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         if (!PagePrivate(page))
 156                 return;
 157
 158         /*
 159          * We can get non-dirty pages here due to races between
 160          * set_page_dirty and truncate_complete_page; just spit out a
 161          * warning, in case we end up with accounting problems later.
 162          */
 163         if (!PageDirty(page))
 164                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 165
 166         ClearPageChecked(page);
 167
 168         dout("%p invalidatepage %p idx %lu full dirty page\n",
 169              inode, page, page->index);
 170
 171         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 172         ceph_put_snap_context(snapc);
 173         page->private = 0;
 174         ClearPagePrivate(page);
 175 }
 176
 177 static int ceph_releasepage(struct page *page, gfp_t g)
 178 {
 179         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 180              page, page->index, PageDirty(page) ? "" : "not ");
 181
 182         /* Can we release the page from the cache? */
 183         if (!ceph_release_fscache_page(page, g))
 184                 return 0;
 185
 186         return !PagePrivate(page);
 187 }
 188
 189 /*
 190  * read a single page, without unlocking it.
 191  */
 192 static int ceph_do_readpage(struct file *filp, struct page *page)
 193 {
 194         struct inode *inode = file_inode(filp);
 195         struct ceph_inode_info *ci = ceph_inode(inode);
 196         struct ceph_osd_client *osdc =
 197                 &ceph_inode_to_client(inode)->client->osdc;
 198         int err = 0;
 199         u64 off = page_offset(page);
 200         u64 len = PAGE_SIZE;
 201
 202         if (off >= i_size_read(inode)) {
 203                 zero_user_segment(page, 0, PAGE_SIZE);
 204                 SetPageUptodate(page);
 205                 return 0;
 206         }
 207
 208         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 209                 /*
 210                  * Uptodate inline data should have been added
 211                  * into page cache while getting Fcr caps.
 212                  */
 213                 if (off == 0)
 214                         return -EINVAL;
 215                 zero_user_segment(page, 0, PAGE_SIZE);
 216                 SetPageUptodate(page);
 217                 return 0;
 218         }
 219
 220         err = ceph_readpage_from_fscache(inode, page);
 221         if (err == 0)
 222                 return -EINPROGRESS;
 223
 224         dout("readpage inode %p file %p page %p index %lu\n",
 225              inode, filp, page, page->index);
 226         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 227                                   off, &len,
 228                                   ci->i_truncate_seq, ci->i_truncate_size,
 229                                   &page, 1, 0);
 230         if (err == -ENOENT)
 231                 err = 0;
 232         if (err < 0) {
 233                 SetPageError(page);
 234                 ceph_fscache_readpage_cancel(inode, page);
 235                 goto out;
 236         }
 237         if (err < PAGE_SIZE)
 238                 /* zero fill remainder of page */
 239                 zero_user_segment(page, err, PAGE_SIZE);
 240         else
 241                 flush_dcache_page(page);
 242
 243         SetPageUptodate(page);
 244         ceph_readpage_to_fscache(inode, page);
 245
 246 out:
 247         return err < 0 ? err : 0;
 248 }
 249
 250 static int ceph_readpage(struct file *filp, struct page *page)
 251 {
 252         int r = ceph_do_readpage(filp, page);
 253         if (r != -EINPROGRESS)
 254                 unlock_page(page);
 255         else
 256                 r = 0;
 257         return r;
 258 }
 259
 260 /*
 261  * Finish an async read(ahead) op.
 262  */
 263 static void finish_read(struct ceph_osd_request *req)
 264 {
 265         struct inode *inode = req->r_inode;
 266         struct ceph_osd_data *osd_data;
 267         int rc = req->r_result <= 0 ? req->r_result : 0;
 268         int bytes = req->r_result >= 0 ? req->r_result : 0;
 269         int num_pages;
 270         int i;
 271
 272         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 273
 274         /* unlock all pages, zeroing any data we didn't read */
 275         osd_data = osd_req_op_extent_osd_data(req, 0);
 276         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 277         num_pages = calc_pages_for((u64)osd_data->alignment,
 278                                         (u64)osd_data->length);
 279         for (i = 0; i < num_pages; i++) {
 280                 struct page *page = osd_data->pages[i];
 281
 282                 if (rc < 0 && rc != -ENOENT) {
 283                         ceph_fscache_readpage_cancel(inode, page);
 284                         goto unlock;
 285                 }
 286                 if (bytes < (int)PAGE_SIZE) {
 287                         /* zero (remainder of) page */
 288                         int s = bytes < 0 ? 0 : bytes;
 289                         zero_user_segment(page, s, PAGE_SIZE);
 290                 }
 291                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 292                      page->index);
 293                 flush_dcache_page(page);
 294                 SetPageUptodate(page);
 295                 ceph_readpage_to_fscache(inode, page);
 296 unlock:
 297                 unlock_page(page);
 298                 put_page(page);
 299                 bytes -= PAGE_SIZE;
 300         }
 301         kfree(osd_data->pages);
 302 }
 303
 304 /*
 305  * start an async read(ahead) operation.  return nr_pages we submitted
 306  * a read for on success, or negative error code.
