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 #include <linux/signal.h>
  12
  13 #include "super.h"
  14 #include "mds_client.h"
  15 #include "cache.h"
  16 #include <linux/ceph/osd_client.h>
  17
  18 /*
  19  * Ceph address space ops.
  20  *
  21  * There are a few funny things going on here.
  22  *
  23  * The page->private field is used to reference a struct
  24  * ceph_snap_context for _every_ dirty page.  This indicates which
  25  * snapshot the page was logically dirtied in, and thus which snap
  26  * context needs to be associated with the osd write during writeback.
  27  *
  28  * Similarly, struct ceph_inode_info maintains a set of counters to
  29  * count dirty pages on the inode.  In the absence of snapshots,
  30  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  31  *
  32  * When a snapshot is taken (that is, when the client receives
  33  * notification that a snapshot was taken), each inode with caps and
  34  * with dirty pages (dirty pages implies there is a cap) gets a new
  35  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  36  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  37  * moved to capsnap->dirty. (Unless a sync write is currently in
  38  * progress.  In that case, the capsnap is said to be "pending", new
  39  * writes cannot start, and the capsnap isn't "finalized" until the
  40  * write completes (or fails) and a final size/mtime for the inode for
  41  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  42  *
  43  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  44  * we look for the first capsnap in i_cap_snaps and write out pages in
  45  * that snap context _only_.  Then we move on to the next capsnap,
  46  * eventually reaching the "live" or "head" context (i.e., pages that
  47  * are not yet snapped) and are writing the most recently dirtied
  48  * pages.
  49  *
  50  * Invalidate and so forth must take care to ensure the dirty page
  51  * accounting is preserved.
  52  */
  53
  54 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  55 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  56         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  57          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  58
  59 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  60 {
  61         if (PagePrivate(page))
  62                 return (void *)page->private;
  63         return NULL;
  64 }
  65
  66 /*
  67  * Dirty a page.  Optimistically adjust accounting, on the assumption
  68  * that we won't race with invalidate.  If we do, readjust.
  69  */
  70 static int ceph_set_page_dirty(struct page *page)
  71 {
  72         struct address_space *mapping = page->mapping;
  73         struct inode *inode;
  74         struct ceph_inode_info *ci;
  75         struct ceph_snap_context *snapc;
  76         int ret;
  77
  78         if (unlikely(!mapping))
  79                 return !TestSetPageDirty(page);
  80
  81         if (PageDirty(page)) {
  82                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  83                      mapping->host, page, page->index);
  84                 BUG_ON(!PagePrivate(page));
  85                 return 0;
  86         }
  87
  88         inode = mapping->host;
  89         ci = ceph_inode(inode);
  90
  91         /* dirty the head */
  92         spin_lock(&ci->i_ceph_lock);
  93         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  94         if (__ceph_have_pending_cap_snap(ci)) {
  95                 struct ceph_cap_snap *capsnap =
  96                                 list_last_entry(&ci->i_cap_snaps,
  97                                                 struct ceph_cap_snap,
  98                                                 ci_item);
  99                 snapc = ceph_get_snap_context(capsnap->context);
 100                 capsnap->dirty_pages++;
 101         } else {
 102                 BUG_ON(!ci->i_head_snapc);
 103                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 104                 ++ci->i_wrbuffer_ref_head;
 105         }
 106         if (ci->i_wrbuffer_ref == 0)
 107                 ihold(inode);
 108         ++ci->i_wrbuffer_ref;
 109         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 110              "snapc %p seq %lld (%d snaps)\n",
 111              mapping->host, page, page->index,
 112              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 113              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 114              snapc, snapc->seq, snapc->num_snaps);
 115         spin_unlock(&ci->i_ceph_lock);
 116
 117         /*
 118          * Reference snap context in page->private.  Also set
 119          * PagePrivate so that we get invalidatepage callback.
 120          */
 121         BUG_ON(PagePrivate(page));
 122         page->private = (unsigned long)snapc;
 123         SetPagePrivate(page);
 124
 125         ret = __set_page_dirty_nobuffers(page);
 126         WARN_ON(!PageLocked(page));
 127         WARN_ON(!page->mapping);
 128
 129         return ret;
 130 }
 131
 132 /*
 133  * If we are truncating the full page (i.e. offset == 0), adjust the
 134  * dirty page counters appropriately.  Only called if there is private
 135  * data on the page.
 136  */
 137 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 138                                 unsigned int length)
 139 {
 140         struct inode *inode;
 141         struct ceph_inode_info *ci;
 142         struct ceph_snap_context *snapc = page_snap_context(page);
 143
 144         inode = page->mapping->host;
 145         ci = ceph_inode(inode);
 146
 147         if (offset != 0 || length != PAGE_SIZE) {
 148                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 149                      inode, page, page->index, offset, length);
 150                 return;
 151         }
 152
 153         ceph_invalidate_fscache_page(inode, page);
 154
 155         WARN_ON(!PageLocked(page));
 156         if (!PagePrivate(page))
 157                 return;
 158
 159         ClearPageChecked(page);
 160
 161         dout("%p invalidatepage %p idx %lu full dirty page\n",
 162              inode, page, page->index);
 163
 164         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 165         ceph_put_snap_context(snapc);
 166         page->private = 0;
 167         ClearPagePrivate(page);
 168 }
 169
 170 static int ceph_releasepage(struct page *page, gfp_t g)
 171 {
 172         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 173              page, page->index, PageDirty(page) ? "" : "not ");
 174
 175         /* Can we release the page from the cache? */
 176         if (!ceph_release_fscache_page(page, g))
 177                 return 0;
 178
 179         return !PagePrivate(page);
 180 }
 181
 182 /*
 183  * read a single page, without unlocking it.
 184  */
 185 static int ceph_do_readpage(struct file *filp, struct page *page)
 186 {
 187         struct inode *inode = file_inode(filp);
 188         struct ceph_inode_info *ci = ceph_inode(inode);
 189         struct ceph_osd_client *osdc =
 190                 &ceph_inode_to_client(inode)->client->osdc;
 191         int err = 0;
 192         u64 off = page_offset(page);
 193         u64 len = PAGE_SIZE;
 194
 195         if (off >= i_size_read(inode)) {
 196                 zero_user_segment(page, 0, PAGE_SIZE);
 197                 SetPageUptodate(page);
 198                 return 0;
 199         }
 200
 201         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 202                 /*
 203                  * Uptodate inline data should have been added
 204                  * into page cache while getting Fcr caps.
 205                  */
 206                 if (off == 0)
 207                         return -EINVAL;
 208                 zero_user_segment(page, 0, PAGE_SIZE);
 209                 SetPageUptodate(page);
 210                 return 0;
 211         }
 212
 213         err = ceph_readpage_from_fscache(inode, page);
 214         if (err == 0)
 215                 return -EINPROGRESS;
 216
 217         dout("readpage inode %p file %p page %p index %lu\n",
 218              inode, filp, page, page->index);
 219         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 220                                   off, &len,
 221                                   ci->i_truncate_seq, ci->i_truncate_size,
 222                                   &page, 1, 0);
 223         if (err == -ENOENT)
 224                 err = 0;
 225         if (err < 0) {
 226                 SetPageError(page);
 227                 ceph_fscache_readpage_cancel(inode, page);
 228                 goto out;
 229         }
 230         if (err < PAGE_SIZE)
 231                 /* zero fill remainder of page */
 232                 zero_user_segment(page, err, PAGE_SIZE);
 233         else
 234                 flush_dcache_page(page);
 235
 236         SetPageUptodate(page);
 237         ceph_readpage_to_fscache(inode, page);
 238
 239 out:
 240         return err < 0 ? err : 0;
 241 }
 242
 243 static int ceph_readpage(struct file *filp, struct page *page)
 244 {
 245         int r = ceph_do_readpage(filp, page);
 246         if (r != -EINPROGRESS)
 247                 unlock_page(page);
 248         else
 249                 r = 0;
 250         return r;
 251 }
 252
 253 /*
 254  * Finish an async read(ahead) op.
