fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11 #include <linux/signal.h>
  12
  13 #include "super.h"
  14 #include "mds_client.h"
  15 #include "cache.h"
  16 #include <linux/ceph/osd_client.h>
  17
  18 /*
  19  * Ceph address space ops.
  20  *
  21  * There are a few funny things going on here.
  22  *
  23  * The page->private field is used to reference a struct
  24  * ceph_snap_context for _every_ dirty page.  This indicates which
  25  * snapshot the page was logically dirtied in, and thus which snap
  26  * context needs to be associated with the osd write during writeback.
  27  *
  28  * Similarly, struct ceph_inode_info maintains a set of counters to
  29  * count dirty pages on the inode.  In the absence of snapshots,
  30  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  31  *
  32  * When a snapshot is taken (that is, when the client receives
  33  * notification that a snapshot was taken), each inode with caps and
  34  * with dirty pages (dirty pages implies there is a cap) gets a new
  35  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  36  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  37  * moved to capsnap->dirty. (Unless a sync write is currently in
  38  * progress.  In that case, the capsnap is said to be "pending", new
  39  * writes cannot start, and the capsnap isn't "finalized" until the
  40  * write completes (or fails) and a final size/mtime for the inode for
  41  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  42  *
  43  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  44  * we look for the first capsnap in i_cap_snaps and write out pages in
  45  * that snap context _only_.  Then we move on to the next capsnap,
  46  * eventually reaching the "live" or "head" context (i.e., pages that
  47  * are not yet snapped) and are writing the most recently dirtied
  48  * pages.
  49  *
  50  * Invalidate and so forth must take care to ensure the dirty page
  51  * accounting is preserved.
  52  */
  53
  54 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  55 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  56         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  57          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  58
  59 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  60 {
  61         if (PagePrivate(page))
  62                 return (void *)page->private;
  63         return NULL;
  64 }
  65
  66 /*
  67  * Dirty a page.  Optimistically adjust accounting, on the assumption
  68  * that we won't race with invalidate.  If we do, readjust.
  69  */
  70 static int ceph_set_page_dirty(struct page *page)
  71 {
  72         struct address_space *mapping = page->mapping;
  73         struct inode *inode;
  74         struct ceph_inode_info *ci;
  75         struct ceph_snap_context *snapc;
  76         int ret;
  77
  78         if (unlikely(!mapping))
  79                 return !TestSetPageDirty(page);
  80
  81         if (PageDirty(page)) {
  82                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  83                      mapping->host, page, page->index);
  84                 BUG_ON(!PagePrivate(page));
  85                 return 0;
  86         }
  87
  88         inode = mapping->host;
  89         ci = ceph_inode(inode);
  90
  91         /* dirty the head */
  92         spin_lock(&ci->i_ceph_lock);
  93         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  94         if (__ceph_have_pending_cap_snap(ci)) {
  95                 struct ceph_cap_snap *capsnap =
  96                                 list_last_entry(&ci->i_cap_snaps,
  97                                                 struct ceph_cap_snap,
  98                                                 ci_item);
  99                 snapc = ceph_get_snap_context(capsnap->context);
 100                 capsnap->dirty_pages++;
 101         } else {
 102                 BUG_ON(!ci->i_head_snapc);
 103                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 104                 ++ci->i_wrbuffer_ref_head;
 105         }
 106         if (ci->i_wrbuffer_ref == 0)
 107                 ihold(inode);
 108         ++ci->i_wrbuffer_ref;
 109         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 110              "snapc %p seq %lld (%d snaps)\n",
 111              mapping->host, page, page->index,
 112              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 113              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 114              snapc, snapc->seq, snapc->num_snaps);
 115         spin_unlock(&ci->i_ceph_lock);
 116
 117         /*
 118          * Reference snap context in page->private.  Also set
 119          * PagePrivate so that we get invalidatepage callback.
 120          */
 121         BUG_ON(PagePrivate(page));
 122         page->private = (unsigned long)snapc;
 123         SetPagePrivate(page);
 124
 125         ret = __set_page_dirty_nobuffers(page);
 126         WARN_ON(!PageLocked(page));
 127         WARN_ON(!page->mapping);
 128
 129         return ret;
 130 }
 131
 132 /*
 133  * If we are truncating the full page (i.e. offset == 0), adjust the
 134  * dirty page counters appropriately.  Only called if there is private
 135  * data on the page.
 136  */
 137 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 138                                 unsigned int length)
 139 {
 140         struct inode *inode;
 141         struct ceph_inode_info *ci;
 142         struct ceph_snap_context *snapc = page_snap_context(page);
 143
 144         inode = page->mapping->host;
 145         ci = ceph_inode(inode);
 146
 147         if (offset != 0 || length != PAGE_SIZE) {
 148                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 149                      inode, page, page->index, offset, length);
 150                 return;
 151         }
 152
 153         ceph_invalidate_fscache_page(inode, page);
 154
 155         WARN_ON(!PageLocked(page));
 156         if (!PagePrivate(page))
 157                 return;
 158
 159         ClearPageChecked(page);
 160
 161         dout("%p invalidatepage %p idx %lu full dirty page\n",
 162              inode, page, page->index);
 163
 164         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 165         ceph_put_snap_context(snapc);
 166         page->private = 0;
 167         ClearPagePrivate(page);
 168 }
 169
 170 static int ceph_releasepage(struct page *page, gfp_t g)
 171 {
 172         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 173              page, page->index, PageDirty(page) ? "" : "not ");
 174
 175         /* Can we release the page from the cache? */
 176         if (!ceph_release_fscache_page(page, g))
 177                 return 0;
 178
 179         return !PagePrivate(page);
 180 }
 181
 182 /*
 183  * read a single page, without unlocking it.
 184  */
 185 static int ceph_do_readpage(struct file *filp, struct page *page)
 186 {
 187         struct inode *inode = file_inode(filp);
 188         struct ceph_inode_info *ci = ceph_inode(inode);
 189         struct ceph_osd_client *osdc =
 190                 &ceph_inode_to_client(inode)->client->osdc;
 191         int err = 0;
 192         u64 off = page_offset(page);
 193         u64 len = PAGE_SIZE;
 194
 195         if (off >= i_size_read(inode)) {
 196                 zero_user_segment(page, 0, PAGE_SIZE);
 197                 SetPageUptodate(page);
 198                 return 0;
 199         }
 200
 201         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 202                 /*
 203                  * Uptodate inline data should have been added
 204                  * into page cache while getting Fcr caps.
 205                  */
 206                 if (off == 0)
 207                         return -EINVAL;
 208                 zero_user_segment(page, 0, PAGE_SIZE);
 209                 SetPageUptodate(page);
 210                 return 0;
 211         }
 212
 213         err = ceph_readpage_from_fscache(inode, page);
 214         if (err == 0)
 215                 return -EINPROGRESS;
 216
 217         dout("readpage inode %p file %p page %p index %lu\n",
 218              inode, filp, page, page->index);
 219         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 220                                   off, &len,
 221                                   ci->i_truncate_seq, ci->i_truncate_size,
 222                                   &page, 1, 0);
 223         if (err == -ENOENT)
 224                 err = 0;
 225         if (err < 0) {
 226                 SetPageError(page);
 227                 ceph_fscache_readpage_cancel(inode, page);
 228                 goto out;
 229         }
 230         if (err < PAGE_SIZE)
 231                 /* zero fill remainder of page */
 232                 zero_user_segment(page, err, PAGE_SIZE);
 233         else
 234                 flush_dcache_page(page);
 235
 236         SetPageUptodate(page);
 237         ceph_readpage_to_fscache(inode, page);
 238
 239 out:
 240         return err < 0 ? err : 0;
 241 }
 242
 243 static int ceph_readpage(struct file *filp, struct page *page)
 244 {
 245         int r = ceph_do_readpage(filp, page);
 246         if (r != -EINPROGRESS)
 247                 unlock_page(page);
 248         else
 249                 r = 0;
 250         return r;
 251 }
 252
 253 /*
 254  * Finish an async read(ahead) op.