 307  */
 308 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 309 {
 310         struct ceph_osd_client *osdc =
 311                 &ceph_inode_to_client(inode)->client->osdc;
 312         struct ceph_inode_info *ci = ceph_inode(inode);
 313         struct page *page = list_entry(page_list->prev, struct page, lru);
 314         struct ceph_vino vino;
 315         struct ceph_osd_request *req;
 316         u64 off;
 317         u64 len;
 318         int i;
 319         struct page **pages;
 320         pgoff_t next_index;
 321         int nr_pages = 0;
 322         int got = 0;
 323         int ret = 0;
 324
 325         if (!current->journal_info) {
 326                 /* caller of readpages does not hold buffer and read caps
 327                  * (fadvise, madvise and readahead cases) */
 328                 int want = CEPH_CAP_FILE_CACHE;
 329                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 330                 if (ret < 0) {
 331                         dout("start_read %p, error getting cap\n", inode);
 332                 } else if (!(got & want)) {
 333                         dout("start_read %p, no cache cap\n", inode);
 334                         ret = 0;
 335                 }
 336                 if (ret <= 0) {
 337                         if (got)
 338                                 ceph_put_cap_refs(ci, got);
 339                         while (!list_empty(page_list)) {
 340                                 page = list_entry(page_list->prev,
 341                                                   struct page, lru);
 342                                 list_del(&page->lru);
 343                                 put_page(page);
 344                         }
 345                         return ret;
 346                 }
 347         }
 348
 349         off = (u64) page_offset(page);
 350
 351         /* count pages */
 352         next_index = page->index;
 353         list_for_each_entry_reverse(page, page_list, lru) {
 354                 if (page->index != next_index)
 355                         break;
 356                 nr_pages++;
 357                 next_index++;
 358                 if (max && nr_pages == max)
 359                         break;
 360         }
 361         len = nr_pages << PAGE_SHIFT;
 362         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 363              off, len);
 364         vino = ceph_vino(inode);
 365         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 366                                     0, 1, CEPH_OSD_OP_READ,
 367                                     CEPH_OSD_FLAG_READ, NULL,
 368                                     ci->i_truncate_seq, ci->i_truncate_size,
 369                                     false);
 370         if (IS_ERR(req)) {
 371                 ret = PTR_ERR(req);
 372                 goto out;
 373         }
 374
 375         /* build page vector */
 376         nr_pages = calc_pages_for(0, len);
 377         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 378         if (!pages) {
 379                 ret = -ENOMEM;
 380                 goto out_put;
 381         }
 382         for (i = 0; i < nr_pages; ++i) {
 383                 page = list_entry(page_list->prev, struct page, lru);
 384                 BUG_ON(PageLocked(page));
 385                 list_del(&page->lru);
 386
 387                 dout("start_read %p adding %p idx %lu\n", inode, page,
 388                      page->index);
 389                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 390                                           GFP_KERNEL)) {
 391                         ceph_fscache_uncache_page(inode, page);
 392                         put_page(page);
 393                         dout("start_read %p add_to_page_cache failed %p\n",
 394                              inode, page);
 395                         nr_pages = i;
 396                         if (nr_pages > 0) {
 397                                 len = nr_pages << PAGE_SHIFT;
 398                                 osd_req_op_extent_update(req, 0, len);
 399                                 break;
 400                         }
 401                         goto out_pages;
 402                 }
 403                 pages[i] = page;
 404         }
 405         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 406         req->r_callback = finish_read;
 407         req->r_inode = inode;
 408
 409         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 410         ret = ceph_osdc_start_request(osdc, req, false);
 411         if (ret < 0)
 412                 goto out_pages;
 413         ceph_osdc_put_request(req);
 414
 415         /* After adding locked pages to page cache, the inode holds cache cap.
 416          * So we can drop our cap refs. */
 417         if (got)
 418                 ceph_put_cap_refs(ci, got);
 419
 420         return nr_pages;
 421
 422 out_pages:
 423         for (i = 0; i < nr_pages; ++i) {
 424                 ceph_fscache_readpage_cancel(inode, pages[i]);
 425                 unlock_page(pages[i]);
 426         }
 427         ceph_put_page_vector(pages, nr_pages, false);
 428 out_put:
 429         ceph_osdc_put_request(req);
 430 out:
 431         if (got)
 432                 ceph_put_cap_refs(ci, got);
 433         return ret;
 434 }
 435
 436
 437 /*
 438  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 439  * the caller (VM) cleans them up.
 440  */
 441 static int ceph_readpages(struct file *file, struct address_space *mapping,
 442                           struct list_head *page_list, unsigned nr_pages)
 443 {
 444         struct inode *inode = file_inode(file);
 445         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 446         int rc = 0;
 447         int max = 0;
 448
 449         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 450                 return -EINVAL;
 451
 452         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 453                                          &nr_pages);
 454
 455         if (rc == 0)
 456                 goto out;
 457
 458         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 459         dout("readpages %p file %p nr_pages %d max %d\n",
 460              inode, file, nr_pages, max);
 461         while (!list_empty(page_list)) {
 462                 rc = start_read(inode, page_list, max);
 463                 if (rc < 0)
 464                         goto out;
 465         }
 466 out:
 467         ceph_fscache_readpages_cancel(inode, page_list);
 468
 469         dout("readpages %p file %p ret %d\n", inode, file, rc);
 470         return rc;
 471 }
 472
 473 /*
 474  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 475  * only snap context we are allowed to write back.
 476  */
 477 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 478                                                     loff_t *snap_size,
 479                                                     u64 *truncate_size,
 480                                                     u32 *truncate_seq)
 481 {
 482         struct ceph_inode_info *ci = ceph_inode(inode);
 483         struct ceph_snap_context *snapc = NULL;
 484         struct ceph_cap_snap *capsnap = NULL;
 485
 486         spin_lock(&ci->i_ceph_lock);
 487         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 488                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 489                      capsnap->context, capsnap->dirty_pages);
 490                 if (capsnap->dirty_pages) {
 491                         snapc = ceph_get_snap_context(capsnap->context);
 492                         if (snap_size)
 493                                 *snap_size = capsnap->size;
 494                         if (truncate_size)
 495                                 *truncate_size = capsnap->truncate_size;
 496                         if (truncate_seq)
 497                                 *truncate_seq = capsnap->truncate_seq;
 498                         break;
 499                 }
 500         }
 501         if (!snapc && ci->i_wrbuffer_ref_head) {
 502                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 503                 dout(" head snapc %p has %d dirty pages\n",
 504                      snapc, ci->i_wrbuffer_ref_head);
 505                 if (truncate_size)
 506                         *truncate_size = ci->i_truncate_size;
 507                 if (truncate_seq)
 508                         *truncate_seq = ci->i_truncate_seq;
 509         }
 510         spin_unlock(&ci->i_ceph_lock);
 511         return snapc;
 512 }
 513
 514 /*
 515  * Write a single page, but leave the page locked.
 516  *
 517  * If we get a write error, set the page error bit, but still adjust the
 518  * dirty page accounting (i.e., page is no longer dirty).
