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         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /* dirty the head */
  91         spin_lock(&ci->i_ceph_lock);
  92         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  93         if (__ceph_have_pending_cap_snap(ci)) {
  94                 struct ceph_cap_snap *capsnap =
  95                                 list_last_entry(&ci->i_cap_snaps,
  96                                                 struct ceph_cap_snap,
  97                                                 ci_item);
  98                 snapc = ceph_get_snap_context(capsnap->context);
  99                 capsnap->dirty_pages++;
 100         } else {
 101                 BUG_ON(!ci->i_head_snapc);
 102                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 103                 ++ci->i_wrbuffer_ref_head;
 104         }
 105         if (ci->i_wrbuffer_ref == 0)
 106                 ihold(inode);
 107         ++ci->i_wrbuffer_ref;
 108         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 109              "snapc %p seq %lld (%d snaps)\n",
 110              mapping->host, page, page->index,
 111              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 112              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 113              snapc, snapc->seq, snapc->num_snaps);
 114         spin_unlock(&ci->i_ceph_lock);
 115
 116         /*
 117          * Reference snap context in page->private.  Also set
 118          * PagePrivate so that we get invalidatepage callback.
 119          */
 120         BUG_ON(PagePrivate(page));
 121         page->private = (unsigned long)snapc;
 122         SetPagePrivate(page);
 123
 124         ret = __set_page_dirty_nobuffers(page);
 125         WARN_ON(!PageLocked(page));
 126         WARN_ON(!page->mapping);
 127
 128         return ret;
 129 }
 130
 131 /*
 132  * If we are truncating the full page (i.e. offset == 0), adjust the
 133  * dirty page counters appropriately.  Only called if there is private
 134  * data on the page.
 135  */
 136 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 137                                 unsigned int length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc = page_snap_context(page);
 142
 143         inode = page->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != PAGE_SIZE) {
 147                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 148                      inode, page, page->index, offset, length);
 149                 return;
 150         }
 151
 152         ceph_invalidate_fscache_page(inode, page);
 153
 154         if (!PagePrivate(page))
 155                 return;
 156
 157         /*
 158          * We can get non-dirty pages here due to races between
 159          * set_page_dirty and truncate_complete_page; just spit out a
 160          * warning, in case we end up with accounting problems later.
 161          */
 162         if (!PageDirty(page))
 163                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 164
 165         ClearPageChecked(page);
 166
 167         dout("%p invalidatepage %p idx %lu full dirty page\n",
 168              inode, page, page->index);
 169
 170         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 171         ceph_put_snap_context(snapc);
 172         page->private = 0;
 173         ClearPagePrivate(page);
 174 }
 175
 176 static int ceph_releasepage(struct page *page, gfp_t g)
 177 {
 178         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 179              page, page->index, PageDirty(page) ? "" : "not ");
 180
 181         /* Can we release the page from the cache? */
 182         if (!ceph_release_fscache_page(page, g))
 183                 return 0;
 184
 185         return !PagePrivate(page);
 186 }
 187
 188 /*
 189  * read a single page, without unlocking it.
 190  */
 191 static int readpage_nounlock(struct file *filp, struct page *page)
 192 {
 193         struct inode *inode = file_inode(filp);
 194         struct ceph_inode_info *ci = ceph_inode(inode);
 195         struct ceph_osd_client *osdc =
 196                 &ceph_inode_to_client(inode)->client->osdc;
 197         int err = 0;
 198         u64 off = page_offset(page);
 199         u64 len = PAGE_SIZE;
 200
 201         if (off >= i_size_read(inode)) {
 202                 zero_user_segment(page, 0, PAGE_SIZE);
 203                 SetPageUptodate(page);
 204                 return 0;
 205         }
 206
 207         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 208                 /*
 209                  * Uptodate inline data should have been added
 210                  * into page cache while getting Fcr caps.
 211                  */
 212                 if (off == 0)
 213                         return -EINVAL;
 214                 zero_user_segment(page, 0, PAGE_SIZE);
 215                 SetPageUptodate(page);
 216                 return 0;
 217         }
 218
 219         err = ceph_readpage_from_fscache(inode, page);
 220         if (err == 0)
 221                 goto out;
 222
 223         dout("readpage inode %p file %p page %p index %lu\n",
 224              inode, filp, page, page->index);
 225         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 226                                   off, &len,
 227                                   ci->i_truncate_seq, ci->i_truncate_size,
 228                                   &page, 1, 0);
 229         if (err == -ENOENT)
 230                 err = 0;
 231         if (err < 0) {
 232                 SetPageError(page);
 233                 ceph_fscache_readpage_cancel(inode, page);
 234                 goto out;
 235         }
 236         if (err < PAGE_SIZE)
 237                 /* zero fill remainder of page */
 238                 zero_user_segment(page, err, PAGE_SIZE);
 239         else
 240                 flush_dcache_page(page);
 241
 242         SetPageUptodate(page);
 243         ceph_readpage_to_fscache(inode, page);
 244
 245 out:
 246         return err < 0 ? err : 0;
 247 }
 248
 249 static int ceph_readpage(struct file *filp, struct page *page)
 250 {
 251         int r = readpage_nounlock(filp, page);
 252         unlock_page(page);
 253         return r;
 254 }
 255
 256 /*
 257  * Finish an async read(ahead) op.