 255  */
 256 static void finish_read(struct ceph_osd_request *req)
 257 {
 258         struct inode *inode = req->r_inode;
 259         struct ceph_osd_data *osd_data;
 260         int rc = req->r_result <= 0 ? req->r_result : 0;
 261         int bytes = req->r_result >= 0 ? req->r_result : 0;
 262         int num_pages;
 263         int i;
 264
 265         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 266
 267         /* unlock all pages, zeroing any data we didn't read */
 268         osd_data = osd_req_op_extent_osd_data(req, 0);
 269         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 270         num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                         (u64)osd_data->length);
 272         for (i = 0; i < num_pages; i++) {
 273                 struct page *page = osd_data->pages[i];
 274
 275                 if (rc < 0 && rc != -ENOENT) {
 276                         ceph_fscache_readpage_cancel(inode, page);
 277                         goto unlock;
 278                 }
 279                 if (bytes < (int)PAGE_SIZE) {
 280                         /* zero (remainder of) page */
 281                         int s = bytes < 0 ? 0 : bytes;
 282                         zero_user_segment(page, s, PAGE_SIZE);
 283                 }
 284                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 285                      page->index);
 286                 flush_dcache_page(page);
 287                 SetPageUptodate(page);
 288                 ceph_readpage_to_fscache(inode, page);
 289 unlock:
 290                 unlock_page(page);
 291                 put_page(page);
 292                 bytes -= PAGE_SIZE;
 293         }
 294         kfree(osd_data->pages);
 295 }
 296
 297 /*
 298  * start an async read(ahead) operation.  return nr_pages we submitted
 299  * a read for on success, or negative error code.
 300  */
 301 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 302 {
 303         struct ceph_osd_client *osdc =
 304                 &ceph_inode_to_client(inode)->client->osdc;
 305         struct ceph_inode_info *ci = ceph_inode(inode);
 306         struct page *page = list_entry(page_list->prev, struct page, lru);
 307         struct ceph_vino vino;
 308         struct ceph_osd_request *req;
 309         u64 off;
 310         u64 len;
 311         int i;
 312         struct page **pages;
 313         pgoff_t next_index;
 314         int nr_pages = 0;
 315         int got = 0;
 316         int ret = 0;
 317
 318         if (!current->journal_info) {
 319                 /* caller of readpages does not hold buffer and read caps
 320                  * (fadvise, madvise and readahead cases) */
 321                 int want = CEPH_CAP_FILE_CACHE;
 322                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 323                 if (ret < 0) {
 324                         dout("start_read %p, error getting cap\n", inode);
 325                 } else if (!(got & want)) {
 326                         dout("start_read %p, no cache cap\n", inode);
 327                         ret = 0;
 328                 }
 329                 if (ret <= 0) {
 330                         if (got)
 331                                 ceph_put_cap_refs(ci, got);
 332                         while (!list_empty(page_list)) {
 333                                 page = list_entry(page_list->prev,
 334                                                   struct page, lru);
 335                                 list_del(&page->lru);
 336                                 put_page(page);
 337                         }
 338                         return ret;
 339                 }
 340         }
 341
 342         off = (u64) page_offset(page);
 343
 344         /* count pages */
 345         next_index = page->index;
 346         list_for_each_entry_reverse(page, page_list, lru) {
 347                 if (page->index != next_index)
 348                         break;
 349                 nr_pages++;
 350                 next_index++;
 351                 if (max && nr_pages == max)
 352                         break;
 353         }
 354         len = nr_pages << PAGE_SHIFT;
 355         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 356              off, len);
 357         vino = ceph_vino(inode);
 358         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 359                                     0, 1, CEPH_OSD_OP_READ,
 360                                     CEPH_OSD_FLAG_READ, NULL,
 361                                     ci->i_truncate_seq, ci->i_truncate_size,
 362                                     false);
 363         if (IS_ERR(req)) {
 364                 ret = PTR_ERR(req);
 365                 goto out;
 366         }
 367
 368         /* build page vector */
 369         nr_pages = calc_pages_for(0, len);
 370         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 371         if (!pages) {
 372                 ret = -ENOMEM;
 373                 goto out_put;
 374         }
 375         for (i = 0; i < nr_pages; ++i) {
 376                 page = list_entry(page_list->prev, struct page, lru);
 377                 BUG_ON(PageLocked(page));
 378                 list_del(&page->lru);
 379
 380                 dout("start_read %p adding %p idx %lu\n", inode, page,
 381                      page->index);
 382                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 383                                           GFP_KERNEL)) {
 384                         ceph_fscache_uncache_page(inode, page);
 385                         put_page(page);
 386                         dout("start_read %p add_to_page_cache failed %p\n",
 387                              inode, page);
 388                         nr_pages = i;
 389                         if (nr_pages > 0) {
 390                                 len = nr_pages << PAGE_SHIFT;
 391                                 osd_req_op_extent_update(req, 0, len);
 392                                 break;
 393                         }
 394                         goto out_pages;
 395                 }
 396                 pages[i] = page;
 397         }
 398         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 399         req->r_callback = finish_read;
 400         req->r_inode = inode;
 401
 402         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 403         ret = ceph_osdc_start_request(osdc, req, false);
 404         if (ret < 0)
 405                 goto out_pages;
 406         ceph_osdc_put_request(req);
 407
 408         /* After adding locked pages to page cache, the inode holds cache cap.
 409          * So we can drop our cap refs. */
 410         if (got)
 411                 ceph_put_cap_refs(ci, got);
 412
 413         return nr_pages;
 414
 415 out_pages:
 416         for (i = 0; i < nr_pages; ++i) {
 417                 ceph_fscache_readpage_cancel(inode, pages[i]);
 418                 unlock_page(pages[i]);
 419         }
 420         ceph_put_page_vector(pages, nr_pages, false);
 421 out_put:
 422         ceph_osdc_put_request(req);
 423 out:
 424         if (got)
 425                 ceph_put_cap_refs(ci, got);
 426         return ret;
 427 }
 428
 429
 430 /*
 431  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 432  * the caller (VM) cleans them up.
 433  */
 434 static int ceph_readpages(struct file *file, struct address_space *mapping,
 435                           struct list_head *page_list, unsigned nr_pages)
 436 {
 437         struct inode *inode = file_inode(file);
 438         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 439         int rc = 0;
 440         int max = 0;
 441
 442         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 443                 return -EINVAL;
 444
 445         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 446                                          &nr_pages);
 447
 448         if (rc == 0)
 449                 goto out;
 450
 451         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 452         dout("readpages %p file %p nr_pages %d max %d\n",
 453              inode, file, nr_pages, max);
 454         while (!list_empty(page_list)) {
 455                 rc = start_read(inode, page_list, max);
 456                 if (rc < 0)
 457                         goto out;
 458         }
 459 out:
 460         ceph_fscache_readpages_cancel(inode, page_list);
 461
 462         dout("readpages %p file %p ret %d\n", inode, file, rc);
 463         return rc;
 464 }
 465
 466 /*
 467  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 468  * only snap context we are allowed to write back.
 469  */
 470 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 471                                                     loff_t *snap_size,
 472                                                     u64 *truncate_size,
 473                                                     u32 *truncate_seq)
 474 {
 475         struct ceph_inode_info *ci = ceph_inode(inode);
 476         struct ceph_snap_context *snapc = NULL;
 477         struct ceph_cap_snap *capsnap = NULL;
 478
 479         spin_lock(&ci->i_ceph_lock);
 480         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 481                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 482                      capsnap->context, capsnap->dirty_pages);
 483                 if (capsnap->dirty_pages) {
 484                         snapc = ceph_get_snap_context(capsnap->context);
 485                         if (snap_size)
 486                                 *snap_size = capsnap->size;
 487                         if (truncate_size)
 488                                 *truncate_size = capsnap->truncate_size;
 489                         if (truncate_seq)
 490                                 *truncate_seq = capsnap->truncate_seq;
 491                         break;
 492                 }
 493         }
 494         if (!snapc && ci->i_wrbuffer_ref_head) {
 495                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 496                 dout(" head snapc %p has %d dirty pages\n",
 497                      snapc, ci->i_wrbuffer_ref_head);
 498                 if (truncate_size)
 499                         *truncate_size = ci->i_truncate_size;
 500                 if (truncate_seq)
 501                         *truncate_seq = ci->i_truncate_seq;
 502         }
 503         spin_unlock(&ci->i_ceph_lock);
 504         return snapc;
 505 }
 506
 507 /*
 508  * Write a single page, but leave the page locked.
