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         if (fsc->mount_options->rsize >= PAGE_SIZE)
 459                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 460                         >> PAGE_SHIFT;
 461
 462         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 463                 file, nr_pages,
 464              max);
 465         while (!list_empty(page_list)) {
 466                 rc = start_read(inode, page_list, max);
 467                 if (rc < 0)
 468                         goto out;
 469         }
 470 out:
 471         ceph_fscache_readpages_cancel(inode, page_list);
 472
 473         dout("readpages %p file %p ret %d\n", inode, file, rc);
 474         return rc;
 475 }
 476
 477 /*
 478  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 479  * only snap context we are allowed to write back.
 480  */
 481 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 482                                                     loff_t *snap_size,
 483                                                     u64 *truncate_size,
 484                                                     u32 *truncate_seq)
 485 {
 486         struct ceph_inode_info *ci = ceph_inode(inode);
 487         struct ceph_snap_context *snapc = NULL;
 488         struct ceph_cap_snap *capsnap = NULL;
 489
 490         spin_lock(&ci->i_ceph_lock);
 491         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 492                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 493                      capsnap->context, capsnap->dirty_pages);
 494                 if (capsnap->dirty_pages) {
 495                         snapc = ceph_get_snap_context(capsnap->context);
 496                         if (snap_size)
 497                                 *snap_size = capsnap->size;
 498                         if (truncate_size)
 499                                 *truncate_size = capsnap->truncate_size;
 500                         if (truncate_seq)
 501                                 *truncate_seq = capsnap->truncate_seq;
 502                         break;
 503                 }
 504         }
 505         if (!snapc && ci->i_wrbuffer_ref_head) {
 506                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 507                 dout(" head snapc %p has %d dirty pages\n",
 508                      snapc, ci->i_wrbuffer_ref_head);
 509                 if (truncate_size)
 510                         *truncate_size = ci->i_truncate_size;
 511                 if (truncate_seq)
 512                         *truncate_seq = ci->i_truncate_seq;
 513         }
 514         spin_unlock(&ci->i_ceph_lock);
 515         return snapc;
 516 }
 517
 518 /*
 519  * Write a single page, but leave the page locked.
 520  *
 521  * If we get a write error, set the page error bit, but still adjust the
 522  * dirty page accounting (i.e., page is no longer dirty).
 523  */
 524 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 525 {
 526         struct inode *inode;
 527         struct ceph_inode_info *ci;
 528         struct ceph_fs_client *fsc;
 529         struct ceph_osd_client *osdc;
 530         struct ceph_snap_context *snapc, *oldest;
 531         loff_t page_off = page_offset(page);
 532         loff_t snap_size = -1;
 533         long writeback_stat;
 534         u64 truncate_size;
 535         u32 truncate_seq;
 536         int err, len = PAGE_SIZE;
 537
 538         dout("writepage %p idx %lu\n", page, page->index);
 539
 540         inode = page->mapping->host;
 541         ci = ceph_inode(inode);
 542         fsc = ceph_inode_to_client(inode);
 543         osdc = &fsc->client->osdc;
 544
 545         /* verify this is a writeable snap context */
 546         snapc = page_snap_context(page);
 547         if (snapc == NULL) {
 548                 dout("writepage %p page %p not dirty?\n", inode, page);
 549                 return 0;
 550         }
 551         oldest = get_oldest_context(inode, &snap_size,
 552                                     &truncate_size, &truncate_seq);
 553         if (snapc->seq > oldest->seq) {
 554                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 555                      inode, page, snapc);
 556                 /* we should only noop if called by kswapd */
 557                 WARN_ON(!(current->flags & PF_MEMALLOC));
 558                 ceph_put_snap_context(oldest);
 559                 redirty_page_for_writepage(wbc, page);
 560                 return 0;
 561         }
 562         ceph_put_snap_context(oldest);
 563
 564         if (snap_size == -1)
 565                 snap_size = i_size_read(inode);
 566
 567         /* is this a partial page at end of file? */
 568         if (page_off >= snap_size) {
 569                 dout("%p page eof %llu\n", page, snap_size);
 570                 return 0;
 571         }
 572
 573         if (snap_size < page_off + len)
 574                 len = snap_size - page_off;
 575
 576         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 577              inode, page, page->index, page_off, len, snapc);
 578
 579         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 580         if (writeback_stat >
 581             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 582                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 583
 584         set_page_writeback(page);
 585         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 586                                    &ci->i_layout, snapc,
 587                                    page_off, len,
 588                                    truncate_seq, truncate_size,
 589                                    &inode->i_mtime, &page, 1);
 590         if (err < 0) {
 591                 struct writeback_control tmp_wbc;
 592                 if (!wbc)
 593                         wbc = &tmp_wbc;
 594                 if (err == -ERESTARTSYS) {
 595                         /* killed by SIGKILL */
 596                         dout("writepage interrupted page %p\n", page);
 597                         redirty_page_for_writepage(wbc, page);
 598                         end_page_writeback(page);
 599                         return err;
 600                 }
 601                 dout("writepage setting page/mapping error %d %p\n",
 602                      err, page);
 603                 SetPageError(page);
 604                 mapping_set_error(&inode->i_data, err);
 605                 wbc->pages_skipped++;
 606         } else {
 607                 dout("writepage cleaned page %p\n", page);
 608                 err = 0;  /* vfs expects us to return 0 */
 609         }
 610         page->private = 0;
 611         ClearPagePrivate(page);
 612         end_page_writeback(page);
 613         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 614         ceph_put_snap_context(snapc);  /* page's reference */
 615         return err;
 616 }
 617
 618 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 619 {
 620         int err;
 621         struct inode *inode = page->mapping->host;
 622         BUG_ON(!inode);
 623         ihold(inode);
 624         err = writepage_nounlock(page, wbc);
 625         if (err == -ERESTARTSYS) {
 626                 /* direct memory reclaimer was killed by SIGKILL. return 0
 627                  * to prevent caller from setting mapping/page error */
 628                 err = 0;
 629         }
 630         unlock_page(page);
 631         iput(inode);
 632         return err;
 633 }
 634
 635 /*
 636  * lame release_pages helper.  release_pages() isn't exported to
 637  * modules.
