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
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75
  76         if (PageDirty(page)) {
  77                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  78                      mapping->host, page, page->index);
  79                 BUG_ON(!PagePrivate(page));
  80                 return 0;
  81         }
  82
  83         inode = mapping->host;
  84         ci = ceph_inode(inode);
  85
  86         /* dirty the head */
  87         spin_lock(&ci->i_ceph_lock);
  88         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  89         if (__ceph_have_pending_cap_snap(ci)) {
  90                 struct ceph_cap_snap *capsnap =
  91                                 list_last_entry(&ci->i_cap_snaps,
  92                                                 struct ceph_cap_snap,
  93                                                 ci_item);
  94                 snapc = ceph_get_snap_context(capsnap->context);
  95                 capsnap->dirty_pages++;
  96         } else {
  97                 BUG_ON(!ci->i_head_snapc);
  98                 snapc = ceph_get_snap_context(ci->i_head_snapc);
  99                 ++ci->i_wrbuffer_ref_head;
 100         }
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         return __set_page_dirty_nobuffers(page);
 121 }
 122
 123 /*
 124  * If we are truncating the full page (i.e. offset == 0), adjust the
 125  * dirty page counters appropriately.  Only called if there is private
 126  * data on the page.
 127  */
 128 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 129                                 unsigned int length)
 130 {
 131         struct inode *inode;
 132         struct ceph_inode_info *ci;
 133         struct ceph_snap_context *snapc = page_snap_context(page);
 134
 135         inode = page->mapping->host;
 136         ci = ceph_inode(inode);
 137
 138         if (offset != 0 || length != PAGE_SIZE) {
 139                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 140                      inode, page, page->index, offset, length);
 141                 return;
 142         }
 143
 144         ceph_invalidate_fscache_page(inode, page);
 145
 146         if (!PagePrivate(page))
 147                 return;
 148
 149         /*
 150          * We can get non-dirty pages here due to races between
 151          * set_page_dirty and truncate_complete_page; just spit out a
 152          * warning, in case we end up with accounting problems later.
 153          */
 154         if (!PageDirty(page))
 155                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 156
 157         ClearPageChecked(page);
 158
 159         dout("%p invalidatepage %p idx %lu full dirty page\n",
 160              inode, page, page->index);
 161
 162         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 163         ceph_put_snap_context(snapc);
 164         page->private = 0;
 165         ClearPagePrivate(page);
 166 }
 167
 168 static int ceph_releasepage(struct page *page, gfp_t g)
 169 {
 170         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 171              page, page->index, PageDirty(page) ? "" : "not ");
 172
 173         /* Can we release the page from the cache? */
 174         if (!ceph_release_fscache_page(page, g))
 175                 return 0;
 176
 177         return !PagePrivate(page);
 178 }
 179
 180 /*
 181  * read a single page, without unlocking it.
 182  */
 183 static int ceph_do_readpage(struct file *filp, struct page *page)
 184 {
 185         struct inode *inode = file_inode(filp);
 186         struct ceph_inode_info *ci = ceph_inode(inode);
 187         struct ceph_osd_client *osdc =
 188                 &ceph_inode_to_client(inode)->client->osdc;
 189         int err = 0;
 190         u64 off = page_offset(page);
 191         u64 len = PAGE_SIZE;
 192
 193         if (off >= i_size_read(inode)) {
 194                 zero_user_segment(page, 0, PAGE_SIZE);
 195                 SetPageUptodate(page);
 196                 return 0;
 197         }
 198
 199         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 200                 /*
 201                  * Uptodate inline data should have been added
 202                  * into page cache while getting Fcr caps.
 203                  */
 204                 if (off == 0)
 205                         return -EINVAL;
 206                 zero_user_segment(page, 0, PAGE_SIZE);
 207                 SetPageUptodate(page);
 208                 return 0;
 209         }
 210
 211         err = ceph_readpage_from_fscache(inode, page);
 212         if (err == 0)
 213                 return -EINPROGRESS;
 214
 215         dout("readpage inode %p file %p page %p index %lu\n",
 216              inode, filp, page, page->index);
 217         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 218                                   off, &len,
 219                                   ci->i_truncate_seq, ci->i_truncate_size,
 220                                   &page, 1, 0);
 221         if (err == -ENOENT)
 222                 err = 0;
 223         if (err < 0) {
 224                 SetPageError(page);
 225                 ceph_fscache_readpage_cancel(inode, page);
 226                 goto out;
 227         }
 228         if (err < PAGE_SIZE)
 229                 /* zero fill remainder of page */
 230                 zero_user_segment(page, err, PAGE_SIZE);
 231         else
 232                 flush_dcache_page(page);
 233
 234         SetPageUptodate(page);
 235         ceph_readpage_to_fscache(inode, page);
 236
 237 out:
 238         return err < 0 ? err : 0;
 239 }
 240
 241 static int ceph_readpage(struct file *filp, struct page *page)
 242 {
 243         int r = ceph_do_readpage(filp, page);
 244         if (r != -EINPROGRESS)
 245                 unlock_page(page);
 246         else
 247                 r = 0;
 248         return r;
 249 }
 250
 251 /*
 252  * Finish an async read(ahead) op.
 253  */
 254 static void finish_read(struct ceph_osd_request *req)
 255 {
 256         struct inode *inode = req->r_inode;
 257         struct ceph_osd_data *osd_data;
 258         int rc = req->r_result <= 0 ? req->r_result : 0;
 259         int bytes = req->r_result >= 0 ? req->r_result : 0;
 260         int num_pages;
 261         int i;
 262
 263         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 264
 265         /* unlock all pages, zeroing any data we didn't read */
 266         osd_data = osd_req_op_extent_osd_data(req, 0);
 267         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 268         num_pages = calc_pages_for((u64)osd_data->alignment,
 269                                         (u64)osd_data->length);
 270         for (i = 0; i < num_pages; i++) {
 271                 struct page *page = osd_data->pages[i];
 272
 273                 if (rc < 0 && rc != -ENOENT) {
 274                         ceph_fscache_readpage_cancel(inode, page);
 275                         goto unlock;
 276                 }
 277                 if (bytes < (int)PAGE_SIZE) {
 278                         /* zero (remainder of) page */
 279                         int s = bytes < 0 ? 0 : bytes;
 280                         zero_user_segment(page, s, PAGE_SIZE);
 281                 }
 282                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 283                      page->index);
 284                 flush_dcache_page(page);
 285                 SetPageUptodate(page);
 286                 ceph_readpage_to_fscache(inode, page);
 287 unlock:
 288                 unlock_page(page);
 289                 put_page(page);
 290                 bytes -= PAGE_SIZE;
 291         }
 292         kfree(osd_data->pages);
 293 }
 294
 295 /*
 296  * start an async read(ahead) operation.  return nr_pages we submitted
 297  * a read for on success, or negative error code.