 519  */
 520 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 521 {
 522         struct inode *inode;
 523         struct ceph_inode_info *ci;
 524         struct ceph_fs_client *fsc;
 525         struct ceph_osd_client *osdc;
 526         struct ceph_snap_context *snapc, *oldest;
 527         loff_t page_off = page_offset(page);
 528         loff_t snap_size = -1;
 529         long writeback_stat;
 530         u64 truncate_size;
 531         u32 truncate_seq;
 532         int err, len = PAGE_SIZE;
 533
 534         dout("writepage %p idx %lu\n", page, page->index);
 535
 536         inode = page->mapping->host;
 537         ci = ceph_inode(inode);
 538         fsc = ceph_inode_to_client(inode);
 539         osdc = &fsc->client->osdc;
 540
 541         /* verify this is a writeable snap context */
 542         snapc = page_snap_context(page);
 543         if (snapc == NULL) {
 544                 dout("writepage %p page %p not dirty?\n", inode, page);
 545                 return 0;
 546         }
 547         oldest = get_oldest_context(inode, &snap_size,
 548                                     &truncate_size, &truncate_seq);
 549         if (snapc->seq > oldest->seq) {
 550                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 551                      inode, page, snapc);
 552                 /* we should only noop if called by kswapd */
 553                 WARN_ON(!(current->flags & PF_MEMALLOC));
 554                 ceph_put_snap_context(oldest);
 555                 redirty_page_for_writepage(wbc, page);
 556                 return 0;
 557         }
 558         ceph_put_snap_context(oldest);
 559
 560         if (snap_size == -1)
 561                 snap_size = i_size_read(inode);
 562
 563         /* is this a partial page at end of file? */
 564         if (page_off >= snap_size) {
 565                 dout("%p page eof %llu\n", page, snap_size);
 566                 return 0;
 567         }
 568
 569         if (snap_size < page_off + len)
 570                 len = snap_size - page_off;
 571
 572         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 573              inode, page, page->index, page_off, len, snapc);
 574
 575         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 576         if (writeback_stat >
 577             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 578                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 579
 580         set_page_writeback(page);
 581         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 582                                    &ci->i_layout, snapc,
 583                                    page_off, len,
 584                                    truncate_seq, truncate_size,
 585                                    &inode->i_mtime, &page, 1);
 586         if (err < 0) {
 587                 struct writeback_control tmp_wbc;
 588                 if (!wbc)
 589                         wbc = &tmp_wbc;
 590                 if (err == -ERESTARTSYS) {
 591                         /* killed by SIGKILL */
 592                         dout("writepage interrupted page %p\n", page);
 593                         redirty_page_for_writepage(wbc, page);
 594                         end_page_writeback(page);
 595                         return err;
 596                 }
 597                 dout("writepage setting page/mapping error %d %p\n",
 598                      err, page);
 599                 SetPageError(page);
 600                 mapping_set_error(&inode->i_data, err);
 601                 wbc->pages_skipped++;
 602         } else {
 603                 dout("writepage cleaned page %p\n", page);
 604                 err = 0;  /* vfs expects us to return 0 */
 605         }
 606         page->private = 0;
 607         ClearPagePrivate(page);
 608         end_page_writeback(page);
 609         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 610         ceph_put_snap_context(snapc);  /* page's reference */
 611         return err;
 612 }
 613
 614 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 615 {
 616         int err;
 617         struct inode *inode = page->mapping->host;
 618         BUG_ON(!inode);
 619         ihold(inode);
 620         err = writepage_nounlock(page, wbc);
 621         if (err == -ERESTARTSYS) {
 622                 /* direct memory reclaimer was killed by SIGKILL. return 0
 623                  * to prevent caller from setting mapping/page error */
 624                 err = 0;
 625         }
 626         unlock_page(page);
 627         iput(inode);
 628         return err;
 629 }
 630
 631 /*
 632  * lame release_pages helper.  release_pages() isn't exported to
 633  * modules.
 634  */
 635 static void ceph_release_pages(struct page **pages, int num)
 636 {
 637         struct pagevec pvec;
 638         int i;
 639
 640         pagevec_init(&pvec, 0);
 641         for (i = 0; i < num; i++) {
 642                 if (pagevec_add(&pvec, pages[i]) == 0)
 643                         pagevec_release(&pvec);
 644         }
 645         pagevec_release(&pvec);
 646 }
 647
 648 /*
 649  * async writeback completion handler.
 650  *
 651  * If we get an error, set the mapping error bit, but not the individual
 652  * page error bits.
 653  */
 654 static void writepages_finish(struct ceph_osd_request *req)
 655 {
 656         struct inode *inode = req->r_inode;
 657         struct ceph_inode_info *ci = ceph_inode(inode);
 658         struct ceph_osd_data *osd_data;
 659         struct page *page;
 660         int num_pages, total_pages = 0;
 661         int i, j;
 662         int rc = req->r_result;
 663         struct ceph_snap_context *snapc = req->r_snapc;
 664         struct address_space *mapping = inode->i_mapping;
 665         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 666         bool remove_page;
 667
 668         dout("writepages_finish %p rc %d\n", inode, rc);
 669         if (rc < 0) {
 670                 mapping_set_error(mapping, rc);
 671                 ceph_set_error_write(ci);
 672         } else {
 673                 ceph_clear_error_write(ci);
 674         }
 675
 676         /*
 677          * We lost the cache cap, need to truncate the page before
 678          * it is unlocked, otherwise we'd truncate it later in the
 679          * page truncation thread, possibly losing some data that
 680          * raced its way in
 681          */
 682         remove_page = !(ceph_caps_issued(ci) &
 683                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 684
 685         /* clean all pages */
 686         for (i = 0; i < req->r_num_ops; i++) {
 687                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 688                         break;
 689
 690                 osd_data = osd_req_op_extent_osd_data(req, i);
 691                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 692                 num_pages = calc_pages_for((u64)osd_data->alignment,
 693                                            (u64)osd_data->length);
 694                 total_pages += num_pages;
 695                 for (j = 0; j < num_pages; j++) {
 696                         page = osd_data->pages[j];
 697                         BUG_ON(!page);
 698                         WARN_ON(!PageUptodate(page));
 699
 700                         if (atomic_long_dec_return(&fsc->writeback_count) <
 701                              CONGESTION_OFF_THRESH(
 702                                         fsc->mount_options->congestion_kb))
 703                                 clear_bdi_congested(inode_to_bdi(inode),
 704                                                     BLK_RW_ASYNC);
 705
 706                         ceph_put_snap_context(page_snap_context(page));
 707                         page->private = 0;
 708                         ClearPagePrivate(page);
 709                         dout("unlocking %p\n", page);
 710                         end_page_writeback(page);
 711
 712                         if (remove_page)
 713                                 generic_error_remove_page(inode->i_mapping,
 714                                                           page);
 715
 716                         unlock_page(page);
 717                 }
 718                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 719                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 720
 721                 ceph_release_pages(osd_data->pages, num_pages);
 722         }
 723
 724         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 725
 726         osd_data = osd_req_op_extent_osd_data(req, 0);
 727         if (osd_data->pages_from_pool)
 728                 mempool_free(osd_data->pages,
 729                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 730         else
 731                 kfree(osd_data->pages);
 732         ceph_osdc_put_request(req);
 733 }
 734
 735 /*
 736  * initiate async writeback
 737  */
 738 static int ceph_writepages_start(struct address_space *mapping,
 739                                  struct writeback_control *wbc)
 740 {
 741         struct inode *inode = mapping->host;
 742         struct ceph_inode_info *ci = ceph_inode(inode);
 743         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 744         struct ceph_vino vino = ceph_vino(inode);
 745         pgoff_t index, start, end;
 746         int range_whole = 0;
 747         int should_loop = 1;
 748         pgoff_t max_pages = 0, max_pages_ever = 0;
 749         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 750         struct pagevec pvec;
 751         int done = 0;
 752         int rc = 0;
 753         unsigned int wsize = i_blocksize(inode);
 754         struct ceph_osd_request *req = NULL;
 755         int do_sync = 0;
 756         loff_t snap_size, i_size;
 757         u64 truncate_size;
 758         u32 truncate_seq;
 759
 760         /*
 761          * Include a 'sync' in the OSD request if this is a data
 762          * integrity write (e.g., O_SYNC write or fsync()), or if our
 763          * cap is being revoked.
