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
  14 #include "super.h"
  15 #include "mds_client.h"
  16 #include "cache.h"
  17 #include <linux/ceph/osd_client.h>
  18 #include <linux/ceph/striper.h>
  19
  20 /*
  21  * Ceph address space ops.
  22  *
  23  * There are a few funny things going on here.
  24  *
  25  * The page->private field is used to reference a struct
  26  * ceph_snap_context for _every_ dirty page.  This indicates which
  27  * snapshot the page was logically dirtied in, and thus which snap
  28  * context needs to be associated with the osd write during writeback.
  29  *
  30  * Similarly, struct ceph_inode_info maintains a set of counters to
  31  * count dirty pages on the inode.  In the absence of snapshots,
  32  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  33  *
  34  * When a snapshot is taken (that is, when the client receives
  35  * notification that a snapshot was taken), each inode with caps and
  36  * with dirty pages (dirty pages implies there is a cap) gets a new
  37  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  38  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  39  * moved to capsnap->dirty. (Unless a sync write is currently in
  40  * progress.  In that case, the capsnap is said to be "pending", new
  41  * writes cannot start, and the capsnap isn't "finalized" until the
  42  * write completes (or fails) and a final size/mtime for the inode for
  43  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  44  *
  45  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  46  * we look for the first capsnap in i_cap_snaps and write out pages in
  47  * that snap context _only_.  Then we move on to the next capsnap,
  48  * eventually reaching the "live" or "head" context (i.e., pages that
  49  * are not yet snapped) and are writing the most recently dirtied
  50  * pages.
  51  *
  52  * Invalidate and so forth must take care to ensure the dirty page
  53  * accounting is preserved.
  54  */
  55
  56 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  57 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  58         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  59          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  60
  61 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  62 {
  63         if (PagePrivate(page))
  64                 return (void *)page->private;
  65         return NULL;
  66 }
  67
  68 /*
  69  * Dirty a page.  Optimistically adjust accounting, on the assumption
  70  * that we won't race with invalidate.  If we do, readjust.
  71  */
  72 static int ceph_set_page_dirty(struct page *page)
  73 {
  74         struct address_space *mapping = page->mapping;
  75         struct inode *inode;
  76         struct ceph_inode_info *ci;
  77         struct ceph_snap_context *snapc;
  78
  79         if (PageDirty(page)) {
  80                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  81                      mapping->host, page, page->index);
  82                 BUG_ON(!PagePrivate(page));
  83                 return 0;
  84         }
  85
  86         inode = mapping->host;
  87         ci = ceph_inode(inode);
  88
  89         /* dirty the head */
  90         spin_lock(&ci->i_ceph_lock);
  91         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  92         if (__ceph_have_pending_cap_snap(ci)) {
  93                 struct ceph_cap_snap *capsnap =
  94                                 list_last_entry(&ci->i_cap_snaps,
  95                                                 struct ceph_cap_snap,
  96                                                 ci_item);
  97                 snapc = ceph_get_snap_context(capsnap->context);
  98                 capsnap->dirty_pages++;
  99         } else {
 100                 BUG_ON(!ci->i_head_snapc);
 101                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 102                 ++ci->i_wrbuffer_ref_head;
 103         }
 104         if (ci->i_wrbuffer_ref == 0)
 105                 ihold(inode);
 106         ++ci->i_wrbuffer_ref;
 107         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 108              "snapc %p seq %lld (%d snaps)\n",
 109              mapping->host, page, page->index,
 110              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 111              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 112              snapc, snapc->seq, snapc->num_snaps);
 113         spin_unlock(&ci->i_ceph_lock);
 114
 115         /*
 116          * Reference snap context in page->private.  Also set
 117          * PagePrivate so that we get invalidatepage callback.
 118          */
 119         BUG_ON(PagePrivate(page));
 120         page->private = (unsigned long)snapc;
 121         SetPagePrivate(page);
 122
 123         return __set_page_dirty_nobuffers(page);
 124 }
 125
 126 /*
 127  * If we are truncating the full page (i.e. offset == 0), adjust the
 128  * dirty page counters appropriately.  Only called if there is private
 129  * data on the page.