 638  */
 639 static void ceph_release_pages(struct page **pages, int num)
 640 {
 641         struct pagevec pvec;
 642         int i;
 643
 644         pagevec_init(&pvec, 0);
 645         for (i = 0; i < num; i++) {
 646                 if (pagevec_add(&pvec, pages[i]) == 0)
 647                         pagevec_release(&pvec);
 648         }
 649         pagevec_release(&pvec);
 650 }
 651
 652 /*
 653  * async writeback completion handler.
 654  *
 655  * If we get an error, set the mapping error bit, but not the individual
 656  * page error bits.
 657  */
 658 static void writepages_finish(struct ceph_osd_request *req)
 659 {
 660         struct inode *inode = req->r_inode;
 661         struct ceph_inode_info *ci = ceph_inode(inode);
 662         struct ceph_osd_data *osd_data;
 663         struct page *page;
 664         int num_pages, total_pages = 0;
 665         int i, j;
 666         int rc = req->r_result;
 667         struct ceph_snap_context *snapc = req->r_snapc;
 668         struct address_space *mapping = inode->i_mapping;
 669         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 670         bool remove_page;
 671
 672         dout("writepages_finish %p rc %d\n", inode, rc);
 673         if (rc < 0) {
 674                 mapping_set_error(mapping, rc);
 675                 ceph_set_error_write(ci);
 676         } else {
 677                 ceph_clear_error_write(ci);
 678         }
 679
 680         /*
 681          * We lost the cache cap, need to truncate the page before
 682          * it is unlocked, otherwise we'd truncate it later in the
 683          * page truncation thread, possibly losing some data that
 684          * raced its way in
 685          */
 686         remove_page = !(ceph_caps_issued(ci) &
 687                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 688
 689         /* clean all pages */
 690         for (i = 0; i < req->r_num_ops; i++) {
 691                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 692                         break;
 693
 694                 osd_data = osd_req_op_extent_osd_data(req, i);
 695                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 696                 num_pages = calc_pages_for((u64)osd_data->alignment,
 697                                            (u64)osd_data->length);
 698                 total_pages += num_pages;
 699                 for (j = 0; j < num_pages; j++) {
 700                         page = osd_data->pages[j];
 701                         BUG_ON(!page);
 702                         WARN_ON(!PageUptodate(page));
 703
 704                         if (atomic_long_dec_return(&fsc->writeback_count) <
 705                              CONGESTION_OFF_THRESH(
 706                                         fsc->mount_options->congestion_kb))
 707                                 clear_bdi_congested(inode_to_bdi(inode),
 708                                                     BLK_RW_ASYNC);
 709
 710                         ceph_put_snap_context(page_snap_context(page));
 711                         page->private = 0;
 712                         ClearPagePrivate(page);
 713                         dout("unlocking %p\n", page);
 714                         end_page_writeback(page);
 715
 716                         if (remove_page)
 717                                 generic_error_remove_page(inode->i_mapping,
 718                                                           page);
 719
 720                         unlock_page(page);
 721                 }
 722                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 723                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 724
 725                 ceph_release_pages(osd_data->pages, num_pages);
 726         }
 727
 728         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 729
 730         osd_data = osd_req_op_extent_osd_data(req, 0);
 731         if (osd_data->pages_from_pool)
 732                 mempool_free(osd_data->pages,
 733                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 734         else
 735                 kfree(osd_data->pages);
 736         ceph_osdc_put_request(req);
 737 }
 738
 739 /*
 740  * initiate async writeback
 741  */
 742 static int ceph_writepages_start(struct address_space *mapping,
 743                                  struct writeback_control *wbc)
 744 {
 745         struct inode *inode = mapping->host;
 746         struct ceph_inode_info *ci = ceph_inode(inode);
 747         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 748         struct ceph_vino vino = ceph_vino(inode);
 749         pgoff_t index, start, end;
 750         int range_whole = 0;
 751         int should_loop = 1;
 752         pgoff_t max_pages = 0, max_pages_ever = 0;
 753         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 754         struct pagevec pvec;
 755         int done = 0;
 756         int rc = 0;
 757         unsigned int wsize = i_blocksize(inode);
 758         struct ceph_osd_request *req = NULL;
 759         int do_sync = 0;
 760         loff_t snap_size, i_size;
 761         u64 truncate_size;
 762         u32 truncate_seq;
 763
 764         /*
 765          * Include a 'sync' in the OSD request if this is a data
 766          * integrity write (e.g., O_SYNC write or fsync()), or if our
 767          * cap is being revoked.
 768          */
 769         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 770                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 771                 do_sync = 1;
 772         dout("writepages_start %p dosync=%d (mode=%s)\n",
 773              inode, do_sync,
 774              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 775              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 776
 777         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 778                 if (ci->i_wrbuffer_ref > 0) {
 779                         pr_warn_ratelimited(
 780                                 "writepage_start %p %lld forced umount\n",
 781                                 inode, ceph_ino(inode));
 782                 }
 783                 mapping_set_error(mapping, -EIO);
 784                 return -EIO; /* we're in a forced umount, don't write! */
 785         }
 786         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 787                 wsize = fsc->mount_options->wsize;
 788         if (wsize < PAGE_SIZE)
 789                 wsize = PAGE_SIZE;
 790         max_pages_ever = wsize >> PAGE_SHIFT;
 791
 792         pagevec_init(&pvec, 0);
 793
 794         /* where to start/end? */
 795         if (wbc->range_cyclic) {
 796                 start = mapping->writeback_index; /* Start from prev offset */
 797                 end = -1;
 798                 dout(" cyclic, start at %lu\n", start);
 799         } else {
 800                 start = wbc->range_start >> PAGE_SHIFT;
 801                 end = wbc->range_end >> PAGE_SHIFT;
 802                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 803                         range_whole = 1;
 804                 should_loop = 0;
 805                 dout(" not cyclic, %lu to %lu\n", start, end);
 806         }
 807         index = start;
 808
 809 retry:
 810         /* find oldest snap context with dirty data */
 811         ceph_put_snap_context(snapc);
 812         snap_size = -1;
 813         snapc = get_oldest_context(inode, &snap_size,
 814                                    &truncate_size, &truncate_seq);
 815         if (!snapc) {
 816                 /* hmm, why does writepages get called when there
 817                    is no dirty data? */
 818                 dout(" no snap context with dirty data?\n");
 819                 goto out;
 820         }
 821         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 822              snapc, snapc->seq, snapc->num_snaps);
 823
 824         i_size = i_size_read(inode);
 825
 826         if (last_snapc && snapc != last_snapc) {
 827                 /* if we switched to a newer snapc, restart our scan at the
 828                  * start of the original file range. */
 829                 dout("  snapc differs from last pass, restarting at %lu\n",
 830                      index);
 831                 index = start;
 832         }
 833         last_snapc = snapc;
 834
 835         while (!done && index <= end) {
 836                 unsigned i;
 837                 int first;
