fs/ceph/addr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/backing-dev.h>
   5 #include <linux/fs.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>    /* generic_writepages */
   9 #include <linux/slab.h>
  10 #include <linux/pagevec.h>
  11 #include <linux/task_io_accounting_ops.h>
  12 #include <linux/signal.h>
  13 #include <linux/iversion.h>
  14
  15 #include "super.h"
  16 #include "mds_client.h"
  17 #include "cache.h"
  18 #include <linux/ceph/osd_client.h>
  19 #include <linux/ceph/striper.h>
  20
  21 /*
  22  * Ceph address space ops.
  23  *
  24  * There are a few funny things going on here.
  25  *
  26  * The page->private field is used to reference a struct
  27  * ceph_snap_context for _every_ dirty page.  This indicates which
  28  * snapshot the page was logically dirtied in, and thus which snap
  29  * context needs to be associated with the osd write during writeback.
  30  *
  31  * Similarly, struct ceph_inode_info maintains a set of counters to
  32  * count dirty pages on the inode.  In the absence of snapshots,
  33  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  34  *
  35  * When a snapshot is taken (that is, when the client receives
  36  * notification that a snapshot was taken), each inode with caps and
  37  * with dirty pages (dirty pages implies there is a cap) gets a new
  38  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  39  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  40  * moved to capsnap->dirty. (Unless a sync write is currently in
  41  * progress.  In that case, the capsnap is said to be "pending", new
  42  * writes cannot start, and the capsnap isn't "finalized" until the
  43  * write completes (or fails) and a final size/mtime for the inode for
  44  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  45  *
  46  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  47  * we look for the first capsnap in i_cap_snaps and write out pages in
  48  * that snap context _only_.  Then we move on to the next capsnap,
  49  * eventually reaching the "live" or "head" context (i.e., pages that
  50  * are not yet snapped) and are writing the most recently dirtied
  51  * pages.
  52  *
  53  * Invalidate and so forth must take care to ensure the dirty page
  54  * accounting is preserved.
  55  */
  56
  57 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  58 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  59         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  60          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  61
  62 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  63 {
  64         if (PagePrivate(page))
  65                 return (void *)page->private;
  66         return NULL;
  67 }
  68
  69 /*
  70  * Dirty a page.  Optimistically adjust accounting, on the assumption
  71  * that we won't race with invalidate.  If we do, readjust.
  72  */
  73 static int ceph_set_page_dirty(struct page *page)
  74 {
  75         struct address_space *mapping = page->mapping;
  76         struct inode *inode;
  77         struct ceph_inode_info *ci;
  78         struct ceph_snap_context *snapc;
  79         int ret;
  80
  81         if (unlikely(!mapping))
  82                 return !TestSetPageDirty(page);
  83
  84         if (PageDirty(page)) {
  85                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  86                      mapping->host, page, page->index);
  87                 BUG_ON(!PagePrivate(page));
  88                 return 0;
  89         }
  90
  91         inode = mapping->host;
  92         ci = ceph_inode(inode);
  93
  94         /* dirty the head */
  95         spin_lock(&ci->i_ceph_lock);
  96         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  97         if (__ceph_have_pending_cap_snap(ci)) {
  98                 struct ceph_cap_snap *capsnap =
  99                                 list_last_entry(&ci->i_cap_snaps,
 100                                                 struct ceph_cap_snap,
 101                                                 ci_item);
 102                 snapc = ceph_get_snap_context(capsnap->context);
 103                 capsnap->dirty_pages++;
 104         } else {
 105                 BUG_ON(!ci->i_head_snapc);
 106                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 107                 ++ci->i_wrbuffer_ref_head;
 108         }
 109         if (ci->i_wrbuffer_ref == 0)
 110                 ihold(inode);
 111         ++ci->i_wrbuffer_ref;
 112         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 113              "snapc %p seq %lld (%d snaps)\n",
 114              mapping->host, page, page->index,
 115              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 116              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 117              snapc, snapc->seq, snapc->num_snaps);
 118         spin_unlock(&ci->i_ceph_lock);
 119
 120         /*
 121          * Reference snap context in page->private.  Also set
 122          * PagePrivate so that we get invalidatepage callback.
 123          */
 124         BUG_ON(PagePrivate(page));
 125         page->private = (unsigned long)snapc;
 126         SetPagePrivate(page);
 127
 128         ret = __set_page_dirty_nobuffers(page);
 129         WARN_ON(!PageLocked(page));
 130         WARN_ON(!page->mapping);
 131
 132         return ret;
 133 }
 134
 135 /*
 136  * If we are truncating the full page (i.e. offset == 0), adjust the
 137  * dirty page counters appropriately.  Only called if there is private
 138  * data on the page.
 139  */
 140 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 141                                 unsigned int length)
 142 {
 143         struct inode *inode;
 144         struct ceph_inode_info *ci;
 145         struct ceph_snap_context *snapc = page_snap_context(page);
 146
 147         inode = page->mapping->host;
 148         ci = ceph_inode(inode);
 149
 150         if (offset != 0 || length != PAGE_SIZE) {
 151                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 152                      inode, page, page->index, offset, length);
 153                 return;
 154         }
 155
 156         ceph_invalidate_fscache_page(inode, page);
 157
 158         WARN_ON(!PageLocked(page));
 159         if (!PagePrivate(page))
 160                 return;
 161
 162         ClearPageChecked(page);
 163
 164         dout("%p invalidatepage %p idx %lu full dirty page\n",
 165              inode, page, page->index);
 166
 167         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 168         ceph_put_snap_context(snapc);
 169         page->private = 0;
 170         ClearPagePrivate(page);
 171 }
 172
 173 static int ceph_releasepage(struct page *page, gfp_t g)
 174 {
 175         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 176              page, page->index, PageDirty(page) ? "" : "not ");
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * read a single page, without unlocking it.
 187  */
 188 static int ceph_do_readpage(struct file *filp, struct page *page)
 189 {
 190         struct inode *inode = file_inode(filp);
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_osd_client *osdc =
 193                 &ceph_inode_to_client(inode)->client->osdc;
 194         int err = 0;
 195         u64 off = page_offset(page);
 196         u64 len = PAGE_SIZE;
 197
 198         if (off >= i_size_read(inode)) {
 199                 zero_user_segment(page, 0, PAGE_SIZE);
 200                 SetPageUptodate(page);
 201                 return 0;
 202         }
 203
 204         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 205                 /*
 206                  * Uptodate inline data should have been added
 207                  * into page cache while getting Fcr caps.