 298  */
 299 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 300 {
 301         struct ceph_osd_client *osdc =
 302                 &ceph_inode_to_client(inode)->client->osdc;
 303         struct ceph_inode_info *ci = ceph_inode(inode);
 304         struct page *page = list_entry(page_list->prev, struct page, lru);
 305         struct ceph_vino vino;
 306         struct ceph_osd_request *req;
 307         u64 off;
 308         u64 len;
 309         int i;
 310         struct page **pages;
 311         pgoff_t next_index;
 312         int nr_pages = 0;
 313         int got = 0;
 314         int ret = 0;
 315
 316         if (!current->journal_info) {
 317                 /* caller of readpages does not hold buffer and read caps
 318                  * (fadvise, madvise and readahead cases) */
 319                 int want = CEPH_CAP_FILE_CACHE;
 320                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 321                 if (ret < 0) {
 322                         dout("start_read %p, error getting cap\n", inode);
 323                 } else if (!(got & want)) {
 324                         dout("start_read %p, no cache cap\n", inode);
 325                         ret = 0;
 326                 }
 327                 if (ret <= 0) {
 328                         if (got)
 329                                 ceph_put_cap_refs(ci, got);
 330                         while (!list_empty(page_list)) {
 331                                 page = list_entry(page_list->prev,
 332                                                   struct page, lru);
 333                                 list_del(&page->lru);
 334                                 put_page(page);
 335                         }
 336                         return ret;
 337                 }
 338         }
 339
 340         off = (u64) page_offset(page);
 341
 342         /* count pages */
 343         next_index = page->index;
 344         list_for_each_entry_reverse(page, page_list, lru) {
 345                 if (page->index != next_index)
 346                         break;
 347                 nr_pages++;
 348                 next_index++;
 349                 if (max && nr_pages == max)
 350                         break;
 351         }
 352         len = nr_pages << PAGE_SHIFT;
 353         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 354              off, len);
 355         vino = ceph_vino(inode);
 356         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 357                                     0, 1, CEPH_OSD_OP_READ,
 358                                     CEPH_OSD_FLAG_READ, NULL,
 359                                     ci->i_truncate_seq, ci->i_truncate_size,
 360                                     false);
 361         if (IS_ERR(req)) {
 362                 ret = PTR_ERR(req);
 363                 goto out;
 364         }
 365
 366         /* build page vector */
 367         nr_pages = calc_pages_for(0, len);
 368         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 369         if (!pages) {
 370                 ret = -ENOMEM;
 371                 goto out_put;
 372         }
 373         for (i = 0; i < nr_pages; ++i) {
 374                 page = list_entry(page_list->prev, struct page, lru);
 375                 BUG_ON(PageLocked(page));
 376                 list_del(&page->lru);
 377
 378                 dout("start_read %p adding %p idx %lu\n", inode, page,
 379                      page->index);
 380                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 381                                           GFP_KERNEL)) {
 382                         ceph_fscache_uncache_page(inode, page);
 383                         put_page(page);
 384                         dout("start_read %p add_to_page_cache failed %p\n",
 385                              inode, page);
 386                         nr_pages = i;
 387                         if (nr_pages > 0) {
 388                                 len = nr_pages << PAGE_SHIFT;
 389                                 osd_req_op_extent_update(req, 0, len);
 390                                 break;
 391                         }
 392                         goto out_pages;
 393                 }
 394                 pages[i] = page;
 395         }
 396         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 397         req->r_callback = finish_read;
 398         req->r_inode = inode;
 399
 400         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 401         ret = ceph_osdc_start_request(osdc, req, false);
 402         if (ret < 0)
 403                 goto out_pages;
 404         ceph_osdc_put_request(req);
 405
 406         /* After adding locked pages to page cache, the inode holds cache cap.
 407          * So we can drop our cap refs. */
 408         if (got)
 409                 ceph_put_cap_refs(ci, got);
 410
 411         return nr_pages;
 412
 413 out_pages:
 414         for (i = 0; i < nr_pages; ++i) {
 415                 ceph_fscache_readpage_cancel(inode, pages[i]);
 416                 unlock_page(pages[i]);
 417         }
 418         ceph_put_page_vector(pages, nr_pages, false);
 419 out_put:
 420         ceph_osdc_put_request(req);
 421 out:
 422         if (got)
 423                 ceph_put_cap_refs(ci, got);
 424         return ret;
 425 }
 426
 427
 428 /*
 429  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 430  * the caller (VM) cleans them up.
 431  */
 432 static int ceph_readpages(struct file *file, struct address_space *mapping,
 433                           struct list_head *page_list, unsigned nr_pages)
 434 {
 435         struct inode *inode = file_inode(file);
 436         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 437         int rc = 0;
 438         int max = 0;
 439
 440         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 441                 return -EINVAL;
 442
 443         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 444                                          &nr_pages);
 445
 446         if (rc == 0)
 447                 goto out;
 448
 449         if (fsc->mount_options->rsize >= PAGE_SIZE)
 450                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 451                         >> PAGE_SHIFT;
 452
 453         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 454                 file, nr_pages,
 455              max);
 456         while (!list_empty(page_list)) {
 457                 rc = start_read(inode, page_list, max);
 458                 if (rc < 0)
 459                         goto out;
 460         }
 461 out:
 462         ceph_fscache_readpages_cancel(inode, page_list);
 463
 464         dout("readpages %p file %p ret %d\n", inode, file, rc);
 465         return rc;
 466 }
 467
 468 /*
 469  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 470  * only snap context we are allowed to write back.
 471  */
 472 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 473                                                     loff_t *snap_size)
 474 {
 475         struct ceph_inode_info *ci = ceph_inode(inode);
 476         struct ceph_snap_context *snapc = NULL;
 477         struct ceph_cap_snap *capsnap = NULL;
 478
 479         spin_lock(&ci->i_ceph_lock);
 480         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 481                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 482                      capsnap->context, capsnap->dirty_pages);
 483                 if (capsnap->dirty_pages) {
 484                         snapc = ceph_get_snap_context(capsnap->context);
 485                         if (snap_size)
 486                                 *snap_size = capsnap->size;
 487                         break;
 488                 }
 489         }
 490         if (!snapc && ci->i_wrbuffer_ref_head) {
 491                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 492                 dout(" head snapc %p has %d dirty pages\n",
 493                      snapc, ci->i_wrbuffer_ref_head);
 494         }
 495         spin_unlock(&ci->i_ceph_lock);
 496         return snapc;
 497 }
 498
 499 /*
 500  * Write a single page, but leave the page locked.
 501  *
 502  * If we get a write error, set the page error bit, but still adjust the
 503  * dirty page accounting (i.e., page is no longer dirty).