 838                 pgoff_t strip_unit_end = 0;
 839                 int num_ops = 0, op_idx;
 840                 int pvec_pages, locked_pages = 0;
 841                 struct page **pages = NULL, **data_pages;
 842                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 843                 struct page *page;
 844                 int want;
 845                 u64 offset = 0, len = 0;
 846
 847                 max_pages = max_pages_ever;
 848
 849 get_more_pages:
 850                 first = -1;
 851                 want = min(end - index,
 852                            min((pgoff_t)PAGEVEC_SIZE,
 853                                max_pages - (pgoff_t)locked_pages) - 1)
 854                         + 1;
 855                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 856                                                 PAGECACHE_TAG_DIRTY,
 857                                                 want);
 858                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 859                 if (!pvec_pages && !locked_pages)
 860                         break;
 861                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 862                         page = pvec.pages[i];
 863                         dout("? %p idx %lu\n", page, page->index);
 864                         if (locked_pages == 0)
 865                                 lock_page(page);  /* first page */
 866                         else if (!trylock_page(page))
 867                                 break;
 868
 869                         /* only dirty pages, or our accounting breaks */
 870                         if (unlikely(!PageDirty(page)) ||
 871                             unlikely(page->mapping != mapping)) {
 872                                 dout("!dirty or !mapping %p\n", page);
 873                                 unlock_page(page);
 874                                 break;
 875                         }
 876                         if (!wbc->range_cyclic && page->index > end) {
 877                                 dout("end of range %p\n", page);
 878                                 done = 1;
 879                                 unlock_page(page);
 880                                 break;
 881                         }
 882                         if (strip_unit_end && (page->index > strip_unit_end)) {
 883                                 dout("end of strip unit %p\n", page);
 884                                 unlock_page(page);
 885                                 break;
 886                         }
 887                         if (wbc->sync_mode != WB_SYNC_NONE) {
 888                                 dout("waiting on writeback %p\n", page);
 889                                 wait_on_page_writeback(page);
 890                         }
 891                         if (page_offset(page) >=
 892                             (snap_size == -1 ? i_size : snap_size)) {
 893                                 dout("%p page eof %llu\n", page,
 894                                      (snap_size == -1 ? i_size : snap_size));
 895                                 done = 1;
 896                                 unlock_page(page);
 897                                 break;
 898                         }
 899                         if (PageWriteback(page)) {
 900                                 dout("%p under writeback\n", page);
 901                                 unlock_page(page);
 902                                 break;
 903                         }
 904
 905                         /* only if matching snap context */
 906                         pgsnapc = page_snap_context(page);
 907                         if (pgsnapc->seq > snapc->seq) {
 908                                 dout("page snapc %p %lld > oldest %p %lld\n",
 909                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 910                                 unlock_page(page);
 911                                 if (!locked_pages)
 912                                         continue; /* keep looking for snap */
 913                                 break;
 914                         }
 915
 916                         if (!clear_page_dirty_for_io(page)) {
 917                                 dout("%p !clear_page_dirty_for_io\n", page);
 918                                 unlock_page(page);
 919                                 break;
 920                         }
 921
 922                         /*
 923                          * We have something to write.  If this is
 924                          * the first locked page this time through,
 925                          * calculate max possinle write size and
 926                          * allocate a page array
 927                          */
 928                         if (locked_pages == 0) {
 929                                 u64 objnum;
 930                                 u64 objoff;
 931
 932                                 /* prepare async write request */
 933                                 offset = (u64)page_offset(page);
 934                                 len = wsize;
 935
 936                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 937                                                                 offset, len,
 938                                                                 &objnum, &objoff,
 939                                                                 &len);
 940                                 if (rc < 0) {
 941                                         unlock_page(page);
 942                                         break;
 943                                 }
 944
 945                                 num_ops = 1 + do_sync;
 946                                 strip_unit_end = page->index +
 947                                         ((len - 1) >> PAGE_SHIFT);
 948
 949                                 BUG_ON(pages);
 950                                 max_pages = calc_pages_for(0, (u64)len);
 951                                 pages = kmalloc(max_pages * sizeof (*pages),
 952                                                 GFP_NOFS);
 953                                 if (!pages) {
 954                                         pool = fsc->wb_pagevec_pool;
 955                                         pages = mempool_alloc(pool, GFP_NOFS);
 956                                         BUG_ON(!pages);
 957                                 }
 958
 959                                 len = 0;
 960                         } else if (page->index !=
 961                                    (offset + len) >> PAGE_SHIFT) {
 962                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 963                                                         CEPH_OSD_MAX_OPS)) {
 964                                         redirty_page_for_writepage(wbc, page);
 965                                         unlock_page(page);
 966                                         break;
 967                                 }
 968
 969                                 num_ops++;
 970                                 offset = (u64)page_offset(page);
 971                                 len = 0;
 972                         }
 973
 974                         /* note position of first page in pvec */
 975                         if (first < 0)
 976                                 first = i;
 977                         dout("%p will write page %p idx %lu\n",
 978                              inode, page, page->index);
 979
 980                         if (atomic_long_inc_return(&fsc->writeback_count) >
 981                             CONGESTION_ON_THRESH(
 982                                     fsc->mount_options->congestion_kb)) {
 983                                 set_bdi_congested(inode_to_bdi(inode),
 984                                                   BLK_RW_ASYNC);
 985                         }
 986
 987                         pages[locked_pages] = page;
 988                         locked_pages++;
 989                         len += PAGE_SIZE;
 990                 }
 991
 992                 /* did we get anything? */
 993                 if (!locked_pages)
 994                         goto release_pvec_pages;
 995                 if (i) {
 996                         int j;
 997                         BUG_ON(!locked_pages || first < 0);
 998
 999                         if (pvec_pages && i == pvec_pages &&
1000                             locked_pages < max_pages) {
