                                 perm->pool_ns,
1799                                                 perm->pool_ns_len);
1800                         if (ret < 0)
1801                                 p = &(*p)->rb_left;
1802                         else if (ret > 0)
1803                                 p = &(*p)->rb_right;
1804                         else {
1805                                 have = perm->perm;
1806                                 break;
1807                         }
1808                 }
1809         }
1810         up_read(&mdsc->pool_perm_rwsem);
1811         if (*p)
1812                 goto out;
1813
1814         if (pool_ns)
1815                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1816                      pool, (int)pool_ns->len, pool_ns->str);
1817         else
1818                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1819
1820         down_write(&mdsc->pool_perm_rwsem);
1821         p = &mdsc->pool_perm_tree.rb_node;
1822         parent = NULL;
1823         while (*p) {
1824                 parent = *p;
1825                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1826                 if (pool < perm->pool)
1827                         p = &(*p)->rb_left;
1828                 else if (pool > perm->pool)
1829                         p = &(*p)->rb_right;
1830                 else {
1831                         int ret = ceph_compare_string(pool_ns,
1832                                                 perm->pool_ns,
1833                                                 perm->pool_ns_len);
1834                         if (ret < 0)
1835                                 p = &(*p)->rb_left;
1836                         else if (ret > 0)
1837                                 p = &(*p)->rb_right;
1838                         else {
1839                                 have = perm->perm;
1840                                 break;
1841                         }
1842                 }
1843         }
1844         if (*p) {
1845                 up_write(&mdsc->pool_perm_rwsem);
1846                 goto out;
1847         }
1848
1849         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1850                                          1, false, GFP_NOFS);
1851         if (!rd_req) {
1852                 err = -ENOMEM;
1853                 goto out_unlock;
1854         }
1855
1856         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1857         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1858         rd_req->r_base_oloc.pool = pool;
1859         if (pool_ns)
1860                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1861         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1862
1863         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1864         if (err)
1865                 goto out_unlock;
1866
1867         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1868                                          1, false, GFP_NOFS);
1869         if (!wr_req) {
1870                 err = -ENOMEM;
1871                 goto out_unlock;
1872         }
1873
1874         wr_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ACK;
1875         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1876         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1877         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1878
1879         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1880         if (err)
1881                 goto out_unlock;
1882
1883         /* one page should be large enough for STAT data */
1884         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1885         if (IS_ERR(pages)) {
1886                 err = PTR_ERR(pages);
1887                 goto out_unlock;
1888         }
1889
1890         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1891                                      0, false, true);
1892         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1893
1894         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1895         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1896
1897         if (!err)
1898                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1899         if (!err2)
1900                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1901
1902         if (err >= 0 || err == -ENOENT)
1903                 have |= POOL_READ;
1904         else if (err != -EPERM)
1905                 goto out_unlock;
1906
1907         if (err2 == 0 || err2 == -EEXIST)
1908                 have |= POOL_WRITE;
1909         else if (err2 != -EPERM) {
1910                 err = err2;
1911                 goto out_unlock;
1912         }
1913
1914         pool_ns_len = pool_ns ? pool_ns->len : 0;
1915         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1916         if (!perm) {
1917                 err = -ENOMEM;
1918                 goto out_unlock;
1919         }
1920
1921         perm->pool = pool;
1922         perm->perm = have;
1923         perm->pool_ns_len = pool_ns_len;
1924         if (pool_ns_len > 0)
1925                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1926         perm->pool_ns[pool_ns_len] = 0;
1927
1928         rb_link_node(&perm->node, parent, p);
1929         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1930         err = 0;
1931 out_unlock:
1932         up_write(&mdsc->pool_perm_rwsem);
1933
1934         ceph_osdc_put_request(rd_req);
1935         ceph_osdc_put_request(wr_req);
1936 out:
1937         if (!err)
1938                 err = have;
1939         if (pool_ns)
1940                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1941                      pool, (int)pool_ns->len, pool_ns->str, err);
1942         else
1943                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1944         return err;
1945 }
1946
1947 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1948 {
1949         s64 pool;
1950         struct ceph_string *pool_ns;
1951         int ret, flags;
1952
1953         if (ci->i_vino.snap != CEPH_NOSNAP) {
1954                 /*
1955                  * Pool permission check needs to write to the first object.
1956                  * But for snapshot, head of the first object may have alread
1957                  * been deleted. Skip check to avoid creating orphan object.
1958                  */
1959                 return 0;
1960         }
1961
1962         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1963                                 NOPOOLPERM))
1964                 return 0;
1965
1966         spin_lock(&ci->i_ceph_lock);
1967         flags = ci->i_ceph_flags;
1968         pool = ci->i_layout.pool_id;
1969         spin_unlock(&ci->i_ceph_lock);
1970 check:
1971         if (flags & CEPH_I_POOL_PERM) {
1972                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1973                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1974                              pool);
1975                         return -EPERM;
1976                 }
1977                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1978                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1979                              pool);
1980                         return -EPERM;
1981                 }
1982                 return 0;
1983         }
1984
1985         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1986         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1987         ceph_put_string(pool_ns);
1988         if (ret < 0)
1989                 return ret;
1990
1991         flags = CEPH_I_POOL_PERM;
1992         if (ret & POOL_READ)
1993                 flags |= CEPH_I_POOL_RD;
1994         if (ret & POOL_WRITE)
1995                 flags |= CEPH_I_POOL_WR;
1996
1997         spin_lock(&ci->i_ceph_lock);
1998         if (pool == ci->i_layout.pool_id &&
1999             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2000                 ci->i_ceph_flags |= flags;
2001         } else {
2002                 pool = ci->i_layout.pool_id;
2003                 flags = ci->i_ceph_flags;
2004         }
2005         spin_unlock(&ci->i_ceph_lock);
2006         goto check;
2007 }
2008
2009 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2010 {
2011         struct ceph_pool_perm *perm;
2012         struct rb_node *n;
2013
2014         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2015                 n = rb_first(&mdsc->pool_perm_tree);
2016                 perm = rb_entry(n, struct ceph_pool_perm, node);
2017                 rb_erase(n, &mdsc->pool_perm_tree);
2018                 kfree(perm);
2019         }
2020 }