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