1001                                 dout("reached end pvec, trying for more\n");
1002                                 pagevec_reinit(&pvec);
1003                                 goto get_more_pages;
1004                         }
1005
1006                         /* shift unused pages over in the pvec...  we
1007                          * will need to release them below. */
1008                         for (j = i; j < pvec_pages; j++) {
1009                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
1010                                 pvec.pages[j-i+first] = pvec.pages[j];
1011                         }
1012                         pvec.nr -= i-first;
1013                 }
1014
1015 new_request:
1016                 offset = page_offset(pages[0]);
1017                 len = wsize;
1018
1019                 req = ceph_osdc_new_request(&fsc->client->osdc,
1020                                         &ci->i_layout, vino,
1021                                         offset, &len, 0, num_ops,
1022                                         CEPH_OSD_OP_WRITE,
1023                                         CEPH_OSD_FLAG_WRITE,
1024                                         snapc, truncate_seq,
1025                                         truncate_size, false);
1026                 if (IS_ERR(req)) {
1027                         req = ceph_osdc_new_request(&fsc->client->osdc,
1028                                                 &ci->i_layout, vino,
1029                                                 offset, &len, 0,
1030                                                 min(num_ops,
1031                                                     CEPH_OSD_SLAB_OPS),
1032                                                 CEPH_OSD_OP_WRITE,
1033                                                 CEPH_OSD_FLAG_WRITE,
1034                                                 snapc, truncate_seq,
1035                                                 truncate_size, true);
1036                         BUG_ON(IS_ERR(req));
1037                 }
1038                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1039                              PAGE_SIZE - offset);
1040
1041                 req->r_callback = writepages_finish;
1042                 req->r_inode = inode;
1043
1044                 /* Format the osd request message and submit the write */
1045                 len = 0;
1046                 data_pages = pages;
1047                 op_idx = 0;
1048                 for (i = 0; i < locked_pages; i++) {
1049                         u64 cur_offset = page_offset(pages[i]);
1050                         if (offset + len != cur_offset) {
1051                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1052                                         break;
1053                                 osd_req_op_extent_dup_last(req, op_idx,
1054                                                            cur_offset - offset);
1055                                 dout("writepages got pages at %llu~%llu\n",
1056                                      offset, len);
1057                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1058                                                         data_pages, len, 0,
1059                                                         !!pool, false);
1060                                 osd_req_op_extent_update(req, op_idx, len);
1061
1062                                 len = 0;
1063                                 offset = cur_offset;
1064                                 data_pages = pages + i;
1065                                 op_idx++;
1066                         }
1067
1068                         set_page_writeback(pages[i]);
1069                         len += PAGE_SIZE;
1070                 }
1071
1072                 if (snap_size != -1) {
1073                         len = min(len, snap_size - offset);
1074                 } else if (i == locked_pages) {
1075                         /* writepages_finish() clears writeback pages
1076                          * according to the data length, so make sure
1077                          * data length covers all locked pages */
1078                         u64 min_len = len + 1 - PAGE_SIZE;
1079                         len = min(len, (u64)i_size_read(inode) - offset);
1080                         len = max(len, min_len);
1081                 }
1082                 dout("writepages got pages at %llu~%llu\n", offset, len);
1083
1084                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1085                                                  0, !!pool, false);
1086                 osd_req_op_extent_update(req, op_idx, len);
1087
1088                 if (do_sync) {
1089                         op_idx++;
1090                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1091                 }
1092                 BUG_ON(op_idx + 1 != req->r_num_ops);
1093
1094                 pool = NULL;
1095                 if (i < locked_pages) {
1096                         BUG_ON(num_ops <= req->r_num_ops);
1097                         num_ops -= req->r_num_ops;
1098                         num_ops += do_sync;
1099                         locked_pages -= i;
1100
1101                         /* allocate new pages array for next request */
1102                         data_pages = pages;
1103                         pages = kmalloc(locked_pages * sizeof (*pages),
1104                                         GFP_NOFS);
1105                         if (!pages) {
1106                                 pool = fsc->wb_pagevec_pool;
1107                                 pages = mempool_alloc(pool, GFP_NOFS);
1108                                 BUG_ON(!pages);
1109                         }
1110                         memcpy(pages, data_pages + i,
1111                                locked_pages * sizeof(*pages));
1112                         memset(data_pages + i, 0,
1113                                locked_pages * sizeof(*pages));
1114                 } else {
1115                         BUG_ON(num_ops != req->r_num_ops);
1116                         index = pages[i - 1]->index + 1;
1117                         /* request message now owns the pages array */
1118                         pages = NULL;
1119                 }
1120
1121                 req->r_mtime = inode->i_mtime;
1122                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1123                 BUG_ON(rc);
1124                 req = NULL;
1125
1126                 wbc->nr_to_write -= i;
1127                 if (pages)
1128                         goto new_request;
1129
1130                 if (wbc->nr_to_write <= 0)
1131                         done = 1;
1132
1133 release_pvec_pages:
1134                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1135                      pvec.nr ? pvec.pages[0] : NULL);
1136                 pagevec_release(&pvec);
1137
1138                 if (locked_pages && !done)
1139                         goto retry;
1140         }
1141
1142         if (should_loop && !done) {
1143                 /* more to do; loop back to beginning of file */
1144                 dout("writepages looping back to beginning of file\n");
1145                 should_loop = 0;
1146                 index = 0;
1147                 goto retry;
1148         }
1149
1150         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1151                 mapping->writeback_index = index;
1152
1153 out:
1154         ceph_osdc_put_request(req);
1155         ceph_put_snap_context(snapc);
1156         dout("writepages done, rc = %d\n", rc);
1157         return rc;
1158 }
1159
1160
1161
1162 /*
1163  * See if a given @snapc is either writeable, or already written.
1164  */
1165 static int context_is_writeable_or_written(struct inode *inode,
1166                                            struct ceph_snap_context *snapc)
1167 {
1168         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL,
1169                                                               NULL, NULL);
1170         int ret = !oldest || snapc->seq <= oldest->seq;
1171
1172         ceph_put_snap_context(oldest);
1173         return ret;
1174 }
1175
1176 /*
1177  * We are only allowed to write into/dirty the page if the page is
1178  * clean, or already dirty within the same snap context.
1179  *
1180  * called with page locked.
1181  * return success with page locked,
1182  * or any failure (incl -EAGAIN) with page unlocked.