 258  */
 259 static void finish_read(struct ceph_osd_request *req)
 260 {
 261         struct inode *inode = req->r_inode;
 262         struct ceph_osd_data *osd_data;
 263         int rc = req->r_result <= 0 ? req->r_result : 0;
 264         int bytes = req->r_result >= 0 ? req->r_result : 0;
 265         int num_pages;
 266         int i;
 267
 268         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 269
 270         /* unlock all pages, zeroing any data we didn't read */
 271         osd_data = osd_req_op_extent_osd_data(req, 0);
 272         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 273         num_pages = calc_pages_for((u64)osd_data->alignment,
 274                                         (u64)osd_data->length);
 275         for (i = 0; i < num_pages; i++) {
 276                 struct page *page = osd_data->pages[i];
 277
 278                 if (rc < 0 && rc != -ENOENT) {
 279                         ceph_fscache_readpage_cancel(inode, page);
 280                         goto unlock;
 281                 }
 282                 if (bytes < (int)PAGE_SIZE) {
 283                         /* zero (remainder of) page */
 284                         int s = bytes < 0 ? 0 : bytes;
 285                         zero_user_segment(page, s, PAGE_SIZE);
 286                 }
 287                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 288                      page->index);
 289                 flush_dcache_page(page);
 290                 SetPageUptodate(page);
 291                 ceph_readpage_to_fscache(inode, page);
 292 unlock:
 293                 unlock_page(page);
 294                 put_page(page);
 295                 bytes -= PAGE_SIZE;
 296         }
 297         kfree(osd_data->pages);
 298 }
 299
 300 /*
 301  * start an async read(ahead) operation.  return nr_pages we submitted
 302  * a read for on success, or negative error code.
 303  */
 304 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 305 {
 306         struct ceph_osd_client *osdc =
 307                 &ceph_inode_to_client(inode)->client->osdc;
 308         struct ceph_inode_info *ci = ceph_inode(inode);
 309         struct page *page = list_entry(page_list->prev, struct page, lru);
 310         struct ceph_vino vino;
 311         struct ceph_osd_request *req;
 312         u64 off;
 313         u64 len;
 314         int i;
 315         struct page **pages;
 316         pgoff_t next_index;
 317         int nr_pages = 0;
 318         int got = 0;
 319         int ret = 0;
 320
 321         if (!current->journal_info) {
 322                 /* caller of readpages does not hold buffer and read caps
 323                  * (fadvise, madvise and readahead cases) */
 324                 int want = CEPH_CAP_FILE_CACHE;
 325                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 326                 if (ret < 0) {
 327                         dout("start_read %p, error getting cap\n", inode);
 328                 } else if (!(got & want)) {
 329                         dout("start_read %p, no cache cap\n", inode);
 330                         ret = 0;
 331                 }
 332                 if (ret <= 0) {
 333                         if (got)
 334                                 ceph_put_cap_refs(ci, got);
 335                         while (!list_empty(page_list)) {
 336                                 page = list_entry(page_list->prev,
 337                                                   struct page, lru);
 338                                 list_del(&page->lru);
 339                                 put_page(page);
 340                         }
 341                         return ret;
 342                 }
 343         }
 344
 345         off = (u64) page_offset(page);
 346
 347         /* count pages */
 348         next_index = page->index;
 349         list_for_each_entry_reverse(page, page_list, lru) {
 350                 if (page->index != next_index)
 351                         break;
 352                 nr_pages++;
 353                 next_index++;
 354                 if (max && nr_pages == max)
 355                         break;
 356         }
 357         len = nr_pages << PAGE_SHIFT;
 358         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 359              off, len);
 360         vino = ceph_vino(inode);
 361         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 362                                     0, 1, CEPH_OSD_OP_READ,
 363                                     CEPH_OSD_FLAG_READ, NULL,
 364                                     ci->i_truncate_seq, ci->i_truncate_size,
 365                                     false);
 366         if (IS_ERR(req)) {
 367                 ret = PTR_ERR(req);
 368                 goto out;
 369         }
 370
 371         /* build page vector */
 372         nr_pages = calc_pages_for(0, len);
 373         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 374         if (!pages) {
 375                 ret = -ENOMEM;
 376                 goto out_put;
 377         }
 378         for (i = 0; i < nr_pages; ++i) {
 379                 page = list_entry(page_list->prev, struct page, lru);
 380                 BUG_ON(PageLocked(page));
 381                 list_del(&page->lru);
 382
 383                 dout("start_read %p adding %p idx %lu\n", inode, page,
 384                      page->index);
 385                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 386                                           GFP_KERNEL)) {
 387                         ceph_fscache_uncache_page(inode, page);
 388                         put_page(page);
 389                         dout("start_read %p add_to_page_cache failed %p\n",
 390                              inode, page);
 391                         nr_pages = i;
 392                         if (nr_pages > 0) {
 393                                 len = nr_pages << PAGE_SHIFT;
 394                                 osd_req_op_extent_update(req, 0, len);
 395                                 break;
 396                         }
 397                         goto out_pages;
 398                 }
 399                 pages[i] = page;
 400         }
 401         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 402         req->r_callback = finish_read;
 403         req->r_inode = inode;
 404
 405         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 406         ret = ceph_osdc_start_request(osdc, req, false);
 407         if (ret < 0)
 408                 goto out_pages;
 409         ceph_osdc_put_request(req);
 410
 411         /* After adding locked pages to page cache, the inode holds cache cap.