 208                  */
 209                 if (off == 0)
 210                         return -EINVAL;
 211                 zero_user_segment(page, 0, PAGE_SIZE);
 212                 SetPageUptodate(page);
 213                 return 0;
 214         }
 215
 216         err = ceph_readpage_from_fscache(inode, page);
 217         if (err == 0)
 218                 return -EINPROGRESS;
 219
 220         dout("readpage inode %p file %p page %p index %lu\n",
 221              inode, filp, page, page->index);
 222         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 223                                   off, &len,
 224                                   ci->i_truncate_seq, ci->i_truncate_size,
 225                                   &page, 1, 0);
 226         if (err == -ENOENT)
 227                 err = 0;
 228         if (err < 0) {
 229                 SetPageError(page);
 230                 ceph_fscache_readpage_cancel(inode, page);
 231                 goto out;
 232         }
 233         if (err < PAGE_SIZE)
 234                 /* zero fill remainder of page */
 235                 zero_user_segment(page, err, PAGE_SIZE);
 236         else
 237                 flush_dcache_page(page);
 238
 239         SetPageUptodate(page);
 240         ceph_readpage_to_fscache(inode, page);
 241
 242 out:
 243         return err < 0 ? err : 0;
 244 }
 245
 246 static int ceph_readpage(struct file *filp, struct page *page)
 247 {
 248         int r = ceph_do_readpage(filp, page);
 249         if (r != -EINPROGRESS)
 250                 unlock_page(page);
 251         else
 252                 r = 0;
 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 ceph_rw_context *rw_ctx,
 305                       struct list_head *page_list, int max)
 306 {
 307         struct ceph_osd_client *osdc =
 308                 &ceph_inode_to_client(inode)->client->osdc;
 309         struct ceph_inode_info *ci = ceph_inode(inode);
 310         struct page *page = lru_to_page(page_list);
 311         struct ceph_vino vino;
 312         struct ceph_osd_request *req;
 313         u64 off;
 314         u64 len;
 315         int i;
 316         struct page **pages;
 317         pgoff_t next_index;
 318         int nr_pages = 0;
 319         int got = 0;
 320         int ret = 0;
 321
 322         if (!rw_ctx) {
 323                 /* caller of readpages does not hold buffer and read caps
 324                  * (fadvise, madvise and readahead cases) */
 325                 int want = CEPH_CAP_FILE_CACHE;
 326                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, true, &got);
 327                 if (ret < 0) {
 328                         dout("start_read %p, error getting cap\n", inode);
 329                 } else if (!(got & want)) {
 330                         dout("start_read %p, no cache cap\n", inode);
 331                         ret = 0;
 332                 }
 333                 if (ret <= 0) {
 334                         if (got)
 335                                 ceph_put_cap_refs(ci, got);
 336                         while (!list_empty(page_list)) {
 337                                 page = lru_to_page(page_list);
 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_array(nr_pages, sizeof(*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         struct ceph_file_info *fi = file->private_data;
 443         struct ceph_rw_context *rw_ctx;
 444         int rc = 0;
 445         int max = 0;
 446
 447         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 448                 return -EINVAL;
 449
 450         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 451                                          &nr_pages);
 452
 453         if (rc == 0)
 454                 goto out;
 455
 456         rw_ctx = ceph_find_rw_context(fi);
 457         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 458         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 459              inode, file, rw_ctx, nr_pages, max);
 460         while (!list_empty(page_list)) {
 461                 rc = start_read(inode, rw_ctx, page_list, max);
 462                 if (rc < 0)
 463                         goto out;
 464         }
 465 out:
 466         ceph_fscache_readpages_cancel(inode, page_list);
 467
 468         dout("readpages %p file %p ret %d\n", inode, file, rc);
 469         return rc;
 470 }
 471
 472 struct ceph_writeback_ctl
 473 {
 474         loff_t i_size;
 475         u64 truncate_size;
 476         u32 truncate_seq;
 477         bool size_stable;
 478         bool head_snapc;
 479 };
 480
 481 /*
 482  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 483  * only snap context we are allowed to write back.
 484  */
 485 static struct ceph_snap_context *
 486 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 487                    struct ceph_snap_context *page_snapc)
 488 {
 489         struct ceph_inode_info *ci = ceph_inode(inode);
 490         struct ceph_snap_context *snapc = NULL;
 491         struct ceph_cap_snap *capsnap = NULL;
 492
 493         spin_lock(&ci->i_ceph_lock);
 494         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 495                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 496                      capsnap->context, capsnap->dirty_pages);
 497                 if (!capsnap->dirty_pages)
 498                         continue;
 499
 500                 /* get i_size, truncate_{seq,size} for page_snapc? */
 501                 if (snapc && capsnap->context != page_snapc)
 502                         continue;
 503
 504                 if (ctl) {
 505                         if (capsnap->writing) {
 506                                 ctl->i_size = i_size_read(inode);
 507                                 ctl->size_stable = false;
 508                         } else {
 509                                 ctl->i_size = capsnap->size;
 510                                 ctl->size_stable = true;
 511                         }
 512                         ctl->truncate_size = capsnap->truncate_size;
 513                         ctl->truncate_seq = capsnap->truncate_seq;
 514                         ctl->head_snapc = false;
 515                 }
 516
 517                 if (snapc)
 518                         break;
 519
 520                 snapc = ceph_get_snap_context(capsnap->context);
 521                 if (!page_snapc ||
 522                     page_snapc == snapc ||
 523                     page_snapc->seq > snapc->seq)
 524                         break;
 525         }
 526         if (!snapc && ci->i_wrbuffer_ref_head) {
 527                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 528                 dout(" head snapc %p has %d dirty pages\n",
 529                      snapc, ci->i_wrbuffer_ref_head);
 530                 if (ctl) {
 531                         ctl->i_size = i_size_read(inode);
 532                         ctl->truncate_size = ci->i_truncate_size;
 533                         ctl->truncate_seq = ci->i_truncate_seq;
 534                         ctl->size_stable = false;
 535                         ctl->head_snapc = true;
 536                 }
 537         }
 538         spin_unlock(&ci->i_ceph_lock);
 539         return snapc;
 540 }
 541
 542 static u64 get_writepages_data_length(struct inode *inode,
 543                                       struct page *page, u64 start)
 544 {
 545         struct ceph_inode_info *ci = ceph_inode(inode);
 546         struct ceph_snap_context *snapc = page_snap_context(page);
 547         struct ceph_cap_snap *capsnap = NULL;
 548         u64 end = i_size_read(inode);
 549
 550         if (snapc != ci->i_head_snapc) {
 551                 bool found = false;
 552                 spin_lock(&ci->i_ceph_lock);
 553                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 554                         if (capsnap->context == snapc) {
 555                                 if (!capsnap->writing)
 556                                         end = capsnap->size;
 557                                 found = true;
 558                                 break;
 559                         }
 560                 }
 561                 spin_unlock(&ci->i_ceph_lock);
 562                 WARN_ON(!found);
 563         }
 564         if (end > page_offset(page) + PAGE_SIZE)
 565                 end = page_offset(page) + PAGE_SIZE;
 566         return end > start ? end - start : 0;
 567 }
 568
 569 /*
 570  * Write a single page, but leave the page locked.
 571  *
 572  * If we get a write error, set the page error bit, but still adjust the
 573  * dirty page accounting (i.e., page is no longer dirty).