 509  *
 510  * If we get a write error, set the page error bit, but still adjust the
 511  * dirty page accounting (i.e., page is no longer dirty).
 512  */
 513 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 514 {
 515         struct inode *inode;
 516         struct ceph_inode_info *ci;
 517         struct ceph_fs_client *fsc;
 518         struct ceph_osd_client *osdc;
 519         struct ceph_snap_context *snapc, *oldest;
 520         loff_t page_off = page_offset(page);
 521         loff_t snap_size = -1;
 522         long writeback_stat;
 523         u64 truncate_size;
 524         u32 truncate_seq;
 525         int err, len = PAGE_SIZE;
 526
 527         dout("writepage %p idx %lu\n", page, page->index);
 528
 529         inode = page->mapping->host;
 530         ci = ceph_inode(inode);
 531         fsc = ceph_inode_to_client(inode);
 532         osdc = &fsc->client->osdc;
 533
 534         /* verify this is a writeable snap context */
 535         snapc = page_snap_context(page);
 536         if (!snapc) {
 537                 dout("writepage %p page %p not dirty?\n", inode, page);
 538                 return 0;
 539         }
 540         oldest = get_oldest_context(inode, &snap_size,
 541                                     &truncate_size, &truncate_seq);
 542         if (snapc->seq > oldest->seq) {
 543                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 544                      inode, page, snapc);
 545                 /* we should only noop if called by kswapd */
 546                 WARN_ON(!(current->flags & PF_MEMALLOC));
 547                 ceph_put_snap_context(oldest);
 548                 redirty_page_for_writepage(wbc, page);
 549                 return 0;
 550         }
 551         ceph_put_snap_context(oldest);
 552
 553         if (snap_size == -1)
 554                 snap_size = i_size_read(inode);
 555
 556         /* is this a partial page at end of file? */
 557         if (page_off >= snap_size) {
 558                 dout("%p page eof %llu\n", page, snap_size);
 559                 return 0;
 560         }
 561
 562         if (snap_size < page_off + len)
 563                 len = snap_size - page_off;
 564
 565         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 566              inode, page, page->index, page_off, len, snapc, snapc->seq);
 567
 568         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 569         if (writeback_stat >
 570             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 571                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 572
 573         set_page_writeback(page);
 574         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 575                                    &ci->i_layout, snapc,
 576                                    page_off, len,
 577                                    truncate_seq, truncate_size,
 578                                    &inode->i_mtime, &page, 1);
 579         if (err < 0) {
 580                 struct writeback_control tmp_wbc;
 581                 if (!wbc)
 582                         wbc = &tmp_wbc;
 583                 if (err == -ERESTARTSYS) {
 584                         /* killed by SIGKILL */
 585                         dout("writepage interrupted page %p\n", page);
 586                         redirty_page_for_writepage(wbc, page);
 587                         end_page_writeback(page);
 588                         return err;
 589                 }
 590                 dout("writepage setting page/mapping error %d %p\n",
 591                      err, page);
 592                 SetPageError(page);
 593                 mapping_set_error(&inode->i_data, err);
 594                 wbc->pages_skipped++;
 595         } else {
 596                 dout("writepage cleaned page %p\n", page);
 597                 err = 0;  /* vfs expects us to return 0 */
 598         }
 599         page->private = 0;
 600         ClearPagePrivate(page);
 601         end_page_writeback(page);
 602         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 603         ceph_put_snap_context(snapc);  /* page's reference */
 604         return err;
 605 }
 606
 607 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 608 {
 609         int err;
 610         struct inode *inode = page->mapping->host;
 611         BUG_ON(!inode);
 612         ihold(inode);
 613         err = writepage_nounlock(page, wbc);
 614         if (err == -ERESTARTSYS) {
 615                 /* direct memory reclaimer was killed by SIGKILL. return 0
 616                  * to prevent caller from setting mapping/page error */
 617                 err = 0;
 618         }
 619         unlock_page(page);
 620         iput(inode);
 621         return err;
 622 }
 623
 624 /*
 625  * lame release_pages helper.  release_pages() isn't exported to
 626  * modules.
 627  */
 628 static void ceph_release_pages(struct page **pages, int num)
 629 {
 630         struct pagevec pvec;
 631         int i;
 632
 633         pagevec_init(&pvec, 0);
 634         for (i = 0; i < num; i++) {
 635                 if (pagevec_add(&pvec, pages[i]) == 0)
 636                         pagevec_release(&pvec);
 637         }
 638         pagevec_release(&pvec);
 639 }
 640
 641 /*
 642  * async writeback completion handler.
 643  *
 644  * If we get an error, set the mapping error bit, but not the individual
 645  * page error bits.
 646  */
 647 static void writepages_finish(struct ceph_osd_request *req)
 648 {
 649         struct inode *inode = req->r_inode;
 650         struct ceph_inode_info *ci = ceph_inode(inode);
 651         struct ceph_osd_data *osd_data;
 652         struct page *page;
 653         int num_pages, total_pages = 0;
 654         int i, j;
 655         int rc = req->r_result;
 656         struct ceph_snap_context *snapc = req->r_snapc;
 657         struct address_space *mapping = inode->i_mapping;
 658         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 659         bool remove_page;
 660
 661         dout("writepages_finish %p rc %d\n", inode, rc);
 662         if (rc < 0) {
 663                 mapping_set_error(mapping, rc);
 664                 ceph_set_error_write(ci);
 665         } else {
 666                 ceph_clear_error_write(ci);
 667         }
 668
 669         /*
 670          * We lost the cache cap, need to truncate the page before
 671          * it is unlocked, otherwise we'd truncate it later in the
 672          * page truncation thread, possibly losing some data that
 673          * raced its way in
 674          */
 675         remove_page = !(ceph_caps_issued(ci) &
 676                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 677
 678         /* clean all pages */
 679         for (i = 0; i < req->r_num_ops; i++) {
 680                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 681                         break;
 682
 683                 osd_data = osd_req_op_extent_osd_data(req, i);
 684                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 685                 num_pages = calc_pages_for((u64)osd_data->alignment,
 686                                            (u64)osd_data->length);
 687                 total_pages += num_pages;
 688                 for (j = 0; j < num_pages; j++) {
 689                         page = osd_data->pages[j];
 690                         BUG_ON(!page);
 691                         WARN_ON(!PageUptodate(page));
 692
 693                         if (atomic_long_dec_return(&fsc->writeback_count) <
 694                              CONGESTION_OFF_THRESH(
 695                                         fsc->mount_options->congestion_kb))
 696                                 clear_bdi_congested(inode_to_bdi(inode),
 697                                                     BLK_RW_ASYNC);
 698
 699                         ceph_put_snap_context(page_snap_context(page));
 700                         page->private = 0;
 701                         ClearPagePrivate(page);
 702                         dout("unlocking %p\n", page);
 703                         end_page_writeback(page);
 704
 705                         if (remove_page)
 706                                 generic_error_remove_page(inode->i_mapping,
 707                                                           page);
 708
 709                         unlock_page(page);
 710                 }
 711                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 712                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 713
 714                 ceph_release_pages(osd_data->pages, num_pages);
 715         }
 716
 717         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 718
 719         osd_data = osd_req_op_extent_osd_data(req, 0);
 720         if (osd_data->pages_from_pool)
 721                 mempool_free(osd_data->pages,
 722                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 723         else
 724                 kfree(osd_data->pages);
 725         ceph_osdc_put_request(req);
 726 }
 727
 728 /*
 729  * initiate async writeback
 730  */
 731 static int ceph_writepages_start(struct address_space *mapping,
 732                                  struct writeback_control *wbc)
 733 {
 734         struct inode *inode = mapping->host;
 735         struct ceph_inode_info *ci = ceph_inode(inode);
 736         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 737         struct ceph_vino vino = ceph_vino(inode);
 738         pgoff_t index, start, end;
 739         int range_whole = 0;
 740         int should_loop = 1;
 741         pgoff_t max_pages = 0, max_pages_ever = 0;
 742         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 743         struct pagevec pvec;
 744         int done = 0;
 745         int rc = 0;
 746         unsigned int wsize = i_blocksize(inode);
 747         struct ceph_osd_request *req = NULL;
 748         loff_t snap_size, i_size;
 749         u64 truncate_size;
 750         u32 truncate_seq;
 751
 752         dout("writepages_start %p (mode=%s)\n", inode,
 753              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 754              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 755