 130  */
 131 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 132                                 unsigned int length)
 133 {
 134         struct inode *inode;
 135         struct ceph_inode_info *ci;
 136         struct ceph_snap_context *snapc = page_snap_context(page);
 137
 138         inode = page->mapping->host;
 139         ci = ceph_inode(inode);
 140
 141         if (offset != 0 || length != PAGE_SIZE) {
 142                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 143                      inode, page, page->index, offset, length);
 144                 return;
 145         }
 146
 147         ceph_invalidate_fscache_page(inode, page);
 148
 149         WARN_ON(!PageLocked(page));
 150         if (!PagePrivate(page))
 151                 return;
 152
 153         ClearPageChecked(page);
 154
 155         dout("%p invalidatepage %p idx %lu full dirty page\n",
 156              inode, page, page->index);
 157
 158         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159         ceph_put_snap_context(snapc);
 160         page->private = 0;
 161         ClearPagePrivate(page);
 162 }
 163
 164 static int ceph_releasepage(struct page *page, gfp_t g)
 165 {
 166         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 167              page, page->index, PageDirty(page) ? "" : "not ");
 168
 169         /* Can we release the page from the cache? */
 170         if (!ceph_release_fscache_page(page, g))
 171                 return 0;
 172
 173         return !PagePrivate(page);
 174 }
 175
 176 /*
 177  * read a single page, without unlocking it.
 178  */
 179 static int ceph_do_readpage(struct file *filp, struct page *page)
 180 {
 181         struct inode *inode = file_inode(filp);
 182         struct ceph_inode_info *ci = ceph_inode(inode);
 183         struct ceph_osd_client *osdc =
 184                 &ceph_inode_to_client(inode)->client->osdc;
 185         int err = 0;
 186         u64 off = page_offset(page);
 187         u64 len = PAGE_SIZE;
 188
 189         if (off >= i_size_read(inode)) {
 190                 zero_user_segment(page, 0, PAGE_SIZE);
 191                 SetPageUptodate(page);
 192                 return 0;
 193         }
 194
 195         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 196                 /*
 197                  * Uptodate inline data should have been added
 198                  * into page cache while getting Fcr caps.
 199                  */
 200                 if (off == 0)
 201                         return -EINVAL;
 202                 zero_user_segment(page, 0, PAGE_SIZE);
 203                 SetPageUptodate(page);
 204                 return 0;
 205         }
 206
 207         err = ceph_readpage_from_fscache(inode, page);
 208         if (err == 0)
 209                 return -EINPROGRESS;
 210
 211         dout("readpage inode %p file %p page %p index %lu\n",
 212              inode, filp, page, page->index);
 213         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 214                                   off, &len,
 215                                   ci->i_truncate_seq, ci->i_truncate_size,
 216                                   &page, 1, 0);
 217         if (err == -ENOENT)
 218                 err = 0;
 219         if (err < 0) {
 220                 SetPageError(page);
 221                 ceph_fscache_readpage_cancel(inode, page);
 222                 goto out;
 223         }
 224         if (err < PAGE_SIZE)
 225                 /* zero fill remainder of page */
 226                 zero_user_segment(page, err, PAGE_SIZE);
 227         else
 228                 flush_dcache_page(page);
 229
 230         SetPageUptodate(page);
 231         ceph_readpage_to_fscache(inode, page);
 232
 233 out:
 234         return err < 0 ? err : 0;
 235 }
 236
 237 static int ceph_readpage(struct file *filp, struct page *page)
 238 {
 239         int r = ceph_do_readpage(filp, page);
 240         if (r != -EINPROGRESS)
 241                 unlock_page(page);
 242         else
 243                 r = 0;
 244         return r;
 245 }
 246
 247 /*
 248  * Finish an async read(ahead) op.