 504  */
 505 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 506 {
 507         struct inode *inode;
 508         struct ceph_inode_info *ci;
 509         struct ceph_fs_client *fsc;
 510         struct ceph_osd_client *osdc;
 511         struct ceph_snap_context *snapc, *oldest;
 512         loff_t page_off = page_offset(page);
 513         loff_t snap_size = -1;
 514         long writeback_stat;
 515         u64 truncate_size;
 516         u32 truncate_seq;
 517         int err = 0, len = PAGE_SIZE;
 518
 519         dout("writepage %p idx %lu\n", page, page->index);
 520
 521         if (!page->mapping || !page->mapping->host) {
 522                 dout("writepage %p - no mapping\n", page);
 523                 return -EFAULT;
 524         }
 525         inode = page->mapping->host;
 526         ci = ceph_inode(inode);
 527         fsc = ceph_inode_to_client(inode);
 528         osdc = &fsc->client->osdc;
 529
 530         /* verify this is a writeable snap context */
 531         snapc = page_snap_context(page);
 532         if (snapc == NULL) {
 533                 dout("writepage %p page %p not dirty?\n", inode, page);
 534                 goto out;
 535         }
 536         oldest = get_oldest_context(inode, &snap_size);
 537         if (snapc->seq > oldest->seq) {
 538                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 539                      inode, page, snapc);
 540                 /* we should only noop if called by kswapd */
 541                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 542                 ceph_put_snap_context(oldest);
 543                 goto out;
 544         }
 545         ceph_put_snap_context(oldest);
 546
 547         spin_lock(&ci->i_ceph_lock);
 548         truncate_seq = ci->i_truncate_seq;
 549         truncate_size = ci->i_truncate_size;
 550         if (snap_size == -1)
 551                 snap_size = i_size_read(inode);
 552         spin_unlock(&ci->i_ceph_lock);
 553
 554         /* is this a partial page at end of file? */
 555         if (page_off >= snap_size) {
 556                 dout("%p page eof %llu\n", page, snap_size);
 557                 goto out;
 558         }
 559         if (snap_size < page_off + len)
 560                 len = snap_size - page_off;
 561
 562         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 563              inode, page, page->index, page_off, len, snapc);
 564
 565         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 566         if (writeback_stat >
 567             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 568                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 569
 570         set_page_writeback(page);
 571         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 572                                    &ci->i_layout, snapc,
 573                                    page_off, len,
 574                                    truncate_seq, truncate_size,
 575                                    &inode->i_mtime, &page, 1);
 576         if (err < 0) {
 577                 struct writeback_control tmp_wbc;
 578                 if (!wbc)
 579                         wbc = &tmp_wbc;
 580                 if (err == -ERESTARTSYS) {
 581                         /* killed by SIGKILL */
 582                         dout("writepage interrupted page %p\n", page);
 583                         redirty_page_for_writepage(wbc, page);
 584                         end_page_writeback(page);
 585                         goto out;
 586                 }
 587                 dout("writepage setting page/mapping error %d %p\n",
 588                      err, page);
 589                 SetPageError(page);
 590                 mapping_set_error(&inode->i_data, err);
 591                 wbc->pages_skipped++;
 592         } else {
 593                 dout("writepage cleaned page %p\n", page);
 594                 err = 0;  /* vfs expects us to return 0 */
 595         }
 596         page->private = 0;
 597         ClearPagePrivate(page);
 598         end_page_writeback(page);
 599         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 600         ceph_put_snap_context(snapc);  /* page's reference */
 601 out:
 602         return err;
 603 }
 604
 605 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 606 {
 607         int err;
 608         struct inode *inode = page->mapping->host;
 609         BUG_ON(!inode);
 610         ihold(inode);
 611         err = writepage_nounlock(page, wbc);
 612         if (err == -ERESTARTSYS) {
 613                 /* direct memory reclaimer was killed by SIGKILL. return 0
 614                  * to prevent caller from setting mapping/page error */
 615                 err = 0;
 616         }
 617         unlock_page(page);
 618         iput(inode);
 619         return err;
 620 }
 621
 622 /*
 623  * lame release_pages helper.  release_pages() isn't exported to
 624  * modules.
 625  */
 626 static void ceph_release_pages(struct page **pages, int num)
 627 {
 628         struct pagevec pvec;
 629         int i;
 630
 631         pagevec_init(&pvec, 0);
 632         for (i = 0; i < num; i++) {
 633                 if (pagevec_add(&pvec, pages[i]) == 0)
 634                         pagevec_release(&pvec);
 635         }
 636         pagevec_release(&pvec);
 637 }
 638
 639 /*
 640  * async writeback completion handler.
 641  *
 642  * If we get an error, set the mapping error bit, but not the individual
 643  * page error bits.
 644  */
 645 static void writepages_finish(struct ceph_osd_request *req)
 646 {
 647         struct inode *inode = req->r_inode;
 648         struct ceph_inode_info *ci = ceph_inode(inode);
 649         struct ceph_osd_data *osd_data;
 650         struct page *page;
 651         int num_pages, total_pages = 0;
 652         int i, j;
 653         int rc = req->r_result;
 654         struct ceph_snap_context *snapc = req->r_snapc;
 655         struct address_space *mapping = inode->i_mapping;
 656         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 657         bool remove_page;
 658
 659         dout("writepages_finish %p rc %d\n", inode, rc);
 660         if (rc < 0)
 661                 mapping_set_error(mapping, rc);
 662
 663         /*
 664          * We lost the cache cap, need to truncate the page before
 665          * it is unlocked, otherwise we'd truncate it later in the
 666          * page truncation thread, possibly losing some data that
 667          * raced its way in
 668          */
 669         remove_page = !(ceph_caps_issued(ci) &
 670                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 671
 672         /* clean all pages */
 673         for (i = 0; i < req->r_num_ops; i++) {
 674                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 675                         break;
 676
 677                 osd_data = osd_req_op_extent_osd_data(req, i);
 678                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 679                 num_pages = calc_pages_for((u64)osd_data->alignment,
 680                                            (u64)osd_data->length);
 681                 total_pages += num_pages;
 682                 for (j = 0; j < num_pages; j++) {
 683                         page = osd_data->pages[j];
 684                         BUG_ON(!page);
 685                         WARN_ON(!PageUptodate(page));
 686
 687                         if (atomic_long_dec_return(&fsc->writeback_count) <
 688                              CONGESTION_OFF_THRESH(
 689                                         fsc->mount_options->congestion_kb))
 690                                 clear_bdi_congested(&fsc->backing_dev_info,
 691                                                     BLK_RW_ASYNC);
 692
 693                         if (rc < 0)
 694                                 SetPageError(page);
 695
 696                         ceph_put_snap_context(page_snap_context(page));
 697                         page->private = 0;
 698                         ClearPagePrivate(page);
 699                         dout("unlocking %p\n", page);
 700                         end_page_writeback(page);
 701
 702                         if (remove_page)
 703                                 generic_error_remove_page(inode->i_mapping,
 704                                                           page);
 705
 706                         unlock_page(page);
 707                 }
 708                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 709                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 710
 711                 ceph_release_pages(osd_data->pages, num_pages);
 712         }
 713
 714         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 715
 716         osd_data = osd_req_op_extent_osd_data(req, 0);
 717         if (osd_data->pages_from_pool)
 718                 mempool_free(osd_data->pages,
 719                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 720         else
 721                 kfree(osd_data->pages);
 722         ceph_osdc_put_request(req);
 723 }
 724
 725 /*
 726  * initiate async writeback
 727  */
 728 static int ceph_writepages_start(struct address_space *mapping,
 729                                  struct writeback_control *wbc)
 730 {
 731         struct inode *inode = mapping->host;
 732         struct ceph_inode_info *ci = ceph_inode(inode);
 733         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 734         struct ceph_vino vino = ceph_vino(inode);
 735         pgoff_t index, start, end;
 736         int range_whole = 0;
 737         int should_loop = 1;
 738         pgoff_t max_pages = 0, max_pages_ever = 0;
 739         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 740         struct pagevec pvec;
 741         int done = 0;
 742         int rc = 0;
 743         unsigned int wsize = i_blocksize(inode);
 744         struct ceph_osd_request *req = NULL;
 745         int do_sync = 0;
 746         loff_t snap_size, i_size;
 747         u64 truncate_size;
 748         u32 truncate_seq;
 749
 750         /*
 751          * Include a 'sync' in the OSD request if this is a data
 752          * integrity write (e.g., O_SYNC write or fsync()), or if our
 753          * cap is being revoked.