 255  */
 256 static void finish_read(struct ceph_osd_request *req)
 257 {
 258         struct inode *inode = req->r_inode;
 259         struct ceph_osd_data *osd_data;
 260         int rc = req->r_result <= 0 ? req->r_result : 0;
 261         int bytes = req->r_result >= 0 ? req->r_result : 0;
 262         int num_pages;
 263         int i;
 264
 265         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 266
 267         /* unlock all pages, zeroing any data we didn't read */
 268         osd_data = osd_req_op_extent_osd_data(req, 0);
 269         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 270         num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                         (u64)osd_data->length);
 272         for (i = 0; i < num_pages; i++) {
 273                 struct page *page = osd_data->pages[i];
 274
 275                 if (rc < 0 && rc != -ENOENT) {
 276                         ceph_fscache_readpage_cancel(inode, page);
 277                         goto unlock;
 278                 }
 279                 if (bytes < (int)PAGE_SIZE) {
 280                         /* zero (remainder of) page */
 281                         int s = bytes < 0 ? 0 : bytes;
 282                         zero_user_segment(page, s, PAGE_SIZE);
 283                 }
 284                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 285                      page->index);
 286                 flush_dcache_page(page);
 287                 SetPageUptodate(page);
 288                 ceph_readpage_to_fscache(inode, page);
 289 unlock:
 290                 unlock_page(page);
 291                 put_page(page);
 292                 bytes -= PAGE_SIZE;
 293         }
 294         kfree(osd_data->pages);
 295 }
 296
 297 /*
 298  * start an async read(ahead) operation.  return nr_pages we submitted
 299  * a read for on success, or negative error code.
 300  */
 301 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 302 {
 303         struct ceph_osd_client *osdc =
 304                 &ceph_inode_to_client(inode)->client->osdc;
 305         struct ceph_inode_info *ci = ceph_inode(inode);
 306         struct page *page = list_entry(page_list->prev, struct page, lru);
 307         struct ceph_vino vino;
 308         struct ceph_osd_request *req;
 309         u64 off;
 310         u64 len;
 311         int i;
 312         struct page **pages;
 313         pgoff_t next_index;
 314         int nr_pages = 0;
 315         int got = 0;
 316         int ret = 0;
 317
 318         if (!current->journal_info) {
 319                 /* caller of readpages does not hold buffer and read caps
 320                  * (fadvise, madvise and readahead cases) */
 321                 int want = CEPH_CAP_FILE_CACHE;
 322                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 323                 if (ret < 0) {
 324                         dout("start_read %p, error getting cap\n", inode);
 325                 } else if (!(got & want)) {
 326                         dout("start_read %p, no cache cap\n", inode);
 327                         ret = 0;
 328                 }
 329                 if (ret <= 0) {
 330                         if (got)
 331                                 ceph_put_cap_refs(ci, got);
 332                         while (!list_empty(page_list)) {
 333                                 page = list_entry(page_list->prev,
 334                                                   struct page, lru);
 335                                 list_del(&page->lru);
 336                                 put_page(page);
 337                         }
 338                         return ret;
 339                 }
 340         }
 341
 342         off = (u64) page_offset(page);
 343
 344         /* count pages */
 345         next_index = page->index;
 346         list_for_each_entry_reverse(page, page_list, lru) {
 347                 if (page->index != next_index)
 348                         break;
 349                 nr_pages++;
 350                 next_index++;
 351                 if (max && nr_pages == max)
 352                         break;
 353         }
 354         len = nr_pages << PAGE_SHIFT;
 355         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 356              off, len);
 357         vino = ceph_vino(inode);
 358         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 359                                     0, 1, CEPH_OSD_OP_READ,
 360                                     CEPH_OSD_FLAG_READ, NULL,
 361                                     ci->i_truncate_seq, ci->i_truncate_size,
 362                                     false);
 363         if (IS_ERR(req)) {
 364                 ret = PTR_ERR(req);
 365                 goto out;
 366         }
 367
 368         /* build page vector */
 369         nr_pages = calc_pages_for(0, len);
 370         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 371         if (!pages) {
 372                 ret = -ENOMEM;
 373                 goto out_put;
 374         }
 375         for (i = 0; i < nr_pages; ++i) {
 376                 page = list_entry(page_list->prev, struct page, lru);
 377                 BUG_ON(PageLocked(page));
 378                 list_del(&page->lru);
 379
 380                 dout("start_read %p adding %p idx %lu\n", inode, page,
 381                      page->index);
 382                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 383                                           GFP_KERNEL)) {
 384                         ceph_fscache_uncache_page(inode, page);
 385                         put_page(page);
 386                         dout("start_read %p add_to_page_cache failed %p\n",
 387                              inode, page);
 388                         nr_pages = i;
 389                         if (nr_pages > 0) {
 390                                 len = nr_pages << PAGE_SHIFT;
 391                                 osd_req_op_extent_update(req, 0, len);
 392                                 break;
 393                         }
 394                         goto out_pages;
 395                 }
 396                 pages[i] = page;
 397         }
 398         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 399         req->r_callback = finish_read;
 400         req->r_inode = inode;
 401
 402         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 403         ret = ceph_osdc_start_request(osdc, req, false);
 404         if (ret < 0)
 405                 goto out_pages;
 406         ceph_osdc_put_request(req);
 407
 408         /* After adding locked pages to page cache, the inode holds cache cap.
 409          * So we can drop our cap refs. */
 410         if (got)
 411                 ceph_put_cap_refs(ci, got);
 412
 413         return nr_pages;
 414
 415 out_pages:
 416         for (i = 0; i < nr_pages; ++i) {
 417                 ceph_fscache_readpage_cancel(inode, pages[i]);
 418                 unlock_page(pages[i]);
 419         }
 420         ceph_put_page_vector(pages, nr_pages, false);
 421 out_put:
 422         ceph_osdc_put_request(req);
 423 out:
 424         if (got)
 425                 ceph_put_cap_refs(ci, got);
 426         return ret;
 427 }
 428
 429
 430 /*
 431  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 432  * the caller (VM) cleans them up.
 433  */
 434 static int ceph_readpages(struct file *file, struct address_space *mapping,
 435                           struct list_head *page_list, unsigned nr_pages)
 436 {
 437         struct inode *inode = file_inode(file);
 438         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 439         int rc = 0;
 440         int max = 0;
 441
 442         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 443                 return -EINVAL;
 444
 445         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 446                                          &nr_pages);
 447
 448         if (rc == 0)
 449                 goto out;
 450
 451         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 452         dout("readpages %p file %p nr_pages %d max %d\n",
 453              inode, file, nr_pages, max);
 454         while (!list_empty(page_list)) {
 455                 rc = start_read(inode, page_list, max);
 456                 if (rc < 0)
 457                         goto out;
 458         }
 459 out:
 460         ceph_fscache_readpages_cancel(inode, page_list);
 461
 462         dout("readpages %p file %p ret %d\n", inode, file, rc);
 463         return rc;
 464 }
 465
 466 struct ceph_writeback_ctl
 467 {
 468         loff_t i_size;
 469         u64 truncate_size;
 470         u32 truncate_seq;
 471         bool size_stable;
 472 };
 473
 474 /*
 475  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 476  * only snap context we are allowed to write back.