1183  */
1184 static int ceph_update_writeable_page(struct file *file,
1185                             loff_t pos, unsigned len,
1186                             struct page *page)
1187 {
1188         struct inode *inode = file_inode(file);
1189         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1190         struct ceph_inode_info *ci = ceph_inode(inode);
1191         loff_t page_off = pos & PAGE_MASK;
1192         int pos_in_page = pos & ~PAGE_MASK;
1193         int end_in_page = pos_in_page + len;
1194         loff_t i_size;
1195         int r;
1196         struct ceph_snap_context *snapc, *oldest;
1197
1198         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1199                 dout(" page %p forced umount\n", page);
1200                 unlock_page(page);
1201                 return -EIO;
1202         }
1203
1204 retry_locked:
1205         /* writepages currently holds page lock, but if we change that later, */
1206         wait_on_page_writeback(page);
1207
1208         snapc = page_snap_context(page);
1209         if (snapc && snapc != ci->i_head_snapc) {
1210                 /*
1211                  * this page is already dirty in another (older) snap
1212                  * context!  is it writeable now?
1213                  */
1214                 oldest = get_oldest_context(inode, NULL, NULL, NULL);
1215
1216                 if (snapc->seq > oldest->seq) {
1217                         ceph_put_snap_context(oldest);
1218                         dout(" page %p snapc %p not current or oldest\n",
1219                              page, snapc);
1220                         /*
1221                          * queue for writeback, and wait for snapc to
1222                          * be writeable or written
1223                          */
1224                         snapc = ceph_get_snap_context(snapc);
1225                         unlock_page(page);
1226                         ceph_queue_writeback(inode);
1227                         r = wait_event_killable(ci->i_cap_wq,
1228                                context_is_writeable_or_written(inode, snapc));
1229                         ceph_put_snap_context(snapc);
1230                         if (r == -ERESTARTSYS)
1231                                 return r;
1232                         return -EAGAIN;
1233                 }
1234                 ceph_put_snap_context(oldest);
1235
1236                 /* yay, writeable, do it now (without dropping page lock) */
1237                 dout(" page %p snapc %p not current, but oldest\n",
1238                      page, snapc);
1239                 if (!clear_page_dirty_for_io(page))
1240                         goto retry_locked;
1241                 r = writepage_nounlock(page, NULL);
1242                 if (r < 0)
1243                         goto fail_unlock;
1244                 goto retry_locked;
1245         }
1246
1247         if (PageUptodate(page)) {
1248                 dout(" page %p already uptodate\n", page);
1249                 return 0;
1250         }
1251
1252         /* full page? */
1253         if (pos_in_page == 0 && len == PAGE_SIZE)
1254                 return 0;
1255
1256         /* past end of file? */
1257         i_size = i_size_read(inode);
1258
1259         if (page_off >= i_size ||
1260             (pos_in_page == 0 && (pos+len) >= i_size &&
1261              end_in_page - pos_in_page != PAGE_SIZE)) {
1262                 dout(" zeroing %p 0 - %d and %d - %d\n",
1263                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1264                 zero_user_segments(page,
1265                                    0, pos_in_page,
1266                                    end_in_page, PAGE_SIZE);
1267                 return 0;
1268         }
1269
1270         /* we need to read it. */
1271         r = ceph_do_readpage(file, page);
1272         if (r < 0) {
1273                 if (r == -EINPROGRESS)
1274                         return -EAGAIN;
1275                 goto fail_unlock;
1276         }
1277         goto retry_locked;
1278 fail_unlock:
1279         unlock_page(page);
1280         return r;
1281 }
1282
1283 /*
1284  * We are only allowed to write into/dirty the page if the page is
1285  * clean, or already dirty within the same snap context.
1286  */
1287 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1288                             loff_t pos, unsigned len, unsigned flags,
1289                             struct page **pagep, void **fsdata)
1290 {
1291         struct inode *inode = file_inode(file);
1292         struct page *page;
1293         pgoff_t index = pos >> PAGE_SHIFT;
1294         int r;
1295
1296         do {
1297                 /* get a page */
1298                 page = grab_cache_page_write_begin(mapping, index, 0);
1299                 if (!page)
1300                         return -ENOMEM;
1301
1302                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1303                      inode, page, (int)pos, (int)len);
1304
1305                 r = ceph_update_writeable_page(file, pos, len, page);
1306                 if (r < 0)
1307                         put_page(page);
1308                 else
1309                         *pagep = page;
1310         } while (r == -EAGAIN);
1311
1312         return r;
1313 }
1314
1315 /*
1316  * we don't do anything in here that simple_write_end doesn't do
1317  * except adjust dirty page accounting
1318  */
1319 static int ceph_write_end(struct file *file, struct address_space *mapping,
1320                           loff_t pos, unsigned len, unsigned copied,
1321                           struct page *page, void *fsdata)
1322 {
1323         struct inode *inode = file_inode(file);
1324         bool check_cap = false;
1325
1326         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1327              inode, page, (int)pos, (int)copied, (int)len);
1328
1329         /* zero the stale part of the page if we did a short copy */
1330         if (!PageUptodate(page)) {
1331                 if (copied < len) {
1332                         copied = 0;
1333                         goto out;
1334                 }
1335                 SetPageUptodate(page);
1336         }
1337
1338         /* did file size increase? */
1339         if (pos+copied > i_size_read(inode))
1340                 check_cap = ceph_inode_set_size(inode, pos+copied);
1341
1342         set_page_dirty(page);
1343
1344 out:
1345         unlock_page(page);
1346         put_page(page);
1347
1348         if (check_cap)
1349                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1350
1351         return copied;
1352 }
1353
1354 /*
1355  * we set .direct_IO to indicate direct io is supported, but since we
1356  * intercept O_DIRECT reads and writes early, this function should
1357  * never get called.
