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         bool head_snapc;
 473 };
 474
 475 /*
 476  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 477  * only snap context we are allowed to write back.
 478  */
 479 static struct ceph_snap_context *
 480 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 481                    struct ceph_snap_context *page_snapc)
 482 {
 483         struct ceph_inode_info *ci = ceph_inode(inode);
 484         struct ceph_snap_context *snapc = NULL;
 485         struct ceph_cap_snap *capsnap = NULL;
 486
 487         spin_lock(&ci->i_ceph_lock);
 488         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 489                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 490                      capsnap->context, capsnap->dirty_pages);
 491                 if (!capsnap->dirty_pages)
 492                         continue;
 493
 494                 /* get i_size, truncate_{seq,size} for page_snapc? */
 495                 if (snapc && capsnap->context != page_snapc)
 496                         continue;
 497
 498                 if (ctl) {
 499                         if (capsnap->writing) {
 500                                 ctl->i_size = i_size_read(inode);
 501                                 ctl->size_stable = false;
 502                         } else {
 503                                 ctl->i_size = capsnap->size;
 504                                 ctl->size_stable = true;
 505                         }
 506                         ctl->truncate_size = capsnap->truncate_size;
 507                         ctl->truncate_seq = capsnap->truncate_seq;
 508                         ctl->head_snapc = false;
 509                 }
 510
 511                 if (snapc)
 512                         break;
 513
 514                 snapc = ceph_get_snap_context(capsnap->context);
 515                 if (!page_snapc ||
 516                     page_snapc == snapc ||
 517                     page_snapc->seq > snapc->seq)
 518                         break;
 519         }
 520         if (!snapc && ci->i_wrbuffer_ref_head) {
 521                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 522                 dout(" head snapc %p has %d dirty pages\n",
 523                      snapc, ci->i_wrbuffer_ref_head);
 524                 if (ctl) {
 525                         ctl->i_size = i_size_read(inode);
 526                         ctl->truncate_size = ci->i_truncate_size;
 527                         ctl->truncate_seq = ci->i_truncate_seq;
 528                         ctl->size_stable = false;
 529                         ctl->head_snapc = true;
 530                 }
 531         }
 532         spin_unlock(&ci->i_ceph_lock);
 533         return snapc;
 534 }
 535
 536 static u64 get_writepages_data_length(struct inode *inode,
 537                                       struct page *page, u64 start)
 538 {
 539         struct ceph_inode_info *ci = ceph_inode(inode);
 540         struct ceph_snap_context *snapc = page_snap_context(page);
 541         struct ceph_cap_snap *capsnap = NULL;
 542         u64 end = i_size_read(inode);
 543
 544         if (snapc != ci->i_head_snapc) {
 545                 bool found = false;
 546                 spin_lock(&ci->i_ceph_lock);
 547                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 548                         if (capsnap->context == snapc) {
 549                                 if (!capsnap->writing)
 550                                         end = capsnap->size;
 551                                 found = true;
 552                                 break;
 553                         }
 554                 }
 555                 spin_unlock(&ci->i_ceph_lock);
 556                 WARN_ON(!found);
 557         }
 558         if (end > page_offset(page) + PAGE_SIZE)
 559                 end = page_offset(page) + PAGE_SIZE;
 560         return end > start ? end - start : 0;
 561 }
 562
 563 /*
 564  * Write a single page, but leave the page locked.
 565  *
 566  * If we get a write error, set the page error bit, but still adjust the
 567  * dirty page accounting (i.e., page is no longer dirty).
 568  */
 569 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 570 {
 571         struct inode *inode;
 572         struct ceph_inode_info *ci;
 573         struct ceph_fs_client *fsc;
 574         struct ceph_snap_context *snapc, *oldest;
 575         loff_t page_off = page_offset(page);
 576         long writeback_stat;
 577         int err, len = PAGE_SIZE;
 578         struct ceph_writeback_ctl ceph_wbc;
 579
 580         dout("writepage %p idx %lu\n", page, page->index);
 581
 582         inode = page->mapping->host;
 583         ci = ceph_inode(inode);
 584         fsc = ceph_inode_to_client(inode);
 585
 586         /* verify this is a writeable snap context */
 587         snapc = page_snap_context(page);
 588         if (!snapc) {
 589                 dout("writepage %p page %p not dirty?\n", inode, page);
 590                 return 0;
 591         }
 592         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 593         if (snapc->seq > oldest->seq) {
 594                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 595                      inode, page, snapc);
 596                 /* we should only noop if called by kswapd */
 597                 WARN_ON(!(current->flags & PF_MEMALLOC));
 598                 ceph_put_snap_context(oldest);
 599                 redirty_page_for_writepage(wbc, page);
 600                 return 0;
 601         }
 602         ceph_put_snap_context(oldest);
 603
 604         /* is this a partial page at end of file? */
 605         if (page_off >= ceph_wbc.i_size) {
 606                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 607                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 608                 return 0;
 609         }
 610
 611         if (ceph_wbc.i_size < page_off + len)
 612                 len = ceph_wbc.i_size - page_off;
 613
 614         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 615              inode, page, page->index, page_off, len, snapc, snapc->seq);
 616
 617         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 618         if (writeback_stat >
 619             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 620                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 621
 622         set_page_writeback(page);
 623         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 624                                    &ci->i_layout, snapc, page_off, len,
 625                                    ceph_wbc.truncate_seq,
 626                                    ceph_wbc.truncate_size,
 627                                    &inode->i_mtime, &page, 1);
 628         if (err < 0) {
 629                 struct writeback_control tmp_wbc;
 630                 if (!wbc)
 631                         wbc = &tmp_wbc;
 632                 if (err == -ERESTARTSYS) {
 633                         /* killed by SIGKILL */
 634                         dout("writepage interrupted page %p\n", page);
 635                         redirty_page_for_writepage(wbc, page);
 636                         end_page_writeback(page);
 637                         return err;
 638                 }
 639                 dout("writepage setting page/mapping error %d %p\n",
 640                      err, page);
 641                 SetPageError(page);
 642                 mapping_set_error(&inode->i_data, err);
 643                 wbc->pages_skipped++;
 644         } else {
 645                 dout("writepage cleaned page %p\n", page);
 646                 err = 0;  /* vfs expects us to return 0 */
 647         }
 648         page->private = 0;
 649         ClearPagePrivate(page);
 650         end_page_writeback(page);
 651         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 652         ceph_put_snap_context(snapc);  /* page's reference */
 653         return err;
 654 }
 655
 656 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 657 {
 658         int err;
 659         struct inode *inode = page->mapping->host;
 660         BUG_ON(!inode);
 661         ihold(inode);
 662         err = writepage_nounlock(page, wbc);
 663         if (err == -ERESTARTSYS) {
 664                 /* direct memory reclaimer was killed by SIGKILL. return 0
 665                  * to prevent caller from setting mapping/page error */
 666                 err = 0;
 667         }
 668         unlock_page(page);
 669         iput(inode);
 670         return err;
 671 }
 672
 673 /*
 674  * lame release_pages helper.  release_pages() isn't exported to
 675  * modules.