 477  */
 478 static struct ceph_snap_context *
 479 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 480                    struct ceph_snap_context *page_snapc)
 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                         continue;
 492
 493                 /* get i_size, truncate_{seq,size} for page_snapc? */
 494                 if (snapc && capsnap->context != page_snapc)
 495                         continue;
 496
 497                 if (ctl) {
 498                         if (capsnap->writing) {
 499                                 ctl->i_size = i_size_read(inode);
 500                                 ctl->size_stable = false;
 501                         } else {
 502                                 ctl->i_size = capsnap->size;
 503                                 ctl->size_stable = true;
 504                         }
 505                         ctl->truncate_size = capsnap->truncate_size;
 506                         ctl->truncate_seq = capsnap->truncate_seq;
 507                 }
 508
 509                 if (snapc)
 510                         break;
 511
 512                 snapc = ceph_get_snap_context(capsnap->context);
 513                 if (!page_snapc ||
 514                     page_snapc == snapc ||
 515                     page_snapc->seq > snapc->seq)
 516                         break;
 517         }
 518         if (!snapc && ci->i_wrbuffer_ref_head) {
 519                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 520                 dout(" head snapc %p has %d dirty pages\n",
 521                      snapc, ci->i_wrbuffer_ref_head);
 522                 if (ctl) {
 523                         ctl->i_size = i_size_read(inode);
 524                         ctl->truncate_size = ci->i_truncate_size;
 525                         ctl->truncate_seq = ci->i_truncate_seq;
 526                         ctl->size_stable = false;
 527                 }
 528         }
 529         spin_unlock(&ci->i_ceph_lock);
 530         return snapc;
 531 }
 532
 533 static u64 get_writepages_data_length(struct inode *inode,
 534                                       struct page *page, u64 start)
 535 {
 536         struct ceph_inode_info *ci = ceph_inode(inode);
 537         struct ceph_snap_context *snapc = page_snap_context(page);
 538         struct ceph_cap_snap *capsnap = NULL;
 539         u64 end = i_size_read(inode);
 540
 541         if (snapc != ci->i_head_snapc) {
 542                 bool found = false;
 543                 spin_lock(&ci->i_ceph_lock);
 544                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 545                         if (capsnap->context == snapc) {
 546                                 if (!capsnap->writing)
 547                                         end = capsnap->size;
 548                                 found = true;
 549                                 break;
 550                         }
 551                 }
 552                 spin_unlock(&ci->i_ceph_lock);
 553                 WARN_ON(!found);
 554         }
 555         if (end > page_offset(page) + PAGE_SIZE)
 556                 end = page_offset(page) + PAGE_SIZE;
 557         return end > start ? end - start : 0;
 558 }
 559
 560 /*
 561  * Write a single page, but leave the page locked.
 562  *
 563  * If we get a write error, set the page error bit, but still adjust the
 564  * dirty page accounting (i.e., page is no longer dirty).
 565  */
 566 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 567 {
 568         struct inode *inode;
 569         struct ceph_inode_info *ci;
 570         struct ceph_fs_client *fsc;
 571         struct ceph_snap_context *snapc, *oldest;
 572         loff_t page_off = page_offset(page);
 573         long writeback_stat;
 574         int err, len = PAGE_SIZE;
 575         struct ceph_writeback_ctl ceph_wbc;
 576
 577         dout("writepage %p idx %lu\n", page, page->index);
 578
 579         inode = page->mapping->host;
 580         ci = ceph_inode(inode);
 581         fsc = ceph_inode_to_client(inode);
 582
 583         /* verify this is a writeable snap context */
 584         snapc = page_snap_context(page);
 585         if (!snapc) {
 586                 dout("writepage %p page %p not dirty?\n", inode, page);
 587                 return 0;
 588         }
 589         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 590         if (snapc->seq > oldest->seq) {
 591                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 592                      inode, page, snapc);
 593                 /* we should only noop if called by kswapd */
 594                 WARN_ON(!(current->flags & PF_MEMALLOC));
 595                 ceph_put_snap_context(oldest);
 596                 redirty_page_for_writepage(wbc, page);
 597                 return 0;
 598         }
 599         ceph_put_snap_context(oldest);
 600
 601         /* is this a partial page at end of file? */
 602         if (page_off >= ceph_wbc.i_size) {
 603                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 604                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 605                 return 0;
 606         }
 607
 608         if (ceph_wbc.i_size < page_off + len)
 609                 len = ceph_wbc.i_size - page_off;
 610
 611         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 612              inode, page, page->index, page_off, len, snapc, snapc->seq);
 613
 614         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 615         if (writeback_stat >
 616             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 617                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 618
 619         set_page_writeback(page);
 620         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 621                                    &ci->i_layout, snapc, page_off, len,
 622                                    ceph_wbc.truncate_seq,
 623                                    ceph_wbc.truncate_size,
 624                                    &inode->i_mtime, &page, 1);
 625         if (err < 0) {
 626                 struct writeback_control tmp_wbc;
 627                 if (!wbc)
 628                         wbc = &tmp_wbc;
 629                 if (err == -ERESTARTSYS) {
 630                         /* killed by SIGKILL */
 631                         dout("writepage interrupted page %p\n", page);
 632                         redirty_page_for_writepage(wbc, page);
 633                         end_page_writeback(page);
 634                         return err;
 635                 }
 636                 dout("writepage setting page/mapping error %d %p\n",
 637                      err, page);
 638                 SetPageError(page);
 639                 mapping_set_error(&inode->i_data, err);
 640                 wbc->pages_skipped++;
 641         } else {
 642                 dout("writepage cleaned page %p\n", page);
 643                 err = 0;  /* vfs expects us to return 0 */
 644         }
 645         page->private = 0;
 646         ClearPagePrivate(page);
 647         end_page_writeback(page);
 648         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 649         ceph_put_snap_context(snapc);  /* page's reference */
 650         return err;
 651 }
 652
 653 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 654 {
 655         int err;
 656         struct inode *inode = page->mapping->host;
 657         BUG_ON(!inode);
 658         ihold(inode);
 659         err = writepage_nounlock(page, wbc);
 660         if (err == -ERESTARTSYS) {
 661                 /* direct memory reclaimer was killed by SIGKILL. return 0
 662                  * to prevent caller from setting mapping/page error */
 663                 err = 0;
 664         }
 665         unlock_page(page);
 666         iput(inode);
 667         return err;
 668 }
 669
 670 /*
 671  * lame release_pages helper.  release_pages() isn't exported to
 672  * modules.
 673  */
 674 static void ceph_release_pages(struct page **pages, int num)
 675 {
 676         struct pagevec pvec;
 677         int i;
 678
 679         pagevec_init(&pvec, 0);
 680         for (i = 0; i < num; i++) {
 681                 if (pagevec_add(&pvec, pages[i]) == 0)
 682                         pagevec_release(&pvec);
 683         }
 684         pagevec_release(&pvec);
 685 }
 686
 687 /*
 688  * async writeback completion handler.
 689  *
 690  * If we get an error, set the mapping error bit, but not the individual
 691  * page error bits.