1358  */
1359 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1360 {
1361         WARN_ON(1);
1362         return -EINVAL;
1363 }
1364
1365 const struct address_space_operations ceph_aops = {
1366         .readpage = ceph_readpage,
1367         .readpages = ceph_readpages,
1368         .writepage = ceph_writepage,
1369         .writepages = ceph_writepages_start,
1370         .write_begin = ceph_write_begin,
1371         .write_end = ceph_write_end,
1372         .set_page_dirty = ceph_set_page_dirty,
1373         .invalidatepage = ceph_invalidatepage,
1374         .releasepage = ceph_releasepage,
1375         .direct_IO = ceph_direct_io,
1376 };
1377
1378 static void ceph_block_sigs(sigset_t *oldset)
1379 {
1380         sigset_t mask;
1381         siginitsetinv(&mask, sigmask(SIGKILL));
1382         sigprocmask(SIG_BLOCK, &mask, oldset);
1383 }
1384
1385 static void ceph_restore_sigs(sigset_t *oldset)
1386 {
1387         sigprocmask(SIG_SETMASK, oldset, NULL);
1388 }
1389
1390 /*
1391  * vm ops
1392  */
1393 static int ceph_filemap_fault(struct vm_fault *vmf)
1394 {
1395         struct vm_area_struct *vma = vmf->vma;
1396         struct inode *inode = file_inode(vma->vm_file);
1397         struct ceph_inode_info *ci = ceph_inode(inode);
1398         struct ceph_file_info *fi = vma->vm_file->private_data;
1399         struct page *pinned_page = NULL;
1400         loff_t off = vmf->pgoff << PAGE_SHIFT;
1401         int want, got, ret;
1402         sigset_t oldset;
1403
1404         ceph_block_sigs(&oldset);
1405
1406         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1407              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1408         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1409                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1410         else
1411                 want = CEPH_CAP_FILE_CACHE;
1412
1413         got = 0;
1414         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1415         if (ret < 0)
1416                 goto out_restore;
1417
1418         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1419              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1420
1421         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1422             ci->i_inline_version == CEPH_INLINE_NONE) {
1423                 current->journal_info = vma->vm_file;
1424                 ret = filemap_fault(vmf);
1425                 current->journal_info = NULL;
1426         } else
1427                 ret = -EAGAIN;
1428
1429         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1430              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1431         if (pinned_page)
1432                 put_page(pinned_page);
1433         ceph_put_cap_refs(ci, got);
1434
1435         if (ret != -EAGAIN)
1436                 goto out_restore;
1437
1438         /* read inline data */
1439         if (off >= PAGE_SIZE) {
1440                 /* does not support inline data > PAGE_SIZE */
1441                 ret = VM_FAULT_SIGBUS;
1442         } else {
1443                 int ret1;
1444                 struct address_space *mapping = inode->i_mapping;
1445                 struct page *page = find_or_create_page(mapping, 0,
1446                                                 mapping_gfp_constraint(mapping,
1447                                                 ~__GFP_FS));
1448                 if (!page) {
1449                         ret = VM_FAULT_OOM;
1450                         goto out_inline;
1451                 }
1452                 ret1 = __ceph_do_getattr(inode, page,
1453                                          CEPH_STAT_CAP_INLINE_DATA, true);
1454                 if (ret1 < 0 || off >= i_size_read(inode)) {
1455                         unlock_page(page);
1456                         put_page(page);
1457                         if (ret1 < 0)
1458                                 ret = ret1;
1459                         else
1460                                 ret = VM_FAULT_SIGBUS;
1461                         goto out_inline;
1462                 }
1463                 if (ret1 < PAGE_SIZE)
1464                         zero_user_segment(page, ret1, PAGE_SIZE);
1465                 else
1466                         flush_dcache_page(page);
1467                 SetPageUptodate(page);
1468                 vmf->page = page;
1469                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1470 out_inline:
1471                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1472                      inode, off, (size_t)PAGE_SIZE, ret);
1473         }
1474 out_restore:
1475         ceph_restore_sigs(&oldset);
1476         if (ret < 0)
1477                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1478
1479         return ret;
1480 }
1481
1482 /*
1483  * Reuse write_begin here for simplicity.
1484  */
1485 static int ceph_page_mkwrite(struct vm_fault *vmf)
1486 {
1487         struct vm_area_struct *vma = vmf->vma;
1488         struct inode *inode = file_inode(vma->vm_file);
1489         struct ceph_inode_info *ci = ceph_inode(inode);
1490         struct ceph_file_info *fi = vma->vm_file->private_data;
1491         struct ceph_cap_flush *prealloc_cf;
1492         struct page *page = vmf->page;
1493         loff_t off = page_offset(page);
1494         loff_t size = i_size_read(inode);
1495         size_t len;
1496         int want, got, ret;
1497         sigset_t oldset;
1498
1499         prealloc_cf = ceph_alloc_cap_flush();
1500         if (!prealloc_cf)
1501                 return VM_FAULT_OOM;
1502
1503         ceph_block_sigs(&oldset);
1504
1505         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1506                 struct page *locked_page = NULL;
1507                 if (off == 0) {
1508                         lock_page(page);
1509                         locked_page = page;
1510                 }
1511                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1512                 if (locked_page)
1513                         unlock_page(locked_page);
1514                 if (ret < 0)
1515                         goto out_free;
1516         }
1517
1518         if (off + PAGE_SIZE <= size)
1519                 len = PAGE_SIZE;
1520         else
1521                 len = size & ~PAGE_MASK;
1522
1523         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1524              inode, ceph_vinop(inode), off, len, size);
1525         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1526                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1527         else
1528                 want = CEPH_CAP_FILE_BUFFER;
1529
1530         got = 0;
1531         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1532                             &got, NULL);
1533         if (ret < 0)
1534                 goto out_free;
1535
1536         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1537              inode, off, len, ceph_cap_string(got));
1538
1539         /* Update time before taking page lock */
1540         file_update_time(vma->vm_file);
1541
1542         do {
1543                 lock_page(page);
1544
1545                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1546                         unlock_page(page);
1547                         ret = VM_FAULT_NOPAGE;
1548                         break;
1549                 }
1550
1551                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1552                 if (ret >= 0) {
1553                         /* success.  we'll keep the page locked. */
1554                         set_page_dirty(page);
1555                         ret = VM_FAULT_LOCKED;
1556                 }
1557         } while (ret == -EAGAIN);
1558
1559         if (ret == VM_FAULT_LOCKED ||
1560             ci->i_inline_version != CEPH_INLINE_NONE) {
1561                 int dirty;
1562                 spin_lock(&ci->i_ceph_lock);
1563                 ci->i_inline_version = CEPH_INLINE_NONE;
1564                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1565                                                &prealloc_cf);
1566                 spin_unlock(&ci->i_ceph_lock);
1567                 if (dirty)
1568                         __mark_inode_dirty(inode, dirty);
1569         }
1570
1571         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1572              inode, off, len, ceph_cap_string(got), ret);
1573         ceph_put_cap_refs(ci, got);
1574 out_free:
1575         ceph_restore_sigs(&oldset);
1576         ceph_free_cap_flush(prealloc_cf);
1577         if (ret < 0)
1578                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1579         return ret;
1580 }
1581