 412          * So we can drop our cap refs. */
 413         if (got)
 414                 ceph_put_cap_refs(ci, got);
 415
 416         return nr_pages;
 417
 418 out_pages:
 419         for (i = 0; i < nr_pages; ++i) {
 420                 ceph_fscache_readpage_cancel(inode, pages[i]);
 421                 unlock_page(pages[i]);
 422         }
 423         ceph_put_page_vector(pages, nr_pages, false);
 424 out_put:
 425         ceph_osdc_put_request(req);
 426 out:
 427         if (got)
 428                 ceph_put_cap_refs(ci, got);
 429         return ret;
 430 }
 431
 432
 433 /*
 434  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 435  * the caller (VM) cleans them up.
 436  */
 437 static int ceph_readpages(struct file *file, struct address_space *mapping,
 438                           struct list_head *page_list, unsigned nr_pages)
 439 {
 440         struct inode *inode = file_inode(file);
 441         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 442         int rc = 0;
 443         int max = 0;
 444
 445         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 446                 return -EINVAL;
 447
 448         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 449                                          &nr_pages);
 450
 451         if (rc == 0)
 452                 goto out;
 453
 454         if (fsc->mount_options->rsize >= PAGE_SIZE)
 455                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 456                         >> PAGE_SHIFT;
 457
 458         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 459                 file, nr_pages,
 460              max);
 461         while (!list_empty(page_list)) {
 462                 rc = start_read(inode, page_list, max);
 463                 if (rc < 0)
 464                         goto out;
 465         }
 466 out:
 467         ceph_fscache_readpages_cancel(inode, page_list);
 468
 469         dout("readpages %p file %p ret %d\n", inode, file, rc);
 470         return rc;
 471 }
 472
 473 /*
 474  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 475  * only snap context we are allowed to write back.
 476  */
 477 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 478                                                     loff_t *snap_size,
 479                                                     u64 *truncate_size,
 480                                                     u32 *truncate_seq)
 481 {
 482         struct ceph_inode_info *ci = ceph_inode(inode);
 483         struct ceph_snap_context *snapc = NULL;
 484         struct ceph_cap_snap *capsnap = NULL;
 485
 486         spin_lock(&ci->i_ceph_lock);
 487         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 488                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 489                      capsnap->context, capsnap->dirty_pages);
 490                 if (capsnap->dirty_pages) {
 491                         snapc = ceph_get_snap_context(capsnap->context);
 492                         if (snap_size)
 493                                 *snap_size = capsnap->size;
 494                         if (truncate_size)
 495                                 *truncate_size = capsnap->truncate_size;
 496                         if (truncate_seq)
 497                                 *truncate_seq = capsnap->truncate_seq;
 498                         break;
 499                 }
 500         }
 501         if (!snapc && ci->i_wrbuffer_ref_head) {
 502                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 503                 dout(" head snapc %p has %d dirty pages\n",
 504                      snapc, ci->i_wrbuffer_ref_head);
 505                 if (truncate_size)
 506                         *truncate_size = ci->i_truncate_size;
 507                 if (truncate_seq)
 508                         *truncate_seq = ci->i_truncate_seq;
 509         }
 510         spin_unlock(&ci->i_ceph_lock);
 511         return snapc;
 512 }
 513
 514 /*
 515  * Write a single page, but leave the page locked.
 516  *
 517  * If we get a write error, set the page error bit, but still adjust the
 518  * dirty page accounting (i.e., page is no longer dirty).
 519  */
 520 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 521 {
 522         struct inode *inode;
 523         struct ceph_inode_info *ci;
 524         struct ceph_fs_client *fsc;
 525         struct ceph_osd_client *osdc;
 526         struct ceph_snap_context *snapc, *oldest;
 527         loff_t page_off = page_offset(page);
 528         loff_t snap_size = -1;
 529         long writeback_stat;
 530         u64 truncate_size;
 531         u32 truncate_seq;
 532         int err = 0, len = PAGE_SIZE;
 533
 534         dout("writepage %p idx %lu\n", page, page->index);
 535
 536         if (!page->mapping || !page->mapping->host) {
 537                 dout("writepage %p - no mapping\n", page);
 538                 return -EFAULT;
 539         }
 540         inode = page->mapping->host;
 541         ci = ceph_inode(inode);
 542         fsc = ceph_inode_to_client(inode);
 543         osdc = &fsc->client->osdc;
 544
 545         /* verify this is a writeable snap context */
 546         snapc = page_snap_context(page);
 547         if (snapc == NULL) {
 548                 dout("writepage %p page %p not dirty?\n", inode, page);
 549                 goto out;
 550         }
 551         oldest = get_oldest_context(inode, &snap_size,
 552                                     &truncate_size, &truncate_seq);
 553         if (snapc->seq > oldest->seq) {
 554                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 555                      inode, page, snapc);
 556                 /* we should only noop if called by kswapd */
 557                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 558                 ceph_put_snap_context(oldest);
 559                 goto out;
 560         }
 561         ceph_put_snap_context(oldest);
 562
 563         if (snap_size == -1)
 564                 snap_size = i_size_read(inode);
 565
 566         /* is this a partial page at end of file? */
 567         if (page_off >= snap_size) {
 568                 dout("%p page eof %llu\n", page, snap_size);
 569                 goto out;
 570         }
 571         if (snap_size < page_off + len)
 572                 len = snap_size - page_off;
 573
 574         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 575              inode, page, page->index, page_off, len, snapc);
 576
 577         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 578         if (writeback_stat >
 579             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 580                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 581
 582         set_page_writeback(page);
 583         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 584                                    &ci->i_layout, snapc,
 585                                    page_off, len,
 586                                    truncate_seq, truncate_size,
 587                                    &inode->i_mtime, &page, 1);
 588         if (err < 0) {
 589                 struct writeback_control tmp_wbc;
 590                 if (!wbc)
 591                         wbc = &tmp_wbc;
 592                 if (err == -ERESTARTSYS) {
 593                         /* killed by SIGKILL */
 594                         dout("writepage interrupted page %p\n", page);
 595                         redirty_page_for_writepage(wbc, page);
 596                         end_page_writeback(page);
 597                         goto out;
 598                 }
 599                 dout("writepage setting page/mapping error %d %p\n",
 600                      err, page);
 601                 SetPageError(page);
 602                 mapping_set_error(&inode->i_data, err);
 603                 wbc->pages_skipped++;
 604         } else {
 605                 dout("writepage cleaned page %p\n", page);
 606                 err = 0;  /* vfs expects us to return 0 */
 607         }
 608         page->private = 0;
 609         ClearPagePrivate(page);
 610         end_page_writeback(page);
 611         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 612         ceph_put_snap_context(snapc);  /* page's reference */
 613 out:
 614         return err;
 615 }
 616
 617 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 618 {
 619         int err;
 620         struct inode *inode = page->mapping->host;
 621         BUG_ON(!inode);
 622         ihold(inode);
 623         err = writepage_nounlock(page, wbc);
 624         if (err == -ERESTARTSYS) {
 625                 /* direct memory reclaimer was killed by SIGKILL. return 0
 626                  * to prevent caller from setting mapping/page error */
 627                 err = 0;
 628         }
 629         unlock_page(page);
 630         iput(inode);
 631         return err;
 632 }
 633
 634 /*
 635  * lame release_pages helper.  release_pages() isn't exported to
 636  * modules.