 754          */
 755         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 756                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 757                 do_sync = 1;
 758         dout("writepages_start %p dosync=%d (mode=%s)\n",
 759              inode, do_sync,
 760              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 761              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 762
 763         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 764                 if (ci->i_wrbuffer_ref > 0) {
 765                         pr_warn_ratelimited(
 766                                 "writepage_start %p %lld forced umount\n",
 767                                 inode, ceph_ino(inode));
 768                 }
 769                 mapping_set_error(mapping, -EIO);
 770                 return -EIO; /* we're in a forced umount, don't write! */
 771         }
 772         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 773                 wsize = fsc->mount_options->wsize;
 774         if (wsize < PAGE_SIZE)
 775                 wsize = PAGE_SIZE;
 776         max_pages_ever = wsize >> PAGE_SHIFT;
 777
 778         pagevec_init(&pvec, 0);
 779
 780         /* where to start/end? */
 781         if (wbc->range_cyclic) {
 782                 start = mapping->writeback_index; /* Start from prev offset */
 783                 end = -1;
 784                 dout(" cyclic, start at %lu\n", start);
 785         } else {
 786                 start = wbc->range_start >> PAGE_SHIFT;
 787                 end = wbc->range_end >> PAGE_SHIFT;
 788                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 789                         range_whole = 1;
 790                 should_loop = 0;
 791                 dout(" not cyclic, %lu to %lu\n", start, end);
 792         }
 793         index = start;
 794
 795 retry:
 796         /* find oldest snap context with dirty data */
 797         ceph_put_snap_context(snapc);
 798         snap_size = -1;
 799         snapc = get_oldest_context(inode, &snap_size);
 800         if (!snapc) {
 801                 /* hmm, why does writepages get called when there
 802                    is no dirty data? */
 803                 dout(" no snap context with dirty data?\n");
 804                 goto out;
 805         }
 806         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 807              snapc, snapc->seq, snapc->num_snaps);
 808
 809         spin_lock(&ci->i_ceph_lock);
 810         truncate_seq = ci->i_truncate_seq;
 811         truncate_size = ci->i_truncate_size;
 812         i_size = i_size_read(inode);
 813         spin_unlock(&ci->i_ceph_lock);
 814
 815         if (last_snapc && snapc != last_snapc) {
 816                 /* if we switched to a newer snapc, restart our scan at the
 817                  * start of the original file range. */
 818                 dout("  snapc differs from last pass, restarting at %lu\n",
 819                      index);
 820                 index = start;
 821         }
 822         last_snapc = snapc;
 823
 824         while (!done && index <= end) {
 825                 unsigned i;
 826                 int first;
 827                 pgoff_t strip_unit_end = 0;
 828                 int num_ops = 0, op_idx;
 829                 int pvec_pages, locked_pages = 0;
 830                 struct page **pages = NULL, **data_pages;
 831                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 832                 struct page *page;
 833                 int want;
 834                 u64 offset = 0, len = 0;
 835
 836                 max_pages = max_pages_ever;
 837
 838 get_more_pages:
 839                 first = -1;
 840                 want = min(end - index,
 841                            min((pgoff_t)PAGEVEC_SIZE,
 842                                max_pages - (pgoff_t)locked_pages) - 1)
 843                         + 1;
 844                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 845                                                 PAGECACHE_TAG_DIRTY,
 846                                                 want);
 847                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 848                 if (!pvec_pages && !locked_pages)
 849                         break;
 850                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 851                         page = pvec.pages[i];
 852                         dout("? %p idx %lu\n", page, page->index);
 853                         if (locked_pages == 0)
 854                                 lock_page(page);  /* first page */
 855                         else if (!trylock_page(page))
 856                                 break;
 857
 858                         /* only dirty pages, or our accounting breaks */
 859                         if (unlikely(!PageDirty(page)) ||
 860                             unlikely(page->mapping != mapping)) {
 861                                 dout("!dirty or !mapping %p\n", page);
 862                                 unlock_page(page);
 863                                 break;
 864                         }
 865                         if (!wbc->range_cyclic && page->index > end) {
 866                                 dout("end of range %p\n", page);
 867                                 done = 1;
 868                                 unlock_page(page);
 869                                 break;
 870                         }
 871                         if (strip_unit_end && (page->index > strip_unit_end)) {
 872                                 dout("end of strip unit %p\n", page);
 873                                 unlock_page(page);
 874                                 break;
 875                         }
 876                         if (wbc->sync_mode != WB_SYNC_NONE) {
 877                                 dout("waiting on writeback %p\n", page);
 878                                 wait_on_page_writeback(page);
 879                         }
 880                         if (page_offset(page) >=
 881                             (snap_size == -1 ? i_size : snap_size)) {
 882                                 dout("%p page eof %llu\n", page,
 883                                      (snap_size == -1 ? i_size : snap_size));
 884                                 done = 1;
 885                                 unlock_page(page);
 886                                 break;
 887                         }
 888                         if (PageWriteback(page)) {
 889                                 dout("%p under writeback\n", page);
 890                                 unlock_page(page);
 891                                 break;
 892                         }
 893
 894                         /* only if matching snap context */
 895                         pgsnapc = page_snap_context(page);
 896                         if (pgsnapc->seq > snapc->seq) {
 897                                 dout("page snapc %p %lld > oldest %p %lld\n",
 898                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 899                                 unlock_page(page);
 900                                 if (!locked_pages)
 901                                         continue; /* keep looking for snap */
 902                                 break;
 903                         }
 904
 905                         if (!clear_page_dirty_for_io(page)) {
 906                                 dout("%p !clear_page_dirty_for_io\n", page);
 907                                 unlock_page(page);
 908                                 break;
 909                         }
 910
 911                         /*
 912                          * We have something to write.  If this is
 913                          * the first locked page this time through,
 914                          * calculate max possinle write size and
 915                          * allocate a page array
 916                          */
 917                         if (locked_pages == 0) {
 918                                 u64 objnum;
 919                                 u64 objoff;
 920
 921                                 /* prepare async write request */
 922                                 offset = (u64)page_offset(page);
 923                                 len = wsize;
 924
 925                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 926                                                                 offset, len,
 927                                                                 &objnum, &objoff,
 928                                                                 &len);
 929                                 if (rc < 0) {
 930                                         unlock_page(page);
 931                                         break;
 932                                 }
 933
 934                                 num_ops = 1 + do_sync;
 935                                 strip_unit_end = page->index +
 936                                         ((len - 1) >> PAGE_SHIFT);
 937
 938                                 BUG_ON(pages);
 939                                 max_pages = calc_pages_for(0, (u64)len);
 940                                 pages = kmalloc(max_pages * sizeof (*pages),
 941                                                 GFP_NOFS);
 942                                 if (!pages) {
 943                                         pool = fsc->wb_pagevec_pool;
 944                                         pages = mempool_alloc(pool, GFP_NOFS);
 945                                         BUG_ON(!pages);
 946                                 }
 947
 948                                 len = 0;
 949                         } else if (page->index !=
 950                                    (offset + len) >> PAGE_SHIFT) {
 951                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 952                                                         CEPH_OSD_MAX_OPS)) {
 953                                         redirty_page_for_writepage(wbc, page);
 954                                         unlock_page(page);
 955                                         break;
 956                                 }
 957
 958                                 num_ops++;
 959                                 offset = (u64)page_offset(page);
 960                                 len = 0;
 961                         }
 962
 963                         /* note position of first page in pvec */
 964                         if (first < 0)
 965                                 first = i;
 966                         dout("%p will write page %p idx %lu\n",
 967                              inode, page, page->index);
 968