1582 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1583                            char *data, size_t len)
1584 {
1585         struct address_space *mapping = inode->i_mapping;
1586         struct page *page;
1587
1588         if (locked_page) {
1589                 page = locked_page;
1590         } else {
1591                 if (i_size_read(inode) == 0)
1592                         return;
1593                 page = find_or_create_page(mapping, 0,
1594                                            mapping_gfp_constraint(mapping,
1595                                            ~__GFP_FS));
1596                 if (!page)
1597                         return;
1598                 if (PageUptodate(page)) {
1599                         unlock_page(page);
1600                         put_page(page);
1601                         return;
1602                 }
1603         }
1604
1605         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1606              inode, ceph_vinop(inode), len, locked_page);
1607
1608         if (len > 0) {
1609                 void *kaddr = kmap_atomic(page);
1610                 memcpy(kaddr, data, len);
1611                 kunmap_atomic(kaddr);
1612         }
1613
1614         if (page != locked_page) {
1615                 if (len < PAGE_SIZE)
1616                         zero_user_segment(page, len, PAGE_SIZE);
1617                 else
1618                         flush_dcache_page(page);
1619
1620                 SetPageUptodate(page);
1621                 unlock_page(page);
1622                 put_page(page);
1623         }
1624 }
1625
1626 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1627 {
1628         struct inode *inode = file_inode(filp);
1629         struct ceph_inode_info *ci = ceph_inode(inode);
1630         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1631         struct ceph_osd_request *req;
1632         struct page *page = NULL;
1633         u64 len, inline_version;
1634         int err = 0;
1635         bool from_pagecache = false;
1636
1637         spin_lock(&ci->i_ceph_lock);
1638         inline_version = ci->i_inline_version;
1639         spin_unlock(&ci->i_ceph_lock);
1640
1641         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1642              inode, ceph_vinop(inode), inline_version);
1643
1644         if (inline_version == 1 || /* initial version, no data */
1645             inline_version == CEPH_INLINE_NONE)
1646                 goto out;
1647
1648         if (locked_page) {
1649                 page = locked_page;
1650                 WARN_ON(!PageUptodate(page));
1651         } else if (ceph_caps_issued(ci) &
1652                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1653                 page = find_get_page(inode->i_mapping, 0);
1654                 if (page) {
1655                         if (PageUptodate(page)) {
1656                                 from_pagecache = true;
1657                                 lock_page(page);
1658                         } else {
1659                                 put_page(page);
1660                                 page = NULL;
1661                         }
1662                 }
1663         }
1664
1665         if (page) {
1666                 len = i_size_read(inode);
1667                 if (len > PAGE_SIZE)
1668                         len = PAGE_SIZE;
1669         } else {
1670                 page = __page_cache_alloc(GFP_NOFS);
1671                 if (!page) {
1672                         err = -ENOMEM;
1673                         goto out;
1674                 }
1675                 err = __ceph_do_getattr(inode, page,
1676                                         CEPH_STAT_CAP_INLINE_DATA, true);
1677                 if (err < 0) {
1678                         /* no inline data */
1679                         if (err == -ENODATA)
1680                                 err = 0;
1681                         goto out;
1682                 }
1683                 len = err;
1684         }
1685
1686         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1687                                     ceph_vino(inode), 0, &len, 0, 1,
1688                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1689                                     NULL, 0, 0, false);
1690         if (IS_ERR(req)) {
1691                 err = PTR_ERR(req);
1692                 goto out;
1693         }
1694
1695         req->r_mtime = inode->i_mtime;
1696         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1697         if (!err)
1698                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1699         ceph_osdc_put_request(req);
1700         if (err < 0)
1701                 goto out;
1702
1703         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1704                                     ceph_vino(inode), 0, &len, 1, 3,
1705                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1706                                     NULL, ci->i_truncate_seq,
1707                                     ci->i_truncate_size, false);
1708         if (IS_ERR(req)) {
1709                 err = PTR_ERR(req);
1710                 goto out;
1711         }
1712
1713         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1714
1715         {
1716                 __le64 xattr_buf = cpu_to_le64(inline_version);
1717                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1718                                             "inline_version", &xattr_buf,
1719                                             sizeof(xattr_buf),
1720                                             CEPH_OSD_CMPXATTR_OP_GT,
1721                                             CEPH_OSD_CMPXATTR_MODE_U64);
1722                 if (err)
1723                         goto out_put;
1724         }
1725
1726         {
1727                 char xattr_buf[32];
1728                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1729                                          "%llu", inline_version);
1730                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1731                                             "inline_version",
1732                                             xattr_buf, xattr_len, 0, 0);
1733                 if (err)
1734                         goto out_put;
1735         }
1736
1737         req->r_mtime = inode->i_mtime;
1738         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1739         if (!err)
1740                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1741 out_put:
1742         ceph_osdc_put_request(req);
1743         if (err == -ECANCELED)
1744                 err = 0;
1745 out:
1746         if (page && page != locked_page) {
1747                 if (from_pagecache) {
1748                         unlock_page(page);
1749                         put_page(page);
1750                 } else
1751                         __free_pages(page, 0);
1752         }
1753
1754         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1755              inode, ceph_vinop(inode), inline_version, err);
1756         return err;
1757 }
1758
1759 static const struct vm_operations_struct ceph_vmops = {
1760         .fault          = ceph_filemap_fault,
1761         .page_mkwrite   = ceph_page_mkwrite,
1762 };
1763
1764 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1765 {
1766         struct address_space *mapping = file->f_mapping;
1767
1768         if (!mapping->a_ops->readpage)
1769                 return -ENOEXEC;
1770         file_accessed(file);
1771         vma->vm_ops = &ceph_vmops;
1772         return 0;
1773 }
1774
1775 enum {
1776         POOL_READ       = 1,
1777         POOL_WRITE      = 2,
1778 };
1779
1780 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1781                                 s64 pool, struct ceph_string *pool_ns)
1782 {
1783         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1784         struct ceph_mds_client *mdsc = fsc->mdsc;
1785         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1786         struct rb_node **p, *parent;
1787         struct ceph_pool_perm *perm;
1788         struct page **pages;
1789         size_t pool_ns_len;
1790         int err = 0, err2 = 0, have = 0;
1791
1792         down_read(&mdsc->pool_perm_rwsem);
1793         p = &mdsc->pool_perm_tree.rb_node;
1794         while (*p) {
1795                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1796                 if (pool < perm->pool)
1797                         p = &(*p)->rb_left;
1798                 else if (pool > perm->pool)
1799                         p = &(*p)->rb_right;
1800                 else {
1801                         int ret = ceph_compare_string(pool_ns,