 764          */
 765         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 766                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 767                 do_sync = 1;
 768         dout("writepages_start %p dosync=%d (mode=%s)\n",
 769              inode, do_sync,
 770              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 771              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 772
 773         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 774                 if (ci->i_wrbuffer_ref > 0) {
 775                         pr_warn_ratelimited(
 776                                 "writepage_start %p %lld forced umount\n",
 777                                 inode, ceph_ino(inode));
 778                 }
 779                 mapping_set_error(mapping, -EIO);
 780                 return -EIO; /* we're in a forced umount, don't write! */
 781         }
 782         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 783                 wsize = fsc->mount_options->wsize;
 784         if (wsize < PAGE_SIZE)
 785                 wsize = PAGE_SIZE;
 786         max_pages_ever = wsize >> PAGE_SHIFT;
 787
 788         pagevec_init(&pvec, 0);
 789
 790         /* where to start/end? */
 791         if (wbc->range_cyclic) {
 792                 start = mapping->writeback_index; /* Start from prev offset */
 793                 end = -1;
 794                 dout(" cyclic, start at %lu\n", start);
 795         } else {
 796                 start = wbc->range_start >> PAGE_SHIFT;
 797                 end = wbc->range_end >> PAGE_SHIFT;
 798                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 799                         range_whole = 1;
 800                 should_loop = 0;
 801                 dout(" not cyclic, %lu to %lu\n", start, end);
 802         }
 803         index = start;
 804
 805 retry:
 806         /* find oldest snap context with dirty data */
 807         ceph_put_snap_context(snapc);
 808         snap_size = -1;
 809         snapc = get_oldest_context(inode, &snap_size,
 810                                    &truncate_size, &truncate_seq);
 811         if (!snapc) {
 812                 /* hmm, why does writepages get called when there
 813                    is no dirty data? */
 814                 dout(" no snap context with dirty data?\n");
 815                 goto out;
 816         }
 817         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 818              snapc, snapc->seq, snapc->num_snaps);
 819
 820         i_size = i_size_read(inode);
 821
 822         if (last_snapc && snapc != last_snapc) {
 823                 /* if we switched to a newer snapc, restart our scan at the
 824                  * start of the original file range. */
 825                 dout("  snapc differs from last pass, restarting at %lu\n",
 826                      index);
 827                 index = start;
 828         }
 829         last_snapc = snapc;
 830
 831         while (!done && index <= end) {
 832                 unsigned i;
 833                 int first;
 834                 pgoff_t strip_unit_end = 0;
 835                 int num_ops = 0, op_idx;
 836                 int pvec_pages, locked_pages = 0;
 837                 struct page **pages = NULL, **data_pages;
 838                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 839                 struct page *page;
 840                 int want;
 841                 u64 offset = 0, len = 0;
 842
 843                 max_pages = max_pages_ever;
 844
 845 get_more_pages:
 846                 first = -1;
 847                 want = min(end - index,
 848                            min((pgoff_t)PAGEVEC_SIZE,
 849                                max_pages - (pgoff_t)locked_pages) - 1)
 850                         + 1;
 851                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 852                                                 PAGECACHE_TAG_DIRTY,
 853                                                 want);
 854                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 855                 if (!pvec_pages && !locked_pages)
 856                         break;
 857                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 858                         page = pvec.pages[i];
 859                         dout("? %p idx %lu\n", page, page->index);
 860                         if (locked_pages == 0)
 861                                 lock_page(page);  /* first page */
 862                         else if (!trylock_page(page))
 863                                 break;
 864
 865                         /* only dirty pages, or our accounting breaks */
 866                         if (unlikely(!PageDirty(page)) ||
 867                             unlikely(page->mapping != mapping)) {
 868                                 dout("!dirty or !mapping %p\n", page);
 869                                 unlock_page(page);
 870                                 break;
 871                         }
 872                         if (!wbc->range_cyclic && page->index > end) {
 873                                 dout("end of range %p\n", page);
 874                                 done = 1;
 875                                 unlock_page(page);
 876                                 break;
 877                         }
 878                         if (strip_unit_end && (page->index > strip_unit_end)) {
 879                                 dout("end of strip unit %p\n", page);
 880                                 unlock_page(page);
 881                                 break;
 882                         }
 883                         if (wbc->sync_mode != WB_SYNC_NONE) {
 884                                 dout("waiting on writeback %p\n", page);
 885                                 wait_on_page_writeback(page);
 886                         }
 887                         if (page_offset(page) >=
 888                             (snap_size == -1 ? i_size : snap_size)) {
 889                                 dout("%p page eof %llu\n", page,
 890                                      (snap_size == -1 ? i_size : snap_size));
 891                                 done = 1;
 892                                 unlock_page(page);
 893                                 break;
 894                         }
 895                         if (PageWriteback(page)) {
 896                                 dout("%p under writeback\n", page);
 897                                 unlock_page(page);
 898                                 break;
 899                         }
 900
 901                         /* only if matching snap context */
 902                         pgsnapc = page_snap_context(page);
 903                         if (pgsnapc->seq > snapc->seq) {
 904                                 dout("page snapc %p %lld > oldest %p %lld\n",
 905                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 906                                 unlock_page(page);
 907                                 if (!locked_pages)
 908                                         continue; /* keep looking for snap */
 909                                 break;
 910                         }
 911
 912                         if (!clear_page_dirty_for_io(page)) {
 913                                 dout("%p !clear_page_dirty_for_io\n", page);
 914                                 unlock_page(page);
 915                                 break;
 916                         }
 917
 918                         /*
 919                          * We have something to write.  If this is
 920                          * the first locked page this time through,
 921                          * calculate max possinle write size and
 922                          * allocate a page array
 923                          */
 924                         if (locked_pages == 0) {
 925                                 u64 objnum;
 926                                 u64 objoff;
 927
 928                                 /* prepare async write request */
 929                                 offset = (u64)page_offset(page);
 930                                 len = wsize;
 931
 932                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 933                                                                 offset, len,
 934                                                                 &objnum, &objoff,
 935                                                                 &len);
 936                                 if (rc < 0) {
 937                                         unlock_page(page);
 938                                         break;
 939                                 }
 940
 941                                 num_ops = 1 + do_sync;
 942                                 strip_unit_end = page->index +
 943                                         ((len - 1) >> PAGE_SHIFT);
 944
 945                                 BUG_ON(pages);
 946                                 max_pages = calc_pages_for(0, (u64)len);
 947                                 pages = kmalloc(max_pages * sizeof (*pages),
 948                                                 GFP_NOFS);
 949                                 if (!pages) {
 950                                         pool = fsc->wb_pagevec_pool;
 951                                         pages = mempool_alloc(pool, GFP_NOFS);
 952                                         BUG_ON(!pages);
 953                                 }
 954
 955                                 len = 0;
 956                         } else if (page->index !=