 574  */
 575 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 576 {
 577         struct inode *inode;
 578         struct ceph_inode_info *ci;
 579         struct ceph_fs_client *fsc;
 580         struct ceph_snap_context *snapc, *oldest;
 581         loff_t page_off = page_offset(page);
 582         int err, len = PAGE_SIZE;
 583         struct ceph_writeback_ctl ceph_wbc;
 584
 585         dout("writepage %p idx %lu\n", page, page->index);
 586
 587         inode = page->mapping->host;
 588         ci = ceph_inode(inode);
 589         fsc = ceph_inode_to_client(inode);
 590
 591         /* verify this is a writeable snap context */
 592         snapc = page_snap_context(page);
 593         if (!snapc) {
 594                 dout("writepage %p page %p not dirty?\n", inode, page);
 595                 return 0;
 596         }
 597         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 598         if (snapc->seq > oldest->seq) {
 599                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 600                      inode, page, snapc);
 601                 /* we should only noop if called by kswapd */
 602                 WARN_ON(!(current->flags & PF_MEMALLOC));
 603                 ceph_put_snap_context(oldest);
 604                 redirty_page_for_writepage(wbc, page);
 605                 return 0;
 606         }
 607         ceph_put_snap_context(oldest);
 608
 609         /* is this a partial page at end of file? */
 610         if (page_off >= ceph_wbc.i_size) {
 611                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 612                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 613                 return 0;
 614         }
 615
 616         if (ceph_wbc.i_size < page_off + len)
 617                 len = ceph_wbc.i_size - page_off;
 618
 619         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 620              inode, page, page->index, page_off, len, snapc, snapc->seq);
 621
 622         if (atomic_long_inc_return(&fsc->writeback_count) >
 623             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 624                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 625
 626         set_page_writeback(page);
 627         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 628                                    &ci->i_layout, snapc, page_off, len,
 629                                    ceph_wbc.truncate_seq,
 630                                    ceph_wbc.truncate_size,
 631                                    &inode->i_mtime, &page, 1);
 632         if (err < 0) {
 633                 struct writeback_control tmp_wbc;
 634                 if (!wbc)
 635                         wbc = &tmp_wbc;
 636                 if (err == -ERESTARTSYS) {
 637                         /* killed by SIGKILL */
 638                         dout("writepage interrupted page %p\n", page);
 639                         redirty_page_for_writepage(wbc, page);
 640                         end_page_writeback(page);
 641                         return err;
 642                 }
 643                 dout("writepage setting page/mapping error %d %p\n",
 644                      err, page);
 645                 SetPageError(page);
 646                 mapping_set_error(&inode->i_data, err);
 647                 wbc->pages_skipped++;
 648         } else {
 649                 dout("writepage cleaned page %p\n", page);
 650                 err = 0;  /* vfs expects us to return 0 */
 651         }
 652         page->private = 0;
 653         ClearPagePrivate(page);
 654         end_page_writeback(page);
 655         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 656         ceph_put_snap_context(snapc);  /* page's reference */
 657
 658         if (atomic_long_dec_return(&fsc->writeback_count) <
 659             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 660                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 661
 662         return err;
 663 }
 664
 665 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 666 {
 667         int err;
 668         struct inode *inode = page->mapping->host;
 669         BUG_ON(!inode);
 670         ihold(inode);
 671         err = writepage_nounlock(page, wbc);
 672         if (err == -ERESTARTSYS) {
 673                 /* direct memory reclaimer was killed by SIGKILL. return 0
 674                  * to prevent caller from setting mapping/page error */
 675                 err = 0;
 676         }
 677         unlock_page(page);
 678         iput(inode);
 679         return err;
 680 }
 681
 682 /*
 683  * lame release_pages helper.  release_pages() isn't exported to
 684  * modules.
 685  */
 686 static void ceph_release_pages(struct page **pages, int num)
 687 {
 688         struct pagevec pvec;
 689         int i;
 690
 691         pagevec_init(&pvec);
 692         for (i = 0; i < num; i++) {
 693                 if (pagevec_add(&pvec, pages[i]) == 0)
 694                         pagevec_release(&pvec);
 695         }
 696         pagevec_release(&pvec);
 697 }
 698
 699 /*
 700  * async writeback completion handler.
 701  *
 702  * If we get an error, set the mapping error bit, but not the individual
 703  * page error bits.
 704  */
 705 static void writepages_finish(struct ceph_osd_request *req)
 706 {
 707         struct inode *inode = req->r_inode;
 708         struct ceph_inode_info *ci = ceph_inode(inode);
 709         struct ceph_osd_data *osd_data;
 710         struct page *page;
 711         int num_pages, total_pages = 0;
 712         int i, j;
 713         int rc = req->r_result;
 714         struct ceph_snap_context *snapc = req->r_snapc;
 715         struct address_space *mapping = inode->i_mapping;
 716         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 717         bool remove_page;
 718
 719         dout("writepages_finish %p rc %d\n", inode, rc);
 720         if (rc < 0) {
 721                 mapping_set_error(mapping, rc);
 722                 ceph_set_error_write(ci);
 723         } else {
 724                 ceph_clear_error_write(ci);
 725         }
 726
 727         /*
 728          * We lost the cache cap, need to truncate the page before
 729          * it is unlocked, otherwise we'd truncate it later in the
 730          * page truncation thread, possibly losing some data that
 731          * raced its way in
 732          */
 733         remove_page = !(ceph_caps_issued(ci) &
 734                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 735
 736         /* clean all pages */
 737         for (i = 0; i < req->r_num_ops; i++) {
 738                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 739                         break;
 740
 741                 osd_data = osd_req_op_extent_osd_data(req, i);
 742                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 743                 num_pages = calc_pages_for((u64)osd_data->alignment,
 744                                            (u64)osd_data->length);
 745                 total_pages += num_pages;
 746                 for (j = 0; j < num_pages; j++) {
 747                         page = osd_data->pages[j];
 748                         BUG_ON(!page);
 749                         WARN_ON(!PageUptodate(page));
 750
 751                         if (atomic_long_dec_return(&fsc->writeback_count) <
 752                              CONGESTION_OFF_THRESH(
 753                                         fsc->mount_options->congestion_kb))
 754                                 clear_bdi_congested(inode_to_bdi(inode),
 755                                                     BLK_RW_ASYNC);
 756
 757                         ceph_put_snap_context(page_snap_context(page));
 758                         page->private = 0;
 759                         ClearPagePrivate(page);
 760                         dout("unlocking %p\n", page);
 761                         end_page_writeback(page);
 762
 763                         if (remove_page)
 764                                 generic_error_remove_page(inode->i_mapping,
 765                                                           page);
 766
 767                         unlock_page(page);
 768                 }
 769                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 770                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 771
 772                 ceph_release_pages(osd_data->pages, num_pages);
 773         }
 774
 775         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 776
 777         osd_data = osd_req_op_extent_osd_data(req, 0);
 778         if (osd_data->pages_from_pool)
 779                 mempool_free(osd_data->pages,
 780                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 781         else
 782                 kfree(osd_data->pages);
 783         ceph_osdc_put_request(req);
 784 }
 785
 786 /*
 787  * initiate async writeback
 788  */
 789 static int ceph_writepages_start(struct address_space *mapping,
 790                                  struct writeback_control *wbc)
 791 {
 792         struct inode *inode = mapping->host;
 793         struct ceph_inode_info *ci = ceph_inode(inode);
 794         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 795         struct ceph_vino vino = ceph_vino(inode);
 796         pgoff_t index, start_index, end = -1;
 797         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 798         struct pagevec pvec;
 799         int rc = 0;
 800         unsigned int wsize = i_blocksize(inode);
 801         struct ceph_osd_request *req = NULL;
 802         struct ceph_writeback_ctl ceph_wbc;
 803         bool should_loop, range_whole = false;
 804         bool done = false;
 805
 806         dout("writepages_start %p (mode=%s)\n", inode,
 807              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 808              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 809
 810         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 811                 if (ci->i_wrbuffer_ref > 0) {
 812                         pr_warn_ratelimited(
 813                                 "writepage_start %p %lld forced umount\n",
 814                                 inode, ceph_ino(inode));
 815                 }
 816                 mapping_set_error(mapping, -EIO);
 817                 return -EIO; /* we're in a forced umount, don't write! */
 818         }
 819         if (fsc->mount_options->wsize < wsize)
 820                 wsize = fsc->mount_options->wsize;
 821
 822         pagevec_init(&pvec);
 823
 824         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 825         index = start_index;
 826
 827 retry:
 828         /* find oldest snap context with dirty data */
 829         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 830         if (!snapc) {
 831                 /* hmm, why does writepages get called when there
 832                    is no dirty data? */
 833                 dout(" no snap context with dirty data?\n");
 834                 goto out;
 835         }
 836         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 837              snapc, snapc->seq, snapc->num_snaps);
 838
 839         should_loop = false;
 840         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 841                 /* where to start/end? */
 842                 if (wbc->range_cyclic) {
 843                         index = start_index;
 844                         end = -1;
 845                         if (index > 0)
 846                                 should_loop = true;
 847                         dout(" cyclic, start at %lu\n", index);
 848                 } else {
 849                         index = wbc->range_start >> PAGE_SHIFT;
 850                         end = wbc->range_end >> PAGE_SHIFT;
 851                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 852                                 range_whole = true;
 853                         dout(" not cyclic, %lu to %lu\n", index, end);
 854                 }
 855         } else if (!ceph_wbc.head_snapc) {
 856                 /* Do not respect wbc->range_{start,end}. Dirty pages
 857                  * in that range can be associated with newer snapc.