 692  */
 693 static void writepages_finish(struct ceph_osd_request *req)
 694 {
 695         struct inode *inode = req->r_inode;
 696         struct ceph_inode_info *ci = ceph_inode(inode);
 697         struct ceph_osd_data *osd_data;
 698         struct page *page;
 699         int num_pages, total_pages = 0;
 700         int i, j;
 701         int rc = req->r_result;
 702         struct ceph_snap_context *snapc = req->r_snapc;
 703         struct address_space *mapping = inode->i_mapping;
 704         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 705         bool remove_page;
 706
 707         dout("writepages_finish %p rc %d\n", inode, rc);
 708         if (rc < 0) {
 709                 mapping_set_error(mapping, rc);
 710                 ceph_set_error_write(ci);
 711         } else {
 712                 ceph_clear_error_write(ci);
 713         }
 714
 715         /*
 716          * We lost the cache cap, need to truncate the page before
 717          * it is unlocked, otherwise we'd truncate it later in the
 718          * page truncation thread, possibly losing some data that
 719          * raced its way in
 720          */
 721         remove_page = !(ceph_caps_issued(ci) &
 722                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 723
 724         /* clean all pages */
 725         for (i = 0; i < req->r_num_ops; i++) {
 726                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 727                         break;
 728
 729                 osd_data = osd_req_op_extent_osd_data(req, i);
 730                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 731                 num_pages = calc_pages_for((u64)osd_data->alignment,
 732                                            (u64)osd_data->length);
 733                 total_pages += num_pages;
 734                 for (j = 0; j < num_pages; j++) {
 735                         page = osd_data->pages[j];
 736                         BUG_ON(!page);
 737                         WARN_ON(!PageUptodate(page));
 738
 739                         if (atomic_long_dec_return(&fsc->writeback_count) <
 740                              CONGESTION_OFF_THRESH(
 741                                         fsc->mount_options->congestion_kb))
 742                                 clear_bdi_congested(inode_to_bdi(inode),
 743                                                     BLK_RW_ASYNC);
 744
 745                         ceph_put_snap_context(page_snap_context(page));
 746                         page->private = 0;
 747                         ClearPagePrivate(page);
 748                         dout("unlocking %p\n", page);
 749                         end_page_writeback(page);
 750
 751                         if (remove_page)
 752                                 generic_error_remove_page(inode->i_mapping,
 753                                                           page);
 754
 755                         unlock_page(page);
 756                 }
 757                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 758                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 759
 760                 ceph_release_pages(osd_data->pages, num_pages);
 761         }
 762
 763         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 764
 765         osd_data = osd_req_op_extent_osd_data(req, 0);
 766         if (osd_data->pages_from_pool)
 767                 mempool_free(osd_data->pages,
 768                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 769         else
 770                 kfree(osd_data->pages);
 771         ceph_osdc_put_request(req);
 772 }
 773
 774 /*
 775  * initiate async writeback
 776  */
 777 static int ceph_writepages_start(struct address_space *mapping,
 778                                  struct writeback_control *wbc)
 779 {
 780         struct inode *inode = mapping->host;
 781         struct ceph_inode_info *ci = ceph_inode(inode);
 782         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 783         struct ceph_vino vino = ceph_vino(inode);
 784         pgoff_t index, start, end;
 785         int range_whole = 0;
 786         int should_loop = 1;
 787         pgoff_t max_pages = 0, max_pages_ever = 0;
 788         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 789         struct pagevec pvec;
 790         int done = 0;
 791         int rc = 0;
 792         unsigned int wsize = i_blocksize(inode);
 793         struct ceph_osd_request *req = NULL;
 794         struct ceph_writeback_ctl ceph_wbc;
 795
 796         dout("writepages_start %p (mode=%s)\n", inode,
 797              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 798              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 799
 800         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 801                 if (ci->i_wrbuffer_ref > 0) {
 802                         pr_warn_ratelimited(
 803                                 "writepage_start %p %lld forced umount\n",
 804                                 inode, ceph_ino(inode));
 805                 }
 806                 mapping_set_error(mapping, -EIO);
 807                 return -EIO; /* we're in a forced umount, don't write! */
 808         }
 809         if (fsc->mount_options->wsize < wsize)
 810                 wsize = fsc->mount_options->wsize;
 811         max_pages_ever = wsize >> PAGE_SHIFT;
 812
 813         pagevec_init(&pvec, 0);
 814
 815         /* where to start/end? */
 816         if (wbc->range_cyclic) {
 817                 start = mapping->writeback_index; /* Start from prev offset */
 818                 end = -1;
 819                 dout(" cyclic, start at %lu\n", start);
 820         } else {
 821                 start = wbc->range_start >> PAGE_SHIFT;
 822                 end = wbc->range_end >> PAGE_SHIFT;
 823                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 824                         range_whole = 1;
 825                 should_loop = 0;
 826                 dout(" not cyclic, %lu to %lu\n", start, end);
 827         }
 828         index = start;
 829
 830 retry:
 831         /* find oldest snap context with dirty data */
 832         ceph_put_snap_context(snapc);
 833         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 834         if (!snapc) {
 835                 /* hmm, why does writepages get called when there
 836                    is no dirty data? */
 837                 dout(" no snap context with dirty data?\n");
 838                 goto out;
 839         }
 840         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 841              snapc, snapc->seq, snapc->num_snaps);
 842
 843         if (last_snapc && snapc != last_snapc) {
 844                 /* if we switched to a newer snapc, restart our scan at the
 845                  * start of the original file range. */
 846                 dout("  snapc differs from last pass, restarting at %lu\n",
 847                      index);
 848                 index = start;
 849         }
 850         last_snapc = snapc;
 851
 852         while (!done && index <= end) {
 853                 unsigned i;
 854                 pgoff_t strip_unit_end = 0;
 855                 int num_ops = 0, op_idx;
 856                 int pvec_pages, locked_pages = 0;
 857                 struct page **pages = NULL, **data_pages;
 858                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 859                 struct page *page;
 860                 int want;
 861                 u64 offset = 0, len = 0;
 862
 863                 max_pages = max_pages_ever;
 864
 865 get_more_pages:
 866                 want = min(end - index,
 867                            min((pgoff_t)PAGEVEC_SIZE,
 868                                max_pages - (pgoff_t)locked_pages) - 1)
 869                         + 1;
 870                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 871                                                 PAGECACHE_TAG_DIRTY,
 872                                                 want);
 873                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 874                 if (!pvec_pages && !locked_pages)
 875                         break;
 876                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 877                         page = pvec.pages[i];
 878                         dout("? %p idx %lu\n", page, page->index);
 879                         if (locked_pages == 0)
 880                                 lock_page(page);  /* first page */
 881                         else if (!trylock_page(page))
 882                                 break;
 883
 884                         /* only dirty pages, or our accounting breaks */
 885                         if (unlikely(!PageDirty(page)) ||
 886                             unlikely(page->mapping != mapping)) {
 887                                 dout("!dirty or !mapping %p\n", page);
 888                                 unlock_page(page);
 889                                 continue;
 890                         }
 891                         if (!wbc->range_cyclic && page->index > end) {
 892                                 dout("end of range %p\n", page);
 893                                 done = 1;
 894                                 unlock_page(page);
 895                                 break;
 896                         }
 897                         if (strip_unit_end && (page->index > strip_unit_end)) {
 898                                 dout("end of strip unit %p\n", page);
 899                                 unlock_page(page);
 900                                 break;
 901                         }
 902                         if (page_offset(page) >= ceph_wbc.i_size) {
 903                                 dout("%p page eof %llu\n",
 904                                      page, ceph_wbc.i_size);
 905                                 done = 1;
 906                                 unlock_page(page);
 907                                 break;
 908                         }
 909                         if (PageWriteback(page)) {
 910                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 911                                         dout("%p under writeback\n", page);
 912                                         unlock_page(page);
 913                                         continue;
 914                                 }
 915                                 dout("waiting on writeback %p\n", page);
 916                                 wait_on_page_writeback(page);
 917                         }
 918
 919                         /* only if matching snap context */
 920                         pgsnapc = page_snap_context(page);
 921                         if (pgsnapc->seq > snapc->seq) {
 922                                 dout("page snapc %p %lld > oldest %p %lld\n",
 923                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 924                                 unlock_page(page);
 925                                 continue;
 926                         }
 927
 928                         if (!clear_page_dirty_for_io(page)) {
 929                                 dout("%p !clear_page_dirty_for_io\n", page);
 930                                 unlock_page(page);
 931                                 continue;
 932                         }
 933
 934                         /*
 935                          * We have something to write.  If this is
 936                          * the first locked page this time through,
 937                          * calculate max possinle write size and
 938                          * allocate a page array
 939                          */
 940                         if (locked_pages == 0) {
 941                                 u64 objnum;
 942                                 u64 objoff;
 943
 944                                 /* prepare async write request */
 945                                 offset = (u64)page_offset(page);
 946                                 len = wsize;
 947
 948                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 949                                                                 offset, len,
 950                                                                 &objnum, &objoff,
 951                                                                 &len);
 952                                 if (rc < 0) {