 637  */
 638 static void ceph_release_pages(struct page **pages, int num)
 639 {
 640         struct pagevec pvec;
 641         int i;
 642
 643         pagevec_init(&pvec, 0);
 644         for (i = 0; i < num; i++) {
 645                 if (pagevec_add(&pvec, pages[i]) == 0)
 646                         pagevec_release(&pvec);
 647         }
 648         pagevec_release(&pvec);
 649 }
 650
 651 /*
 652  * async writeback completion handler.
 653  *
 654  * If we get an error, set the mapping error bit, but not the individual
 655  * page error bits.
 656  */
 657 static void writepages_finish(struct ceph_osd_request *req)
 658 {
 659         struct inode *inode = req->r_inode;
 660         struct ceph_inode_info *ci = ceph_inode(inode);
 661         struct ceph_osd_data *osd_data;
 662         struct page *page;
 663         int num_pages, total_pages = 0;
 664         int i, j;
 665         int rc = req->r_result;
 666         struct ceph_snap_context *snapc = req->r_snapc;
 667         struct address_space *mapping = inode->i_mapping;
 668         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 669         bool remove_page;
 670
 671         dout("writepages_finish %p rc %d\n", inode, rc);
 672         if (rc < 0)
 673                 mapping_set_error(mapping, rc);
 674
 675         /*
 676          * We lost the cache cap, need to truncate the page before
 677          * it is unlocked, otherwise we'd truncate it later in the
 678          * page truncation thread, possibly losing some data that
 679          * raced its way in
 680          */
 681         remove_page = !(ceph_caps_issued(ci) &
 682                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 683
 684         /* clean all pages */
 685         for (i = 0; i < req->r_num_ops; i++) {
 686                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 687                         break;
 688
 689                 osd_data = osd_req_op_extent_osd_data(req, i);
 690                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 691                 num_pages = calc_pages_for((u64)osd_data->alignment,
 692                                            (u64)osd_data->length);
 693                 total_pages += num_pages;
 694                 for (j = 0; j < num_pages; j++) {
 695                         page = osd_data->pages[j];
 696                         BUG_ON(!page);
 697                         WARN_ON(!PageUptodate(page));
 698
 699                         if (atomic_long_dec_return(&fsc->writeback_count) <
 700                              CONGESTION_OFF_THRESH(
 701                                         fsc->mount_options->congestion_kb))
 702                                 clear_bdi_congested(&fsc->backing_dev_info,
 703                                                     BLK_RW_ASYNC);
 704
 705                         if (rc < 0)
 706                                 SetPageError(page);
 707
 708                         ceph_put_snap_context(page_snap_context(page));
 709                         page->private = 0;
 710                         ClearPagePrivate(page);
 711                         dout("unlocking %p\n", page);
 712                         end_page_writeback(page);
 713
 714                         if (remove_page)
 715                                 generic_error_remove_page(inode->i_mapping,
 716                                                           page);
 717
 718                         unlock_page(page);
 719                 }
 720                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 721                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 722
 723                 ceph_release_pages(osd_data->pages, num_pages);
 724         }
 725
 726         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 727
 728         osd_data = osd_req_op_extent_osd_data(req, 0);
 729         if (osd_data->pages_from_pool)
 730                 mempool_free(osd_data->pages,
 731                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 732         else
 733                 kfree(osd_data->pages);
 734         ceph_osdc_put_request(req);
 735 }
 736
 737 /*
 738  * initiate async writeback
 739  */
 740 static int ceph_writepages_start(struct address_space *mapping,
 741                                  struct writeback_control *wbc)
 742 {
 743         struct inode *inode = mapping->host;
 744         struct ceph_inode_info *ci = ceph_inode(inode);
 745         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 746         struct ceph_vino vino = ceph_vino(inode);
 747         pgoff_t index, start, end;
 748         int range_whole = 0;
 749         int should_loop = 1;
 750         pgoff_t max_pages = 0, max_pages_ever = 0;
 751         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 752         struct pagevec pvec;
 753         int done = 0;
 754         int rc = 0;
 755         unsigned int wsize = i_blocksize(inode);
 756         struct ceph_osd_request *req = NULL;
 757         int do_sync = 0;
 758         loff_t snap_size, i_size;
 759         u64 truncate_size;
 760         u32 truncate_seq;
 761
 762         /*
 763          * Include a 'sync' in the OSD request if this is a data
 764          * integrity write (e.g., O_SYNC write or fsync()), or if our
 765          * cap is being revoked.