1802                                                 perm->pool_ns,
1803                                                 perm->pool_ns_len);
1804                         if (ret < 0)
1805                                 p = &(*p)->rb_left;
1806                         else if (ret > 0)
1807                                 p = &(*p)->rb_right;
1808                         else {
1809                                 have = perm->perm;
1810                                 break;
1811                         }
1812                 }
1813         }
1814         up_read(&mdsc->pool_perm_rwsem);
1815         if (*p)
1816                 goto out;
1817
1818         if (pool_ns)
1819                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1820                      pool, (int)pool_ns->len, pool_ns->str);
1821         else
1822                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1823
1824         down_write(&mdsc->pool_perm_rwsem);
1825         p = &mdsc->pool_perm_tree.rb_node;
1826         parent = NULL;
1827         while (*p) {
1828                 parent = *p;
1829                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1830                 if (pool < perm->pool)
1831                         p = &(*p)->rb_left;
1832                 else if (pool > perm->pool)
1833                         p = &(*p)->rb_right;
1834                 else {
1835                         int ret = ceph_compare_string(pool_ns,
1836                                                 perm->pool_ns,
1837                                                 perm->pool_ns_len);
1838                         if (ret < 0)
1839                                 p = &(*p)->rb_left;
1840                         else if (ret > 0)
1841                                 p = &(*p)->rb_right;
1842                         else {
1843                                 have = perm->perm;
1844                                 break;
1845                         }
1846                 }
1847         }
1848         if (*p) {
1849                 up_write(&mdsc->pool_perm_rwsem);
1850                 goto out;
1851         }
1852
1853         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1854                                          1, false, GFP_NOFS);
1855         if (!rd_req) {
1856                 err = -ENOMEM;
1857                 goto out_unlock;
1858         }
1859
1860         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1861         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1862         rd_req->r_base_oloc.pool = pool;
1863         if (pool_ns)
1864                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1865         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1866
1867         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1868         if (err)
1869                 goto out_unlock;
1870
1871         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1872                                          1, false, GFP_NOFS);
1873         if (!wr_req) {
1874                 err = -ENOMEM;
1875                 goto out_unlock;
1876         }
1877
1878         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1879         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1880         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1881         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1882
1883         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1884         if (err)
1885                 goto out_unlock;
1886
1887         /* one page should be large enough for STAT data */
1888         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1889         if (IS_ERR(pages)) {
1890                 err = PTR_ERR(pages);
1891                 goto out_unlock;
1892         }
1893
1894         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1895                                      0, false, true);
1896         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1897
1898         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1899         wr_req->r_abort_on_full = true;
1900         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1901
1902         if (!err)
1903                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1904         if (!err2)
1905                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1906
1907         if (err >= 0 || err == -ENOENT)
1908                 have |= POOL_READ;
1909         else if (err != -EPERM)
1910                 goto out_unlock;
1911
1912         if (err2 == 0 || err2 == -EEXIST)
1913                 have |= POOL_WRITE;
1914         else if (err2 != -EPERM) {
1915                 err = err2;
1916                 goto out_unlock;
1917         }
1918
1919         pool_ns_len = pool_ns ? pool_ns->len : 0;
1920         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1921         if (!perm) {
1922                 err = -ENOMEM;
1923                 goto out_unlock;
1924         }
1925
1926         perm->pool = pool;
1927         perm->perm = have;
1928         perm->pool_ns_len = pool_ns_len;
1929         if (pool_ns_len > 0)
1930                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1931         perm->pool_ns[pool_ns_len] = 0;
1932
1933         rb_link_node(&perm->node, parent, p);
1934         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1935         err = 0;
1936 out_unlock:
1937         up_write(&mdsc->pool_perm_rwsem);
1938
1939         ceph_osdc_put_request(rd_req);
1940         ceph_osdc_put_request(wr_req);
1941 out:
1942         if (!err)
1943                 err = have;
1944         if (pool_ns)
1945                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1946                      pool, (int)pool_ns->len, pool_ns->str, err);
1947         else
1948                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1949         return err;
1950 }
1951
1952 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1953 {
1954         s64 pool;
1955         struct ceph_string *pool_ns;
1956         int ret, flags;
1957
1958         if (ci->i_vino.snap != CEPH_NOSNAP) {
1959                 /*
1960                  * Pool permission check needs to write to the first object.
1961                  * But for snapshot, head of the first object may have alread
1962                  * been deleted. Skip check to avoid creating orphan object.
1963                  */
1964                 return 0;
1965         }
1966
1967         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1968                                 NOPOOLPERM))
1969                 return 0;
1970
1971         spin_lock(&ci->i_ceph_lock);
1972         flags = ci->i_ceph_flags;
1973         pool = ci->i_layout.pool_id;
1974         spin_unlock(&ci->i_ceph_lock);
1975 check:
1976         if (flags & CEPH_I_POOL_PERM) {
1977                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1978                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1979                              pool);
1980                         return -EPERM;
1981                 }
1982                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1983                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1984                              pool);
1985                         return -EPERM;
1986                 }
1987                 return 0;
1988         }
1989
1990         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1991         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1992         ceph_put_string(pool_ns);
1993         if (ret < 0)
1994                 return ret;
1995
1996         flags = CEPH_I_POOL_PERM;
1997         if (ret & POOL_READ)
1998                 flags |= CEPH_I_POOL_RD;
1999         if (ret & POOL_WRITE)
2000                 flags |= CEPH_I_POOL_WR;
2001
2002         spin_lock(&ci->i_ceph_lock);
2003         if (pool == ci->i_layout.pool_id &&
2004             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2005                 ci->i_ceph_flags |= flags;
2006         } else {
2007                 pool = ci->i_layout.pool_id;
2008                 flags = ci->i_ceph_flags;
2009         }
2010         spin_unlock(&ci->i_ceph_lock);
2011         goto check;
2012 }
2013
2014 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2015 {
2016         struct ceph_pool_perm *perm;
2017         struct rb_node *n;
2018
2019         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2020                 n = rb_first(&mdsc->pool_perm_tree);
2021                 perm = rb_entry(n, struct ceph_pool_perm, node);
2022                 rb_erase(n, &mdsc->pool_perm_tree);
2023                 kfree(perm);
2024         }
2025 }