 676  */
 677 static void ceph_release_pages(struct page **pages, int num)
 678 {
 679         struct pagevec pvec;
 680         int i;
 681
 682         pagevec_init(&pvec, 0);
 683         for (i = 0; i < num; i++) {
 684                 if (pagevec_add(&pvec, pages[i]) == 0)
 685                         pagevec_release(&pvec);
 686         }
 687         pagevec_release(&pvec);
 688 }
 689
 690 /*
 691  * async writeback completion handler.
 692  *
 693  * If we get an error, set the mapping error bit, but not the individual
 694  * page error bits.
 695  */
 696 static void writepages_finish(struct ceph_osd_request *req)
 697 {
 698         struct inode *inode = req->r_inode;
 699         struct ceph_inode_info *ci = ceph_inode(inode);
 700         struct ceph_osd_data *osd_data;
 701         struct page *page;
 702         int num_pages, total_pages = 0;
 703         int i, j;
 704         int rc = req->r_result;
 705         struct ceph_snap_context *snapc = req->r_snapc;
 706         struct address_space *mapping = inode->i_mapping;
 707         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 708         bool remove_page;
 709
 710         dout("writepages_finish %p rc %d\n", inode, rc);
 711         if (rc < 0) {
 712                 mapping_set_error(mapping, rc);
 713                 ceph_set_error_write(ci);
 714         } else {
 715                 ceph_clear_error_write(ci);
 716         }
 717
 718         /*
 719          * We lost the cache cap, need to truncate the page before
 720          * it is unlocked, otherwise we'd truncate it later in the
 721          * page truncation thread, possibly losing some data that
 722          * raced its way in
 723          */
 724         remove_page = !(ceph_caps_issued(ci) &
 725                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 726
 727         /* clean all pages */
 728         for (i = 0; i < req->r_num_ops; i++) {
 729                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 730                         break;
 731
 732                 osd_data = osd_req_op_extent_osd_data(req, i);
 733                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 734                 num_pages = calc_pages_for((u64)osd_data->alignment,
 735                                            (u64)osd_data->length);
 736                 total_pages += num_pages;
 737                 for (j = 0; j < num_pages; j++) {
 738                         page = osd_data->pages[j];
 739                         BUG_ON(!page);
 740                         WARN_ON(!PageUptodate(page));
 741
 742                         if (atomic_long_dec_return(&fsc->writeback_count) <
 743                              CONGESTION_OFF_THRESH(
 744                                         fsc->mount_options->congestion_kb))
 745                                 clear_bdi_congested(inode_to_bdi(inode),
 746                                                     BLK_RW_ASYNC);
 747
 748                         ceph_put_snap_context(page_snap_context(page));
 749                         page->private = 0;
 750                         ClearPagePrivate(page);
 751                         dout("unlocking %p\n", page);
 752                         end_page_writeback(page);
 753
 754                         if (remove_page)
 755                                 generic_error_remove_page(inode->i_mapping,
 756                                                           page);
 757
 758                         unlock_page(page);
 759                 }
 760                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 761                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 762
 763                 ceph_release_pages(osd_data->pages, num_pages);
 764         }
 765
 766         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 767
 768         osd_data = osd_req_op_extent_osd_data(req, 0);
 769         if (osd_data->pages_from_pool)
 770                 mempool_free(osd_data->pages,
 771                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 772         else
 773                 kfree(osd_data->pages);
 774         ceph_osdc_put_request(req);
 775 }
 776
 777 /*
 778  * initiate async writeback
 779  */
 780 static int ceph_writepages_start(struct address_space *mapping,
 781                                  struct writeback_control *wbc)
 782 {
 783         struct inode *inode = mapping->host;
 784         struct ceph_inode_info *ci = ceph_inode(inode);
 785         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 786         struct ceph_vino vino = ceph_vino(inode);
 787         pgoff_t index, start_index, end = -1;
 788         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 789         struct pagevec pvec;
 790         int rc = 0;
 791         unsigned int wsize = i_blocksize(inode);
 792         struct ceph_osd_request *req = NULL;
 793         struct ceph_writeback_ctl ceph_wbc;
 794         bool should_loop, range_whole = false;
 795         bool stop, done = false;
 796
 797         dout("writepages_start %p (mode=%s)\n", inode,
 798              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 799              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 800
 801         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 802                 if (ci->i_wrbuffer_ref > 0) {
 803                         pr_warn_ratelimited(
 804                                 "writepage_start %p %lld forced umount\n",
 805                                 inode, ceph_ino(inode));
 806                 }
 807                 mapping_set_error(mapping, -EIO);
 808                 return -EIO; /* we're in a forced umount, don't write! */
 809         }
 810         if (fsc->mount_options->wsize < wsize)
 811                 wsize = fsc->mount_options->wsize;
 812
 813         pagevec_init(&pvec, 0);
 814
 815         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 816         index = start_index;
 817
 818 retry:
 819         /* find oldest snap context with dirty data */
 820         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 821         if (!snapc) {
 822                 /* hmm, why does writepages get called when there
 823                    is no dirty data? */
 824                 dout(" no snap context with dirty data?\n");
 825                 goto out;
 826         }
 827         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 828              snapc, snapc->seq, snapc->num_snaps);
 829
 830         should_loop = false;
 831         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 832                 /* where to start/end? */
 833                 if (wbc->range_cyclic) {
 834                         index = start_index;
 835                         end = -1;
 836                         if (index > 0)
 837                                 should_loop = true;
 838                         dout(" cyclic, start at %lu\n", index);
 839                 } else {
 840                         index = wbc->range_start >> PAGE_SHIFT;
 841                         end = wbc->range_end >> PAGE_SHIFT;
 842                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 843                                 range_whole = true;
 844                         dout(" not cyclic, %lu to %lu\n", index, end);
 845                 }
 846         } else if (!ceph_wbc.head_snapc) {
 847                 /* Do not respect wbc->range_{start,end}. Dirty pages
 848                  * in that range can be associated with newer snapc.