 756         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 757                 if (ci->i_wrbuffer_ref > 0) {
 758                         pr_warn_ratelimited(
 759                                 "writepage_start %p %lld forced umount\n",
 760                                 inode, ceph_ino(inode));
 761                 }
 762                 mapping_set_error(mapping, -EIO);
 763                 return -EIO; /* we're in a forced umount, don't write! */
 764         }
 765         if (fsc->mount_options->wsize < wsize)
 766                 wsize = fsc->mount_options->wsize;
 767         max_pages_ever = wsize >> PAGE_SHIFT;
 768
 769         pagevec_init(&pvec, 0);
 770
 771         /* where to start/end? */
 772         if (wbc->range_cyclic) {
 773                 start = mapping->writeback_index; /* Start from prev offset */
 774                 end = -1;
 775                 dout(" cyclic, start at %lu\n", start);
 776         } else {
 777                 start = wbc->range_start >> PAGE_SHIFT;
 778                 end = wbc->range_end >> PAGE_SHIFT;
 779                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 780                         range_whole = 1;
 781                 should_loop = 0;
 782                 dout(" not cyclic, %lu to %lu\n", start, end);
 783         }
 784         index = start;
 785
 786 retry:
 787         /* find oldest snap context with dirty data */
 788         ceph_put_snap_context(snapc);
 789         snap_size = -1;
 790         snapc = get_oldest_context(inode, &snap_size,
 791                                    &truncate_size, &truncate_seq);
 792         if (!snapc) {
 793                 /* hmm, why does writepages get called when there
 794                    is no dirty data? */
 795                 dout(" no snap context with dirty data?\n");
 796                 goto out;
 797         }
 798         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 799              snapc, snapc->seq, snapc->num_snaps);
 800
 801         i_size = i_size_read(inode);
 802
 803         if (last_snapc && snapc != last_snapc) {
 804                 /* if we switched to a newer snapc, restart our scan at the
 805                  * start of the original file range. */
 806                 dout("  snapc differs from last pass, restarting at %lu\n",
 807                      index);
 808                 index = start;
 809         }
 810         last_snapc = snapc;
 811
 812         while (!done && index <= end) {
 813                 unsigned i;
 814                 int first;
 815                 pgoff_t strip_unit_end = 0;
 816                 int num_ops = 0, op_idx;
 817                 int pvec_pages, locked_pages = 0;
 818                 struct page **pages = NULL, **data_pages;
 819                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 820                 struct page *page;
 821                 int want;
 822                 u64 offset = 0, len = 0;
 823
 824                 max_pages = max_pages_ever;
 825
 826 get_more_pages:
 827                 first = -1;
 828                 want = min(end - index,
 829                            min((pgoff_t)PAGEVEC_SIZE,
 830                                max_pages - (pgoff_t)locked_pages) - 1)
 831                         + 1;
 832                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 833                                                 PAGECACHE_TAG_DIRTY,
 834                                                 want);
 835                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 836                 if (!pvec_pages && !locked_pages)
 837                         break;
 838                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 839                         page = pvec.pages[i];
 840                         dout("? %p idx %lu\n", page, page->index);
 841                         if (locked_pages == 0)
 842                                 lock_page(page);  /* first page */
 843                         else if (!trylock_page(page))
 844                                 break;
 845
 846                         /* only dirty pages, or our accounting breaks */
 847                         if (unlikely(!PageDirty(page)) ||
 848                             unlikely(page->mapping != mapping)) {
 849                                 dout("!dirty or !mapping %p\n", page);
 850                                 unlock_page(page);
 851                                 break;
 852                         }
 853                         if (!wbc->range_cyclic && page->index > end) {
 854                                 dout("end of range %p\n", page);
 855                                 done = 1;
 856                                 unlock_page(page);
 857                                 break;
 858                         }
 859                         if (strip_unit_end && (page->index > strip_unit_end)) {
 860                                 dout("end of strip unit %p\n", page);
 861                                 unlock_page(page);
 862                                 break;
 863                         }
 864                         if (wbc->sync_mode != WB_SYNC_NONE) {
 865                                 dout("waiting on writeback %p\n", page);
 866                                 wait_on_page_writeback(page);
 867                         }
 868                         if (page_offset(page) >=
 869                             (snap_size == -1 ? i_size : snap_size)) {
 870                                 dout("%p page eof %llu\n", page,
 871                                      (snap_size == -1 ? i_size : snap_size));
 872                                 done = 1;
 873                                 unlock_page(page);
 874                                 break;
 875                         }
 876                         if (PageWriteback(page)) {
 877                                 dout("%p under writeback\n", page);
 878                                 unlock_page(page);
 879                                 break;
 880                         }
 881
 882                         /* only if matching snap context */
 883                         pgsnapc = page_snap_context(page);
 884                         if (pgsnapc->seq > snapc->seq) {
 885                                 dout("page snapc %p %lld > oldest %p %lld\n",
 886                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 887                                 unlock_page(page);
 888                                 if (!locked_pages)
 889                                         continue; /* keep looking for snap */
 890                                 break;
 891                         }
 892
 893                         if (!clear_page_dirty_for_io(page)) {
 894                                 dout("%p !clear_page_dirty_for_io\n", page);
 895                                 unlock_page(page);
 896                                 break;
 897                         }
 898
 899                         /*
 900                          * We have something to write.  If this is
 901                          * the first locked page this time through,
 902                          * calculate max possinle write size and
 903                          * allocate a page array
 904                          */
 905                         if (locked_pages == 0) {
 906                                 u64 objnum;
 907                                 u64 objoff;
 908
 909                                 /* prepare async write request */
 910                                 offset = (u64)page_offset(page);
 911                                 len = wsize;
 912
 913                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 914                                                                 offset, len,
 915                                                                 &objnum, &objoff,
 916                                                                 &len);
 917                                 if (rc < 0) {
 918                                         unlock_page(page);
 919                                         break;
 920                                 }
 921
 922                                 num_ops = 1;
 923                                 strip_unit_end = page->index +
 924                                         ((len - 1) >> PAGE_SHIFT);
 925
 926                                 BUG_ON(pages);
 927                                 max_pages = calc_pages_for(0, (u64)len);
 928                                 pages = kmalloc(max_pages * sizeof (*pages),
 929                                                 GFP_NOFS);
 930                                 if (!pages) {
 931                                         pool = fsc->wb_pagevec_pool;
 932                                         pages = mempool_alloc(pool, GFP_NOFS);
 933                                         BUG_ON(!pages);
 934                                 }
 935
 936                                 len = 0;
 937                         } else if (page->index !=
 938                                    (offset + len) >> PAGE_SHIFT) {
 939                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 940                                                         CEPH_OSD_MAX_OPS)) {
 941                                         redirty_page_for_writepage(wbc, page);
 942                                         unlock_page(page);
 943                                         break;
 944                                 }
 945
 946                                 num_ops++;
 947                                 offset = (u64)page_offset(page);
 948                                 len = 0;
 949                         }
 950
 951                         /* note position of first page in pvec */
 952                         if (first < 0)
 953                                 first = i;
 954                         dout("%p will write page %p idx %lu\n",
 955                              inode, page, page->index);
 956
 957                         if (atomic_long_inc_return(&fsc->writeback_count) >
 958                             CONGESTION_ON_THRESH(
 959                                     fsc->mount_options->congestion_kb)) {