 969                         if (atomic_long_inc_return(&fsc->writeback_count) >
 970                             CONGESTION_ON_THRESH(
 971                                     fsc->mount_options->congestion_kb)) {
 972                                 set_bdi_congested(&fsc->backing_dev_info,
 973                                                   BLK_RW_ASYNC);
 974                         }
 975
 976                         pages[locked_pages] = page;
 977                         locked_pages++;
 978                         len += PAGE_SIZE;
 979                 }
 980
 981                 /* did we get anything? */
 982                 if (!locked_pages)
 983                         goto release_pvec_pages;
 984                 if (i) {
 985                         int j;
 986                         BUG_ON(!locked_pages || first < 0);
 987
 988                         if (pvec_pages && i == pvec_pages &&
 989                             locked_pages < max_pages) {
 990                                 dout("reached end pvec, trying for more\n");
 991                                 pagevec_reinit(&pvec);
 992                                 goto get_more_pages;
 993                         }
 994
 995                         /* shift unused pages over in the pvec...  we
 996                          * will need to release them below. */
 997                         for (j = i; j < pvec_pages; j++) {
 998                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
 999                                 pvec.pages[j-i+first] = pvec.pages[j];
1000                         }
1001                         pvec.nr -= i-first;
1002                 }
1003
1004 new_request:
1005                 offset = page_offset(pages[0]);
1006                 len = wsize;
1007
1008                 req = ceph_osdc_new_request(&fsc->client->osdc,
1009                                         &ci->i_layout, vino,
1010                                         offset, &len, 0, num_ops,
1011                                         CEPH_OSD_OP_WRITE,
1012                                         CEPH_OSD_FLAG_WRITE |
1013                                         CEPH_OSD_FLAG_ONDISK,
1014                                         snapc, truncate_seq,
1015                                         truncate_size, false);
1016                 if (IS_ERR(req)) {
1017                         req = ceph_osdc_new_request(&fsc->client->osdc,
1018                                                 &ci->i_layout, vino,
1019                                                 offset, &len, 0,
1020                                                 min(num_ops,
1021                                                     CEPH_OSD_SLAB_OPS),
1022                                                 CEPH_OSD_OP_WRITE,
1023                                                 CEPH_OSD_FLAG_WRITE |
1024                                                 CEPH_OSD_FLAG_ONDISK,
1025                                                 snapc, truncate_seq,
1026                                                 truncate_size, true);
1027                         BUG_ON(IS_ERR(req));
1028                 }
1029                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1030                              PAGE_SIZE - offset);
1031
1032                 req->r_callback = writepages_finish;
1033                 req->r_inode = inode;
1034
1035                 /* Format the osd request message and submit the write */
1036                 len = 0;
1037                 data_pages = pages;
1038                 op_idx = 0;
1039                 for (i = 0; i < locked_pages; i++) {
1040                         u64 cur_offset = page_offset(pages[i]);
1041                         if (offset + len != cur_offset) {
1042                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1043                                         break;
1044                                 osd_req_op_extent_dup_last(req, op_idx,
1045                                                            cur_offset - offset);
1046                                 dout("writepages got pages at %llu~%llu\n",
1047                                      offset, len);
1048                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1049                                                         data_pages, len, 0,
1050                                                         !!pool, false);
1051                                 osd_req_op_extent_update(req, op_idx, len);
1052
1053                                 len = 0;
1054                                 offset = cur_offset;
1055                                 data_pages = pages + i;
1056                                 op_idx++;
1057                         }
1058
1059                         set_page_writeback(pages[i]);
1060                         len += PAGE_SIZE;
1061                 }
1062
1063                 if (snap_size != -1) {
1064                         len = min(len, snap_size - offset);
1065                 } else if (i == locked_pages) {
1066                         /* writepages_finish() clears writeback pages
1067                          * according to the data length, so make sure
1068                          * data length covers all locked pages */
1069                         u64 min_len = len + 1 - PAGE_SIZE;
1070                         len = min(len, (u64)i_size_read(inode) - offset);
1071                         len = max(len, min_len);
1072                 }
1073                 dout("writepages got pages at %llu~%llu\n", offset, len);
1074
1075                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1076                                                  0, !!pool, false);
1077                 osd_req_op_extent_update(req, op_idx, len);
1078
1079                 if (do_sync) {
1080                         op_idx++;
1081                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1082                 }
1083                 BUG_ON(op_idx + 1 != req->r_num_ops);
1084
1085                 pool = NULL;
1086                 if (i < locked_pages) {
1087                         BUG_ON(num_ops <= req->r_num_ops);
1088                         num_ops -= req->r_num_ops;
1089                         num_ops += do_sync;
1090                         locked_pages -= i;
1091
1092                         /* allocate new pages array for next request */
1093                         data_pages = pages;
1094                         pages = kmalloc(locked_pages * sizeof (*pages),
1095                                         GFP_NOFS);
1096                         if (!pages) {
1097                                 pool = fsc->wb_pagevec_pool;
1098                                 pages = mempool_alloc(pool, GFP_NOFS);
1099                                 BUG_ON(!pages);
1100                         }
1101                         memcpy(pages, data_pages + i,
1102                                locked_pages * sizeof(*pages));
1103                         memset(data_pages + i, 0,
1104                                locked_pages * sizeof(*pages));
1105                 } else {
1106                         BUG_ON(num_ops != req->r_num_ops);
1107                         index = pages[i - 1]->index + 1;
1108                         /* request message now owns the pages array */
1109                         pages = NULL;
1110                 }
1111
1112                 req->r_mtime = inode->i_mtime;
1113                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1114                 BUG_ON(rc);
1115                 req = NULL;
1116
1117                 wbc->nr_to_write -= i;
1118                 if (pages)
1119                         goto new_request;
1120
1121                 if (wbc->nr_to_write <= 0)
1122                         done = 1;
1123
1124 release_pvec_pages:
1125                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1126                      pvec.nr ? pvec.pages[0] : NULL);
1127                 pagevec_release(&pvec);
1128
1129                 if (locked_pages && !done)
1130                         goto retry;
1131         }
1132
1133         if (should_loop && !done) {
1134                 /* more to do; loop back to beginning of file */
1135                 dout("writepages looping back to beginning of file\n");
1136                 should_loop = 0;
1137                 index = 0;
1138                 goto retry;
1139         }
1140
1141         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1142                 mapping->writeback_index = index;
1143
1144 out:
1145         ceph_osdc_put_request(req);
1146         ceph_put_snap_context(snapc);
1147         dout("writepages done, rc = %d\n", rc);
1148         return rc;
1149 }
1150
1151
1152
1153 /*
1154  * See if a given @snapc is either writeable, or already written.
1155  */
1156 static int context_is_writeable_or_written(struct inode *inode,
1157                                            struct ceph_snap_context *snapc)
1158 {
1159         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1160         int ret = !oldest || snapc->seq <= oldest->seq;
1161
1162         ceph_put_snap_context(oldest);
1163         return ret;
1164 }
1165
1166 /*
1167  * We are only allowed to write into/dirty the page if the page is
1168  * clean, or already dirty within the same snap context.
1169  *
1170  * called with page locked.
1171  * return success with page locked,
1172  * or any failure (incl -EAGAIN) with page unlocked.
1173  */
1174 static int ceph_update_writeable_page(struct file *file,
1175                             loff_t pos, unsigned len,
1176                             struct page *page)
1177 {
1178         struct inode *inode = file_inode(file);
1179         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1180         struct ceph_inode_info *ci = ceph_inode(inode);
1181         loff_t page_off = pos & PAGE_MASK;
1182         int pos_in_page = pos & ~PAGE_MASK;
1183         int end_in_page = pos_in_page + len;
1184         loff_t i_size;
1185         int r;
1186         struct ceph_snap_context *snapc, *oldest;
1187
1188         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1189                 dout(" page %p forced umount\n", page);
1190                 unlock_page(page);
1191                 return -EIO;
1192         }
1193
1194 retry_locked:
1195         /* writepages currently holds page lock, but if we change that later, */
1196         wait_on_page_writeback(page);
1197
1198         snapc = page_snap_context(page);
1199         if (snapc && snapc != ci->i_head_snapc) {
1200                 /*
1201                  * this page is already dirty in another (older) snap
1202                  * context!  is it writeable now?