 849                  * They are not writeable until we write all dirty pages
 850                  * associated with 'snapc' get written */
 851                 if (index > 0 || wbc->sync_mode != WB_SYNC_NONE)
 852                         should_loop = true;
 853                 dout(" non-head snapc, range whole\n");
 854         }
 855
 856         ceph_put_snap_context(last_snapc);
 857         last_snapc = snapc;
 858
 859         stop = false;
 860         while (!stop && index <= end) {
 861                 int num_ops = 0, op_idx;
 862                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 863                 struct page **pages = NULL, **data_pages;
 864                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 865                 struct page *page;
 866                 pgoff_t strip_unit_end = 0;
 867                 u64 offset = 0, len = 0;
 868
 869                 max_pages = wsize >> PAGE_SHIFT;
 870
 871 get_more_pages:
 872                 pvec_pages = min_t(unsigned, PAGEVEC_SIZE,
 873                                    max_pages - locked_pages);
 874                 if (end - index < (u64)(pvec_pages - 1))
 875                         pvec_pages = (unsigned)(end - index) + 1;
 876
 877                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 878                                                 PAGECACHE_TAG_DIRTY,
 879                                                 pvec_pages);
 880                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 881                 if (!pvec_pages && !locked_pages)
 882                         break;
 883                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 884                         page = pvec.pages[i];
 885                         dout("? %p idx %lu\n", page, page->index);
 886                         if (locked_pages == 0)
 887                                 lock_page(page);  /* first page */
 888                         else if (!trylock_page(page))
 889                                 break;
 890
 891                         /* only dirty pages, or our accounting breaks */
 892                         if (unlikely(!PageDirty(page)) ||
 893                             unlikely(page->mapping != mapping)) {
 894                                 dout("!dirty or !mapping %p\n", page);
 895                                 unlock_page(page);
 896                                 continue;
 897                         }
 898                         if (page->index > end) {
 899                                 dout("end of range %p\n", page);
 900                                 /* can't be range_cyclic (1st pass) because
 901                                  * end == -1 in that case. */
 902                                 stop = true;
 903                                 if (ceph_wbc.head_snapc)
 904                                         done = true;
 905                                 unlock_page(page);
 906                                 break;
 907                         }
 908                         if (strip_unit_end && (page->index > strip_unit_end)) {
 909                                 dout("end of strip unit %p\n", page);
 910                                 unlock_page(page);
 911                                 break;
 912                         }
 913                         if (page_offset(page) >= ceph_wbc.i_size) {
 914                                 dout("%p page eof %llu\n",
 915                                      page, ceph_wbc.i_size);
 916                                 /* not done if range_cyclic */
 917                                 stop = true;
 918                                 unlock_page(page);
 919                                 break;
 920                         }
 921                         if (PageWriteback(page)) {
 922                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 923                                         dout("%p under writeback\n", page);
 924                                         unlock_page(page);
 925                                         continue;
 926                                 }
 927                                 dout("waiting on writeback %p\n", page);
 928                                 wait_on_page_writeback(page);
 929                         }
 930
 931                         /* only if matching snap context */
 932                         pgsnapc = page_snap_context(page);
 933                         if (pgsnapc != snapc) {
 934                                 dout("page snapc %p %lld != oldest %p %lld\n",
 935                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 936                                 unlock_page(page);
 937                                 continue;
 938                         }
 939
 940                         if (!clear_page_dirty_for_io(page)) {
 941                                 dout("%p !clear_page_dirty_for_io\n", page);
 942                                 unlock_page(page);
 943                                 continue;
 944                         }
 945
 946                         /*
 947                          * We have something to write.  If this is
 948                          * the first locked page this time through,
 949                          * calculate max possinle write size and
 950                          * allocate a page array
 951                          */
 952                         if (locked_pages == 0) {
 953                                 u64 objnum;
 954                                 u64 objoff;
 955
 956                                 /* prepare async write request */
 957                                 offset = (u64)page_offset(page);
 958                                 len = wsize;
 959
 960                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 961                                                                 offset, len,
 962                                                                 &objnum, &objoff,
 963                                                                 &len);
 964                                 if (rc < 0) {
 965                                         unlock_page(page);
 966                                         break;
 967                                 }
 968
 969                                 num_ops = 1;
 970                                 strip_unit_end = page->index +
 971                                         ((len - 1) >> PAGE_SHIFT);
 972
 973                                 BUG_ON(pages);
 974                                 max_pages = calc_pages_for(0, (u64)len);
 975                                 pages = kmalloc(max_pages * sizeof (*pages),
 976                                                 GFP_NOFS);
 977                                 if (!pages) {
 978                                         pool = fsc->wb_pagevec_pool;
 979                                         pages = mempool_alloc(pool, GFP_NOFS);
 980                                         BUG_ON(!pages);
 981                                 }
 982
 983                                 len = 0;
 984                         } else if (page->index !=
 985                                    (offset + len) >> PAGE_SHIFT) {
 986                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 987                                                         CEPH_OSD_MAX_OPS)) {
 988                                         redirty_page_for_writepage(wbc, page);
 989                                         unlock_page(page);
 990                                         break;
 991                                 }
 992
 993                                 num_ops++;
 994                                 offset = (u64)page_offset(page);
 995                                 len = 0;
 996                         }
 997
 998                         /* note position of first page in pvec */
 999                         dout("%p will write page %p idx %lu\n",
1000                              inode, page, page->index);
1001
1002                         if (atomic_long_inc_return(&fsc->writeback_count) >
1003                             CONGESTION_ON_THRESH(
1004                                     fsc->mount_options->congestion_kb)) {
1005                                 set_bdi_congested(inode_to_bdi(inode),
1006                                                   BLK_RW_ASYNC);
1007                         }
1008
1009
1010                         pages[locked_pages++] = page;
1011                         pvec.pages[i] = NULL;
1012
1013                         len += PAGE_SIZE;
1014                 }
1015
1016                 /* did we get anything? */
1017                 if (!locked_pages)
1018                         goto release_pvec_pages;
1019                 if (i) {
1020                         unsigned j, n = 0;
1021                         /* shift unused page to beginning of pvec */
1022                         for (j = 0; j < pvec_pages; j++) {
1023                                 if (!pvec.pages[j])
1024                                         continue;
1025                                 if (n < j)