 953                                         unlock_page(page);
 954                                         break;
 955                                 }
 956
 957                                 num_ops = 1;
 958                                 strip_unit_end = page->index +
 959                                         ((len - 1) >> PAGE_SHIFT);
 960
 961                                 BUG_ON(pages);
 962                                 max_pages = calc_pages_for(0, (u64)len);
 963                                 pages = kmalloc(max_pages * sizeof (*pages),
 964                                                 GFP_NOFS);
 965                                 if (!pages) {
 966                                         pool = fsc->wb_pagevec_pool;
 967                                         pages = mempool_alloc(pool, GFP_NOFS);
 968                                         BUG_ON(!pages);
 969                                 }
 970
 971                                 len = 0;
 972                         } else if (page->index !=
 973                                    (offset + len) >> PAGE_SHIFT) {
 974                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 975                                                         CEPH_OSD_MAX_OPS)) {
 976                                         redirty_page_for_writepage(wbc, page);
 977                                         unlock_page(page);
 978                                         break;
 979                                 }
 980
 981                                 num_ops++;
 982                                 offset = (u64)page_offset(page);
 983                                 len = 0;
 984                         }
 985
 986                         /* note position of first page in pvec */
 987                         dout("%p will write page %p idx %lu\n",
 988                              inode, page, page->index);
 989
 990                         if (atomic_long_inc_return(&fsc->writeback_count) >
 991                             CONGESTION_ON_THRESH(
 992                                     fsc->mount_options->congestion_kb)) {
 993                                 set_bdi_congested(inode_to_bdi(inode),
 994                                                   BLK_RW_ASYNC);
 995                         }
 996
 997
 998                         pages[locked_pages++] = page;
 999                         pvec.pages[i] = NULL;
1000
1001                         len += PAGE_SIZE;
1002                 }
1003
1004                 /* did we get anything? */
1005                 if (!locked_pages)
1006                         goto release_pvec_pages;
1007                 if (i) {
1008                         unsigned j, n = 0;
1009                         /* shift unused page to beginning of pvec */
1010                         for (j = 0; j < pvec_pages; j++) {
1011                                 if (!pvec.pages[j])
1012                                         continue;
1013                                 if (n < j)
1014                                         pvec.pages[n] = pvec.pages[j];
1015                                 n++;
1016                         }
1017                         pvec.nr = n;
1018
1019                         if (pvec_pages && i == pvec_pages &&
1020                             locked_pages < max_pages) {
1021                                 dout("reached end pvec, trying for more\n");
1022                                 pagevec_release(&pvec);
1023                                 goto get_more_pages;
1024                         }
1025                 }
1026
1027 new_request:
1028                 offset = page_offset(pages[0]);
1029                 len = wsize;
1030
1031                 req = ceph_osdc_new_request(&fsc->client->osdc,
1032                                         &ci->i_layout, vino,
1033                                         offset, &len, 0, num_ops,
1034                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1035                                         snapc, ceph_wbc.truncate_seq,
1036                                         ceph_wbc.truncate_size, false);
1037                 if (IS_ERR(req)) {
1038                         req = ceph_osdc_new_request(&fsc->client->osdc,
1039                                                 &ci->i_layout, vino,
1040                                                 offset, &len, 0,
1041                                                 min(num_ops,
1042                                                     CEPH_OSD_SLAB_OPS),
1043                                                 CEPH_OSD_OP_WRITE,
1044                                                 CEPH_OSD_FLAG_WRITE,
1045                                                 snapc, ceph_wbc.truncate_seq,
1046                                                 ceph_wbc.truncate_size, true);
1047                         BUG_ON(IS_ERR(req));
1048                 }
1049                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1050                              PAGE_SIZE - offset);
1051
1052                 req->r_callback = writepages_finish;
1053                 req->r_inode = inode;
1054
1055                 /* Format the osd request message and submit the write */
1056                 len = 0;
1057                 data_pages = pages;
1058                 op_idx = 0;
1059                 for (i = 0; i < locked_pages; i++) {
1060                         u64 cur_offset = page_offset(pages[i]);
1061                         if (offset + len != cur_offset) {
1062                                 if (op_idx + 1 == req->r_num_ops)
1063                                         break;
1064                                 osd_req_op_extent_dup_last(req, op_idx,
1065                                                            cur_offset - offset);
1066                                 dout("writepages got pages at %llu~%llu\n",
1067                                      offset, len);
1068                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1069                                                         data_pages, len, 0,
1070                                                         !!pool, false);
1071                                 osd_req_op_extent_update(req, op_idx, len);
1072
1073                                 len = 0;
1074                                 offset = cur_offset;
1075                                 data_pages = pages + i;
1076                                 op_idx++;
1077                         }
1078
1079                         set_page_writeback(pages[i]);
1080                         len += PAGE_SIZE;
1081                 }
1082
1083                 if (ceph_wbc.size_stable) {
1084                         len = min(len, ceph_wbc.i_size - offset);
1085                 } else if (i == locked_pages) {
1086                         /* writepages_finish() clears writeback pages
1087                          * according to the data length, so make sure
1088                          * data length covers all locked pages */
1089                         u64 min_len = len + 1 - PAGE_SIZE;
1090                         len = get_writepages_data_length(inode, pages[i - 1],
1091                                                          offset);
1092                         len = max(len, min_len);
1093                 }
1094                 dout("writepages got pages at %llu~%llu\n", offset, len);
1095
1096                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1097                                                  0, !!pool, false);
1098                 osd_req_op_extent_update(req, op_idx, len);
1099
1100                 BUG_ON(op_idx + 1 != req->r_num_ops);
1101
1102                 pool = NULL;
1103                 if (i < locked_pages) {
1104                         BUG_ON(num_ops <= req->r_num_ops);
1105                         num_ops -= req->r_num_ops;
1106                         locked_pages -= i;
1107
1108                         /* allocate new pages array for next request */
1109                         data_pages = pages;
1110                         pages = kmalloc(locked_pages * sizeof (*pages),
1111                                         GFP_NOFS);
1112                         if (!pages) {
1113                                 pool = fsc->wb_pagevec_pool;
1114                                 pages = mempool_alloc(pool, GFP_NOFS);
1115                                 BUG_ON(!pages);
1116                         }
1117                         memcpy(pages, data_pages + i,
1118                                locked_pages * sizeof(*pages));
1119                         memset(data_pages + i, 0,
1120                                locked_pages * sizeof(*pages));
1121                 } else {
1122                         BUG_ON(num_ops != req->r_num_ops);
1123                         index = pages[i - 1]->index + 1;
1124                         /* request message now owns the pages array */
1125                         pages = NULL;
1126                 }
1127
1128                 req->r_mtime = inode->i_mtime;
1129                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1130                 BUG_ON(rc);
1131                 req = NULL;
1132
1133                 wbc->nr_to_write -= i;
1134                 if (pages)
1135                         goto new_request;
1136
1137                 if (wbc->nr_to_write <= 0)
1138                         done = 1;
1139
1140 release_pvec_pages:
1141                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1142                      pvec.nr ? pvec.pages[0] : NULL);
1143                 pagevec_release(&pvec);
1144
1145                 if (locked_pages && !done)
1146                         goto retry;
1147         }
1148
1149         if (should_loop && !done) {
1150                 /* more to do; loop back to beginning of file */
1151                 dout("writepages looping back to beginning of file\n");
1152                 should_loop = 0;
1153                 index = 0;
1154                 goto retry;
1155         }
1156
1157         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1158                 mapping->writeback_index = index;
1159
1160 out:
1161         ceph_osdc_put_request(req);
1162         ceph_put_snap_context(snapc);
1163         dout("writepages done, rc = %d\n", rc);
1164         return rc;
1165 }
1166
1167
1168
1169 /*
1170  * See if a given @snapc is either writeable, or already written.
1171  */
1172 static int context_is_writeable_or_written(struct inode *inode,
1173                                            struct ceph_snap_context *snapc)
1174 {
1175         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1176         int ret = !oldest || snapc->seq <= oldest->seq;
1177
1178         ceph_put_snap_context(oldest);
1179         return ret;
1180 }
1181
1182 /*
1183  * We are only allowed to write into/dirty the page if the page is
1184  * clean, or already dirty within the same snap context.
1185  *
1186  * called with page locked.
1187  * return success with page locked,
1188  * or any failure (incl -EAGAIN) with page unlocked.
1189  */
1190 static int ceph_update_writeable_page(struct file *file,
1191                             loff_t pos, unsigned len,
1192                             struct page *page)
1193 {
1194         struct inode *inode = file_inode(file);
1195         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1196         struct ceph_inode_info *ci = ceph_inode(inode);
1197         loff_t page_off = pos & PAGE_MASK;
1198         int pos_in_page = pos & ~PAGE_MASK;
1199         int end_in_page = pos_in_page + len;
1200         loff_t i_size;
1201         int r;
1202         struct ceph_snap_context *snapc, *oldest;
1203
1204         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1205                 dout(" page %p forced umount\n", page);
1206                 unlock_page(page);
1207                 return -EIO;
1208         }
1209
1210 retry_locked:
1211         /* writepages currently holds page lock, but if we change that later, */
1212         wait_on_page_writeback(page);
1213
1214         snapc = page_snap_context(page);
1215         if (snapc && snapc != ci->i_head_snapc) {
1216                 /*
1217                  * this page is already dirty in another (older) snap
1218                  * context!  is it writeable now?