 858                  * They are not writeable until we write all dirty pages
 859                  * associated with 'snapc' get written */
 860                 if (index > 0)
 861                         should_loop = true;
 862                 dout(" non-head snapc, range whole\n");
 863         }
 864
 865         ceph_put_snap_context(last_snapc);
 866         last_snapc = snapc;
 867
 868         while (!done && index <= end) {
 869                 int num_ops = 0, op_idx;
 870                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 871                 struct page **pages = NULL, **data_pages;
 872                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 873                 struct page *page;
 874                 pgoff_t strip_unit_end = 0;
 875                 u64 offset = 0, len = 0;
 876
 877                 max_pages = wsize >> PAGE_SHIFT;
 878
 879 get_more_pages:
 880                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 881                                                 end, PAGECACHE_TAG_DIRTY,
 882                                                 max_pages - locked_pages);
 883                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 884                 if (!pvec_pages && !locked_pages)
 885                         break;
 886                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 887                         page = pvec.pages[i];
 888                         dout("? %p idx %lu\n", page, page->index);
 889                         if (locked_pages == 0)
 890                                 lock_page(page);  /* first page */
 891                         else if (!trylock_page(page))
 892                                 break;
 893
 894                         /* only dirty pages, or our accounting breaks */
 895                         if (unlikely(!PageDirty(page)) ||
 896                             unlikely(page->mapping != mapping)) {
 897                                 dout("!dirty or !mapping %p\n", page);
 898                                 unlock_page(page);
 899                                 continue;
 900                         }
 901                         /* only if matching snap context */
 902                         pgsnapc = page_snap_context(page);
 903                         if (pgsnapc != snapc) {
 904                                 dout("page snapc %p %lld != oldest %p %lld\n",
 905                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 906                                 if (!should_loop &&
 907                                     !ceph_wbc.head_snapc &&
 908                                     wbc->sync_mode != WB_SYNC_NONE)
 909                                         should_loop = true;
 910                                 unlock_page(page);
 911                                 continue;
 912                         }
 913                         if (page_offset(page) >= ceph_wbc.i_size) {
 914                                 dout("%p page eof %llu\n",
 915                                      page, ceph_wbc.i_size);
 916                                 if ((ceph_wbc.size_stable ||
 917                                     page_offset(page) >= i_size_read(inode)) &&
 918                                     clear_page_dirty_for_io(page))
 919                                         mapping->a_ops->invalidatepage(page,
 920                                                                 0, PAGE_SIZE);
 921                                 unlock_page(page);
 922                                 continue;
 923                         }
 924                         if (strip_unit_end && (page->index > strip_unit_end)) {
 925                                 dout("end of strip unit %p\n", page);
 926                                 unlock_page(page);
 927                                 break;
 928                         }
 929                         if (PageWriteback(page)) {
 930                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 931                                         dout("%p under writeback\n", page);
 932                                         unlock_page(page);
 933                                         continue;
 934                                 }
 935                                 dout("waiting on writeback %p\n", page);
 936                                 wait_on_page_writeback(page);
 937                         }
 938
 939                         if (!clear_page_dirty_for_io(page)) {
 940                                 dout("%p !clear_page_dirty_for_io\n", page);
 941                                 unlock_page(page);
 942                                 continue;
 943                         }
 944
 945                         /*
 946                          * We have something to write.  If this is
 947                          * the first locked page this time through,
 948                          * calculate max possinle write size and
 949                          * allocate a page array
 950                          */
 951                         if (locked_pages == 0) {
 952                                 u64 objnum;
 953                                 u64 objoff;
 954                                 u32 xlen;
 955
 956                                 /* prepare async write request */
 957                                 offset = (u64)page_offset(page);
 958                                 ceph_calc_file_object_mapping(&ci->i_layout,
 959                                                               offset, wsize,
 960                                                               &objnum, &objoff,
 961                                                               &xlen);
 962                                 len = xlen;
 963
 964                                 num_ops = 1;
 965                                 strip_unit_end = page->index +
 966                                         ((len - 1) >> PAGE_SHIFT);
 967
 968                                 BUG_ON(pages);
 969                                 max_pages = calc_pages_for(0, (u64)len);
 970                                 pages = kmalloc_array(max_pages,
 971                                                       sizeof(*pages),
 972                                                       GFP_NOFS);
 973                                 if (!pages) {
 974                                         pool = fsc->wb_pagevec_pool;
 975                                         pages = mempool_alloc(pool, GFP_NOFS);
 976                                         BUG_ON(!pages);
 977                                 }
 978
 979                                 len = 0;
 980                         } else if (page->index !=
 981                                    (offset + len) >> PAGE_SHIFT) {
 982                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 983                                                         CEPH_OSD_MAX_OPS)) {
 984                                         redirty_page_for_writepage(wbc, page);
 985                                         unlock_page(page);
 986                                         break;
 987                                 }
 988
 989                                 num_ops++;
 990                                 offset = (u64)page_offset(page);
 991                                 len = 0;
 992                         }
 993
 994                         /* note position of first page in pvec */
 995                         dout("%p will write page %p idx %lu\n",
 996                              inode, page, page->index);
 997
 998                         if (atomic_long_inc_return(&fsc->writeback_count) >
 999                             CONGESTION_ON_THRESH(
1000                                     fsc->mount_options->congestion_kb)) {
1001                                 set_bdi_congested(inode_to_bdi(inode),
1002                                                   BLK_RW_ASYNC);
1003                         }
1004
1005
1006                         pages[locked_pages++] = page;
1007                         pvec.pages[i] = NULL;
1008
1009                         len += PAGE_SIZE;
1010                 }
1011
1012                 /* did we get anything? */
1013                 if (!locked_pages)
1014                         goto release_pvec_pages;
1015                 if (i) {
1016                         unsigned j, n = 0;
1017                         /* shift unused page to beginning of pvec */
1018                         for (j = 0; j < pvec_pages; j++) {
1019                                 if (!pvec.pages[j])
1020                                         continue;
1021                                 if (n < j)
1022                                         pvec.pages[n] = pvec.pages[j];
1023                                 n++;
1024                         }
1025                         pvec.nr = n;
1026
1027                         if (pvec_pages && i == pvec_pages &&
1028                             locked_pages < max_pages) {
1029                                 dout("reached end pvec, trying for more\n");
1030                                 pagevec_release(&pvec);
1031                                 goto get_more_pages;
1032                         }
1033                 }
1034
1035 new_request:
1036                 offset = page_offset(pages[0]);
1037                 len = wsize;
1038
1039                 req = ceph_osdc_new_request(&fsc->client->osdc,
1040                                         &ci->i_layout, vino,
1041                                         offset, &len, 0, num_ops,
1042                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1043                                         snapc, ceph_wbc.truncate_seq,
1044                                         ceph_wbc.truncate_size, false);
1045                 if (IS_ERR(req)) {
1046                         req = ceph_osdc_new_request(&fsc->client->osdc,
1047                                                 &ci->i_layout, vino,
1048                                                 offset, &len, 0,
1049                                                 min(num_ops,
1050                                                     CEPH_OSD_SLAB_OPS),
1051                                                 CEPH_OSD_OP_WRITE,
1052                                                 CEPH_OSD_FLAG_WRITE,
1053                                                 snapc, ceph_wbc.truncate_seq,
1054                                                 ceph_wbc.truncate_size, true);
1055                         BUG_ON(IS_ERR(req));
1056                 }
1057                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1058                              PAGE_SIZE - offset);
1059
1060                 req->r_callback = writepages_finish;
1061                 req->r_inode = inode;
1062
1063                 /* Format the osd request message and submit the write */
1064                 len = 0;
1065                 data_pages = pages;
1066                 op_idx = 0;
1067                 for (i = 0; i < locked_pages; i++) {
1068                         u64 cur_offset = page_offset(pages[i]);
1069                         if (offset + len != cur_offset) {
1070                                 if (op_idx + 1 == req->r_num_ops)
1071                                         break;