 957                                    (offset + len) >> PAGE_SHIFT) {
 958                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 959                                                         CEPH_OSD_MAX_OPS)) {
 960                                         redirty_page_for_writepage(wbc, page);
 961                                         unlock_page(page);
 962                                         break;
 963                                 }
 964
 965                                 num_ops++;
 966                                 offset = (u64)page_offset(page);
 967                                 len = 0;
 968                         }
 969
 970                         /* note position of first page in pvec */
 971                         if (first < 0)
 972                                 first = i;
 973                         dout("%p will write page %p idx %lu\n",
 974                              inode, page, page->index);
 975
 976                         if (atomic_long_inc_return(&fsc->writeback_count) >
 977                             CONGESTION_ON_THRESH(
 978                                     fsc->mount_options->congestion_kb)) {
 979                                 set_bdi_congested(inode_to_bdi(inode),
 980                                                   BLK_RW_ASYNC);
 981                         }
 982
 983                         pages[locked_pages] = page;
 984                         locked_pages++;
 985                         len += PAGE_SIZE;
 986                 }
 987
 988                 /* did we get anything? */
 989                 if (!locked_pages)
 990                         goto release_pvec_pages;
 991                 if (i) {
 992                         int j;
 993                         BUG_ON(!locked_pages || first < 0);
 994
 995                         if (pvec_pages && i == pvec_pages &&
 996                             locked_pages < max_pages) {
 997                                 dout("reached end pvec, trying for more\n");
 998                                 pagevec_reinit(&pvec);
 999                                 goto get_more_pages;
1000                         }
1001
1002                         /* shift unused pages over in the pvec...  we
1003                          * will need to release them below. */
1004                         for (j = i; j < pvec_pages; j++) {
1005                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
1006                                 pvec.pages[j-i+first] = pvec.pages[j];
1007                         }
1008                         pvec.nr -= i-first;
1009                 }
1010
1011 new_request:
1012                 offset = page_offset(pages[0]);
1013                 len = wsize;
1014
1015                 req = ceph_osdc_new_request(&fsc->client->osdc,
1016                                         &ci->i_layout, vino,
1017                                         offset, &len, 0, num_ops,
1018                                         CEPH_OSD_OP_WRITE,
1019                                         CEPH_OSD_FLAG_WRITE,
1020                                         snapc, truncate_seq,
1021                                         truncate_size, false);
1022                 if (IS_ERR(req)) {
1023                         req = ceph_osdc_new_request(&fsc->client->osdc,
1024                                                 &ci->i_layout, vino,
1025                                                 offset, &len, 0,
1026                                                 min(num_ops,
1027                                                     CEPH_OSD_SLAB_OPS),
1028                                                 CEPH_OSD_OP_WRITE,
1029                                                 CEPH_OSD_FLAG_WRITE,
1030                                                 snapc, truncate_seq,
1031                                                 truncate_size, true);
1032                         BUG_ON(IS_ERR(req));
1033                 }
1034                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1035                              PAGE_SIZE - offset);
1036
1037                 req->r_callback = writepages_finish;
1038                 req->r_inode = inode;
1039
1040                 /* Format the osd request message and submit the write */
1041                 len = 0;
1042                 data_pages = pages;
1043                 op_idx = 0;
1044                 for (i = 0; i < locked_pages; i++) {
1045                         u64 cur_offset = page_offset(pages[i]);
1046                         if (offset + len != cur_offset) {
1047                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1048                                         break;
1049                                 osd_req_op_extent_dup_last(req, op_idx,
1050                                                            cur_offset - offset);
1051                                 dout("writepages got pages at %llu~%llu\n",
1052                                      offset, len);
1053                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1054                                                         data_pages, len, 0,
1055                                                         !!pool, false);
1056                                 osd_req_op_extent_update(req, op_idx, len);
1057
1058                                 len = 0;
1059                                 offset = cur_offset;
1060                                 data_pages = pages + i;
1061                                 op_idx++;
1062                         }
1063
1064                         set_page_writeback(pages[i]);
1065                         len += PAGE_SIZE;
1066                 }
1067
1068                 if (snap_size != -1) {
1069                         len = min(len, snap_size - offset);
1070                 } else if (i == locked_pages) {
1071                         /* writepages_finish() clears writeback pages
1072                          * according to the data length, so make sure
1073                          * data length covers all locked pages */
1074                         u64 min_len = len + 1 - PAGE_SIZE;
1075                         len = min(len, (u64)i_size_read(inode) - offset);
1076                         len = max(len, min_len);
1077                 }
1078                 dout("writepages got pages at %llu~%llu\n", offset, len);
1079
1080                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1081                                                  0, !!pool, false);
1082                 osd_req_op_extent_update(req, op_idx, len);
1083
1084                 if (do_sync) {
1085                         op_idx++;
1086                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1087                 }
1088                 BUG_ON(op_idx + 1 != req->r_num_ops);
1089
1090                 pool = NULL;
1091                 if (i < locked_pages) {
1092                         BUG_ON(num_ops <= req->r_num_ops);
1093                         num_ops -= req->r_num_ops;
1094                         num_ops += do_sync;
1095                         locked_pages -= i;
1096
1097                         /* allocate new pages array for next request */
1098                         data_pages = pages;
1099                         pages = kmalloc(locked_pages * sizeof (*pages),
1100                                         GFP_NOFS);
1101                         if (!pages) {
1102                                 pool = fsc->wb_pagevec_pool;
1103                                 pages = mempool_alloc(pool, GFP_NOFS);
1104                                 BUG_ON(!pages);
1105                         }
1106                         memcpy(pages, data_pages + i,
1107                                locked_pages * sizeof(*pages));
1108                         memset(data_pages + i, 0,
1109                                locked_pages * sizeof(*pages));
1110                 } else {
1111                         BUG_ON(num_ops != req->r_num_ops);
1112                         index = pages[i - 1]->index + 1;
1113                         /* request message now owns the pages array */
1114                         pages = NULL;
1115                 }
1116
1117                 req->r_mtime = inode->i_mtime;
1118                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1119                 BUG_ON(rc);
1120                 req = NULL;
1121
1122                 wbc->nr_to_write -= i;
1123                 if (pages)
1124                         goto new_request;
1125
1126                 if (wbc->nr_to_write <= 0)
1127                         done = 1;
1128
1129 release_pvec_pages:
1130                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1131                      pvec.nr ? pvec.pages[0] : NULL);
1132                 pagevec_release(&pvec);
1133
1134                 if (locked_pages && !done)
1135                         goto retry;
1136         }
1137
1138         if (should_loop && !done) {
1139                 /* more to do; loop back to beginning of file */
1140                 dout("writepages looping back to beginning of file\n");
1141                 should_loop = 0;
1142                 index = 0;
1143                 goto retry;
1144         }
1145
1146         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1147                 mapping->writeback_index = index;
1148
1149 out:
1150         ceph_osdc_put_request(req);
1151         ceph_put_snap_context(snapc);
1152         dout("writepages done, rc = %d\n", rc);
1153         return rc;
1154 }
1155
1156
1157
1158 /*
1159  * See if a given @snapc is either writeable, or already written.