 766          */
 767         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 768                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 769                 do_sync = 1;
 770         dout("writepages_start %p dosync=%d (mode=%s)\n",
 771              inode, do_sync,
 772              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 773              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 774
 775         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 776                 if (ci->i_wrbuffer_ref > 0) {
 777                         pr_warn_ratelimited(
 778                                 "writepage_start %p %lld forced umount\n",
 779                                 inode, ceph_ino(inode));
 780                 }
 781                 mapping_set_error(mapping, -EIO);
 782                 return -EIO; /* we're in a forced umount, don't write! */
 783         }
 784         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 785                 wsize = fsc->mount_options->wsize;
 786         if (wsize < PAGE_SIZE)
 787                 wsize = PAGE_SIZE;
 788         max_pages_ever = wsize >> PAGE_SHIFT;
 789
 790         pagevec_init(&pvec, 0);
 791
 792         /* where to start/end? */
 793         if (wbc->range_cyclic) {
 794                 start = mapping->writeback_index; /* Start from prev offset */
 795                 end = -1;
 796                 dout(" cyclic, start at %lu\n", start);
 797         } else {
 798                 start = wbc->range_start >> PAGE_SHIFT;
 799                 end = wbc->range_end >> PAGE_SHIFT;
 800                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 801                         range_whole = 1;
 802                 should_loop = 0;
 803                 dout(" not cyclic, %lu to %lu\n", start, end);
 804         }
 805         index = start;
 806
 807 retry:
 808         /* find oldest snap context with dirty data */
 809         ceph_put_snap_context(snapc);
 810         snap_size = -1;
 811         snapc = get_oldest_context(inode, &snap_size,
 812                                    &truncate_size, &truncate_seq);
 813         if (!snapc) {
 814                 /* hmm, why does writepages get called when there
 815                    is no dirty data? */
 816                 dout(" no snap context with dirty data?\n");
 817                 goto out;
 818         }
 819         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 820              snapc, snapc->seq, snapc->num_snaps);
 821
 822         i_size = i_size_read(inode);
 823
 824         if (last_snapc && snapc != last_snapc) {
 825                 /* if we switched to a newer snapc, restart our scan at the
 826                  * start of the original file range. */
 827                 dout("  snapc differs from last pass, restarting at %lu\n",
 828                      index);
 829                 index = start;
 830         }
 831         last_snapc = snapc;
 832
 833         while (!done && index <= end) {
 834                 unsigned i;
 835                 int first;
 836                 pgoff_t strip_unit_end = 0;
 837                 int num_ops = 0, op_idx;
 838                 int pvec_pages, locked_pages = 0;
 839                 struct page **pages = NULL, **data_pages;
 840                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 841                 struct page *page;
 842                 int want;
 843                 u64 offset = 0, len = 0;
 844
 845                 max_pages = max_pages_ever;
 846
 847 get_more_pages:
 848                 first = -1;
 849                 want = min(end - index,
 850                            min((pgoff_t)PAGEVEC_SIZE,
 851                                max_pages - (pgoff_t)locked_pages) - 1)
 852                         + 1;
 853                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 854                                                 PAGECACHE_TAG_DIRTY,
 855                                                 want);
 856                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 857                 if (!pvec_pages && !locked_pages)
 858                         break;
 859                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 860                         page = pvec.pages[i];
 861                         dout("? %p idx %lu\n", page, page->index);
 862                         if (locked_pages == 0)
 863                                 lock_page(page);  /* first page */
 864                         else if (!trylock_page(page))
 865                                 break;
 866
 867                         /* only dirty pages, or our accounting breaks */
 868                         if (unlikely(!PageDirty(page)) ||
 869                             unlikely(page->mapping != mapping)) {
 870                                 dout("!dirty or !mapping %p\n", page);
 871                                 unlock_page(page);
 872                                 break;
 873                         }
 874                         if (!wbc->range_cyclic && page->index > end) {
 875                                 dout("end of range %p\n", page);
 876                                 done = 1;
 877                                 unlock_page(page);
 878                                 break;
 879                         }
 880                         if (strip_unit_end && (page->index > strip_unit_end)) {
 881                                 dout("end of strip unit %p\n", page);
 882                                 unlock_page(page);
 883                                 break;
 884                         }
 885                         if (wbc->sync_mode != WB_SYNC_NONE) {
 886                                 dout("waiting on writeback %p\n", page);
 887                                 wait_on_page_writeback(page);
 888                         }
 889                         if (page_offset(page) >=
 890                             (snap_size == -1 ? i_size : snap_size)) {
 891                                 dout("%p page eof %llu\n", page,
 892                                      (snap_size == -1 ? i_size : snap_size));
 893                                 done = 1;
 894                                 unlock_page(page);
 895                                 break;
 896                         }
 897                         if (PageWriteback(page)) {
 898                                 dout("%p under writeback\n", page);
 899                                 unlock_page(page);
 900                                 break;
 901                         }
 902
 903                         /* only if matching snap context */
 904                         pgsnapc = page_snap_context(page);
 905                         if (pgsnapc->seq > snapc->seq) {
 906                                 dout("page snapc %p %lld > oldest %p %lld\n",
 907                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 908                                 unlock_page(page);