1072                                 osd_req_op_extent_dup_last(req, op_idx,
1073                                                            cur_offset - offset);
1074                                 dout("writepages got pages at %llu~%llu\n",
1075                                      offset, len);
1076                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1077                                                         data_pages, len, 0,
1078                                                         !!pool, false);
1079                                 osd_req_op_extent_update(req, op_idx, len);
1080
1081                                 len = 0;
1082                                 offset = cur_offset;
1083                                 data_pages = pages + i;
1084                                 op_idx++;
1085                         }
1086
1087                         set_page_writeback(pages[i]);
1088                         len += PAGE_SIZE;
1089                 }
1090
1091                 if (ceph_wbc.size_stable) {
1092                         len = min(len, ceph_wbc.i_size - offset);
1093                 } else if (i == locked_pages) {
1094                         /* writepages_finish() clears writeback pages
1095                          * according to the data length, so make sure
1096                          * data length covers all locked pages */
1097                         u64 min_len = len + 1 - PAGE_SIZE;
1098                         len = get_writepages_data_length(inode, pages[i - 1],
1099                                                          offset);
1100                         len = max(len, min_len);
1101                 }
1102                 dout("writepages got pages at %llu~%llu\n", offset, len);
1103
1104                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1105                                                  0, !!pool, false);
1106                 osd_req_op_extent_update(req, op_idx, len);
1107
1108                 BUG_ON(op_idx + 1 != req->r_num_ops);
1109
1110                 pool = NULL;
1111                 if (i < locked_pages) {
1112                         BUG_ON(num_ops <= req->r_num_ops);
1113                         num_ops -= req->r_num_ops;
1114                         locked_pages -= i;
1115
1116                         /* allocate new pages array for next request */
1117                         data_pages = pages;
1118                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1119                                               GFP_NOFS);
1120                         if (!pages) {
1121                                 pool = fsc->wb_pagevec_pool;
1122                                 pages = mempool_alloc(pool, GFP_NOFS);
1123                                 BUG_ON(!pages);
1124                         }
1125                         memcpy(pages, data_pages + i,
1126                                locked_pages * sizeof(*pages));
1127                         memset(data_pages + i, 0,
1128                                locked_pages * sizeof(*pages));
1129                 } else {
1130                         BUG_ON(num_ops != req->r_num_ops);
1131                         index = pages[i - 1]->index + 1;
1132                         /* request message now owns the pages array */
1133                         pages = NULL;
1134                 }
1135
1136                 req->r_mtime = inode->i_mtime;
1137                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1138                 BUG_ON(rc);
1139                 req = NULL;
1140
1141                 wbc->nr_to_write -= i;
1142                 if (pages)
1143                         goto new_request;
1144
1145                 /*
1146                  * We stop writing back only if we are not doing
1147                  * integrity sync. In case of integrity sync we have to
1148                  * keep going until we have written all the pages
1149                  * we tagged for writeback prior to entering this loop.
1150                  */
1151                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1152                         done = true;
1153
1154 release_pvec_pages:
1155                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1156                      pvec.nr ? pvec.pages[0] : NULL);
1157                 pagevec_release(&pvec);
1158         }
1159
1160         if (should_loop && !done) {
1161                 /* more to do; loop back to beginning of file */
1162                 dout("writepages looping back to beginning of file\n");
1163                 end = start_index - 1; /* OK even when start_index == 0 */
1164
1165                 /* to write dirty pages associated with next snapc,
1166                  * we need to wait until current writes complete */
1167                 if (wbc->sync_mode != WB_SYNC_NONE &&
1168                     start_index == 0 && /* all dirty pages were checked */
1169                     !ceph_wbc.head_snapc) {
1170                         struct page *page;
1171                         unsigned i, nr;
1172                         index = 0;
1173                         while ((index <= end) &&
1174                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1175                                                 PAGECACHE_TAG_WRITEBACK))) {
1176                                 for (i = 0; i < nr; i++) {
1177                                         page = pvec.pages[i];
1178                                         if (page_snap_context(page) != snapc)
1179                                                 continue;
1180                                         wait_on_page_writeback(page);
1181                                 }
1182                                 pagevec_release(&pvec);
1183                                 cond_resched();
1184                         }
1185                 }
1186
1187                 start_index = 0;
1188                 index = 0;
1189                 goto retry;
1190         }
1191
1192         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1193                 mapping->writeback_index = index;
1194
1195 out:
1196         ceph_osdc_put_request(req);
1197         ceph_put_snap_context(last_snapc);
1198         dout("writepages dend - startone, rc = %d\n", rc);
1199         return rc;
1200 }
1201
1202
1203
1204 /*
1205  * See if a given @snapc is either writeable, or already written.
1206  */
1207 static int context_is_writeable_or_written(struct inode *inode,
1208                                            struct ceph_snap_context *snapc)
1209 {
1210         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1211         int ret = !oldest || snapc->seq <= oldest->seq;
1212
1213         ceph_put_snap_context(oldest);
1214         return ret;
1215 }
1216
1217 /*
1218  * We are only allowed to write into/dirty the page if the page is
1219  * clean, or already dirty within the same snap context.
1220  *
1221  * called with page locked.
1222  * return success with page locked,
1223  * or any failure (incl -EAGAIN) with page unlocked.
1224  */
1225 static int ceph_update_writeable_page(struct file *file,
1226                             loff_t pos, unsigned len,
1227                             struct page *page)
1228 {
1229         struct inode *inode = file_inode(file);
1230         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1231         struct ceph_inode_info *ci = ceph_inode(inode);
1232         loff_t page_off = pos & PAGE_MASK;
1233         int pos_in_page = pos & ~PAGE_MASK;
1234         int end_in_page = pos_in_page + len;
1235         loff_t i_size;
1236         int r;
1237         struct ceph_snap_context *snapc, *oldest;
1238
1239         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1240                 dout(" page %p forced umount\n", page);
1241                 unlock_page(page);
1242                 return -EIO;
1243         }
1244
1245 retry_locked:
1246         /* writepages currently holds page lock, but if we change that later, */
1247         wait_on_page_writeback(page);
1248
1249         snapc = page_snap_context(page);
1250         if (snapc && snapc != ci->i_head_snapc) {
1251                 /*
1252                  * this page is already dirty in another (older) snap
1253                  * context!  is it writeable now?
1254                  */
1255                 oldest = get_oldest_context(inode, NULL, NULL);
1256                 if (snapc->seq > oldest->seq) {
1257                         ceph_put_snap_context(oldest);
1258                         dout(" page %p snapc %p not current or oldest\n",
1259                              page, snapc);
1260                         /*
1261                          * queue for writeback, and wait for snapc to
1262                          * be writeable or written
1263                          */
1264                         snapc = ceph_get_snap_context(snapc);
1265                         unlock_page(page);
1266                         ceph_queue_writeback(inode);
1267                         r = wait_event_killable(ci->i_cap_wq,
1268                                context_is_writeable_or_written(inode, snapc));
1269                         ceph_put_snap_context(snapc);
1270                         if (r == -ERESTARTSYS)
1271                                 return r;
1272                         return -EAGAIN;
1273                 }
1274                 ceph_put_snap_context(oldest);
1275
1276                 /* yay, writeable, do it now (without dropping page lock) */
1277                 dout(" page %p snapc %p not current, but oldest\n",
1278                      page, snapc);
1279                 if (!clear_page_dirty_for_io(page))
1280                         goto retry_locked;
1281                 r = writepage_nounlock(page, NULL);
1282                 if (r < 0)
1283                         goto fail_unlock;
1284                 goto retry_locked;
1285         }
1286
1287         if (PageUptodate(page)) {
1288                 dout(" page %p already uptodate\n", page);
1289                 return 0;
1290         }
1291
1292         /* full page? */
1293         if (pos_in_page == 0 && len == PAGE_SIZE)
1294                 return 0;
1295
1296         /* past end of file? */
1297         i_size = i_size_read(inode);
1298
1299         if (page_off >= i_size ||
1300             (pos_in_page == 0 && (pos+len) >= i_size &&
1301              end_in_page - pos_in_page != PAGE_SIZE)) {
1302                 dout(" zeroing %p 0 - %d and %d - %d\n",
1303                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1304                 zero_user_segments(page,
1305                                    0, pos_in_page,
1306                                    end_in_page, PAGE_SIZE);
1307                 return 0;
1308         }
1309
1310         /* we need to read it. */
1311         r = ceph_do_readpage(file, page);
1312         if (r < 0) {
1313                 if (r == -EINPROGRESS)
1314                         return -EAGAIN;
1315                 goto fail_unlock;
1316         }
1317         goto retry_locked;
1318 fail_unlock:
1319         unlock_page(page);
1320         return r;
1321 }
1322
1323 /*
1324  * We are only allowed to write into/dirty the page if the page is
1325  * clean, or already dirty within the same snap context.