1026                                         pvec.pages[n] = pvec.pages[j];
1027                                 n++;
1028                         }
1029                         pvec.nr = n;
1030
1031                         if (pvec_pages && i == pvec_pages &&
1032                             locked_pages < max_pages) {
1033                                 dout("reached end pvec, trying for more\n");
1034                                 pagevec_release(&pvec);
1035                                 goto get_more_pages;
1036                         }
1037                 }
1038
1039 new_request:
1040                 offset = page_offset(pages[0]);
1041                 len = wsize;
1042
1043                 req = ceph_osdc_new_request(&fsc->client->osdc,
1044                                         &ci->i_layout, vino,
1045                                         offset, &len, 0, num_ops,
1046                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1047                                         snapc, ceph_wbc.truncate_seq,
1048                                         ceph_wbc.truncate_size, false);
1049                 if (IS_ERR(req)) {
1050                         req = ceph_osdc_new_request(&fsc->client->osdc,
1051                                                 &ci->i_layout, vino,
1052                                                 offset, &len, 0,
1053                                                 min(num_ops,
1054                                                     CEPH_OSD_SLAB_OPS),
1055                                                 CEPH_OSD_OP_WRITE,
1056                                                 CEPH_OSD_FLAG_WRITE,
1057                                                 snapc, ceph_wbc.truncate_seq,
1058                                                 ceph_wbc.truncate_size, true);
1059                         BUG_ON(IS_ERR(req));
1060                 }
1061                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1062                              PAGE_SIZE - offset);
1063
1064                 req->r_callback = writepages_finish;
1065                 req->r_inode = inode;
1066
1067                 /* Format the osd request message and submit the write */
1068                 len = 0;
1069                 data_pages = pages;
1070                 op_idx = 0;
1071                 for (i = 0; i < locked_pages; i++) {
1072                         u64 cur_offset = page_offset(pages[i]);
1073                         if (offset + len != cur_offset) {
1074                                 if (op_idx + 1 == req->r_num_ops)
1075                                         break;
1076                                 osd_req_op_extent_dup_last(req, op_idx,
1077                                                            cur_offset - offset);
1078                                 dout("writepages got pages at %llu~%llu\n",
1079                                      offset, len);
1080                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1081                                                         data_pages, len, 0,
1082                                                         !!pool, false);
1083                                 osd_req_op_extent_update(req, op_idx, len);
1084
1085                                 len = 0;
1086                                 offset = cur_offset;
1087                                 data_pages = pages + i;
1088                                 op_idx++;
1089                         }
1090
1091                         set_page_writeback(pages[i]);
1092                         len += PAGE_SIZE;
1093                 }
1094
1095                 if (ceph_wbc.size_stable) {
1096                         len = min(len, ceph_wbc.i_size - offset);
1097                 } else if (i == locked_pages) {
1098                         /* writepages_finish() clears writeback pages
1099                          * according to the data length, so make sure
1100                          * data length covers all locked pages */
1101                         u64 min_len = len + 1 - PAGE_SIZE;
1102                         len = get_writepages_data_length(inode, pages[i - 1],
1103                                                          offset);
1104                         len = max(len, min_len);
1105                 }
1106                 dout("writepages got pages at %llu~%llu\n", offset, len);
1107
1108                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1109                                                  0, !!pool, false);
1110                 osd_req_op_extent_update(req, op_idx, len);
1111
1112                 BUG_ON(op_idx + 1 != req->r_num_ops);
1113
1114                 pool = NULL;
1115                 if (i < locked_pages) {
1116                         BUG_ON(num_ops <= req->r_num_ops);
1117                         num_ops -= req->r_num_ops;
1118                         locked_pages -= i;
1119
1120                         /* allocate new pages array for next request */
1121                         data_pages = pages;
1122                         pages = kmalloc(locked_pages * sizeof (*pages),
1123                                         GFP_NOFS);
1124                         if (!pages) {
1125                                 pool = fsc->wb_pagevec_pool;
1126                                 pages = mempool_alloc(pool, GFP_NOFS);
1127                                 BUG_ON(!pages);
1128                         }
1129                         memcpy(pages, data_pages + i,
1130                                locked_pages * sizeof(*pages));
1131                         memset(data_pages + i, 0,
1132                                locked_pages * sizeof(*pages));
1133                 } else {
1134                         BUG_ON(num_ops != req->r_num_ops);
1135                         index = pages[i - 1]->index + 1;
1136                         /* request message now owns the pages array */
1137                         pages = NULL;
1138                 }
1139
1140                 req->r_mtime = inode->i_mtime;
1141                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1142                 BUG_ON(rc);
1143                 req = NULL;
1144
1145                 wbc->nr_to_write -= i;
1146                 if (pages)
1147                         goto new_request;
1148
1149                 /*
1150                  * We stop writing back only if we are not doing
1151                  * integrity sync. In case of integrity sync we have to
1152                  * keep going until we have written all the pages
1153                  * we tagged for writeback prior to entering this loop.
1154                  */
1155                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1156                         done = stop = true;
1157
1158 release_pvec_pages:
1159                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1160                      pvec.nr ? pvec.pages[0] : NULL);
1161                 pagevec_release(&pvec);
1162         }
1163
1164         if (should_loop && !done) {
1165                 /* more to do; loop back to beginning of file */
1166                 dout("writepages looping back to beginning of file\n");
1167                 end = start_index - 1; /* OK even when start_index == 0 */
1168
1169                 /* to write dirty pages associated with next snapc,
1170                  * we need to wait until current writes complete */
1171                 if (wbc->sync_mode != WB_SYNC_NONE &&
1172                     start_index == 0 && /* all dirty pages were checked */
1173                     !ceph_wbc.head_snapc) {
1174                         struct page *page;
1175                         unsigned i, nr;
1176                         index = 0;
1177                         while ((index <= end) &&
1178                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1179                                                         PAGECACHE_TAG_WRITEBACK,
1180                                                         PAGEVEC_SIZE))) {
1181                                 for (i = 0; i < nr; i++) {
1182                                         page = pvec.pages[i];
1183                                         if (page_snap_context(page) != snapc)
1184                                                 continue;
1185                                         wait_on_page_writeback(page);
1186                                 }
1187                                 pagevec_release(&pvec);
1188                                 cond_resched();
1189                         }
1190                 }
1191
1192                 start_index = 0;
1193                 index = 0;
1194                 goto retry;
1195         }
1196
1197         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1198                 mapping->writeback_index = index;
1199
1200 out:
1201         ceph_osdc_put_request(req);
1202         ceph_put_snap_context(last_snapc);
1203         dout("writepages dend - startone, rc = %d\n", rc);
1204         return rc;
1205 }
1206
1207
1208
1209 /*
1210  * See if a given @snapc is either writeable, or already written.