 960                                 set_bdi_congested(inode_to_bdi(inode),
 961                                                   BLK_RW_ASYNC);
 962                         }
 963
 964                         pages[locked_pages] = page;
 965                         locked_pages++;
 966                         len += PAGE_SIZE;
 967                 }
 968
 969                 /* did we get anything? */
 970                 if (!locked_pages)
 971                         goto release_pvec_pages;
 972                 if (i) {
 973                         int j;
 974                         BUG_ON(!locked_pages || first < 0);
 975
 976                         if (pvec_pages && i == pvec_pages &&
 977                             locked_pages < max_pages) {
 978                                 dout("reached end pvec, trying for more\n");
 979                                 pagevec_reinit(&pvec);
 980                                 goto get_more_pages;
 981                         }
 982
 983                         /* shift unused pages over in the pvec...  we
 984                          * will need to release them below. */
 985                         for (j = i; j < pvec_pages; j++) {
 986                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
 987                                 pvec.pages[j-i+first] = pvec.pages[j];
 988                         }
 989                         pvec.nr -= i-first;
 990                 }
 991
 992 new_request:
 993                 offset = page_offset(pages[0]);
 994                 len = wsize;
 995
 996                 req = ceph_osdc_new_request(&fsc->client->osdc,
 997                                         &ci->i_layout, vino,
 998                                         offset, &len, 0, num_ops,
 999                                         CEPH_OSD_OP_WRITE,
1000                                         CEPH_OSD_FLAG_WRITE,
1001                                         snapc, truncate_seq,
1002                                         truncate_size, false);
1003                 if (IS_ERR(req)) {
1004                         req = ceph_osdc_new_request(&fsc->client->osdc,
1005                                                 &ci->i_layout, vino,
1006                                                 offset, &len, 0,
1007                                                 min(num_ops,
1008                                                     CEPH_OSD_SLAB_OPS),
1009                                                 CEPH_OSD_OP_WRITE,
1010                                                 CEPH_OSD_FLAG_WRITE,
1011                                                 snapc, truncate_seq,
1012                                                 truncate_size, true);
1013                         BUG_ON(IS_ERR(req));
1014                 }
1015                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1016                              PAGE_SIZE - offset);
1017
1018                 req->r_callback = writepages_finish;
1019                 req->r_inode = inode;
1020
1021                 /* Format the osd request message and submit the write */
1022                 len = 0;
1023                 data_pages = pages;
1024                 op_idx = 0;
1025                 for (i = 0; i < locked_pages; i++) {
1026                         u64 cur_offset = page_offset(pages[i]);
1027                         if (offset + len != cur_offset) {
1028                                 if (op_idx + 1 == req->r_num_ops)
1029                                         break;
1030                                 osd_req_op_extent_dup_last(req, op_idx,
1031                                                            cur_offset - offset);
1032                                 dout("writepages got pages at %llu~%llu\n",
1033                                      offset, len);
1034                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1035                                                         data_pages, len, 0,
1036                                                         !!pool, false);
1037                                 osd_req_op_extent_update(req, op_idx, len);
1038
1039                                 len = 0;
1040                                 offset = cur_offset;
1041                                 data_pages = pages + i;
1042                                 op_idx++;
1043                         }
1044
1045                         set_page_writeback(pages[i]);
1046                         len += PAGE_SIZE;
1047                 }
1048
1049                 if (snap_size != -1) {
1050                         len = min(len, snap_size - offset);
1051                 } else if (i == locked_pages) {
1052                         /* writepages_finish() clears writeback pages
1053                          * according to the data length, so make sure
1054                          * data length covers all locked pages */
1055                         u64 min_len = len + 1 - PAGE_SIZE;
1056                         len = min(len, (u64)i_size_read(inode) - offset);
1057                         len = max(len, min_len);
1058                 }
1059                 dout("writepages got pages at %llu~%llu\n", offset, len);
1060
1061                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1062                                                  0, !!pool, false);
1063                 osd_req_op_extent_update(req, op_idx, len);
1064
1065                 BUG_ON(op_idx + 1 != req->r_num_ops);
1066
1067                 pool = NULL;
1068                 if (i < locked_pages) {
1069                         BUG_ON(num_ops <= req->r_num_ops);
1070                         num_ops -= req->r_num_ops;
1071                         locked_pages -= i;
1072
1073                         /* allocate new pages array for next request */
1074                         data_pages = pages;
1075                         pages = kmalloc(locked_pages * sizeof (*pages),
1076                                         GFP_NOFS);
1077                         if (!pages) {
1078                                 pool = fsc->wb_pagevec_pool;
1079                                 pages = mempool_alloc(pool, GFP_NOFS);
1080                                 BUG_ON(!pages);
1081                         }
1082                         memcpy(pages, data_pages + i,
1083                                locked_pages * sizeof(*pages));
1084                         memset(data_pages + i, 0,
1085                                locked_pages * sizeof(*pages));
1086                 } else {
1087                         BUG_ON(num_ops != req->r_num_ops);
1088                         index = pages[i - 1]->index + 1;
1089                         /* request message now owns the pages array */
1090                         pages = NULL;
1091                 }
1092
1093                 req->r_mtime = inode->i_mtime;
1094                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1095                 BUG_ON(rc);
1096                 req = NULL;
1097
1098                 wbc->nr_to_write -= i;
1099                 if (pages)
1100                         goto new_request;
1101
1102                 if (wbc->nr_to_write <= 0)
1103                         done = 1;
1104
1105 release_pvec_pages:
1106                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1107                      pvec.nr ? pvec.pages[0] : NULL);
1108                 pagevec_release(&pvec);
1109
1110                 if (locked_pages && !done)
1111                         goto retry;
1112         }
1113
1114         if (should_loop && !done) {
1115                 /* more to do; loop back to beginning of file */
1116                 dout("writepages looping back to beginning of file\n");
1117                 should_loop = 0;
1118                 index = 0;
1119                 goto retry;
1120         }
1121
1122         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1123                 mapping->writeback_index = index;
1124
1125 out:
1126         ceph_osdc_put_request(req);
1127         ceph_put_snap_context(snapc);
1128         dout("writepages done, rc = %d\n", rc);
1129         return rc;
1130 }
1131
1132
1133
1134 /*
1135  * See if a given @snapc is either writeable, or already written.
1136  */
1137 static int context_is_writeable_or_written(struct inode *inode,
1138                                            struct ceph_snap_context *snapc)
1139 {
1140         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL,
1141                                                               NULL, NULL);
1142         int ret = !oldest || snapc->seq <= oldest->seq;
1143
1144         ceph_put_snap_context(oldest);
1145         return ret;
1146 }
1147
1148 /*
1149  * We are only allowed to write into/dirty the page if the page is
1150  * clean, or already dirty within the same snap context.
1151  *
1152  * called with page locked.
1153  * return success with page locked,
1154  * or any failure (incl -EAGAIN) with page unlocked.
1155  */
1156 static int ceph_update_writeable_page(struct file *file,
1157                             loff_t pos, unsigned len,
1158                             struct page *page)
1159 {
1160         struct inode *inode = file_inode(file);
1161         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1162         struct ceph_inode_info *ci = ceph_inode(inode);
1163         loff_t page_off = pos & PAGE_MASK;
1164         int pos_in_page = pos & ~PAGE_MASK;
1165         int end_in_page = pos_in_page + len;
1166         loff_t i_size;
1167         int r;
1168         struct ceph_snap_context *snapc, *oldest;
1169
1170         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1171                 dout(" page %p forced umount\n", page);
1172                 unlock_page(page);
1173                 return -EIO;
1174         }
1175
1176 retry_locked:
1177         /* writepages currently holds page lock, but if we change that later, */
1178         wait_on_page_writeback(page);
1179
1180         snapc = page_snap_context(page);
1181         if (snapc && snapc != ci->i_head_snapc) {
1182                 /*
1183                  * this page is already dirty in another (older) snap
1184                  * context!  is it writeable now?