1219                  */
1220                 oldest = get_oldest_context(inode, NULL, NULL);
1221                 if (snapc->seq > oldest->seq) {
1222                         ceph_put_snap_context(oldest);
1223                         dout(" page %p snapc %p not current or oldest\n",
1224                              page, snapc);
1225                         /*
1226                          * queue for writeback, and wait for snapc to
1227                          * be writeable or written
1228                          */
1229                         snapc = ceph_get_snap_context(snapc);
1230                         unlock_page(page);
1231                         ceph_queue_writeback(inode);
1232                         r = wait_event_killable(ci->i_cap_wq,
1233                                context_is_writeable_or_written(inode, snapc));
1234                         ceph_put_snap_context(snapc);
1235                         if (r == -ERESTARTSYS)
1236                                 return r;
1237                         return -EAGAIN;
1238                 }
1239                 ceph_put_snap_context(oldest);
1240
1241                 /* yay, writeable, do it now (without dropping page lock) */
1242                 dout(" page %p snapc %p not current, but oldest\n",
1243                      page, snapc);
1244                 if (!clear_page_dirty_for_io(page))
1245                         goto retry_locked;
1246                 r = writepage_nounlock(page, NULL);
1247                 if (r < 0)
1248                         goto fail_unlock;
1249                 goto retry_locked;
1250         }
1251
1252         if (PageUptodate(page)) {
1253                 dout(" page %p already uptodate\n", page);
1254                 return 0;
1255         }
1256
1257         /* full page? */
1258         if (pos_in_page == 0 && len == PAGE_SIZE)
1259                 return 0;
1260
1261         /* past end of file? */
1262         i_size = i_size_read(inode);
1263
1264         if (page_off >= i_size ||
1265             (pos_in_page == 0 && (pos+len) >= i_size &&
1266              end_in_page - pos_in_page != PAGE_SIZE)) {
1267                 dout(" zeroing %p 0 - %d and %d - %d\n",
1268                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1269                 zero_user_segments(page,
1270                                    0, pos_in_page,
1271                                    end_in_page, PAGE_SIZE);
1272                 return 0;
1273         }
1274
1275         /* we need to read it. */
1276         r = ceph_do_readpage(file, page);
1277         if (r < 0) {
1278                 if (r == -EINPROGRESS)
1279                         return -EAGAIN;
1280                 goto fail_unlock;
1281         }
1282         goto retry_locked;
1283 fail_unlock:
1284         unlock_page(page);
1285         return r;
1286 }
1287
1288 /*
1289  * We are only allowed to write into/dirty the page if the page is
1290  * clean, or already dirty within the same snap context.
1291  */
1292 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1293                             loff_t pos, unsigned len, unsigned flags,
1294                             struct page **pagep, void **fsdata)
1295 {
1296         struct inode *inode = file_inode(file);
1297         struct page *page;
1298         pgoff_t index = pos >> PAGE_SHIFT;
1299         int r;
1300
1301         do {
1302                 /* get a page */
1303                 page = grab_cache_page_write_begin(mapping, index, 0);
1304                 if (!page)
1305                         return -ENOMEM;
1306
1307                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1308                      inode, page, (int)pos, (int)len);
1309
1310                 r = ceph_update_writeable_page(file, pos, len, page);
1311                 if (r < 0)
1312                         put_page(page);
1313                 else
1314                         *pagep = page;
1315         } while (r == -EAGAIN);
1316
1317         return r;
1318 }
1319
1320 /*
1321  * we don't do anything in here that simple_write_end doesn't do
1322  * except adjust dirty page accounting
1323  */
1324 static int ceph_write_end(struct file *file, struct address_space *mapping,
1325                           loff_t pos, unsigned len, unsigned copied,
1326                           struct page *page, void *fsdata)
1327 {
1328         struct inode *inode = file_inode(file);
1329         bool check_cap = false;
1330
1331         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1332              inode, page, (int)pos, (int)copied, (int)len);
1333
1334         /* zero the stale part of the page if we did a short copy */
1335         if (!PageUptodate(page)) {
1336                 if (copied < len) {
1337                         copied = 0;
1338                         goto out;
1339                 }
1340                 SetPageUptodate(page);
1341         }
1342
1343         /* did file size increase? */
1344         if (pos+copied > i_size_read(inode))
1345                 check_cap = ceph_inode_set_size(inode, pos+copied);
1346
1347         set_page_dirty(page);
1348
1349 out:
1350         unlock_page(page);
1351         put_page(page);
1352
1353         if (check_cap)
1354                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1355
1356         return copied;
1357 }
1358
1359 /*
1360  * we set .direct_IO to indicate direct io is supported, but since we
1361  * intercept O_DIRECT reads and writes early, this function should
1362  * never get called.
1363  */
1364 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1365 {
1366         WARN_ON(1);
1367         return -EINVAL;
1368 }
1369
1370 const struct address_space_operations ceph_aops = {
1371         .readpage = ceph_readpage,
1372         .readpages = ceph_readpages,
1373         .writepage = ceph_writepage,
1374         .writepages = ceph_writepages_start,
1375         .write_begin = ceph_write_begin,
1376         .write_end = ceph_write_end,
1377         .set_page_dirty = ceph_set_page_dirty,
1378         .invalidatepage = ceph_invalidatepage,
1379         .releasepage = ceph_releasepage,
1380         .direct_IO = ceph_direct_io,
1381 };
1382
1383 static void ceph_block_sigs(sigset_t *oldset)
1384 {
1385         sigset_t mask;
1386         siginitsetinv(&mask, sigmask(SIGKILL));
1387         sigprocmask(SIG_BLOCK, &mask, oldset);
1388 }
1389
1390 static void ceph_restore_sigs(sigset_t *oldset)
1391 {
1392         sigprocmask(SIG_SETMASK, oldset, NULL);
1393 }
1394
1395 /*
1396  * vm ops
1397  */
1398 static int ceph_filemap_fault(struct vm_fault *vmf)
1399 {
1400         struct vm_area_struct *vma = vmf->vma;
1401         struct inode *inode = file_inode(vma->vm_file);
1402         struct ceph_inode_info *ci = ceph_inode(inode);
1403         struct ceph_file_info *fi = vma->vm_file->private_data;
1404         struct page *pinned_page = NULL;
1405         loff_t off = vmf->pgoff << PAGE_SHIFT;
1406         int want, got, ret;
1407         sigset_t oldset;
1408
1409         ceph_block_sigs(&oldset);
1410
1411         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1412              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1413         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1414                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1415         else
1416                 want = CEPH_CAP_FILE_CACHE;
1417
1418         got = 0;
1419         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1420         if (ret < 0)
1421                 goto out_restore;
1422
1423         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1424              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1425
1426         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1427             ci->i_inline_version == CEPH_INLINE_NONE) {
1428                 current->journal_info = vma->vm_file;
1429                 ret = filemap_fault(vmf);
1430                 current->journal_info = NULL;
1431         } else
1432                 ret = -EAGAIN;
1433
1434         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1435              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1436         if (pinned_page)
1437                 put_page(pinned_page);
1438         ceph_put_cap_refs(ci, got);
1439
1440         if (ret != -EAGAIN)
1441                 goto out_restore;
1442
1443         /* read inline data */
1444         if (off >= PAGE_SIZE) {
1445                 /* does not support inline data > PAGE_SIZE */
1446                 ret = VM_FAULT_SIGBUS;
1447         } else {
1448                 int ret1;
1449                 struct address_space *mapping = inode->i_mapping;
1450                 struct page *page = find_or_create_page(mapping, 0,
1451                                                 mapping_gfp_constraint(mapping,
1452                                                 ~__GFP_FS));
1453                 if (!page) {
1454                         ret = VM_FAULT_OOM;
1455                         goto out_inline;
1456                 }
1457                 ret1 = __ceph_do_getattr(inode, page,
1458                                          CEPH_STAT_CAP_INLINE_DATA, true);
1459                 if (ret1 < 0 || off >= i_size_read(inode)) {
1460                         unlock_page(page);
1461                         put_page(page);
1462                         if (ret1 < 0)
1463                                 ret = ret1;
1464                         else
1465                                 ret = VM_FAULT_SIGBUS;
1466                         goto out_inline;
1467                 }
1468                 if (ret1 < PAGE_SIZE)
1469                         zero_user_segment(page, ret1, PAGE_SIZE);
1470                 else
1471                         flush_dcache_page(page);
1472                 SetPageUptodate(page);
1473                 vmf->page = page;
1474                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1475 out_inline:
1476                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1477                      inode, off, (size_t)PAGE_SIZE, ret);
1478         }
1479 out_restore:
1480         ceph_restore_sigs(&oldset);
1481         if (ret < 0)
1482                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1483
1484         return ret;
1485 }
1486
1487 /*
1488  * Reuse write_begin here for simplicity.