1211  */
1212 static int context_is_writeable_or_written(struct inode *inode,
1213                                            struct ceph_snap_context *snapc)
1214 {
1215         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1216         int ret = !oldest || snapc->seq <= oldest->seq;
1217
1218         ceph_put_snap_context(oldest);
1219         return ret;
1220 }
1221
1222 /*
1223  * We are only allowed to write into/dirty the page if the page is
1224  * clean, or already dirty within the same snap context.
1225  *
1226  * called with page locked.
1227  * return success with page locked,
1228  * or any failure (incl -EAGAIN) with page unlocked.
1229  */
1230 static int ceph_update_writeable_page(struct file *file,
1231                             loff_t pos, unsigned len,
1232                             struct page *page)
1233 {
1234         struct inode *inode = file_inode(file);
1235         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1236         struct ceph_inode_info *ci = ceph_inode(inode);
1237         loff_t page_off = pos & PAGE_MASK;
1238         int pos_in_page = pos & ~PAGE_MASK;
1239         int end_in_page = pos_in_page + len;
1240         loff_t i_size;
1241         int r;
1242         struct ceph_snap_context *snapc, *oldest;
1243
1244         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1245                 dout(" page %p forced umount\n", page);
1246                 unlock_page(page);
1247                 return -EIO;
1248         }
1249
1250 retry_locked:
1251         /* writepages currently holds page lock, but if we change that later, */
1252         wait_on_page_writeback(page);
1253
1254         snapc = page_snap_context(page);
1255         if (snapc && snapc != ci->i_head_snapc) {
1256                 /*
1257                  * this page is already dirty in another (older) snap
1258                  * context!  is it writeable now?
1259                  */
1260                 oldest = get_oldest_context(inode, NULL, NULL);
1261                 if (snapc->seq > oldest->seq) {
1262                         ceph_put_snap_context(oldest);
1263                         dout(" page %p snapc %p not current or oldest\n",
1264                              page, snapc);
1265                         /*
1266                          * queue for writeback, and wait for snapc to
1267                          * be writeable or written
1268                          */
1269                         snapc = ceph_get_snap_context(snapc);
1270                         unlock_page(page);
1271                         ceph_queue_writeback(inode);
1272                         r = wait_event_killable(ci->i_cap_wq,
1273                                context_is_writeable_or_written(inode, snapc));
1274                         ceph_put_snap_context(snapc);
1275                         if (r == -ERESTARTSYS)
1276                                 return r;
1277                         return -EAGAIN;
1278                 }
1279                 ceph_put_snap_context(oldest);
1280
1281                 /* yay, writeable, do it now (without dropping page lock) */
1282                 dout(" page %p snapc %p not current, but oldest\n",
1283                      page, snapc);
1284                 if (!clear_page_dirty_for_io(page))
1285                         goto retry_locked;
1286                 r = writepage_nounlock(page, NULL);
1287                 if (r < 0)
1288                         goto fail_unlock;
1289                 goto retry_locked;
1290         }
1291
1292         if (PageUptodate(page)) {
1293                 dout(" page %p already uptodate\n", page);
1294                 return 0;
1295         }
1296
1297         /* full page? */
1298         if (pos_in_page == 0 && len == PAGE_SIZE)
1299                 return 0;
1300
1301         /* past end of file? */
1302         i_size = i_size_read(inode);
1303
1304         if (page_off >= i_size ||
1305             (pos_in_page == 0 && (pos+len) >= i_size &&
1306              end_in_page - pos_in_page != PAGE_SIZE)) {
1307                 dout(" zeroing %p 0 - %d and %d - %d\n",
1308                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1309                 zero_user_segments(page,
1310                                    0, pos_in_page,
1311                                    end_in_page, PAGE_SIZE);
1312                 return 0;
1313         }
1314
1315         /* we need to read it. */
1316         r = ceph_do_readpage(file, page);
1317         if (r < 0) {
1318                 if (r == -EINPROGRESS)
1319                         return -EAGAIN;
1320                 goto fail_unlock;
1321         }
1322         goto retry_locked;
1323 fail_unlock:
1324         unlock_page(page);
1325         return r;
1326 }
1327
1328 /*
1329  * We are only allowed to write into/dirty the page if the page is
1330  * clean, or already dirty within the same snap context.
1331  */
1332 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1333                             loff_t pos, unsigned len, unsigned flags,
1334                             struct page **pagep, void **fsdata)
1335 {
1336         struct inode *inode = file_inode(file);
1337         struct page *page;
1338         pgoff_t index = pos >> PAGE_SHIFT;
1339         int r;
1340
1341         do {
1342                 /* get a page */
1343                 page = grab_cache_page_write_begin(mapping, index, 0);
1344                 if (!page)
1345                         return -ENOMEM;
1346
1347                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1348                      inode, page, (int)pos, (int)len);
1349
1350                 r = ceph_update_writeable_page(file, pos, len, page);
1351                 if (r < 0)
1352                         put_page(page);
1353                 else
1354                         *pagep = page;
1355         } while (r == -EAGAIN);
1356
1357         return r;
1358 }
1359
1360 /*
1361  * we don't do anything in here that simple_write_end doesn't do
1362  * except adjust dirty page accounting
1363  */
1364 static int ceph_write_end(struct file *file, struct address_space *mapping,
1365                           loff_t pos, unsigned len, unsigned copied,
1366                           struct page *page, void *fsdata)
1367 {
1368         struct inode *inode = file_inode(file);
1369         bool check_cap = false;
1370
1371         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1372              inode, page, (int)pos, (int)copied, (int)len);
1373
1374         /* zero the stale part of the page if we did a short copy */
1375         if (!PageUptodate(page)) {
1376                 if (copied < len) {
1377                         copied = 0;
1378                         goto out;
1379                 }
1380                 SetPageUptodate(page);
1381         }
1382
1383         /* did file size increase? */
1384         if (pos+copied > i_size_read(inode))
1385                 check_cap = ceph_inode_set_size(inode, pos+copied);
1386
1387         set_page_dirty(page);
1388
1389 out:
1390         unlock_page(page);
1391         put_page(page);
1392
1393         if (check_cap)
1394                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1395
1396         return copied;
1397 }
1398
1399 /*
1400  * we set .direct_IO to indicate direct io is supported, but since we
1401  * intercept O_DIRECT reads and writes early, this function should
1402  * never get called.