1185                  */
1186                 oldest = get_oldest_context(inode, NULL, NULL, NULL);
1187
1188                 if (snapc->seq > oldest->seq) {
1189                         ceph_put_snap_context(oldest);
1190                         dout(" page %p snapc %p not current or oldest\n",
1191                              page, snapc);
1192                         /*
1193                          * queue for writeback, and wait for snapc to
1194                          * be writeable or written
1195                          */
1196                         snapc = ceph_get_snap_context(snapc);
1197                         unlock_page(page);
1198                         ceph_queue_writeback(inode);
1199                         r = wait_event_killable(ci->i_cap_wq,
1200                                context_is_writeable_or_written(inode, snapc));
1201                         ceph_put_snap_context(snapc);
1202                         if (r == -ERESTARTSYS)
1203                                 return r;
1204                         return -EAGAIN;
1205                 }
1206                 ceph_put_snap_context(oldest);
1207
1208                 /* yay, writeable, do it now (without dropping page lock) */
1209                 dout(" page %p snapc %p not current, but oldest\n",
1210                      page, snapc);
1211                 if (!clear_page_dirty_for_io(page))
1212                         goto retry_locked;
1213                 r = writepage_nounlock(page, NULL);
1214                 if (r < 0)
1215                         goto fail_unlock;
1216                 goto retry_locked;
1217         }
1218
1219         if (PageUptodate(page)) {
1220                 dout(" page %p already uptodate\n", page);
1221                 return 0;
1222         }
1223
1224         /* full page? */
1225         if (pos_in_page == 0 && len == PAGE_SIZE)
1226                 return 0;
1227
1228         /* past end of file? */
1229         i_size = i_size_read(inode);
1230
1231         if (page_off >= i_size ||
1232             (pos_in_page == 0 && (pos+len) >= i_size &&
1233              end_in_page - pos_in_page != PAGE_SIZE)) {
1234                 dout(" zeroing %p 0 - %d and %d - %d\n",
1235                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1236                 zero_user_segments(page,
1237                                    0, pos_in_page,
1238                                    end_in_page, PAGE_SIZE);
1239                 return 0;
1240         }
1241
1242         /* we need to read it. */
1243         r = ceph_do_readpage(file, page);
1244         if (r < 0) {
1245                 if (r == -EINPROGRESS)
1246                         return -EAGAIN;
1247                 goto fail_unlock;
1248         }
1249         goto retry_locked;
1250 fail_unlock:
1251         unlock_page(page);
1252         return r;
1253 }
1254
1255 /*
1256  * We are only allowed to write into/dirty the page if the page is
1257  * clean, or already dirty within the same snap context.
1258  */
1259 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1260                             loff_t pos, unsigned len, unsigned flags,
1261                             struct page **pagep, void **fsdata)
1262 {
1263         struct inode *inode = file_inode(file);
1264         struct page *page;
1265         pgoff_t index = pos >> PAGE_SHIFT;
1266         int r;
1267
1268         do {
1269                 /* get a page */
1270                 page = grab_cache_page_write_begin(mapping, index, 0);
1271                 if (!page)
1272                         return -ENOMEM;
1273
1274                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1275                      inode, page, (int)pos, (int)len);
1276
1277                 r = ceph_update_writeable_page(file, pos, len, page);
1278                 if (r < 0)
1279                         put_page(page);
1280                 else
1281                         *pagep = page;
1282         } while (r == -EAGAIN);
1283
1284         return r;
1285 }
1286
1287 /*
1288  * we don't do anything in here that simple_write_end doesn't do
1289  * except adjust dirty page accounting
1290  */
1291 static int ceph_write_end(struct file *file, struct address_space *mapping,
1292                           loff_t pos, unsigned len, unsigned copied,
1293                           struct page *page, void *fsdata)
1294 {
1295         struct inode *inode = file_inode(file);
1296         bool check_cap = false;
1297
1298         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1299              inode, page, (int)pos, (int)copied, (int)len);
1300
1301         /* zero the stale part of the page if we did a short copy */
1302         if (!PageUptodate(page)) {
1303                 if (copied < len) {
1304                         copied = 0;
1305                         goto out;
1306                 }
1307                 SetPageUptodate(page);
1308         }
1309
1310         /* did file size increase? */
1311         if (pos+copied > i_size_read(inode))
1312                 check_cap = ceph_inode_set_size(inode, pos+copied);
1313
1314         set_page_dirty(page);
1315
1316 out:
1317         unlock_page(page);
1318         put_page(page);
1319
1320         if (check_cap)
1321                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1322
1323         return copied;
1324 }
1325
1326 /*
1327  * we set .direct_IO to indicate direct io is supported, but since we
1328  * intercept O_DIRECT reads and writes early, this function should
1329  * never get called.
1330  */
1331 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1332 {
1333         WARN_ON(1);
1334         return -EINVAL;
1335 }
1336
1337 const struct address_space_operations ceph_aops = {
1338         .readpage = ceph_readpage,
1339         .readpages = ceph_readpages,
1340         .writepage = ceph_writepage,
1341         .writepages = ceph_writepages_start,
1342         .write_begin = ceph_write_begin,
1343         .write_end = ceph_write_end,
1344         .set_page_dirty = ceph_set_page_dirty,
1345         .invalidatepage = ceph_invalidatepage,
1346         .releasepage = ceph_releasepage,
1347         .direct_IO = ceph_direct_io,
1348 };
1349
1350 static void ceph_block_sigs(sigset_t *oldset)
1351 {
1352         sigset_t mask;
1353         siginitsetinv(&mask, sigmask(SIGKILL));
1354         sigprocmask(SIG_BLOCK, &mask, oldset);
1355 }
1356
1357 static void ceph_restore_sigs(sigset_t *oldset)
1358 {
1359         sigprocmask(SIG_SETMASK, oldset, NULL);
1360 }
1361
1362 /*
1363  * vm ops
1364  */
1365 static int ceph_filemap_fault(struct vm_fault *vmf)
1366 {
1367         struct vm_area_struct *vma = vmf->vma;
1368         struct inode *inode = file_inode(vma->vm_file);
1369         struct ceph_inode_info *ci = ceph_inode(inode);
1370         struct ceph_file_info *fi = vma->vm_file->private_data;
1371         struct page *pinned_page = NULL;
1372         loff_t off = vmf->pgoff << PAGE_SHIFT;
1373         int want, got, ret;
1374         sigset_t oldset;
1375
1376         ceph_block_sigs(&oldset);
1377
1378         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1379              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1380         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1381                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1382         else
1383                 want = CEPH_CAP_FILE_CACHE;
1384
1385         got = 0;
1386         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1387         if (ret < 0)
1388                 goto out_restore;
1389
1390         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1391              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1392
1393         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1394             ci->i_inline_version == CEPH_INLINE_NONE) {
1395                 current->journal_info = vma->vm_file;
1396                 ret = filemap_fault(vmf);
1397                 current->journal_info = NULL;
1398         } else
1399                 ret = -EAGAIN;
1400
1401         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1402              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1403         if (pinned_page)
1404                 put_page(pinned_page);
1405         ceph_put_cap_refs(ci, got);
1406
1407         if (ret != -EAGAIN)
1408                 goto out_restore;
1409
1410         /* read inline data */
1411         if (off >= PAGE_SIZE) {
1412                 /* does not support inline data > PAGE_SIZE */
1413                 ret = VM_FAULT_SIGBUS;
1414         } else {
1415                 int ret1;
1416                 struct address_space *mapping = inode->i_mapping;
1417                 struct page *page = find_or_create_page(mapping, 0,
1418                                                 mapping_gfp_constraint(mapping,
1419                                                 ~__GFP_FS));
1420                 if (!page) {
1421                         ret = VM_FAULT_OOM;
1422                         goto out_inline;
1423                 }
1424                 ret1 = __ceph_do_getattr(inode, page,
1425                                          CEPH_STAT_CAP_INLINE_DATA, true);
1426                 if (ret1 < 0 || off >= i_size_read(inode)) {
1427                         unlock_page(page);
1428                         put_page(page);
1429                         if (ret1 < 0)
1430                                 ret = ret1;
1431                         else
1432                                 ret = VM_FAULT_SIGBUS;
1433                         goto out_inline;
1434                 }
1435                 if (ret1 < PAGE_SIZE)
1436                         zero_user_segment(page, ret1, PAGE_SIZE);
1437                 else
1438                         flush_dcache_page(page);
1439                 SetPageUptodate(page);
1440                 vmf->page = page;
1441                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1442 out_inline:
1443                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1444                      inode, off, (size_t)PAGE_SIZE, ret);
1445         }
1446 out_restore:
1447         ceph_restore_sigs(&oldset);
1448         if (ret < 0)
1449                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1450
1451         return ret;
1452 }
1453
1454 /*
1455  * Reuse write_begin here for simplicity.