1203                  */
1204                 oldest = get_oldest_context(inode, NULL);
1205
1206                 if (snapc->seq > oldest->seq) {
1207                         ceph_put_snap_context(oldest);
1208                         dout(" page %p snapc %p not current or oldest\n",
1209                              page, snapc);
1210                         /*
1211                          * queue for writeback, and wait for snapc to
1212                          * be writeable or written
1213                          */
1214                         snapc = ceph_get_snap_context(snapc);
1215                         unlock_page(page);
1216                         ceph_queue_writeback(inode);
1217                         r = wait_event_killable(ci->i_cap_wq,
1218                                context_is_writeable_or_written(inode, snapc));
1219                         ceph_put_snap_context(snapc);
1220                         if (r == -ERESTARTSYS)
1221                                 return r;
1222                         return -EAGAIN;
1223                 }
1224                 ceph_put_snap_context(oldest);
1225
1226                 /* yay, writeable, do it now (without dropping page lock) */
1227                 dout(" page %p snapc %p not current, but oldest\n",
1228                      page, snapc);
1229                 if (!clear_page_dirty_for_io(page))
1230                         goto retry_locked;
1231                 r = writepage_nounlock(page, NULL);
1232                 if (r < 0)
1233                         goto fail_unlock;
1234                 goto retry_locked;
1235         }
1236
1237         if (PageUptodate(page)) {
1238                 dout(" page %p already uptodate\n", page);
1239                 return 0;
1240         }
1241
1242         /* full page? */
1243         if (pos_in_page == 0 && len == PAGE_SIZE)
1244                 return 0;
1245
1246         /* past end of file? */
1247         i_size = i_size_read(inode);
1248
1249         if (page_off >= i_size ||
1250             (pos_in_page == 0 && (pos+len) >= i_size &&
1251              end_in_page - pos_in_page != PAGE_SIZE)) {
1252                 dout(" zeroing %p 0 - %d and %d - %d\n",
1253                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1254                 zero_user_segments(page,
1255                                    0, pos_in_page,
1256                                    end_in_page, PAGE_SIZE);
1257                 return 0;
1258         }
1259
1260         /* we need to read it. */
1261         r = ceph_do_readpage(file, page);
1262         if (r < 0) {
1263                 if (r == -EINPROGRESS)
1264                         return -EAGAIN;
1265                 goto fail_unlock;
1266         }
1267         goto retry_locked;
1268 fail_unlock:
1269         unlock_page(page);
1270         return r;
1271 }
1272
1273 /*
1274  * We are only allowed to write into/dirty the page if the page is
1275  * clean, or already dirty within the same snap context.
1276  */
1277 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1278                             loff_t pos, unsigned len, unsigned flags,
1279                             struct page **pagep, void **fsdata)
1280 {
1281         struct inode *inode = file_inode(file);
1282         struct page *page;
1283         pgoff_t index = pos >> PAGE_SHIFT;
1284         int r;
1285
1286         do {
1287                 /* get a page */
1288                 page = grab_cache_page_write_begin(mapping, index, 0);
1289                 if (!page)
1290                         return -ENOMEM;
1291
1292                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1293                      inode, page, (int)pos, (int)len);
1294
1295                 r = ceph_update_writeable_page(file, pos, len, page);
1296                 if (r < 0)
1297                         put_page(page);
1298                 else
1299                         *pagep = page;
1300         } while (r == -EAGAIN);
1301
1302         return r;
1303 }
1304
1305 /*
1306  * we don't do anything in here that simple_write_end doesn't do
1307  * except adjust dirty page accounting
1308  */
1309 static int ceph_write_end(struct file *file, struct address_space *mapping,
1310                           loff_t pos, unsigned len, unsigned copied,
1311                           struct page *page, void *fsdata)
1312 {
1313         struct inode *inode = file_inode(file);
1314         unsigned from = pos & (PAGE_SIZE - 1);
1315         int check_cap = 0;
1316
1317         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1318              inode, page, (int)pos, (int)copied, (int)len);
1319
1320         /* zero the stale part of the page if we did a short copy */
1321         if (copied < len)
1322                 zero_user_segment(page, from+copied, len);
1323
1324         /* did file size increase? */
1325         if (pos+copied > i_size_read(inode))
1326                 check_cap = ceph_inode_set_size(inode, pos+copied);
1327
1328         if (!PageUptodate(page))
1329                 SetPageUptodate(page);
1330
1331         set_page_dirty(page);
1332
1333         unlock_page(page);
1334         put_page(page);
1335
1336         if (check_cap)
1337                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1338
1339         return copied;
1340 }
1341
1342 /*
1343  * we set .direct_IO to indicate direct io is supported, but since we
1344  * intercept O_DIRECT reads and writes early, this function should
1345  * never get called.
1346  */
1347 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1348 {
1349         WARN_ON(1);
1350         return -EINVAL;
1351 }
1352
1353 const struct address_space_operations ceph_aops = {
1354         .readpage = ceph_readpage,
1355         .readpages = ceph_readpages,
1356         .writepage = ceph_writepage,
1357         .writepages = ceph_writepages_start,
1358         .write_begin = ceph_write_begin,
1359         .write_end = ceph_write_end,
1360         .set_page_dirty = ceph_set_page_dirty,
1361         .invalidatepage = ceph_invalidatepage,
1362         .releasepage = ceph_releasepage,
1363         .direct_IO = ceph_direct_io,
1364 };
1365
1366 static void ceph_block_sigs(sigset_t *oldset)
1367 {
1368         sigset_t mask;
1369         siginitsetinv(&mask, sigmask(SIGKILL));
1370         sigprocmask(SIG_BLOCK, &mask, oldset);
1371 }
1372
1373 static void ceph_restore_sigs(sigset_t *oldset)
1374 {
1375         sigprocmask(SIG_SETMASK, oldset, NULL);
1376 }
1377
1378 /*
1379  * vm ops
1380  */
1381 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1382 {
1383         struct inode *inode = file_inode(vma->vm_file);
1384         struct ceph_inode_info *ci = ceph_inode(inode);
1385         struct ceph_file_info *fi = vma->vm_file->private_data;
1386         struct page *pinned_page = NULL;
1387         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1388         int want, got, ret;
1389         sigset_t oldset;
1390
1391         ceph_block_sigs(&oldset);
1392
1393         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1394              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1395         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1396                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1397         else
1398                 want = CEPH_CAP_FILE_CACHE;
1399
1400         got = 0;
1401         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1402         if (ret < 0)
1403                 goto out_restore;
1404
1405         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1406              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1407
1408         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1409             ci->i_inline_version == CEPH_INLINE_NONE) {
1410                 current->journal_info = vma->vm_file;
1411                 ret = filemap_fault(vma, vmf);
1412                 current->journal_info = NULL;
1413         } else
1414                 ret = -EAGAIN;
1415
1416         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1417              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1418         if (pinned_page)
1419                 put_page(pinned_page);
1420         ceph_put_cap_refs(ci, got);
1421
1422         if (ret != -EAGAIN)
1423                 goto out_restore;
1424
1425         /* read inline data */
1426         if (off >= PAGE_SIZE) {
1427                 /* does not support inline data > PAGE_SIZE */
1428                 ret = VM_FAULT_SIGBUS;
1429         } else {
1430                 int ret1;
1431                 struct address_space *mapping = inode->i_mapping;
1432                 struct page *page = find_or_create_page(mapping, 0,
1433                                                 mapping_gfp_constraint(mapping,
1434                                                 ~__GFP_FS));
1435                 if (!page) {
1436                         ret = VM_FAULT_OOM;
1437                         goto out_inline;
1438                 }
1439                 ret1 = __ceph_do_getattr(inode, page,
1440                                          CEPH_STAT_CAP_INLINE_DATA, true);
1441                 if (ret1 < 0 || off >= i_size_read(inode)) {
1442                         unlock_page(page);
1443                         put_page(page);
1444                         if (ret1 < 0)
1445                                 ret = ret1;
1446                         else
1447                                 ret = VM_FAULT_SIGBUS;
1448                         goto out_inline;
1449                 }
1450                 if (ret1 < PAGE_SIZE)
1451                         zero_user_segment(page, ret1, PAGE_SIZE);
1452                 else
1453                         flush_dcache_page(page);
1454                 SetPageUptodate(page);
1455                 vmf->page = page;
1456                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1457 out_inline:
1458                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1459                      inode, off, (size_t)PAGE_SIZE, ret);
1460         }
1461 out_restore:
1462         ceph_restore_sigs(&oldset);
1463         if (ret < 0)
1464                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1465
1466         return ret;
1467 }
1468
1469 /*
1470  * Reuse write_begin here for simplicity.