1403  */
1404 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1405 {
1406         WARN_ON(1);
1407         return -EINVAL;
1408 }
1409
1410 const struct address_space_operations ceph_aops = {
1411         .readpage = ceph_readpage,
1412         .readpages = ceph_readpages,
1413         .writepage = ceph_writepage,
1414         .writepages = ceph_writepages_start,
1415         .write_begin = ceph_write_begin,
1416         .write_end = ceph_write_end,
1417         .set_page_dirty = ceph_set_page_dirty,
1418         .invalidatepage = ceph_invalidatepage,
1419         .releasepage = ceph_releasepage,
1420         .direct_IO = ceph_direct_io,
1421 };
1422
1423 static void ceph_block_sigs(sigset_t *oldset)
1424 {
1425         sigset_t mask;
1426         siginitsetinv(&mask, sigmask(SIGKILL));
1427         sigprocmask(SIG_BLOCK, &mask, oldset);
1428 }
1429
1430 static void ceph_restore_sigs(sigset_t *oldset)
1431 {
1432         sigprocmask(SIG_SETMASK, oldset, NULL);
1433 }
1434
1435 /*
1436  * vm ops
1437  */
1438 static int ceph_filemap_fault(struct vm_fault *vmf)
1439 {
1440         struct vm_area_struct *vma = vmf->vma;
1441         struct inode *inode = file_inode(vma->vm_file);
1442         struct ceph_inode_info *ci = ceph_inode(inode);
1443         struct ceph_file_info *fi = vma->vm_file->private_data;
1444         struct page *pinned_page = NULL;
1445         loff_t off = vmf->pgoff << PAGE_SHIFT;
1446         int want, got, ret;
1447         sigset_t oldset;
1448
1449         ceph_block_sigs(&oldset);
1450
1451         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1452              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1453         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1454                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1455         else
1456                 want = CEPH_CAP_FILE_CACHE;
1457
1458         got = 0;
1459         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1460         if (ret < 0)
1461                 goto out_restore;
1462
1463         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1464              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1465
1466         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1467             ci->i_inline_version == CEPH_INLINE_NONE) {
1468                 current->journal_info = vma->vm_file;
1469                 ret = filemap_fault(vmf);
1470                 current->journal_info = NULL;
1471         } else
1472                 ret = -EAGAIN;
1473
1474         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1475              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1476         if (pinned_page)
1477                 put_page(pinned_page);
1478         ceph_put_cap_refs(ci, got);
1479
1480         if (ret != -EAGAIN)
1481                 goto out_restore;
1482
1483         /* read inline data */
1484         if (off >= PAGE_SIZE) {
1485                 /* does not support inline data > PAGE_SIZE */
1486                 ret = VM_FAULT_SIGBUS;
1487         } else {
1488                 int ret1;
1489                 struct address_space *mapping = inode->i_mapping;
1490                 struct page *page = find_or_create_page(mapping, 0,
1491                                                 mapping_gfp_constraint(mapping,
1492                                                 ~__GFP_FS));
1493                 if (!page) {
1494                         ret = VM_FAULT_OOM;
1495                         goto out_inline;
1496                 }
1497                 ret1 = __ceph_do_getattr(inode, page,
1498                                          CEPH_STAT_CAP_INLINE_DATA, true);
1499                 if (ret1 < 0 || off >= i_size_read(inode)) {
1500                         unlock_page(page);
1501                         put_page(page);
1502                         if (ret1 < 0)
1503                                 ret = ret1;
1504                         else
1505                                 ret = VM_FAULT_SIGBUS;
1506                         goto out_inline;
1507                 }
1508                 if (ret1 < PAGE_SIZE)
1509                         zero_user_segment(page, ret1, PAGE_SIZE);
1510                 else
1511                         flush_dcache_page(page);
1512                 SetPageUptodate(page);
1513                 vmf->page = page;
1514                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1515 out_inline:
1516                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1517                      inode, off, (size_t)PAGE_SIZE, ret);
1518         }
1519 out_restore:
1520         ceph_restore_sigs(&oldset);
1521         if (ret < 0)
1522                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1523
1524         return ret;
1525 }
1526
1527 /*
1528  * Reuse write_begin here for simplicity.
1529  */
1530 static int ceph_page_mkwrite(struct vm_fault *vmf)
1531 {
1532         struct vm_area_struct *vma = vmf->vma;
1533         struct inode *inode = file_inode(vma->vm_file);
1534         struct ceph_inode_info *ci = ceph_inode(inode);
1535         struct ceph_file_info *fi = vma->vm_file->private_data;
1536         struct ceph_cap_flush *prealloc_cf;
1537         struct page *page = vmf->page;
1538         loff_t off = page_offset(page);
1539         loff_t size = i_size_read(inode);
1540         size_t len;
1541         int want, got, ret;
1542         sigset_t oldset;
1543
1544         prealloc_cf = ceph_alloc_cap_flush();
1545         if (!prealloc_cf)
1546                 return VM_FAULT_OOM;
1547
1548         ceph_block_sigs(&oldset);
1549
1550         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1551                 struct page *locked_page = NULL;
1552                 if (off == 0) {
1553                         lock_page(page);
1554                         locked_page = page;
1555                 }
1556                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1557                 if (locked_page)
1558                         unlock_page(locked_page);
1559                 if (ret < 0)
1560                         goto out_free;
1561         }
1562
1563         if (off + PAGE_SIZE <= size)
1564                 len = PAGE_SIZE;
1565         else
1566                 len = size & ~PAGE_MASK;
1567
1568         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1569              inode, ceph_vinop(inode), off, len, size);
1570         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1571                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1572         else
1573                 want = CEPH_CAP_FILE_BUFFER;
1574
1575         got = 0;
1576         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1577                             &got, NULL);
1578         if (ret < 0)
1579                 goto out_free;
1580
1581         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1582              inode, off, len, ceph_cap_string(got));
1583
1584         /* Update time before taking page lock */
1585         file_update_time(vma->vm_file);
1586
1587         do {
1588                 lock_page(page);
1589
1590                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1591                         unlock_page(page);
1592                         ret = VM_FAULT_NOPAGE;
1593                         break;
1594                 }
1595
1596                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1597                 if (ret >= 0) {
1598                         /* success.  we'll keep the page locked. */
1599                         set_page_dirty(page);
1600                         ret = VM_FAULT_LOCKED;
1601                 }
1602         } while (ret == -EAGAIN);
1603
1604         if (ret == VM_FAULT_LOCKED ||
1605             ci->i_inline_version != CEPH_INLINE_NONE) {
1606                 int dirty;
1607                 spin_lock(&ci->i_ceph_lock);
1608                 ci->i_inline_version = CEPH_INLINE_NONE;
1609                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1610                                                &prealloc_cf);
1611                 spin_unlock(&ci->i_ceph_lock);
1612                 if (dirty)
1613                         __mark_inode_dirty(inode, dirty);
1614         }
1615
1616         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1617              inode, off, len, ceph_cap_string(got), ret);
1618         ceph_put_cap_refs(ci, got);
1619 out_free:
1620         ceph_restore_sigs(&oldset);
1621         ceph_free_cap_flush(prealloc_cf);
1622         if (ret < 0)
1623                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1624         return ret;
1625 }
1626