1489  */
1490 static int ceph_page_mkwrite(struct vm_fault *vmf)
1491 {
1492         struct vm_area_struct *vma = vmf->vma;
1493         struct inode *inode = file_inode(vma->vm_file);
1494         struct ceph_inode_info *ci = ceph_inode(inode);
1495         struct ceph_file_info *fi = vma->vm_file->private_data;
1496         struct ceph_cap_flush *prealloc_cf;
1497         struct page *page = vmf->page;
1498         loff_t off = page_offset(page);
1499         loff_t size = i_size_read(inode);
1500         size_t len;
1501         int want, got, ret;
1502         sigset_t oldset;
1503
1504         prealloc_cf = ceph_alloc_cap_flush();
1505         if (!prealloc_cf)
1506                 return VM_FAULT_OOM;
1507
1508         ceph_block_sigs(&oldset);
1509
1510         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1511                 struct page *locked_page = NULL;
1512                 if (off == 0) {
1513                         lock_page(page);
1514                         locked_page = page;
1515                 }
1516                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1517                 if (locked_page)
1518                         unlock_page(locked_page);
1519                 if (ret < 0)
1520                         goto out_free;
1521         }
1522
1523         if (off + PAGE_SIZE <= size)
1524                 len = PAGE_SIZE;
1525         else
1526                 len = size & ~PAGE_MASK;
1527
1528         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1529              inode, ceph_vinop(inode), off, len, size);
1530         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1531                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1532         else
1533                 want = CEPH_CAP_FILE_BUFFER;
1534
1535         got = 0;
1536         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1537                             &got, NULL);
1538         if (ret < 0)
1539                 goto out_free;
1540
1541         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1542              inode, off, len, ceph_cap_string(got));
1543
1544         /* Update time before taking page lock */
1545         file_update_time(vma->vm_file);
1546
1547         do {
1548                 lock_page(page);
1549
1550                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1551                         unlock_page(page);
1552                         ret = VM_FAULT_NOPAGE;
1553                         break;
1554                 }
1555
1556                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1557                 if (ret >= 0) {
1558                         /* success.  we'll keep the page locked. */
1559                         set_page_dirty(page);
1560                         ret = VM_FAULT_LOCKED;
1561                 }
1562         } while (ret == -EAGAIN);
1563
1564         if (ret == VM_FAULT_LOCKED ||
1565             ci->i_inline_version != CEPH_INLINE_NONE) {
1566                 int dirty;
1567                 spin_lock(&ci->i_ceph_lock);
1568                 ci->i_inline_version = CEPH_INLINE_NONE;
1569                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1570                                                &prealloc_cf);
1571                 spin_unlock(&ci->i_ceph_lock);
1572                 if (dirty)
1573                         __mark_inode_dirty(inode, dirty);
1574         }
1575
1576         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1577              inode, off, len, ceph_cap_string(got), ret);
1578         ceph_put_cap_refs(ci, got);
1579 out_free:
1580         ceph_restore_sigs(&oldset);
1581         ceph_free_cap_flush(prealloc_cf);
1582         if (ret < 0)
1583                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1584         return ret;
1585 }
1586
1587 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1588                            char *data, size_t len)
1589 {
1590         struct address_space *mapping = inode->i_mapping;
1591         struct page *page;
1592
1593         if (locked_page) {
1594                 page = locked_page;
1595         } else {
1596                 if (i_size_read(inode) == 0)
1597                         return;
1598                 page = find_or_create_page(mapping, 0,
1599                                            mapping_gfp_constraint(mapping,
1600                                            ~__GFP_FS));
1601                 if (!page)
1602                         return;
1603                 if (PageUptodate(page)) {
1604                         unlock_page(page);
1605                         put_page(page);
1606                         return;
1607                 }
1608         }
1609
1610         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1611              inode, ceph_vinop(inode), len, locked_page);
1612
1613         if (len > 0) {
1614                 void *kaddr = kmap_atomic(page);
1615                 memcpy(kaddr, data, len);
1616                 kunmap_atomic(kaddr);
1617         }
1618
1619         if (page != locked_page) {
1620                 if (len < PAGE_SIZE)
1621                         zero_user_segment(page, len, PAGE_SIZE);
1622                 else
1623                         flush_dcache_page(page);
1624
1625                 SetPageUptodate(page);
1626                 unlock_page(page);
1627                 put_page(page);
1628         }
1629 }
1630
1631 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1632 {
1633         struct inode *inode = file_inode(filp);
1634         struct ceph_inode_info *ci = ceph_inode(inode);
1635         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1636         struct ceph_osd_request *req;
1637         struct page *page = NULL;
1638         u64 len, inline_version;
1639         int err = 0;
1640         bool from_pagecache = false;
1641
1642         spin_lock(&ci->i_ceph_lock);
1643         inline_version = ci->i_inline_version;
1644         spin_unlock(&ci->i_ceph_lock);
1645
1646         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1647              inode, ceph_vinop(inode), inline_version);
1648
1649         if (inline_version == 1 || /* initial version, no data */
1650             inline_version == CEPH_INLINE_NONE)
1651                 goto out;
1652
1653         if (locked_page) {
1654                 page = locked_page;
1655                 WARN_ON(!PageUptodate(page));
1656         } else if (ceph_caps_issued(ci) &
1657                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1658                 page = find_get_page(inode->i_mapping, 0);
1659                 if (page) {
1660                         if (PageUptodate(page)) {
1661                                 from_pagecache = true;
1662                                 lock_page(page);
1663                         } else {
1664                                 put_page(page);
1665                                 page = NULL;
1666                         }
1667                 }
1668         }
1669
1670         if (page) {
1671                 len = i_size_read(inode);
1672                 if (len > PAGE_SIZE)
1673                         len = PAGE_SIZE;
1674         } else {
1675                 page = __page_cache_alloc(GFP_NOFS);
1676                 if (!page) {
1677                         err = -ENOMEM;
1678                         goto out;
1679                 }
1680                 err = __ceph_do_getattr(inode, page,
1681                                         CEPH_STAT_CAP_INLINE_DATA, true);
1682                 if (err < 0) {
1683                         /* no inline data */
1684                         if (err == -ENODATA)
1685                                 err = 0;
1686                         goto out;
1687                 }
1688                 len = err;
1689         }
1690
1691         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1692                                     ceph_vino(inode), 0, &len, 0, 1,
1693                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1694                                     NULL, 0, 0, false);
1695         if (IS_ERR(req)) {
1696                 err = PTR_ERR(req);
1697                 goto out;
1698         }
1699
1700         req->r_mtime = inode->i_mtime;
1701         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1702         if (!err)
1703                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1704         ceph_osdc_put_request(req);
1705         if (err < 0)
1706                 goto out;
1707
1708         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1709                                     ceph_vino(inode), 0, &len, 1, 3,
1710                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1711                                     NULL, ci->i_truncate_seq,
1712                                     ci->i_truncate_size, false);
1713         if (IS_ERR(req)) {
1714                 err = PTR_ERR(req);
1715                 goto out;
1716         }
1717
1718         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1719
1720         {
1721                 __le64 xattr_buf = cpu_to_le64(inline_version);
1722                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1723                                             "inline_version", &xattr_buf,
1724                                             sizeof(xattr_buf),
1725                                             CEPH_OSD_CMPXATTR_OP_GT,
1726                                             CEPH_OSD_CMPXATTR_MODE_U64);
1727                 if (err)
1728                         goto out_put;
1729         }
1730
1731         {
1732                 char xattr_buf[32];
1733                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1734                                          "%llu", inline_version);
1735                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1736                                             "inline_version",
1737                                             xattr_buf, xattr_len, 0, 0);
1738                 if (err)
1739                         goto out_put;
1740         }
1741
1742         req->r_mtime = inode->i_mtime;
1743         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1744         if (!err)
1745                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1746 out_put:
1747         ceph_osdc_put_request(req);
1748         if (err == -ECANCELED)
1749                 err = 0;
1750 out:
1751         if (page && page != locked_page) {
1752                 if (from_pagecache) {
1753                         unlock_page(page);