1326  */
1327 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1328                             loff_t pos, unsigned len, unsigned flags,
1329                             struct page **pagep, void **fsdata)
1330 {
1331         struct inode *inode = file_inode(file);
1332         struct page *page;
1333         pgoff_t index = pos >> PAGE_SHIFT;
1334         int r;
1335
1336         do {
1337                 /* get a page */
1338                 page = grab_cache_page_write_begin(mapping, index, 0);
1339                 if (!page)
1340                         return -ENOMEM;
1341
1342                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1343                      inode, page, (int)pos, (int)len);
1344
1345                 r = ceph_update_writeable_page(file, pos, len, page);
1346                 if (r < 0)
1347                         put_page(page);
1348                 else
1349                         *pagep = page;
1350         } while (r == -EAGAIN);
1351
1352         return r;
1353 }
1354
1355 /*
1356  * we don't do anything in here that simple_write_end doesn't do
1357  * except adjust dirty page accounting
1358  */
1359 static int ceph_write_end(struct file *file, struct address_space *mapping,
1360                           loff_t pos, unsigned len, unsigned copied,
1361                           struct page *page, void *fsdata)
1362 {
1363         struct inode *inode = file_inode(file);
1364         bool check_cap = false;
1365
1366         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1367              inode, page, (int)pos, (int)copied, (int)len);
1368
1369         /* zero the stale part of the page if we did a short copy */
1370         if (!PageUptodate(page)) {
1371                 if (copied < len) {
1372                         copied = 0;
1373                         goto out;
1374                 }
1375                 SetPageUptodate(page);
1376         }
1377
1378         /* did file size increase? */
1379         if (pos+copied > i_size_read(inode))
1380                 check_cap = ceph_inode_set_size(inode, pos+copied);
1381
1382         set_page_dirty(page);
1383
1384 out:
1385         unlock_page(page);
1386         put_page(page);
1387
1388         if (check_cap)
1389                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1390
1391         return copied;
1392 }
1393
1394 /*
1395  * we set .direct_IO to indicate direct io is supported, but since we
1396  * intercept O_DIRECT reads and writes early, this function should
1397  * never get called.
1398  */
1399 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1400 {
1401         WARN_ON(1);
1402         return -EINVAL;
1403 }
1404
1405 const struct address_space_operations ceph_aops = {
1406         .readpage = ceph_readpage,
1407         .readpages = ceph_readpages,
1408         .writepage = ceph_writepage,
1409         .writepages = ceph_writepages_start,
1410         .write_begin = ceph_write_begin,
1411         .write_end = ceph_write_end,
1412         .set_page_dirty = ceph_set_page_dirty,
1413         .invalidatepage = ceph_invalidatepage,
1414         .releasepage = ceph_releasepage,
1415         .direct_IO = ceph_direct_io,
1416 };
1417
1418 static void ceph_block_sigs(sigset_t *oldset)
1419 {
1420         sigset_t mask;
1421         siginitsetinv(&mask, sigmask(SIGKILL));
1422         sigprocmask(SIG_BLOCK, &mask, oldset);
1423 }
1424
1425 static void ceph_restore_sigs(sigset_t *oldset)
1426 {
1427         sigprocmask(SIG_SETMASK, oldset, NULL);
1428 }
1429
1430 /*
1431  * vm ops
1432  */
1433 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1434 {
1435         struct vm_area_struct *vma = vmf->vma;
1436         struct inode *inode = file_inode(vma->vm_file);
1437         struct ceph_inode_info *ci = ceph_inode(inode);
1438         struct ceph_file_info *fi = vma->vm_file->private_data;
1439         struct page *pinned_page = NULL;
1440         loff_t off = vmf->pgoff << PAGE_SHIFT;
1441         int want, got, err;
1442         sigset_t oldset;
1443         vm_fault_t ret = VM_FAULT_SIGBUS;
1444
1445         ceph_block_sigs(&oldset);
1446
1447         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1448              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1449         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1450                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1451         else
1452                 want = CEPH_CAP_FILE_CACHE;
1453
1454         got = 0;
1455         err = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1456         if (err < 0)
1457                 goto out_restore;
1458
1459         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1460              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1461
1462         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1463             ci->i_inline_version == CEPH_INLINE_NONE) {
1464                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1465                 ceph_add_rw_context(fi, &rw_ctx);
1466                 ret = filemap_fault(vmf);
1467                 ceph_del_rw_context(fi, &rw_ctx);
1468                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1469                         inode, off, (size_t)PAGE_SIZE,
1470                                 ceph_cap_string(got), ret);
1471         } else
1472                 err = -EAGAIN;
1473
1474         if (pinned_page)
1475                 put_page(pinned_page);
1476         ceph_put_cap_refs(ci, got);
1477
1478         if (err != -EAGAIN)
1479                 goto out_restore;
1480
1481         /* read inline data */
1482         if (off >= PAGE_SIZE) {
1483                 /* does not support inline data > PAGE_SIZE */
1484                 ret = VM_FAULT_SIGBUS;
1485         } else {
1486                 struct address_space *mapping = inode->i_mapping;
1487                 struct page *page = find_or_create_page(mapping, 0,
1488                                                 mapping_gfp_constraint(mapping,
1489                                                 ~__GFP_FS));
1490                 if (!page) {
1491                         ret = VM_FAULT_OOM;
1492                         goto out_inline;
1493                 }
1494                 err = __ceph_do_getattr(inode, page,
1495                                          CEPH_STAT_CAP_INLINE_DATA, true);
1496                 if (err < 0 || off >= i_size_read(inode)) {
1497                         unlock_page(page);
1498                         put_page(page);
1499                         ret = vmf_error(err);
1500                         goto out_inline;
1501                 }
1502                 if (err < PAGE_SIZE)
1503                         zero_user_segment(page, err, PAGE_SIZE);
1504                 else
1505                         flush_dcache_page(page);
1506                 SetPageUptodate(page);
1507                 vmf->page = page;
1508                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1509 out_inline:
1510                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1511                      inode, off, (size_t)PAGE_SIZE, ret);
1512         }
1513 out_restore:
1514         ceph_restore_sigs(&oldset);
1515         if (err < 0)
1516                 ret = vmf_error(err);
1517
1518         return ret;
1519 }
1520
1521 /*
1522  * Reuse write_begin here for simplicity.