1627 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1628                            char *data, size_t len)
1629 {
1630         struct address_space *mapping = inode->i_mapping;
1631         struct page *page;
1632
1633         if (locked_page) {
1634                 page = locked_page;
1635         } else {
1636                 if (i_size_read(inode) == 0)
1637                         return;
1638                 page = find_or_create_page(mapping, 0,
1639                                            mapping_gfp_constraint(mapping,
1640                                            ~__GFP_FS));
1641                 if (!page)
1642                         return;
1643                 if (PageUptodate(page)) {
1644                         unlock_page(page);
1645                         put_page(page);
1646                         return;
1647                 }
1648         }
1649
1650         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1651              inode, ceph_vinop(inode), len, locked_page);
1652
1653         if (len > 0) {
1654                 void *kaddr = kmap_atomic(page);
1655                 memcpy(kaddr, data, len);
1656                 kunmap_atomic(kaddr);
1657         }
1658
1659         if (page != locked_page) {
1660                 if (len < PAGE_SIZE)
1661                         zero_user_segment(page, len, PAGE_SIZE);
1662                 else
1663                         flush_dcache_page(page);
1664
1665                 SetPageUptodate(page);
1666                 unlock_page(page);
1667                 put_page(page);
1668         }
1669 }
1670
1671 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1672 {
1673         struct inode *inode = file_inode(filp);
1674         struct ceph_inode_info *ci = ceph_inode(inode);
1675         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1676         struct ceph_osd_request *req;
1677         struct page *page = NULL;
1678         u64 len, inline_version;
1679         int err = 0;
1680         bool from_pagecache = false;
1681
1682         spin_lock(&ci->i_ceph_lock);
1683         inline_version = ci->i_inline_version;
1684         spin_unlock(&ci->i_ceph_lock);
1685
1686         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1687              inode, ceph_vinop(inode), inline_version);
1688
1689         if (inline_version == 1 || /* initial version, no data */
1690             inline_version == CEPH_INLINE_NONE)
1691                 goto out;
1692
1693         if (locked_page) {
1694                 page = locked_page;
1695                 WARN_ON(!PageUptodate(page));
1696         } else if (ceph_caps_issued(ci) &
1697                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1698                 page = find_get_page(inode->i_mapping, 0);
1699                 if (page) {
1700                         if (PageUptodate(page)) {
1701                                 from_pagecache = true;
1702                                 lock_page(page);
1703                         } else {
1704                                 put_page(page);
1705                                 page = NULL;
1706                         }
1707                 }
1708         }
1709
1710         if (page) {
1711                 len = i_size_read(inode);
1712                 if (len > PAGE_SIZE)
1713                         len = PAGE_SIZE;
1714         } else {
1715                 page = __page_cache_alloc(GFP_NOFS);
1716                 if (!page) {
1717                         err = -ENOMEM;
1718                         goto out;
1719                 }
1720                 err = __ceph_do_getattr(inode, page,
1721                                         CEPH_STAT_CAP_INLINE_DATA, true);
1722                 if (err < 0) {
1723                         /* no inline data */
1724                         if (err == -ENODATA)
1725                                 err = 0;
1726                         goto out;
1727                 }
1728                 len = err;
1729         }
1730
1731         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1732                                     ceph_vino(inode), 0, &len, 0, 1,
1733                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1734                                     NULL, 0, 0, false);
1735         if (IS_ERR(req)) {
1736                 err = PTR_ERR(req);
1737                 goto out;
1738         }
1739
1740         req->r_mtime = inode->i_mtime;
1741         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1742         if (!err)
1743                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1744         ceph_osdc_put_request(req);
1745         if (err < 0)
1746                 goto out;
1747
1748         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1749                                     ceph_vino(inode), 0, &len, 1, 3,
1750                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1751                                     NULL, ci->i_truncate_seq,
1752                                     ci->i_truncate_size, false);
1753         if (IS_ERR(req)) {
1754                 err = PTR_ERR(req);
1755                 goto out;
1756         }
1757
1758         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1759
1760         {
1761                 __le64 xattr_buf = cpu_to_le64(inline_version);
1762                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1763                                             "inline_version", &xattr_buf,
1764                                             sizeof(xattr_buf),
1765                                             CEPH_OSD_CMPXATTR_OP_GT,
1766                                             CEPH_OSD_CMPXATTR_MODE_U64);
1767                 if (err)
1768                         goto out_put;
1769         }
1770
1771         {
1772                 char xattr_buf[32];
1773                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1774                                          "%llu", inline_version);
1775                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1776                                             "inline_version",
1777                                             xattr_buf, xattr_len, 0, 0);
1778                 if (err)
1779                         goto out_put;
1780         }
1781
1782         req->r_mtime = inode->i_mtime;
1783         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1784         if (!err)
1785                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1786 out_put:
1787         ceph_osdc_put_request(req);
1788         if (err == -ECANCELED)
1789                 err = 0;
1790 out:
1791         if (page && page != locked_page) {
1792                 if (from_pagecache) {
1793                         unlock_page(page);
1794                         put_page(page);
1795                 } else
1796                         __free_pages(page, 0);
1797         }
1798
1799         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1800              inode, ceph_vinop(inode), inline_version, err);
1801         return err;
1802 }
1803
1804 static const struct vm_operations_struct ceph_vmops = {
1805         .fault          = ceph_filemap_fault,
1806         .page_mkwrite   = ceph_page_mkwrite,
1807 };
1808
1809 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1810 {
1811         struct address_space *mapping = file->f_mapping;
1812
1813         if (!mapping->a_ops->readpage)
1814                 return -ENOEXEC;
1815         file_accessed(file);
1816         vma->vm_ops = &ceph_vmops;
1817         return 0;
1818 }
1819
1820 enum {
1821         POOL_READ       = 1,
1822         POOL_WRITE      = 2,
1823 };
1824
1825 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1826                                 s64 pool, struct ceph_string *pool_ns)
1827 {
1828         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1829         struct ceph_mds_client *mdsc = fsc->mdsc;
1830         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1831         struct rb_node **p, *parent;
1832         struct ceph_pool_perm *perm;
1833         struct page **pages;
1834         size_t pool_ns_len;
1835         int err = 0, err2 = 0, have = 0;
1836
1837         down_read(&mdsc->pool_perm_rwsem);
1838         p = &mdsc->pool_perm_tree.rb_node;
1839         while (*p) {
1840                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1841                 if (pool < perm->pool)
1842                         p = &(*p)->rb_left;
1843                 else if (pool > perm->pool)
1844                         p = &(*p)->rb_right;
1845                 else {
1846                         int ret = ceph_compare_string(pool_ns,