1456  */
1457 static int ceph_page_mkwrite(struct vm_fault *vmf)
1458 {
1459         struct vm_area_struct *vma = vmf->vma;
1460         struct inode *inode = file_inode(vma->vm_file);
1461         struct ceph_inode_info *ci = ceph_inode(inode);
1462         struct ceph_file_info *fi = vma->vm_file->private_data;
1463         struct ceph_cap_flush *prealloc_cf;
1464         struct page *page = vmf->page;
1465         loff_t off = page_offset(page);
1466         loff_t size = i_size_read(inode);
1467         size_t len;
1468         int want, got, ret;
1469         sigset_t oldset;
1470
1471         prealloc_cf = ceph_alloc_cap_flush();
1472         if (!prealloc_cf)
1473                 return VM_FAULT_OOM;
1474
1475         ceph_block_sigs(&oldset);
1476
1477         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1478                 struct page *locked_page = NULL;
1479                 if (off == 0) {
1480                         lock_page(page);
1481                         locked_page = page;
1482                 }
1483                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1484                 if (locked_page)
1485                         unlock_page(locked_page);
1486                 if (ret < 0)
1487                         goto out_free;
1488         }
1489
1490         if (off + PAGE_SIZE <= size)
1491                 len = PAGE_SIZE;
1492         else
1493                 len = size & ~PAGE_MASK;
1494
1495         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1496              inode, ceph_vinop(inode), off, len, size);
1497         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1498                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1499         else
1500                 want = CEPH_CAP_FILE_BUFFER;
1501
1502         got = 0;
1503         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1504                             &got, NULL);
1505         if (ret < 0)
1506                 goto out_free;
1507
1508         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1509              inode, off, len, ceph_cap_string(got));
1510
1511         /* Update time before taking page lock */
1512         file_update_time(vma->vm_file);
1513
1514         do {
1515                 lock_page(page);
1516
1517                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1518                         unlock_page(page);
1519                         ret = VM_FAULT_NOPAGE;
1520                         break;
1521                 }
1522
1523                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1524                 if (ret >= 0) {
1525                         /* success.  we'll keep the page locked. */
1526                         set_page_dirty(page);
1527                         ret = VM_FAULT_LOCKED;
1528                 }
1529         } while (ret == -EAGAIN);
1530
1531         if (ret == VM_FAULT_LOCKED ||
1532             ci->i_inline_version != CEPH_INLINE_NONE) {
1533                 int dirty;
1534                 spin_lock(&ci->i_ceph_lock);
1535                 ci->i_inline_version = CEPH_INLINE_NONE;
1536                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1537                                                &prealloc_cf);
1538                 spin_unlock(&ci->i_ceph_lock);
1539                 if (dirty)
1540                         __mark_inode_dirty(inode, dirty);
1541         }
1542
1543         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1544              inode, off, len, ceph_cap_string(got), ret);
1545         ceph_put_cap_refs(ci, got);
1546 out_free:
1547         ceph_restore_sigs(&oldset);
1548         ceph_free_cap_flush(prealloc_cf);
1549         if (ret < 0)
1550                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1551         return ret;
1552 }
1553
1554 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1555                            char *data, size_t len)
1556 {
1557         struct address_space *mapping = inode->i_mapping;
1558         struct page *page;
1559
1560         if (locked_page) {
1561                 page = locked_page;
1562         } else {
1563                 if (i_size_read(inode) == 0)
1564                         return;
1565                 page = find_or_create_page(mapping, 0,
1566                                            mapping_gfp_constraint(mapping,
1567                                            ~__GFP_FS));
1568                 if (!page)
1569                         return;
1570                 if (PageUptodate(page)) {
1571                         unlock_page(page);
1572                         put_page(page);
1573                         return;
1574                 }
1575         }
1576
1577         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1578              inode, ceph_vinop(inode), len, locked_page);
1579
1580         if (len > 0) {
1581                 void *kaddr = kmap_atomic(page);
1582                 memcpy(kaddr, data, len);
1583                 kunmap_atomic(kaddr);
1584         }
1585
1586         if (page != locked_page) {
1587                 if (len < PAGE_SIZE)
1588                         zero_user_segment(page, len, PAGE_SIZE);
1589                 else
1590                         flush_dcache_page(page);
1591
1592                 SetPageUptodate(page);
1593                 unlock_page(page);
1594                 put_page(page);
1595         }
1596 }
1597
1598 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1599 {
1600         struct inode *inode = file_inode(filp);
1601         struct ceph_inode_info *ci = ceph_inode(inode);
1602         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1603         struct ceph_osd_request *req;
1604         struct page *page = NULL;
1605         u64 len, inline_version;
1606         int err = 0;
1607         bool from_pagecache = false;
1608
1609         spin_lock(&ci->i_ceph_lock);
1610         inline_version = ci->i_inline_version;
1611         spin_unlock(&ci->i_ceph_lock);
1612
1613         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1614              inode, ceph_vinop(inode), inline_version);
1615
1616         if (inline_version == 1 || /* initial version, no data */
1617             inline_version == CEPH_INLINE_NONE)
1618                 goto out;
1619
1620         if (locked_page) {
1621                 page = locked_page;
1622                 WARN_ON(!PageUptodate(page));
1623         } else if (ceph_caps_issued(ci) &
1624                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1625                 page = find_get_page(inode->i_mapping, 0);
1626                 if (page) {
1627                         if (PageUptodate(page)) {
1628                                 from_pagecache = true;
1629                                 lock_page(page);
1630                         } else {
1631                                 put_page(page);
1632                                 page = NULL;
1633                         }
1634                 }
1635         }
1636
1637         if (page) {
1638                 len = i_size_read(inode);
1639                 if (len > PAGE_SIZE)
1640                         len = PAGE_SIZE;
1641         } else {
1642                 page = __page_cache_alloc(GFP_NOFS);
1643                 if (!page) {
1644                         err = -ENOMEM;
1645                         goto out;
1646                 }
1647                 err = __ceph_do_getattr(inode, page,
1648                                         CEPH_STAT_CAP_INLINE_DATA, true);
1649                 if (err < 0) {
1650                         /* no inline data */
1651                         if (err == -ENODATA)
1652                                 err = 0;
1653                         goto out;
1654                 }
1655                 len = err;
1656         }
1657
1658         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1659                                     ceph_vino(inode), 0, &len, 0, 1,
1660                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1661                                     NULL, 0, 0, false);
1662         if (IS_ERR(req)) {
1663                 err = PTR_ERR(req);
1664                 goto out;
1665         }
1666
1667         req->r_mtime = inode->i_mtime;
1668         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1669         if (!err)
1670                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1671         ceph_osdc_put_request(req);
1672         if (err < 0)
1673                 goto out;
1674
1675         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1676                                     ceph_vino(inode), 0, &len, 1, 3,
1677                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1678                                     NULL, ci->i_truncate_seq,
1679                                     ci->i_truncate_size, false);
1680         if (IS_ERR(req)) {
1681                 err = PTR_ERR(req);
1682                 goto out;
1683         }
1684
1685         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1686
1687         {
1688                 __le64 xattr_buf = cpu_to_le64(inline_version);
1689                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1690                                             "inline_version", &xattr_buf,
1691                                             sizeof(xattr_buf),
1692                                             CEPH_OSD_CMPXATTR_OP_GT,
1693                                             CEPH_OSD_CMPXATTR_MODE_U64);
1694                 if (err)
1695                         goto out_put;
1696         }
1697
1698         {
1699                 char xattr_buf[32];
1700                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1701                                          "%llu", inline_version);
1702                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1703                                             "inline_version",
1704                                             xattr_buf, xattr_len, 0, 0);
1705                 if (err)
1706                         goto out_put;
1707         }
1708
1709         req->r_mtime = inode->i_mtime;
1710         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1711         if (!err)
1712                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1713 out_put:
1714         ceph_osdc_put_request(req);
1715         if (err == -ECANCELED)
1716                 err = 0;
1717 out:
1718         if (page && page != locked_page) {
1719                 if (from_pagecache) {
1720                         unlock_page(page);