1523  */
1524 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1525 {
1526         struct vm_area_struct *vma = vmf->vma;
1527         struct inode *inode = file_inode(vma->vm_file);
1528         struct ceph_inode_info *ci = ceph_inode(inode);
1529         struct ceph_file_info *fi = vma->vm_file->private_data;
1530         struct ceph_cap_flush *prealloc_cf;
1531         struct page *page = vmf->page;
1532         loff_t off = page_offset(page);
1533         loff_t size = i_size_read(inode);
1534         size_t len;
1535         int want, got, err;
1536         sigset_t oldset;
1537         vm_fault_t ret = VM_FAULT_SIGBUS;
1538
1539         prealloc_cf = ceph_alloc_cap_flush();
1540         if (!prealloc_cf)
1541                 return VM_FAULT_OOM;
1542
1543         ceph_block_sigs(&oldset);
1544
1545         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1546                 struct page *locked_page = NULL;
1547                 if (off == 0) {
1548                         lock_page(page);
1549                         locked_page = page;
1550                 }
1551                 err = ceph_uninline_data(vma->vm_file, locked_page);
1552                 if (locked_page)
1553                         unlock_page(locked_page);
1554                 if (err < 0)
1555                         goto out_free;
1556         }
1557
1558         if (off + PAGE_SIZE <= size)
1559                 len = PAGE_SIZE;
1560         else
1561                 len = size & ~PAGE_MASK;
1562
1563         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1564              inode, ceph_vinop(inode), off, len, size);
1565         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1566                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1567         else
1568                 want = CEPH_CAP_FILE_BUFFER;
1569
1570         got = 0;
1571         err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1572                             &got, NULL);
1573         if (err < 0)
1574                 goto out_free;
1575
1576         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1577              inode, off, len, ceph_cap_string(got));
1578
1579         /* Update time before taking page lock */
1580         file_update_time(vma->vm_file);
1581         inode_inc_iversion_raw(inode);
1582
1583         do {
1584                 lock_page(page);
1585
1586                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1587                         unlock_page(page);
1588                         ret = VM_FAULT_NOPAGE;
1589                         break;
1590                 }
1591
1592                 err = ceph_update_writeable_page(vma->vm_file, off, len, page);
1593                 if (err >= 0) {
1594                         /* success.  we'll keep the page locked. */
1595                         set_page_dirty(page);
1596                         ret = VM_FAULT_LOCKED;
1597                 }
1598         } while (err == -EAGAIN);
1599
1600         if (ret == VM_FAULT_LOCKED ||
1601             ci->i_inline_version != CEPH_INLINE_NONE) {
1602                 int dirty;
1603                 spin_lock(&ci->i_ceph_lock);
1604                 ci->i_inline_version = CEPH_INLINE_NONE;
1605                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1606                                                &prealloc_cf);
1607                 spin_unlock(&ci->i_ceph_lock);
1608                 if (dirty)
1609                         __mark_inode_dirty(inode, dirty);
1610         }
1611
1612         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1613              inode, off, len, ceph_cap_string(got), ret);
1614         ceph_put_cap_refs(ci, got);
1615 out_free:
1616         ceph_restore_sigs(&oldset);
1617         ceph_free_cap_flush(prealloc_cf);
1618         if (err < 0)
1619                 ret = vmf_error(err);
1620         return ret;
1621 }
1622
1623 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1624                            char *data, size_t len)
1625 {
1626         struct address_space *mapping = inode->i_mapping;
1627         struct page *page;
1628
1629         if (locked_page) {
1630                 page = locked_page;
1631         } else {
1632                 if (i_size_read(inode) == 0)
1633                         return;
1634                 page = find_or_create_page(mapping, 0,
1635                                            mapping_gfp_constraint(mapping,
1636                                            ~__GFP_FS));
1637                 if (!page)
1638                         return;
1639                 if (PageUptodate(page)) {
1640                         unlock_page(page);
1641                         put_page(page);
1642                         return;
1643                 }
1644         }
1645
1646         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1647              inode, ceph_vinop(inode), len, locked_page);
1648
1649         if (len > 0) {
1650                 void *kaddr = kmap_atomic(page);
1651                 memcpy(kaddr, data, len);
1652                 kunmap_atomic(kaddr);
1653         }
1654
1655         if (page != locked_page) {
1656                 if (len < PAGE_SIZE)
1657                         zero_user_segment(page, len, PAGE_SIZE);
1658                 else
1659                         flush_dcache_page(page);
1660
1661                 SetPageUptodate(page);
1662                 unlock_page(page);
1663                 put_page(page);
1664         }
1665 }
1666
1667 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1668 {
1669         struct inode *inode = file_inode(filp);
1670         struct ceph_inode_info *ci = ceph_inode(inode);
1671         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1672         struct ceph_osd_request *req;
1673         struct page *page = NULL;
1674         u64 len, inline_version;
1675         int err = 0;
1676         bool from_pagecache = false;
1677
1678         spin_lock(&ci->i_ceph_lock);
1679         inline_version = ci->i_inline_version;
1680         spin_unlock(&ci->i_ceph_lock);
1681
1682         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1683              inode, ceph_vinop(inode), inline_version);
1684
1685         if (inline_version == 1 || /* initial version, no data */
1686             inline_version == CEPH_INLINE_NONE)
1687                 goto out;
1688
1689         if (locked_page) {
1690                 page = locked_page;
1691                 WARN_ON(!PageUptodate(page));
1692         } else if (ceph_caps_issued(ci) &
1693                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1694                 page = find_get_page(inode->i_mapping, 0);
1695                 if (page) {
1696                         if (PageUptodate(page)) {
1697                                 from_pagecache = true;
1698                                 lock_page(page);
1699                         } else {
1700                                 put_page(page);
1701                                 page = NULL;
1702                         }
1703                 }
1704         }
1705
1706         if (page) {
1707                 len = i_size_read(inode);
1708                 if (len > PAGE_SIZE)
1709                         len = PAGE_SIZE;
1710         } else {
1711                 page = __page_cache_alloc(GFP_NOFS);
1712                 if (!page) {
1713                         err = -ENOMEM;
1714                         goto out;
1715                 }
1716                 err = __ceph_do_getattr(inode, page,
1717                                         CEPH_STAT_CAP_INLINE_DATA, true);
1718                 if (err < 0) {
1719                         /* no inline data */
1720                         if (err == -ENODATA)
1721                                 err = 0;
1722                         goto out;
1723                 }
1724                 len = err;
1725         }
1726
1727         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1728                                     ceph_vino(inode), 0, &len, 0, 1,
1729                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1730                                     NULL, 0, 0, false);
1731         if (IS_ERR(req)) {
1732                 err = PTR_ERR(req);
1733                 goto out;
1734         }
1735
1736         req->r_mtime = inode->i_mtime;
1737         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1738         if (!err)
1739                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1740         ceph_osdc_put_request(req);
1741         if (err < 0)
1742                 goto out;
1743
1744         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1745                                     ceph_vino(inode), 0, &len, 1, 3,
1746                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1747                                     NULL, ci->i_truncate_seq,
1748                                     ci->i_truncate_size, false);
1749         if (IS_ERR(req)) {
1750                 err = PTR_ERR(req);
1751                 goto out;
1752         }
1753
1754         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1755
1756         {
1757                 __le64 xattr_buf = cpu_to_le64(inline_version);
1758                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1759                                             "inline_version", &xattr_buf,
1760                                             sizeof(xattr_buf),
1761                                             CEPH_OSD_CMPXATTR_OP_GT,
1762                                             CEPH_OSD_CMPXATTR_MODE_U64);
1763                 if (err)
1764                         goto out_put;
1765         }
1766
1767         {
1768                 char xattr_buf[32];
1769                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1770                                          "%llu", inline_version);
1771                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1772                                             "inline_version",
1773                                             xattr_buf, xattr_len, 0, 0);
1774                 if (err)
1775                         goto out_put;
1776         }
1777
1778         req->r_mtime = inode->i_mtime;
1779         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1780         if (!err)
1781                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1782 out_put:
1783         ceph_osdc_put_request(req);
1784         if (err == -ECANCELED)
1785                 err = 0;
1786 out:
1787         if (page && page != locked_page) {
1788                 if (from_pagecache) {