1847                                                 perm->pool_ns,
1848                                                 perm->pool_ns_len);
1849                         if (ret < 0)
1850                                 p = &(*p)->rb_left;
1851                         else if (ret > 0)
1852                                 p = &(*p)->rb_right;
1853                         else {
1854                                 have = perm->perm;
1855                                 break;
1856                         }
1857                 }
1858         }
1859         up_read(&mdsc->pool_perm_rwsem);
1860         if (*p)
1861                 goto out;
1862
1863         if (pool_ns)
1864                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1865                      pool, (int)pool_ns->len, pool_ns->str);
1866         else
1867                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1868
1869         down_write(&mdsc->pool_perm_rwsem);
1870         p = &mdsc->pool_perm_tree.rb_node;
1871         parent = NULL;
1872         while (*p) {
1873                 parent = *p;
1874                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1875                 if (pool < perm->pool)
1876                         p = &(*p)->rb_left;
1877                 else if (pool > perm->pool)
1878                         p = &(*p)->rb_right;
1879                 else {
1880                         int ret = ceph_compare_string(pool_ns,
1881                                                 perm->pool_ns,
1882                                                 perm->pool_ns_len);
1883                         if (ret < 0)
1884                                 p = &(*p)->rb_left;
1885                         else if (ret > 0)
1886                                 p = &(*p)->rb_right;
1887                         else {
1888                                 have = perm->perm;
1889                                 break;
1890                         }
1891                 }
1892         }
1893         if (*p) {
1894                 up_write(&mdsc->pool_perm_rwsem);
1895                 goto out;
1896         }
1897
1898         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1899                                          1, false, GFP_NOFS);
1900         if (!rd_req) {
1901                 err = -ENOMEM;
1902                 goto out_unlock;
1903         }
1904
1905         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1906         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1907         rd_req->r_base_oloc.pool = pool;
1908         if (pool_ns)
1909                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1910         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1911
1912         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1913         if (err)
1914                 goto out_unlock;
1915
1916         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1917                                          1, false, GFP_NOFS);
1918         if (!wr_req) {
1919                 err = -ENOMEM;
1920                 goto out_unlock;
1921         }
1922
1923         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1924         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1925         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1926         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1927
1928         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1929         if (err)
1930                 goto out_unlock;
1931
1932         /* one page should be large enough for STAT data */
1933         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1934         if (IS_ERR(pages)) {
1935                 err = PTR_ERR(pages);
1936                 goto out_unlock;
1937         }
1938
1939         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1940                                      0, false, true);
1941         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1942
1943         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1944         wr_req->r_abort_on_full = true;
1945         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1946
1947         if (!err)
1948                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1949         if (!err2)
1950                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1951
1952         if (err >= 0 || err == -ENOENT)
1953                 have |= POOL_READ;
1954         else if (err != -EPERM)
1955                 goto out_unlock;
1956
1957         if (err2 == 0 || err2 == -EEXIST)
1958                 have |= POOL_WRITE;
1959         else if (err2 != -EPERM) {
1960                 err = err2;
1961                 goto out_unlock;
1962         }
1963
1964         pool_ns_len = pool_ns ? pool_ns->len : 0;
1965         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1966         if (!perm) {
1967                 err = -ENOMEM;
1968                 goto out_unlock;
1969         }
1970
1971         perm->pool = pool;
1972         perm->perm = have;
1973         perm->pool_ns_len = pool_ns_len;
1974         if (pool_ns_len > 0)
1975                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1976         perm->pool_ns[pool_ns_len] = 0;
1977
1978         rb_link_node(&perm->node, parent, p);
1979         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1980         err = 0;
1981 out_unlock:
1982         up_write(&mdsc->pool_perm_rwsem);
1983
1984         ceph_osdc_put_request(rd_req);
1985         ceph_osdc_put_request(wr_req);
1986 out:
1987         if (!err)
1988                 err = have;
1989         if (pool_ns)
1990                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1991                      pool, (int)pool_ns->len, pool_ns->str, err);
1992         else
1993                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1994         return err;
1995 }
1996
1997 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1998 {
1999         s64 pool;
2000         struct ceph_string *pool_ns;
2001         int ret, flags;
2002
2003         if (ci->i_vino.snap != CEPH_NOSNAP) {
2004                 /*
2005                  * Pool permission check needs to write to the first object.
2006                  * But for snapshot, head of the first object may have alread
2007                  * been deleted. Skip check to avoid creating orphan object.
2008                  */
2009                 return 0;
2010         }
2011
2012         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2013                                 NOPOOLPERM))
2014                 return 0;
2015
2016         spin_lock(&ci->i_ceph_lock);
2017         flags = ci->i_ceph_flags;
2018         pool = ci->i_layout.pool_id;
2019         spin_unlock(&ci->i_ceph_lock);
2020 check:
2021         if (flags & CEPH_I_POOL_PERM) {
2022                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2023                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2024                              pool);
2025                         return -EPERM;
2026                 }
2027                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2028                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2029                              pool);
2030                         return -EPERM;
2031                 }
2032                 return 0;
2033         }
2034
2035         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2036         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2037         ceph_put_string(pool_ns);
2038         if (ret < 0)
2039                 return ret;
2040
2041         flags = CEPH_I_POOL_PERM;
2042         if (ret & POOL_READ)
2043                 flags |= CEPH_I_POOL_RD;
2044         if (ret & POOL_WRITE)
2045                 flags |= CEPH_I_POOL_WR;
2046
2047         spin_lock(&ci->i_ceph_lock);
2048         if (pool == ci->i_layout.pool_id &&
2049             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2050                 ci->i_ceph_flags |= flags;
2051         } else {
2052                 pool = ci->i_layout.pool_id;
2053                 flags = ci->i_ceph_flags;
2054         }
2055         spin_unlock(&ci->i_ceph_lock);
2056         goto check;
2057 }
2058
2059 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2060 {
2061         struct ceph_pool_perm *perm;
2062         struct rb_node *n;
2063
2064         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2065                 n = rb_first(&mdsc->pool_perm_tree);
2066                 perm = rb_entry(n, struct ceph_pool_perm, node);
2067                 rb_erase(n, &mdsc->pool_perm_tree);
2068                 kfree(perm);
2069         }
2070 }