1721                         put_page(page);
1722                 } else
1723                         __free_pages(page, 0);
1724         }
1725
1726         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1727              inode, ceph_vinop(inode), inline_version, err);
1728         return err;
1729 }
1730
1731 static const struct vm_operations_struct ceph_vmops = {
1732         .fault          = ceph_filemap_fault,
1733         .page_mkwrite   = ceph_page_mkwrite,
1734 };
1735
1736 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1737 {
1738         struct address_space *mapping = file->f_mapping;
1739
1740         if (!mapping->a_ops->readpage)
1741                 return -ENOEXEC;
1742         file_accessed(file);
1743         vma->vm_ops = &ceph_vmops;
1744         return 0;
1745 }
1746
1747 enum {
1748         POOL_READ       = 1,
1749         POOL_WRITE      = 2,
1750 };
1751
1752 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1753                                 s64 pool, struct ceph_string *pool_ns)
1754 {
1755         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1756         struct ceph_mds_client *mdsc = fsc->mdsc;
1757         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1758         struct rb_node **p, *parent;
1759         struct ceph_pool_perm *perm;
1760         struct page **pages;
1761         size_t pool_ns_len;
1762         int err = 0, err2 = 0, have = 0;
1763
1764         down_read(&mdsc->pool_perm_rwsem);
1765         p = &mdsc->pool_perm_tree.rb_node;
1766         while (*p) {
1767                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1768                 if (pool < perm->pool)
1769                         p = &(*p)->rb_left;
1770                 else if (pool > perm->pool)
1771                         p = &(*p)->rb_right;
1772                 else {
1773                         int ret = ceph_compare_string(pool_ns,
1774                                                 perm->pool_ns,
1775                                                 perm->pool_ns_len);
1776                         if (ret < 0)
1777                                 p = &(*p)->rb_left;
1778                         else if (ret > 0)
1779                                 p = &(*p)->rb_right;
1780                         else {
1781                                 have = perm->perm;
1782                                 break;
1783                         }
1784                 }
1785         }
1786         up_read(&mdsc->pool_perm_rwsem);
1787         if (*p)
1788                 goto out;
1789
1790         if (pool_ns)
1791                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1792                      pool, (int)pool_ns->len, pool_ns->str);
1793         else
1794                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1795
1796         down_write(&mdsc->pool_perm_rwsem);
1797         p = &mdsc->pool_perm_tree.rb_node;
1798         parent = NULL;
1799         while (*p) {
1800                 parent = *p;
1801                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1802                 if (pool < perm->pool)
1803                         p = &(*p)->rb_left;
1804                 else if (pool > perm->pool)
1805                         p = &(*p)->rb_right;
1806                 else {
1807                         int ret = ceph_compare_string(pool_ns,
1808                                                 perm->pool_ns,
1809                                                 perm->pool_ns_len);
1810                         if (ret < 0)
1811                                 p = &(*p)->rb_left;
1812                         else if (ret > 0)
1813                                 p = &(*p)->rb_right;
1814                         else {
1815                                 have = perm->perm;
1816                                 break;
1817                         }
1818                 }
1819         }
1820         if (*p) {
1821                 up_write(&mdsc->pool_perm_rwsem);
1822                 goto out;
1823         }
1824
1825         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1826                                          1, false, GFP_NOFS);
1827         if (!rd_req) {
1828                 err = -ENOMEM;
1829                 goto out_unlock;
1830         }
1831
1832         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1833         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1834         rd_req->r_base_oloc.pool = pool;
1835         if (pool_ns)
1836                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1837         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1838
1839         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1840         if (err)
1841                 goto out_unlock;
1842
1843         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1844                                          1, false, GFP_NOFS);
1845         if (!wr_req) {
1846                 err = -ENOMEM;
1847                 goto out_unlock;
1848         }
1849
1850         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1851         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1852         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1853         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1854
1855         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1856         if (err)
1857                 goto out_unlock;
1858
1859         /* one page should be large enough for STAT data */
1860         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1861         if (IS_ERR(pages)) {
1862                 err = PTR_ERR(pages);
1863                 goto out_unlock;
1864         }
1865
1866         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1867                                      0, false, true);
1868         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1869
1870         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1871         wr_req->r_abort_on_full = true;
1872         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1873
1874         if (!err)
1875                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1876         if (!err2)
1877                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1878
1879         if (err >= 0 || err == -ENOENT)
1880                 have |= POOL_READ;
1881         else if (err != -EPERM)
1882                 goto out_unlock;
1883
1884         if (err2 == 0 || err2 == -EEXIST)
1885                 have |= POOL_WRITE;
1886         else if (err2 != -EPERM) {
1887                 err = err2;
1888                 goto out_unlock;
1889         }
1890
1891         pool_ns_len = pool_ns ? pool_ns->len : 0;
1892         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1893         if (!perm) {
1894                 err = -ENOMEM;
1895                 goto out_unlock;
1896         }
1897
1898         perm->pool = pool;
1899         perm->perm = have;
1900         perm->pool_ns_len = pool_ns_len;
1901         if (pool_ns_len > 0)
1902                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1903         perm->pool_ns[pool_ns_len] = 0;
1904
1905         rb_link_node(&perm->node, parent, p);
1906         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1907         err = 0;
1908 out_unlock:
1909         up_write(&mdsc->pool_perm_rwsem);
1910
1911         ceph_osdc_put_request(rd_req);
1912         ceph_osdc_put_request(wr_req);
1913 out:
1914         if (!err)
1915                 err = have;
1916         if (pool_ns)
1917                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1918                      pool, (int)pool_ns->len, pool_ns->str, err);
1919         else
1920                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1921         return err;
1922 }
1923
1924 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1925 {
1926         s64 pool;
1927         struct ceph_string *pool_ns;
1928         int ret, flags;
1929
1930         if (ci->i_vino.snap != CEPH_NOSNAP) {
1931                 /*
1932                  * Pool permission check needs to write to the first object.
1933                  * But for snapshot, head of the first object may have alread
1934                  * been deleted. Skip check to avoid creating orphan object.
1935                  */
1936                 return 0;
1937         }
1938
1939         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1940                                 NOPOOLPERM))
1941                 return 0;
1942
1943         spin_lock(&ci->i_ceph_lock);
1944         flags = ci->i_ceph_flags;
1945         pool = ci->i_layout.pool_id;
1946         spin_unlock(&ci->i_ceph_lock);
1947 check:
1948         if (flags & CEPH_I_POOL_PERM) {
1949                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1950                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1951                              pool);
1952                         return -EPERM;
1953                 }
1954                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1955                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1956                              pool);
1957                         return -EPERM;
1958                 }
1959                 return 0;
1960         }
1961
1962         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1963         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1964         ceph_put_string(pool_ns);
1965         if (ret < 0)
1966                 return ret;
1967
1968         flags = CEPH_I_POOL_PERM;
1969         if (ret & POOL_READ)
1970                 flags |= CEPH_I_POOL_RD;
1971         if (ret & POOL_WRITE)
1972                 flags |= CEPH_I_POOL_WR;
1973
1974         spin_lock(&ci->i_ceph_lock);
1975         if (pool == ci->i_layout.pool_id &&
1976             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
1977                 ci->i_ceph_flags |= flags;
1978         } else {
1979                 pool = ci->i_layout.pool_id;
1980                 flags = ci->i_ceph_flags;
1981         }
1982         spin_unlock(&ci->i_ceph_lock);
1983         goto check;
1984 }
1985
1986 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1987 {
1988         struct ceph_pool_perm *perm;
1989         struct rb_node *n;
1990
1991         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1992                 n = rb_first(&mdsc->pool_perm_tree);
1993                 perm = rb_entry(n, struct ceph_pool_perm, node);
1994                 rb_erase(n, &mdsc->pool_perm_tree);
1995                 kfree(perm);
1996         }
1997 }