1789                         unlock_page(page);
1790                         put_page(page);
1791                 } else
1792                         __free_pages(page, 0);
1793         }
1794
1795         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1796              inode, ceph_vinop(inode), inline_version, err);
1797         return err;
1798 }
1799
1800 static const struct vm_operations_struct ceph_vmops = {
1801         .fault          = ceph_filemap_fault,
1802         .page_mkwrite   = ceph_page_mkwrite,
1803 };
1804
1805 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1806 {
1807         struct address_space *mapping = file->f_mapping;
1808
1809         if (!mapping->a_ops->readpage)
1810                 return -ENOEXEC;
1811         file_accessed(file);
1812         vma->vm_ops = &ceph_vmops;
1813         return 0;
1814 }
1815
1816 enum {
1817         POOL_READ       = 1,
1818         POOL_WRITE      = 2,
1819 };
1820
1821 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1822                                 s64 pool, struct ceph_string *pool_ns)
1823 {
1824         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1825         struct ceph_mds_client *mdsc = fsc->mdsc;
1826         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1827         struct rb_node **p, *parent;
1828         struct ceph_pool_perm *perm;
1829         struct page **pages;
1830         size_t pool_ns_len;
1831         int err = 0, err2 = 0, have = 0;
1832
1833         down_read(&mdsc->pool_perm_rwsem);
1834         p = &mdsc->pool_perm_tree.rb_node;
1835         while (*p) {
1836                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1837                 if (pool < perm->pool)
1838                         p = &(*p)->rb_left;
1839                 else if (pool > perm->pool)
1840                         p = &(*p)->rb_right;
1841                 else {
1842                         int ret = ceph_compare_string(pool_ns,
1843                                                 perm->pool_ns,
1844                                                 perm->pool_ns_len);
1845                         if (ret < 0)
1846                                 p = &(*p)->rb_left;
1847                         else if (ret > 0)
1848                                 p = &(*p)->rb_right;
1849                         else {
1850                                 have = perm->perm;
1851                                 break;
1852                         }
1853                 }
1854         }
1855         up_read(&mdsc->pool_perm_rwsem);
1856         if (*p)
1857                 goto out;
1858
1859         if (pool_ns)
1860                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1861                      pool, (int)pool_ns->len, pool_ns->str);
1862         else
1863                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1864
1865         down_write(&mdsc->pool_perm_rwsem);
1866         p = &mdsc->pool_perm_tree.rb_node;
1867         parent = NULL;
1868         while (*p) {
1869                 parent = *p;
1870                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1871                 if (pool < perm->pool)
1872                         p = &(*p)->rb_left;
1873                 else if (pool > perm->pool)
1874                         p = &(*p)->rb_right;
1875                 else {
1876                         int ret = ceph_compare_string(pool_ns,
1877                                                 perm->pool_ns,
1878                                                 perm->pool_ns_len);
1879                         if (ret < 0)
1880                                 p = &(*p)->rb_left;
1881                         else if (ret > 0)
1882                                 p = &(*p)->rb_right;
1883                         else {
1884                                 have = perm->perm;
1885                                 break;
1886                         }
1887                 }
1888         }
1889         if (*p) {
1890                 up_write(&mdsc->pool_perm_rwsem);
1891                 goto out;
1892         }
1893
1894         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1895                                          1, false, GFP_NOFS);
1896         if (!rd_req) {
1897                 err = -ENOMEM;
1898                 goto out_unlock;
1899         }
1900
1901         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1902         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1903         rd_req->r_base_oloc.pool = pool;
1904         if (pool_ns)
1905                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1906         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1907
1908         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1909         if (err)
1910                 goto out_unlock;
1911
1912         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1913                                          1, false, GFP_NOFS);
1914         if (!wr_req) {
1915                 err = -ENOMEM;
1916                 goto out_unlock;
1917         }
1918
1919         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1920         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1921         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1922         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1923
1924         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1925         if (err)
1926                 goto out_unlock;
1927
1928         /* one page should be large enough for STAT data */
1929         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1930         if (IS_ERR(pages)) {
1931                 err = PTR_ERR(pages);
1932                 goto out_unlock;
1933         }
1934
1935         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1936                                      0, false, true);
1937         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1938
1939         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1940         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1941
1942         if (!err)
1943                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1944         if (!err2)
1945                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1946
1947         if (err >= 0 || err == -ENOENT)
1948                 have |= POOL_READ;
1949         else if (err != -EPERM)
1950                 goto out_unlock;
1951
1952         if (err2 == 0 || err2 == -EEXIST)
1953                 have |= POOL_WRITE;
1954         else if (err2 != -EPERM) {
1955                 err = err2;
1956                 goto out_unlock;
1957         }
1958
1959         pool_ns_len = pool_ns ? pool_ns->len : 0;
1960         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1961         if (!perm) {
1962                 err = -ENOMEM;
1963                 goto out_unlock;
1964         }
1965
1966         perm->pool = pool;
1967         perm->perm = have;
1968         perm->pool_ns_len = pool_ns_len;
1969         if (pool_ns_len > 0)
1970                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1971         perm->pool_ns[pool_ns_len] = 0;
1972
1973         rb_link_node(&perm->node, parent, p);
1974         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1975         err = 0;
1976 out_unlock:
1977         up_write(&mdsc->pool_perm_rwsem);
1978
1979         ceph_osdc_put_request(rd_req);
1980         ceph_osdc_put_request(wr_req);
1981 out:
1982         if (!err)
1983                 err = have;
1984         if (pool_ns)
1985                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1986                      pool, (int)pool_ns->len, pool_ns->str, err);
1987         else
1988                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1989         return err;
1990 }
1991
1992 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1993 {
1994         s64 pool;
1995         struct ceph_string *pool_ns;
1996         int ret, flags;
1997
1998         if (ci->i_vino.snap != CEPH_NOSNAP) {
1999                 /*
2000                  * Pool permission check needs to write to the first object.
2001                  * But for snapshot, head of the first object may have alread
2002                  * been deleted. Skip check to avoid creating orphan object.
2003                  */
2004                 return 0;
2005         }
2006
2007         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2008                                 NOPOOLPERM))
2009                 return 0;
2010
2011         spin_lock(&ci->i_ceph_lock);
2012         flags = ci->i_ceph_flags;
2013         pool = ci->i_layout.pool_id;
2014         spin_unlock(&ci->i_ceph_lock);
2015 check:
2016         if (flags & CEPH_I_POOL_PERM) {
2017                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2018                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2019                              pool);
2020                         return -EPERM;
2021                 }
2022                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2023                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2024                              pool);
2025                         return -EPERM;
2026                 }
2027                 return 0;
2028         }
2029
2030         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2031         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2032         ceph_put_string(pool_ns);
2033         if (ret < 0)
2034                 return ret;
2035
2036         flags = CEPH_I_POOL_PERM;
2037         if (ret & POOL_READ)
2038                 flags |= CEPH_I_POOL_RD;
2039         if (ret & POOL_WRITE)
2040                 flags |= CEPH_I_POOL_WR;
2041
2042         spin_lock(&ci->i_ceph_lock);
2043         if (pool == ci->i_layout.pool_id &&
2044             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2045                 ci->i_ceph_flags |= flags;
2046         } else {
2047                 pool = ci->i_layout.pool_id;
2048                 flags = ci->i_ceph_flags;
2049         }
2050         spin_unlock(&ci->i_ceph_lock);
2051         goto check;
2052 }
2053
2054 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2055 {
2056         struct ceph_pool_perm *perm;
2057         struct rb_node *n;
2058
2059         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2060                 n = rb_first(&mdsc->pool_perm_tree);
2061                 perm = rb_entry(n, struct ceph_pool_perm, node);
2062                 rb_erase(n, &mdsc->pool_perm_tree);
2063                 kfree(perm);
2064         }
2065 }