1160  */
1161 static int context_is_writeable_or_written(struct inode *inode,
1162                                            struct ceph_snap_context *snapc)
1163 {
1164         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL,
1165                                                               NULL, NULL);
1166         int ret = !oldest || snapc->seq <= oldest->seq;
1167
1168         ceph_put_snap_context(oldest);
1169         return ret;
1170 }
1171
1172 /*
1173  * We are only allowed to write into/dirty the page if the page is
1174  * clean, or already dirty within the same snap context.
1175  *
1176  * called with page locked.
1177  * return success with page locked,
1178  * or any failure (incl -EAGAIN) with page unlocked.
1179  */
1180 static int ceph_update_writeable_page(struct file *file,
1181                             loff_t pos, unsigned len,
1182                             struct page *page)
1183 {
1184         struct inode *inode = file_inode(file);
1185         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1186         struct ceph_inode_info *ci = ceph_inode(inode);
1187         loff_t page_off = pos & PAGE_MASK;
1188         int pos_in_page = pos & ~PAGE_MASK;
1189         int end_in_page = pos_in_page + len;
1190         loff_t i_size;
1191         int r;
1192         struct ceph_snap_context *snapc, *oldest;
1193
1194         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1195                 dout(" page %p forced umount\n", page);
1196                 unlock_page(page);
1197                 return -EIO;
1198         }
1199
1200 retry_locked:
1201         /* writepages currently holds page lock, but if we change that later, */
1202         wait_on_page_writeback(page);
1203
1204         snapc = page_snap_context(page);
1205         if (snapc && snapc != ci->i_head_snapc) {
1206                 /*
1207                  * this page is already dirty in another (older) snap
1208                  * context!  is it writeable now?
1209                  */
1210                 oldest = get_oldest_context(inode, NULL, NULL, NULL);
1211
1212                 if (snapc->seq > oldest->seq) {
1213                         ceph_put_snap_context(oldest);
1214                         dout(" page %p snapc %p not current or oldest\n",
1215                              page, snapc);
1216                         /*
1217                          * queue for writeback, and wait for snapc to
1218                          * be writeable or written
1219                          */
1220                         snapc = ceph_get_snap_context(snapc);
1221                         unlock_page(page);
1222                         ceph_queue_writeback(inode);
1223                         r = wait_event_killable(ci->i_cap_wq,
1224                                context_is_writeable_or_written(inode, snapc));
1225                         ceph_put_snap_context(snapc);
1226                         if (r == -ERESTARTSYS)
1227                                 return r;
1228                         return -EAGAIN;
1229                 }
1230                 ceph_put_snap_context(oldest);
1231
1232                 /* yay, writeable, do it now (without dropping page lock) */
1233                 dout(" page %p snapc %p not current, but oldest\n",
1234                      page, snapc);
1235                 if (!clear_page_dirty_for_io(page))
1236                         goto retry_locked;
1237                 r = writepage_nounlock(page, NULL);
1238                 if (r < 0)
1239                         goto fail_unlock;
1240                 goto retry_locked;
1241         }
1242
1243         if (PageUptodate(page)) {
1244                 dout(" page %p already uptodate\n", page);
1245                 return 0;
1246         }
1247
1248         /* full page? */
1249         if (pos_in_page == 0 && len == PAGE_SIZE)
1250                 return 0;
1251
1252         /* past end of file? */
1253         i_size = i_size_read(inode);
1254
1255         if (page_off >= i_size ||
1256             (pos_in_page == 0 && (pos+len) >= i_size &&
1257              end_in_page - pos_in_page != PAGE_SIZE)) {
1258                 dout(" zeroing %p 0 - %d and %d - %d\n",
1259                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1260                 zero_user_segments(page,
1261                                    0, pos_in_page,
1262                                    end_in_page, PAGE_SIZE);
1263                 return 0;
1264         }
1265
1266         /* we need to read it. */
1267         r = ceph_do_readpage(file, page);
1268         if (r < 0) {
1269                 if (r == -EINPROGRESS)
1270                         return -EAGAIN;
1271                 goto fail_unlock;
1272         }
1273         goto retry_locked;
1274 fail_unlock:
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         bool check_cap = false;
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_fault *vmf)
1390 {
1391         struct vm_area_struct *vma = vmf->vma;
1392         struct inode *inode = file_inode(vma->vm_file);
1393         struct ceph_inode_info *ci = ceph_inode(inode);
1394         struct ceph_file_info *fi = vma->vm_file->private_data;
1395         struct page *pinned_page = NULL;
1396         loff_t off = vmf->pgoff << PAGE_SHIFT;
1397         int want, got, ret;
1398         sigset_t oldset;
1399
1400         ceph_block_sigs(&oldset);
1401
1402         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1403              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1404         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1405                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1406         else
1407                 want = CEPH_CAP_FILE_CACHE;
1408
1409         got = 0;
1410         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1411         if (ret < 0)
1412                 goto out_restore;
1413
1414         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1415              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1416
1417         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1418             ci->i_inline_version == CEPH_INLINE_NONE) {
1419                 current->journal_info = vma->vm_file;
1420                 ret = filemap_fault(vmf);
1421                 current->journal_info = NULL;
1422         } else
1423                 ret = -EAGAIN;
1424
1425         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1426              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1427         if (pinned_page)
1428                 put_page(pinned_page);
1429         ceph_put_cap_refs(ci, got);
1430
1431         if (ret != -EAGAIN)
1432                 goto out_restore;
1433
1434         /* read inline data */
1435         if (off >= PAGE_SIZE) {
1436                 /* does not support inline data > PAGE_SIZE */
1437                 ret = VM_FAULT_SIGBUS;
1438         } else {
1439                 int ret1;
1440                 struct address_space *mapping = inode->i_mapping;
1441                 struct page *page = find_or_create_page(mapping, 0,
1442                                                 mapping_gfp_constraint(mapping,
1443                                                 ~__GFP_FS));
1444                 if (!page) {
1445                         ret = VM_FAULT_OOM;
1446                         goto out_inline;
1447                 }