1471  */
1472 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1473 {
1474         struct inode *inode = file_inode(vma->vm_file);
1475         struct ceph_inode_info *ci = ceph_inode(inode);
1476         struct ceph_file_info *fi = vma->vm_file->private_data;
1477         struct ceph_cap_flush *prealloc_cf;
1478         struct page *page = vmf->page;
1479         loff_t off = page_offset(page);
1480         loff_t size = i_size_read(inode);
1481         size_t len;
1482         int want, got, ret;
1483         sigset_t oldset;
1484
1485         prealloc_cf = ceph_alloc_cap_flush();
1486         if (!prealloc_cf)
1487                 return VM_FAULT_OOM;
1488
1489         ceph_block_sigs(&oldset);
1490
1491         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1492                 struct page *locked_page = NULL;
1493                 if (off == 0) {
1494                         lock_page(page);
1495                         locked_page = page;
1496                 }
1497                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1498                 if (locked_page)
1499                         unlock_page(locked_page);
1500                 if (ret < 0)
1501                         goto out_free;
1502         }
1503
1504         if (off + PAGE_SIZE <= size)
1505                 len = PAGE_SIZE;
1506         else
1507                 len = size & ~PAGE_MASK;
1508
1509         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1510              inode, ceph_vinop(inode), off, len, size);
1511         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1512                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1513         else
1514                 want = CEPH_CAP_FILE_BUFFER;
1515
1516         got = 0;
1517         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1518                             &got, NULL);
1519         if (ret < 0)
1520                 goto out_free;
1521
1522         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1523              inode, off, len, ceph_cap_string(got));
1524
1525         /* Update time before taking page lock */
1526         file_update_time(vma->vm_file);
1527
1528         do {
1529                 lock_page(page);
1530
1531                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1532                         unlock_page(page);
1533                         ret = VM_FAULT_NOPAGE;
1534                         break;
1535                 }
1536
1537                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1538                 if (ret >= 0) {
1539                         /* success.  we'll keep the page locked. */
1540                         set_page_dirty(page);
1541                         ret = VM_FAULT_LOCKED;
1542                 }
1543         } while (ret == -EAGAIN);
1544
1545         if (ret == VM_FAULT_LOCKED ||
1546             ci->i_inline_version != CEPH_INLINE_NONE) {
1547                 int dirty;
1548                 spin_lock(&ci->i_ceph_lock);
1549                 ci->i_inline_version = CEPH_INLINE_NONE;
1550                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1551                                                &prealloc_cf);
1552                 spin_unlock(&ci->i_ceph_lock);
1553                 if (dirty)
1554                         __mark_inode_dirty(inode, dirty);
1555         }
1556
1557         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1558              inode, off, len, ceph_cap_string(got), ret);
1559         ceph_put_cap_refs(ci, got);
1560 out_free:
1561         ceph_restore_sigs(&oldset);
1562         ceph_free_cap_flush(prealloc_cf);
1563         if (ret < 0)
1564                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1565         return ret;
1566 }
1567
1568 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1569                            char *data, size_t len)
1570 {
1571         struct address_space *mapping = inode->i_mapping;
1572         struct page *page;
1573
1574         if (locked_page) {
1575                 page = locked_page;
1576         } else {
1577                 if (i_size_read(inode) == 0)
1578                         return;
1579                 page = find_or_create_page(mapping, 0,
1580                                            mapping_gfp_constraint(mapping,
1581                                            ~__GFP_FS));
1582                 if (!page)
1583                         return;
1584                 if (PageUptodate(page)) {
1585                         unlock_page(page);
1586                         put_page(page);
1587                         return;
1588                 }
1589         }
1590
1591         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1592              inode, ceph_vinop(inode), len, locked_page);
1593
1594         if (len > 0) {
1595                 void *kaddr = kmap_atomic(page);
1596                 memcpy(kaddr, data, len);
1597                 kunmap_atomic(kaddr);
1598         }
1599
1600         if (page != locked_page) {
1601                 if (len < PAGE_SIZE)
1602                         zero_user_segment(page, len, PAGE_SIZE);
1603                 else
1604                         flush_dcache_page(page);
1605
1606                 SetPageUptodate(page);
1607                 unlock_page(page);
1608                 put_page(page);
1609         }
1610 }
1611
1612 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1613 {
1614         struct inode *inode = file_inode(filp);
1615         struct ceph_inode_info *ci = ceph_inode(inode);
1616         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1617         struct ceph_osd_request *req;
1618         struct page *page = NULL;
1619         u64 len, inline_version;
1620         int err = 0;
1621         bool from_pagecache = false;
1622
1623         spin_lock(&ci->i_ceph_lock);
1624         inline_version = ci->i_inline_version;
1625         spin_unlock(&ci->i_ceph_lock);
1626
1627         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1628              inode, ceph_vinop(inode), inline_version);
1629
1630         if (inline_version == 1 || /* initial version, no data */
1631             inline_version == CEPH_INLINE_NONE)
1632                 goto out;
1633
1634         if (locked_page) {
1635                 page = locked_page;
1636                 WARN_ON(!PageUptodate(page));
1637         } else if (ceph_caps_issued(ci) &
1638                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1639                 page = find_get_page(inode->i_mapping, 0);
1640                 if (page) {
1641                         if (PageUptodate(page)) {
1642                                 from_pagecache = true;
1643                                 lock_page(page);
1644                         } else {
1645                                 put_page(page);
1646                                 page = NULL;
1647                         }
1648                 }
1649         }
1650
1651         if (page) {
1652                 len = i_size_read(inode);
1653                 if (len > PAGE_SIZE)
1654                         len = PAGE_SIZE;
1655         } else {
1656                 page = __page_cache_alloc(GFP_NOFS);
1657                 if (!page) {
1658                         err = -ENOMEM;
1659                         goto out;
1660                 }
1661                 err = __ceph_do_getattr(inode, page,
1662                                         CEPH_STAT_CAP_INLINE_DATA, true);
1663                 if (err < 0) {
1664                         /* no inline data */
1665                         if (err == -ENODATA)
1666                                 err = 0;
1667                         goto out;
1668                 }
1669                 len = err;
1670         }
1671
1672         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1673                                     ceph_vino(inode), 0, &len, 0, 1,
1674                                     CEPH_OSD_OP_CREATE,
1675                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1676                                     NULL, 0, 0, false);
1677         if (IS_ERR(req)) {
1678                 err = PTR_ERR(req);
1679                 goto out;
1680         }
1681
1682         req->r_mtime = inode->i_mtime;
1683         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1684         if (!err)
1685                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1686         ceph_osdc_put_request(req);
1687         if (err < 0)
1688                 goto out;
1689
1690         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1691                                     ceph_vino(inode), 0, &len, 1, 3,
1692                                     CEPH_OSD_OP_WRITE,
1693                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1694                                     NULL, ci->i_truncate_seq,
1695                                     ci->i_truncate_size, false);
1696         if (IS_ERR(req)) {
1697                 err = PTR_ERR(req);
1698                 goto out;
1699         }
1700
1701         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1702
1703         {
1704                 __le64 xattr_buf = cpu_to_le64(inline_version);
1705                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1706                                             "inline_version", &xattr_buf,
1707                                             sizeof(xattr_buf),
1708                                             CEPH_OSD_CMPXATTR_OP_GT,
1709                                             CEPH_OSD_CMPXATTR_MODE_U64);
1710                 if (err)
1711                         goto out_put;
1712         }
1713
1714         {
1715                 char xattr_buf[32];
1716                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1717                                          "%llu", inline_version);
1718                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1719                                             "inline_version",
1720                                             xattr_buf, xattr_len, 0, 0);
1721                 if (err)
1722                         goto out_put;
1723         }
1724
1725         req->r_mtime = inode->i_mtime;
1726         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1727         if (!err)
1728                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1729 out_put:
1730         ceph_osdc_put_request(req);
1731         if (err == -ECANCELED)
1732                 err = 0;
1733 out:
1734         if (page && page != locked_page) {
1735                 if (from_pagecache) {
1736                         unlock_page(page);
1737                         put_page(page);
1738                 } else
1739                         __free_pages(page, 0);
1740         }
1741
1742         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1743              inode, ceph_vinop(inode), inline_version, err);
1744         return err;
1745 }
1746
1747 static const struct vm_operations_struct ceph_vmops = {
1748         .fault          = ceph_filemap_fault,
1749         .page_mkwrite   = ceph_page_mkwrite,
1750 };
1751
1752 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1753 {
1754         struct address_space *mapping = file->f_mapping;
1755
1756         if (!mapping->a_ops->readpage)
1757                 return -ENOEXEC;
1758         file_accessed(file);
1759         vma->vm_ops = &ceph_vmops;
1760         return 0;
1761 }
1762
1763 enum {
1764         POOL_READ       = 1,
1765         POOL_WRITE      = 2,
1766 };
1767
1768 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1769                                 s64 pool, struct ceph_string *pool_ns)
1770 {
1771         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1772         struct ceph_mds_client *mdsc = fsc->mdsc;
1773         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1774         struct rb_node **p, *parent;
1775         struct ceph_pool_perm *perm;
1776         struct page **pages;
1777         size_t pool_ns_len;
1778         int err = 0, err2 = 0, have = 0;
1779
1780         down_read(&mdsc->pool_perm_rwsem);
1781         p = &mdsc->pool_perm_tree.rb_node;
1782         while (*p) {
1783                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1784                 if (pool < perm->pool)
1785                         p = &(*p)->rb_left;
1786                 else if (pool > perm->pool)
1787                         p = &(*p)->rb_right;
1788                 else {
1789                         int ret = ceph_compare_string(pool_ns,
1790                                                 perm->pool_ns,
1791                                                 perm->pool_ns_len);
1792                         if (ret < 0)
1793                                 p = &(*p)->rb_left;
1794                         else if (ret > 0)
1795                                 p = &(*p)->rb_right;
1796                         else {
1797                                 have = perm->perm;
1798                                 break;
1799                         }
1800                 }
1801         }
1802         up_read(&mdsc->pool_perm_rwsem);
1803         if (*p)
1804                 goto out;
1805
1806         if (pool_ns)
1807                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1808                      pool, (int)pool_ns->len, pool_ns->str);
1809         else
1810                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1811
1812         down_write(&mdsc->pool_perm_rwsem);
1813         p = &mdsc->pool_perm_tree.rb_node;
1814         parent = NULL;
1815         while (*p) {
1816                 parent = *p;
1817                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1818                 if (pool < perm->pool)
1819                         p = &(*p)->rb_left;
1820                 else if (pool > perm->pool)
1821                         p = &(*p)->rb_right;
1822                 else {
1823                         int ret = ceph_compare_string(pool_ns,
1824                                                 perm->pool_ns,
1825                                                 perm->pool_ns_len);
1826                         if (ret < 0)
1827                                 p = &(*p)->rb_left;
1828                         else if (ret > 0)
1829                                 p = &(*p)->rb_right;
1830                         else {
1831                                 have = perm->perm;
1832                                 break;
1833                         }
1834                 }
1835         }
1836         if (*p) {
1837                 up_write(&mdsc->pool_perm_rwsem);
1838                 goto out;
1839         }
1840
1841         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1842                                          1, false, GFP_NOFS);
1843         if (!rd_req) {
1844                 err = -ENOMEM;
1845                 goto out_unlock;
1846         }
1847
1848         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1849         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1850         rd_req->r_base_oloc.pool = pool;
1851         if (pool_ns)
1852                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1853         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1854
1855         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1856         if (err)
1857                 goto out_unlock;
1858
1859         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1860                                          1, false, GFP_NOFS);
1861         if (!wr_req) {
1862                 err = -ENOMEM;
1863                 goto out_unlock;
1864         }
1865
1866         wr_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ACK;
1867         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1868         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1869         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1870
1871         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1872         if (err)
1873                 goto out_unlock;
1874
1875         /* one page should be large enough for STAT data */
1876         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1877         if (IS_ERR(pages)) {
1878                 err = PTR_ERR(pages);
1879                 goto out_unlock;
1880         }
1881
1882         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1883                                      0, false, true);
1884         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1885
1886         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1887         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1888
1889         if (!err)
1890                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1891         if (!err2)
1892                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1893
1894         if (err >= 0 || err == -ENOENT)
1895                 have |= POOL_READ;
1896         else if (err != -EPERM)
1897                 goto out_unlock;
1898
1899         if (err2 == 0 || err2 == -EEXIST)
1900                 have |= POOL_WRITE;
1901         else if (err2 != -EPERM) {
1902                 err = err2;
1903                 goto out_unlock;
1904         }
1905
1906         pool_ns_len = pool_ns ? pool_ns->len : 0;
1907         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1908         if (!perm) {
1909                 err = -ENOMEM;
1910                 goto out_unlock;
1911         }
1912
1913         perm->pool = pool;
1914         perm->perm = have;
1915         perm->pool_ns_len = pool_ns_len;
1916         if (pool_ns_len > 0)
1917                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1918         perm->pool_ns[pool_ns_len] = 0;
1919
1920         rb_link_node(&perm->node, parent, p);
1921         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1922         err = 0;
1923 out_unlock:
1924         up_write(&mdsc->pool_perm_rwsem);
1925
1926         ceph_osdc_put_request(rd_req);
1927         ceph_osdc_put_request(wr_req);
1928 out:
1929         if (!err)
1930                 err = have;
1931         if (pool_ns)
1932                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1933                      pool, (int)pool_ns->len, pool_ns->str, err);
1934         else
1935                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1936         return err;
1937 }
1938
1939 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1940 {
1941         s64 pool;
1942         struct ceph_string *pool_ns;
1943         int ret, flags;
1944
1945         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1946                                 NOPOOLPERM))
1947                 return 0;
1948
1949         spin_lock(&ci->i_ceph_lock);
1950         flags = ci->i_ceph_flags;
1951         pool = ci->i_layout.pool_id;
1952         spin_unlock(&ci->i_ceph_lock);
1953 check:
1954         if (flags & CEPH_I_POOL_PERM) {
1955                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1956                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1957                              pool);
1958                         return -EPERM;
1959                 }
1960                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1961                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1962                              pool);
1963                         return -EPERM;
1964                 }
1965                 return 0;
1966         }
1967
1968         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1969         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1970         ceph_put_string(pool_ns);
1971         if (ret < 0)
1972                 return ret;
1973
1974         flags = CEPH_I_POOL_PERM;
1975         if (ret & POOL_READ)
1976                 flags |= CEPH_I_POOL_RD;
1977         if (ret & POOL_WRITE)
1978                 flags |= CEPH_I_POOL_WR;
1979
1980         spin_lock(&ci->i_ceph_lock);
1981         if (pool == ci->i_layout.pool_id &&
1982             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
1983                 ci->i_ceph_flags |= flags;
1984         } else {
1985                 pool = ci->i_layout.pool_id;
1986                 flags = ci->i_ceph_flags;
1987         }
1988         spin_unlock(&ci->i_ceph_lock);
1989         goto check;
1990 }
1991
1992 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1993 {
1994         struct ceph_pool_perm *perm;
1995         struct rb_node *n;
1996
1997         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1998                 n = rb_first(&mdsc->pool_perm_tree);
1999                 perm = rb_entry(n, struct ceph_pool_perm, node);
2000                 rb_erase(n, &mdsc->pool_perm_tree);
2001                 kfree(perm);
2002         }
2003 }