 909                                 if (!locked_pages)
 910                                         continue; /* keep looking for snap */
 911                                 break;
 912                         }
 913
 914                         if (!clear_page_dirty_for_io(page)) {
 915                                 dout("%p !clear_page_dirty_for_io\n", page);
 916                                 unlock_page(page);
 917                                 break;
 918                         }
 919
 920                         /*
 921                          * We have something to write.  If this is
 922                          * the first locked page this time through,
 923                          * calculate max possinle write size and
 924                          * allocate a page array
 925                          */
 926                         if (locked_pages == 0) {
 927                                 u64 objnum;
 928                                 u64 objoff;
 929
 930                                 /* prepare async write request */
 931                                 offset = (u64)page_offset(page);
 932                                 len = wsize;
 933
 934                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 935                                                                 offset, len,
 936                                                                 &objnum, &objoff,
 937                                                                 &len);
 938                                 if (rc < 0) {
 939                                         unlock_page(page);
 940                                         break;
 941                                 }
 942
 943                                 num_ops = 1 + do_sync;
 944                                 strip_unit_end = page->index +
 945                                         ((len - 1) >> PAGE_SHIFT);
 946
 947                                 BUG_ON(pages);
 948                                 max_pages = calc_pages_for(0, (u64)len);
 949                                 pages = kmalloc(max_pages * sizeof (*pages),
 950                                                 GFP_NOFS);
 951                                 if (!pages) {
 952                                         pool = fsc->wb_pagevec_pool;
 953                                         pages = mempool_alloc(pool, GFP_NOFS);
 954                                         BUG_ON(!pages);
 955                                 }
 956
 957                                 len = 0;
 958                         } else if (page->index !=
 959                                    (offset + len) >> PAGE_SHIFT) {
 960                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 961                                                         CEPH_OSD_MAX_OPS)) {
 962                                         redirty_page_for_writepage(wbc, page);
 963                                         unlock_page(page);
 964                                         break;
 965                                 }
 966
 967                                 num_ops++;
 968                                 offset = (u64)page_offset(page);
 969                                 len = 0;
 970                         }
 971
 972                         /* note position of first page in pvec */
 973                         if (first < 0)
 974                                 first = i;
 975                         dout("%p will write page %p idx %lu\n",
 976                              inode, page, page->index);
 977
 978                         if (atomic_long_inc_return(&fsc->writeback_count) >
 979                             CONGESTION_ON_THRESH(
 980                                     fsc->mount_options->congestion_kb)) {
 981                                 set_bdi_congested(&fsc->backing_dev_info,
 982                                                   BLK_RW_ASYNC);
 983                         }
 984
 985                         pages[locked_pages] = page;
 986                         locked_pages++;
 987                         len += PAGE_SIZE;
 988                 }
 989
 990                 /* did we get anything? */
 991                 if (!locked_pages)
 992                         goto release_pvec_pages;
 993                 if (i) {
 994                         int j;
 995                         BUG_ON(!locked_pages || first < 0);
 996
 997                         if (pvec_pages && i == pvec_pages &&
 998                             locked_pages < max_pages) {
 999                                 dout("reached end pvec, trying for more\n");
1000                                 pagevec_reinit(&pvec);
1001                                 goto get_more_pages;
1002                         }
1003
1004                         /* shift unused pages over in the pvec...  we
1005                          * will need to release them below. */
1006                         for (j = i; j < pvec_pages; j++) {
1007                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
1008                                 pvec.pages[j-i+first] = pvec.pages[j];
1009                         }
1010                         pvec.nr -= i-first;
1011                 }
1012
1013 new_request:
1014                 offset = page_offset(pages[0]);
1015                 len = wsize;
1016
1017                 req = ceph_osdc_new_request(&fsc->client->osdc,
1018                                         &ci->i_layout, vino,
1019                                         offset, &len, 0, num_ops,
1020                                         CEPH_OSD_OP_WRITE,
1021                                         CEPH_OSD_FLAG_WRITE,
1022                                         snapc, truncate_seq,
1023                                         truncate_size, false);
1024                 if (IS_ERR(req)) {
1025                         req = ceph_osdc_new_request(&fsc->client->osdc,
1026                                                 &ci->i_layout, vino,
1027                                                 offset, &len, 0,
1028                                                 min(num_ops,
1029                                                     CEPH_OSD_SLAB_OPS),
1030                                                 CEPH_OSD_OP_WRITE,
1031                                                 CEPH_OSD_FLAG_WRITE,
1032                                                 snapc, truncate_seq,
1033                                                 truncate_size, true);
1034                         BUG_ON(IS_ERR(req));
1035                 }
1036                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1037                              PAGE_SIZE - offset);
1038
1039                 req->r_callback = writepages_finish;
1040                 req->r_inode = inode;
1041
1042                 /* Format the osd request message and submit the write */
1043                 len = 0;
1044                 data_pages = pages;
1045                 op_idx = 0;