1448                 ret1 = __ceph_do_getattr(inode, page,
1449                                          CEPH_STAT_CAP_INLINE_DATA, true);
1450                 if (ret1 < 0 || off >= i_size_read(inode)) {
1451                         unlock_page(page);
1452                         put_page(page);
1453                         if (ret1 < 0)
1454                                 ret = ret1;
1455                         else
1456                                 ret = VM_FAULT_SIGBUS;
1457                         goto out_inline;
1458                 }
1459                 if (ret1 < PAGE_SIZE)
1460                         zero_user_segment(page, ret1, PAGE_SIZE);
1461                 else
1462                         flush_dcache_page(page);
1463                 SetPageUptodate(page);
1464                 vmf->page = page;
1465                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1466 out_inline:
1467                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1468                      inode, off, (size_t)PAGE_SIZE, ret);
1469         }
1470 out_restore:
1471         ceph_restore_sigs(&oldset);
1472         if (ret < 0)
1473                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1474
1475         return ret;
1476 }
1477
1478 /*
1479  * Reuse write_begin here for simplicity.
1480  */
1481 static int ceph_page_mkwrite(struct vm_fault *vmf)
1482 {
1483         struct vm_area_struct *vma = vmf->vma;
1484         struct inode *inode = file_inode(vma->vm_file);
1485         struct ceph_inode_info *ci = ceph_inode(inode);
1486         struct ceph_file_info *fi = vma->vm_file->private_data;
1487         struct ceph_cap_flush *prealloc_cf;
1488         struct page *page = vmf->page;
1489         loff_t off = page_offset(page);
1490         loff_t size = i_size_read(inode);
1491         size_t len;
1492         int want, got, ret;
1493         sigset_t oldset;
1494
1495         prealloc_cf = ceph_alloc_cap_flush();
1496         if (!prealloc_cf)
1497                 return VM_FAULT_OOM;
1498
1499         ceph_block_sigs(&oldset);
1500
1501         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1502                 struct page *locked_page = NULL;
1503                 if (off == 0) {
1504                         lock_page(page);
1505                         locked_page = page;
1506                 }
1507                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1508                 if (locked_page)
1509                         unlock_page(locked_page);
1510                 if (ret < 0)
1511                         goto out_free;
1512         }
1513
1514         if (off + PAGE_SIZE <= size)
1515                 len = PAGE_SIZE;
1516         else
1517                 len = size & ~PAGE_MASK;
1518
1519         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1520              inode, ceph_vinop(inode), off, len, size);
1521         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1522                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1523         else
1524                 want = CEPH_CAP_FILE_BUFFER;
1525
1526         got = 0;
1527         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1528                             &got, NULL);
1529         if (ret < 0)
1530                 goto out_free;
1531
1532         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1533              inode, off, len, ceph_cap_string(got));
1534
1535         /* Update time before taking page lock */
1536         file_update_time(vma->vm_file);
1537
1538         do {
1539                 lock_page(page);
1540
1541                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1542                         unlock_page(page);
1543                         ret = VM_FAULT_NOPAGE;
1544                         break;
1545                 }
1546
1547                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1548                 if (ret >= 0) {
1549                         /* success.  we'll keep the page locked. */
1550                         set_page_dirty(page);
1551                         ret = VM_FAULT_LOCKED;
1552                 }
1553         } while (ret == -EAGAIN);
1554
1555         if (ret == VM_FAULT_LOCKED ||
1556             ci->i_inline_version != CEPH_INLINE_NONE) {
1557                 int dirty;
1558                 spin_lock(&ci->i_ceph_lock);
1559                 ci->i_inline_version = CEPH_INLINE_NONE;
1560                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1561                                                &prealloc_cf);
1562                 spin_unlock(&ci->i_ceph_lock);
1563                 if (dirty)
1564                         __mark_inode_dirty(inode, dirty);
1565         }
1566
1567         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1568              inode, off, len, ceph_cap_string(got), ret);
1569         ceph_put_cap_refs(ci, got);
1570 out_free:
1571         ceph_restore_sigs(&oldset);
1572         ceph_free_cap_flush(prealloc_cf);
1573         if (ret < 0)
1574                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1575         return ret;
1576 }
1577
1578 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1579                            char *data, size_t len)
1580 {
1581         struct address_space *mapping = inode->i_mapping;
1582         struct page *page;
1583
1584         if (locked_page) {
1585                 page = locked_page;
1586         } else {
1587                 if (i_size_read(inode) == 0)
1588                         return;
1589                 page = find_or_create_page(mapping, 0,
1590                                            mapping_gfp_constraint(mapping,
1591                                            ~__GFP_FS));
1592                 if (!page)
1593                         return;
1594                 if (PageUptodate(page)) {
1595                         unlock_page(page);
1596                         put_page(page);
1597                         return;
1598                 }
1599         }
1600
1601         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1602              inode, ceph_vinop(inode), len, locked_page);
1603
1604         if (len > 0) {
1605                 void *kaddr = kmap_atomic(page);
1606                 memcpy(kaddr, data, len);
1607                 kunmap_atomic(kaddr);
1608         }
1609
1610         if (page != locked_page) {
1611                 if (len < PAGE_SIZE)
1612                         zero_user_segment(page, len, PAGE_SIZE);
1613                 else
1614                         flush_dcache_page(page);
1615
1616                 SetPageUptodate(page);
1617                 unlock_page(page);
1618                 put_page(page);
1619         }
1620 }
1621
1622 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1623 {
1624         struct inode *inode = file_inode(filp);
1625         struct ceph_inode_info *ci = ceph_inode(inode);
1626         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1627         struct ceph_osd_request *req;
1628         struct page *page = NULL;
1629         u64 len, inline_version;
1630         int err = 0;
1631         bool from_pagecache = false;
1632
1633         spin_lock(&ci->i_ceph_lock);
1634         inline_version = ci->i_inline_version;
1635         spin_unlock(&ci->i_ceph_lock);
1636
1637         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1638              inode, ceph_vinop(inode), inline_version);