1754                         put_page(page);
1755                 } else
1756                         __free_pages(page, 0);
1757         }
1758
1759         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1760              inode, ceph_vinop(inode), inline_version, err);
1761         return err;
1762 }
1763
1764 static const struct vm_operations_struct ceph_vmops = {
1765         .fault          = ceph_filemap_fault,
1766         .page_mkwrite   = ceph_page_mkwrite,
1767 };
1768
1769 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1770 {
1771         struct address_space *mapping = file->f_mapping;
1772
1773         if (!mapping->a_ops->readpage)
1774                 return -ENOEXEC;
1775         file_accessed(file);
1776         vma->vm_ops = &ceph_vmops;
1777         return 0;
1778 }
1779
1780 enum {
1781         POOL_READ       = 1,
1782         POOL_WRITE      = 2,
1783 };
1784
1785 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1786                                 s64 pool, struct ceph_string *pool_ns)
1787 {
1788         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1789         struct ceph_mds_client *mdsc = fsc->mdsc;
1790         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1791         struct rb_node **p, *parent;
1792         struct ceph_pool_perm *perm;
1793         struct page **pages;
1794         size_t pool_ns_len;
1795         int err = 0, err2 = 0, have = 0;
1796
1797         down_read(&mdsc->pool_perm_rwsem);
1798         p = &mdsc->pool_perm_tree.rb_node;
1799         while (*p) {
1800                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1801                 if (pool < perm->pool)
1802                         p = &(*p)->rb_left;
1803                 else if (pool > perm->pool)
1804                         p = &(*p)->rb_right;
1805                 else {
1806                         int ret = ceph_compare_string(pool_ns,
1807                                                 perm->pool_ns,
1808                                                 perm->pool_ns_len);
1809                         if (ret < 0)
1810                                 p = &(*p)->rb_left;
1811                         else if (ret > 0)
1812                                 p = &(*p)->rb_right;
1813                         else {
1814                                 have = perm->perm;
1815                                 break;
1816                         }
1817                 }
1818         }
1819         up_read(&mdsc->pool_perm_rwsem);
1820         if (*p)
1821                 goto out;
1822
1823         if (pool_ns)
1824                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1825                      pool, (int)pool_ns->len, pool_ns->str);
1826         else
1827                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1828
1829         down_write(&mdsc->pool_perm_rwsem);
1830         p = &mdsc->pool_perm_tree.rb_node;
1831         parent = NULL;
1832         while (*p) {
1833                 parent = *p;
1834                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1835                 if (pool < perm->pool)
1836                         p = &(*p)->rb_left;
1837                 else if (pool > perm->pool)
1838                         p = &(*p)->rb_right;
1839                 else {
1840                         int ret = ceph_compare_string(pool_ns,
1841                                                 perm->pool_ns,
1842                                                 perm->pool_ns_len);
1843                         if (ret < 0)
1844                                 p = &(*p)->rb_left;
1845                         else if (ret > 0)
1846                                 p = &(*p)->rb_right;
1847                         else {
1848                                 have = perm->perm;
1849                                 break;
1850                         }
1851                 }
1852         }
1853         if (*p) {
1854                 up_write(&mdsc->pool_perm_rwsem);
1855                 goto out;
1856         }
1857
1858         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1859                                          1, false, GFP_NOFS);
1860         if (!rd_req) {
1861                 err = -ENOMEM;
1862                 goto out_unlock;
1863         }
1864
1865         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1866         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1867         rd_req->r_base_oloc.pool = pool;
1868         if (pool_ns)
1869                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1870         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1871
1872         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1873         if (err)
1874                 goto out_unlock;
1875
1876         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1877                                          1, false, GFP_NOFS);
1878         if (!wr_req) {
1879                 err = -ENOMEM;
1880                 goto out_unlock;
1881         }
1882
1883         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1884         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1885         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1886         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1887
1888         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1889         if (err)
1890                 goto out_unlock;
1891
1892         /* one page should be large enough for STAT data */
1893         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1894         if (IS_ERR(pages)) {
1895                 err = PTR_ERR(pages);
1896                 goto out_unlock;
1897         }
1898
1899         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1900                                      0, false, true);
1901         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1902
1903         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1904         wr_req->r_abort_on_full = true;
1905         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1906
1907         if (!err)
1908                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1909         if (!err2)
1910                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1911
1912         if (err >= 0 || err == -ENOENT)
1913                 have |= POOL_READ;
1914         else if (err != -EPERM)
1915                 goto out_unlock;
1916
1917         if (err2 == 0 || err2 == -EEXIST)
1918                 have |= POOL_WRITE;
1919         else if (err2 != -EPERM) {
1920                 err = err2;
1921                 goto out_unlock;
1922         }
1923
1924         pool_ns_len = pool_ns ? pool_ns->len : 0;
1925         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1926         if (!perm) {
1927                 err = -ENOMEM;
1928                 goto out_unlock;
1929         }
1930
1931         perm->pool = pool;
1932         perm->perm = have;
1933         perm->pool_ns_len = pool_ns_len;
1934         if (pool_ns_len > 0)
1935                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1936         perm->pool_ns[pool_ns_len] = 0;
1937
1938         rb_link_node(&perm->node, parent, p);
1939         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1940         err = 0;
1941 out_unlock:
1942         up_write(&mdsc->pool_perm_rwsem);
1943
1944         ceph_osdc_put_request(rd_req);
1945         ceph_osdc_put_request(wr_req);
1946 out:
1947         if (!err)
1948                 err = have;
1949         if (pool_ns)
1950                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1951                      pool, (int)pool_ns->len, pool_ns->str, err);
1952         else
1953                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1954         return err;
1955 }
1956
1957 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1958 {
1959         s64 pool;
1960         struct ceph_string *pool_ns;
1961         int ret, flags;
1962
1963         if (ci->i_vino.snap != CEPH_NOSNAP) {
1964                 /*
1965                  * Pool permission check needs to write to the first object.
1966                  * But for snapshot, head of the first object may have alread
1967                  * been deleted. Skip check to avoid creating orphan object.
1968                  */
1969                 return 0;
1970         }
1971
1972         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1973                                 NOPOOLPERM))
1974                 return 0;
1975
1976         spin_lock(&ci->i_ceph_lock);
1977         flags = ci->i_ceph_flags;
1978         pool = ci->i_layout.pool_id;
1979         spin_unlock(&ci->i_ceph_lock);
1980 check:
1981         if (flags & CEPH_I_POOL_PERM) {
1982                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1983                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1984                              pool);
1985                         return -EPERM;
1986                 }
1987                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1988                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1989                              pool);
1990                         return -EPERM;
1991                 }
1992                 return 0;
1993         }
1994
1995         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1996         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1997         ceph_put_string(pool_ns);
1998         if (ret < 0)
1999                 return ret;
2000
2001         flags = CEPH_I_POOL_PERM;
2002         if (ret & POOL_READ)
2003                 flags |= CEPH_I_POOL_RD;
2004         if (ret & POOL_WRITE)
2005                 flags |= CEPH_I_POOL_WR;
2006
2007         spin_lock(&ci->i_ceph_lock);
2008         if (pool == ci->i_layout.pool_id &&
2009             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2010                 ci->i_ceph_flags |= flags;
2011         } else {
2012                 pool = ci->i_layout.pool_id;
2013                 flags = ci->i_ceph_flags;
2014         }
2015         spin_unlock(&ci->i_ceph_lock);
2016         goto check;
2017 }
2018
2019 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2020 {
2021         struct ceph_pool_perm *perm;
2022         struct rb_node *n;
2023
2024         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2025                 n = rb_first(&mdsc->pool_perm_tree);
2026                 perm = rb_entry(n, struct ceph_pool_perm, node);
2027                 rb_erase(n, &mdsc->pool_perm_tree);
2028                 kfree(perm);
2029         }
2030 }