1046                 for (i = 0; i < locked_pages; i++) {
1047                         u64 cur_offset = page_offset(pages[i]);
1048                         if (offset + len != cur_offset) {
1049                                 if (op_idx + do_sync + 1 == req->r_num_ops)
1050                                         break;
1051                                 osd_req_op_extent_dup_last(req, op_idx,
1052                                                            cur_offset - offset);
1053                                 dout("writepages got pages at %llu~%llu\n",
1054                                      offset, len);
1055                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1056                                                         data_pages, len, 0,
1057                                                         !!pool, false);
1058                                 osd_req_op_extent_update(req, op_idx, len);
1059
1060                                 len = 0;
1061                                 offset = cur_offset;
1062                                 data_pages = pages + i;
1063                                 op_idx++;
1064                         }
1065
1066                         set_page_writeback(pages[i]);
1067                         len += PAGE_SIZE;
1068                 }
1069
1070                 if (snap_size != -1) {
1071                         len = min(len, snap_size - offset);
1072                 } else if (i == locked_pages) {
1073                         /* writepages_finish() clears writeback pages
1074                          * according to the data length, so make sure
1075                          * data length covers all locked pages */
1076                         u64 min_len = len + 1 - PAGE_SIZE;
1077                         len = min(len, (u64)i_size_read(inode) - offset);
1078                         len = max(len, min_len);
1079                 }
1080                 dout("writepages got pages at %llu~%llu\n", offset, len);
1081
1082                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1083                                                  0, !!pool, false);
1084                 osd_req_op_extent_update(req, op_idx, len);
1085
1086                 if (do_sync) {
1087                         op_idx++;
1088                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
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                         num_ops += do_sync;
1097                         locked_pages -= i;
1098
1099                         /* allocate new pages array for next request */
1100                         data_pages = pages;
1101                         pages = kmalloc(locked_pages * sizeof (*pages),
1102                                         GFP_NOFS);
1103                         if (!pages) {
1104                                 pool = fsc->wb_pagevec_pool;
1105                                 pages = mempool_alloc(pool, GFP_NOFS);
1106                                 BUG_ON(!pages);
1107                         }
1108                         memcpy(pages, data_pages + i,
1109                                locked_pages * sizeof(*pages));
1110                         memset(data_pages + i, 0,
1111                                locked_pages * sizeof(*pages));
1112                 } else {
1113                         BUG_ON(num_ops != req->r_num_ops);
1114                         index = pages[i - 1]->index + 1;
1115                         /* request message now owns the pages array */
1116                         pages = NULL;
1117                 }
1118
1119                 req->r_mtime = inode->i_mtime;
1120                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1121                 BUG_ON(rc);
1122                 req = NULL;
1123
1124                 wbc->nr_to_write -= i;
1125                 if (pages)
1126                         goto new_request;
1127
1128                 if (wbc->nr_to_write <= 0)
1129                         done = 1;
1130
1131 release_pvec_pages:
1132                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1133                      pvec.nr ? pvec.pages[0] : NULL);
1134                 pagevec_release(&pvec);
1135
1136                 if (locked_pages && !done)
1137                         goto retry;
1138         }
1139
1140         if (should_loop && !done) {
1141                 /* more to do; loop back to beginning of file */
1142                 dout("writepages looping back to beginning of file\n");
1143                 should_loop = 0;
1144                 index = 0;
1145                 goto retry;
1146         }
1147
1148         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1149                 mapping->writeback_index = index;
1150
1151 out:
1152         ceph_osdc_put_request(req);
1153         ceph_put_snap_context(snapc);
1154         dout("writepages done, rc = %d\n", rc);
1155         return rc;
1156 }
1157
1158
1159
1160 /*
1161  * See if a given @snapc is either writeable, or already written.
1162  */
1163 static int context_is_writeable_or_written(struct inode *inode,
1164                                            struct ceph_snap_context *snapc)
1165 {
1166         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL,
1167                                                               NULL, NULL);
1168         int ret = !oldest || snapc->seq <= oldest->seq;
1169
1170         ceph_put_snap_context(oldest);
1171         return ret;
1172 }
1173
1174 /*
1175  * We are only allowed to write into/dirty the page if the page is
1176  * clean, or already dirty within the same snap context.
1177  *
1178  * called with page locked.
1179  * return success with page locked,
1180  * or any failure (incl -EAGAIN) with page unlocked.
1181  */
1182 static int ceph_update_writeable_page(struct file *file,
1183                             loff_t pos, unsigned len,
1184                             struct page *page)
1185 {
1186         struct inode *inode = file_inode(file);
1187         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1188         struct ceph_inode_info *ci = ceph_inode(inode);
1189         loff_t page_off = pos & PAGE_MASK;
1190         int pos_in_page = pos & ~PAGE_MASK;
1191         int end_in_page = pos_in_page + len;
1192         loff_t i_size;
1193         int r;
1194         struct ceph_snap_context *snapc, *oldest;
1195
1196         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1197                 dout(" page %p forced umount\n", page);
1198                 unlock_page(page);
1199                 return -EIO;
1200         }
1201
1202 retry_locked:
1203         /* writepages currently holds page lock, but if we change that later, */
1204         wait_on_page_writeback(page);
1205
1206         snapc = page_snap_context(page);
1207         if (snapc && snapc != ci->i_head_snapc) {
1208                 /*
1209                  * this page is already dirty in another (older) snap
1210                  * context!  is it writeable now?