1639
1640         if (inline_version == 1 || /* initial version, no data */
1641             inline_version == CEPH_INLINE_NONE)
1642                 goto out;
1643
1644         if (locked_page) {
1645                 page = locked_page;
1646                 WARN_ON(!PageUptodate(page));
1647         } else if (ceph_caps_issued(ci) &
1648                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1649                 page = find_get_page(inode->i_mapping, 0);
1650                 if (page) {
1651                         if (PageUptodate(page)) {
1652                                 from_pagecache = true;
1653                                 lock_page(page);
1654                         } else {
1655                                 put_page(page);
1656                                 page = NULL;
1657                         }
1658                 }
1659         }
1660
1661         if (page) {
1662                 len = i_size_read(inode);
1663                 if (len > PAGE_SIZE)
1664                         len = PAGE_SIZE;
1665         } else {
1666                 page = __page_cache_alloc(GFP_NOFS);
1667                 if (!page) {
1668                         err = -ENOMEM;
1669                         goto out;
1670                 }
1671                 err = __ceph_do_getattr(inode, page,
1672                                         CEPH_STAT_CAP_INLINE_DATA, true);
1673                 if (err < 0) {
1674                         /* no inline data */
1675                         if (err == -ENODATA)
1676                                 err = 0;
1677                         goto out;
1678                 }
1679                 len = err;
1680         }
1681
1682         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1683                                     ceph_vino(inode), 0, &len, 0, 1,
1684                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1685                                     NULL, 0, 0, false);
1686         if (IS_ERR(req)) {
1687                 err = PTR_ERR(req);
1688                 goto out;
1689         }
1690
1691         req->r_mtime = inode->i_mtime;
1692         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1693         if (!err)
1694                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1695         ceph_osdc_put_request(req);
1696         if (err < 0)
1697                 goto out;
1698
1699         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1700                                     ceph_vino(inode), 0, &len, 1, 3,
1701                                     CEPH_OSD_OP_WRITE, 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;
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         wr_req->r_abort_on_full = true;
1896         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1897
1898         if (!err)
1899                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1900         if (!err2)
1901                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1902
1903         if (err >= 0 || err == -ENOENT)
1904                 have |= POOL_READ;
1905         else if (err != -EPERM)
1906                 goto out_unlock;
1907
1908         if (err2 == 0 || err2 == -EEXIST)
1909                 have |= POOL_WRITE;
1910         else if (err2 != -EPERM) {
1911                 err = err2;
1912                 goto out_unlock;
1913         }
1914
1915         pool_ns_len = pool_ns ? pool_ns->len : 0;
1916         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1917         if (!perm) {
1918                 err = -ENOMEM;
1919                 goto out_unlock;
1920         }
1921
1922         perm->pool = pool;
1923         perm->perm = have;
1924         perm->pool_ns_len = pool_ns_len;
1925         if (pool_ns_len > 0)
1926                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1927         perm->pool_ns[pool_ns_len] = 0;
1928
1929         rb_link_node(&perm->node, parent, p);
1930         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1931         err = 0;
1932 out_unlock:
1933         up_write(&mdsc->pool_perm_rwsem);
1934
1935         ceph_osdc_put_request(rd_req);
1936         ceph_osdc_put_request(wr_req);
1937 out:
1938         if (!err)
1939                 err = have;
1940         if (pool_ns)
1941                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1942                      pool, (int)pool_ns->len, pool_ns->str, err);
1943         else
1944                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1945         return err;
1946 }
1947
1948 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1949 {
1950         s64 pool;
1951         struct ceph_string *pool_ns;
1952         int ret, flags;
1953
1954         if (ci->i_vino.snap != CEPH_NOSNAP) {
1955                 /*
1956                  * Pool permission check needs to write to the first object.
1957                  * But for snapshot, head of the first object may have alread
1958                  * been deleted. Skip check to avoid creating orphan object.
1959                  */
1960                 return 0;
1961         }
1962
1963         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1964                                 NOPOOLPERM))
1965                 return 0;
1966
1967         spin_lock(&ci->i_ceph_lock);
1968         flags = ci->i_ceph_flags;
1969         pool = ci->i_layout.pool_id;
1970         spin_unlock(&ci->i_ceph_lock);
1971 check:
1972         if (flags & CEPH_I_POOL_PERM) {
1973                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1974                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1975                              pool);
1976                         return -EPERM;
1977                 }
1978                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1979                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1980                              pool);
1981                         return -EPERM;
1982                 }
1983                 return 0;
1984         }
1985
1986         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1987         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1988         ceph_put_string(pool_ns);
1989         if (ret < 0)
1990                 return ret;
1991
1992         flags = CEPH_I_POOL_PERM;
1993         if (ret & POOL_READ)
1994                 flags |= CEPH_I_POOL_RD;
1995         if (ret & POOL_WRITE)
1996                 flags |= CEPH_I_POOL_WR;
1997
1998         spin_lock(&ci->i_ceph_lock);
1999         if (pool == ci->i_layout.pool_id &&
2000             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2001                 ci->i_ceph_flags |= flags;
2002         } else {
2003                 pool = ci->i_layout.pool_id;
2004                 flags = ci->i_ceph_flags;
2005         }
2006         spin_unlock(&ci->i_ceph_lock);
2007         goto check;
2008 }
2009
2010 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2011 {
2012         struct ceph_pool_perm *perm;
2013         struct rb_node *n;
2014
2015         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2016                 n = rb_first(&mdsc->pool_perm_tree);
2017                 perm = rb_entry(n, struct ceph_pool_perm, node);
2018                 rb_erase(n, &mdsc->pool_perm_tree);
2019                 kfree(perm);
2020         }
2021 }