1211                  */
1212                 oldest = get_oldest_context(inode, NULL, NULL, NULL);
1213
1214                 if (snapc->seq > oldest->seq) {
1215                         ceph_put_snap_context(oldest);
1216                         dout(" page %p snapc %p not current or oldest\n",
1217                              page, snapc);
1218                         /*
1219                          * queue for writeback, and wait for snapc to
1220                          * be writeable or written
1221                          */
1222                         snapc = ceph_get_snap_context(snapc);
1223                         unlock_page(page);
1224                         ceph_queue_writeback(inode);
1225                         r = wait_event_killable(ci->i_cap_wq,
1226                                context_is_writeable_or_written(inode, snapc));
1227                         ceph_put_snap_context(snapc);
1228                         if (r == -ERESTARTSYS)
1229                                 return r;
1230                         return -EAGAIN;
1231                 }
1232                 ceph_put_snap_context(oldest);
1233
1234                 /* yay, writeable, do it now (without dropping page lock) */
1235                 dout(" page %p snapc %p not current, but oldest\n",
1236                      page, snapc);
1237                 if (!clear_page_dirty_for_io(page))
1238                         goto retry_locked;
1239                 r = writepage_nounlock(page, NULL);
1240                 if (r < 0)
1241                         goto fail_nosnap;
1242                 goto retry_locked;
1243         }
1244
1245         if (PageUptodate(page)) {
1246                 dout(" page %p already uptodate\n", page);
1247                 return 0;
1248         }
1249
1250         /* full page? */
1251         if (pos_in_page == 0 && len == PAGE_SIZE)
1252                 return 0;
1253
1254         /* past end of file? */
1255         i_size = i_size_read(inode);
1256
1257         if (page_off >= i_size ||
1258             (pos_in_page == 0 && (pos+len) >= i_size &&
1259              end_in_page - pos_in_page != PAGE_SIZE)) {
1260                 dout(" zeroing %p 0 - %d and %d - %d\n",
1261                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1262                 zero_user_segments(page,
1263                                    0, pos_in_page,
1264                                    end_in_page, PAGE_SIZE);
1265                 return 0;
1266         }
1267
1268         /* we need to read it. */
1269         r = readpage_nounlock(file, page);
1270         if (r < 0)
1271                 goto fail_nosnap;
1272         goto retry_locked;
1273 fail_nosnap:
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         int check_cap = 0;
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         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1895
1896         if (!err)
1897                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1898         if (!err2)
1899                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1900
1901         if (err >= 0 || err == -ENOENT)
1902                 have |= POOL_READ;
1903         else if (err != -EPERM)
1904                 goto out_unlock;
1905
1906         if (err2 == 0 || err2 == -EEXIST)
1907                 have |= POOL_WRITE;
1908         else if (err2 != -EPERM) {
1909                 err = err2;
1910                 goto out_unlock;
1911         }
1912
1913         pool_ns_len = pool_ns ? pool_ns->len : 0;
1914         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1915         if (!perm) {
1916                 err = -ENOMEM;
1917                 goto out_unlock;
1918         }
1919
1920         perm->pool = pool;
1921         perm->perm = have;
1922         perm->pool_ns_len = pool_ns_len;
1923         if (pool_ns_len > 0)
1924                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1925         perm->pool_ns[pool_ns_len] = 0;
1926
1927         rb_link_node(&perm->node, parent, p);
1928         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1929         err = 0;
1930 out_unlock:
1931         up_write(&mdsc->pool_perm_rwsem);
1932
1933         ceph_osdc_put_request(rd_req);
1934         ceph_osdc_put_request(wr_req);
1935 out:
1936         if (!err)
1937                 err = have;
1938         if (pool_ns)
1939                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1940                      pool, (int)pool_ns->len, pool_ns->str, err);
1941         else
1942                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1943         return err;
1944 }
1945
1946 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1947 {
1948         s64 pool;
1949         struct ceph_string *pool_ns;
1950         int ret, flags;
1951
1952         if (ci->i_vino.snap != CEPH_NOSNAP) {
1953                 /*
1954                  * Pool permission check needs to write to the first object.
1955                  * But for snapshot, head of the first object may have alread
1956                  * been deleted. Skip check to avoid creating orphan object.
1957                  */
1958                 return 0;
1959         }
1960
1961         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1962                                 NOPOOLPERM))
1963                 return 0;
1964
1965         spin_lock(&ci->i_ceph_lock);
1966         flags = ci->i_ceph_flags;
1967         pool = ci->i_layout.pool_id;
1968         spin_unlock(&ci->i_ceph_lock);
1969 check:
1970         if (flags & CEPH_I_POOL_PERM) {
1971                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1972                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1973                              pool);
1974                         return -EPERM;
1975                 }
1976                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1977                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1978                              pool);
1979                         return -EPERM;
1980                 }
1981                 return 0;
1982         }
1983
1984         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1985         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1986         ceph_put_string(pool_ns);
1987         if (ret < 0)
1988                 return ret;
1989
1990         flags = CEPH_I_POOL_PERM;
1991         if (ret & POOL_READ)
1992                 flags |= CEPH_I_POOL_RD;
1993         if (ret & POOL_WRITE)
1994                 flags |= CEPH_I_POOL_WR;
1995
1996         spin_lock(&ci->i_ceph_lock);
1997         if (pool == ci->i_layout.pool_id &&
1998             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
1999                 ci->i_ceph_flags |= flags;
2000         } else {
2001                 pool = ci->i_layout.pool_id;
2002                 flags = ci->i_ceph_flags;
2003         }
2004         spin_unlock(&ci->i_ceph_lock);
2005         goto check;
2006 }
2007
2008 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2009 {
2010         struct ceph_pool_perm *perm;
2011         struct rb_node *n;
2012
2013         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2014                 n = rb_first(&mdsc->pool_perm_tree);
2015                 perm = rb_entry(n, struct ceph_pool_perm, node);
2016                 rb_erase(n, &mdsc->pool_perm_tree);
2017                 kfree(perm);
2018         }
2019 }