 249  */
 250 static void finish_read(struct ceph_osd_request *req)
 251 {
 252         struct inode *inode = req->r_inode;
 253         struct ceph_osd_data *osd_data;
 254         int rc = req->r_result <= 0 ? req->r_result : 0;
 255         int bytes = req->r_result >= 0 ? req->r_result : 0;
 256         int num_pages;
 257         int i;
 258
 259         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 260
 261         /* unlock all pages, zeroing any data we didn't read */
 262         osd_data = osd_req_op_extent_osd_data(req, 0);
 263         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 264         num_pages = calc_pages_for((u64)osd_data->alignment,
 265                                         (u64)osd_data->length);
 266         for (i = 0; i < num_pages; i++) {
 267                 struct page *page = osd_data->pages[i];
 268
 269                 if (rc < 0 && rc != -ENOENT) {
 270                         ceph_fscache_readpage_cancel(inode, page);
 271                         goto unlock;
 272                 }
 273                 if (bytes < (int)PAGE_SIZE) {
 274                         /* zero (remainder of) page */
 275                         int s = bytes < 0 ? 0 : bytes;
 276                         zero_user_segment(page, s, PAGE_SIZE);
 277                 }
 278                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 279                      page->index);
 280                 flush_dcache_page(page);
 281                 SetPageUptodate(page);
 282                 ceph_readpage_to_fscache(inode, page);
 283 unlock:
 284                 unlock_page(page);
 285                 put_page(page);
 286                 bytes -= PAGE_SIZE;
 287         }
 288         kfree(osd_data->pages);
 289 }
 290
 291 /*
 292  * start an async read(ahead) operation.  return nr_pages we submitted
 293  * a read for on success, or negative error code.
 294  */
 295 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
 296                       struct list_head *page_list, int max)
 297 {
 298         struct ceph_osd_client *osdc =
 299                 &ceph_inode_to_client(inode)->client->osdc;
 300         struct ceph_inode_info *ci = ceph_inode(inode);
 301         struct page *page = list_entry(page_list->prev, struct page, lru);
 302         struct ceph_vino vino;
 303         struct ceph_osd_request *req;
 304         u64 off;
 305         u64 len;
 306         int i;
 307         struct page **pages;
 308         pgoff_t next_index;
 309         int nr_pages = 0;
 310         int got = 0;
 311         int ret = 0;
 312
 313         if (!rw_ctx) {
 314                 /* caller of readpages does not hold buffer and read caps
 315                  * (fadvise, madvise and readahead cases) */
 316                 int want = CEPH_CAP_FILE_CACHE;
 317                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 318                 if (ret < 0) {
 319                         dout("start_read %p, error getting cap\n", inode);
 320                 } else if (!(got & want)) {
 321                         dout("start_read %p, no cache cap\n", inode);
 322                         ret = 0;
 323                 }
 324                 if (ret <= 0) {
 325                         if (got)
 326                                 ceph_put_cap_refs(ci, got);
 327                         while (!list_empty(page_list)) {
 328                                 page = list_entry(page_list->prev,
 329                                                   struct page, lru);
 330                                 list_del(&page->lru);
 331                                 put_page(page);
 332                         }
 333                         return ret;
 334                 }
 335         }
 336
 337         off = (u64) page_offset(page);
 338
 339         /* count pages */
 340         next_index = page->index;
 341         list_for_each_entry_reverse(page, page_list, lru) {
 342                 if (page->index != next_index)
 343                         break;
 344                 nr_pages++;
 345                 next_index++;
 346                 if (max && nr_pages == max)
 347                         break;
 348         }
 349         len = nr_pages << PAGE_SHIFT;
 350         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 351              off, len);
 352         vino = ceph_vino(inode);
 353         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 354                                     0, 1, CEPH_OSD_OP_READ,
 355                                     CEPH_OSD_FLAG_READ, NULL,
 356                                     ci->i_truncate_seq, ci->i_truncate_size,
 357                                     false);
 358         if (IS_ERR(req)) {
 359                 ret = PTR_ERR(req);
 360                 goto out;
 361         }
 362
 363         /* build page vector */
 364         nr_pages = calc_pages_for(0, len);
 365         pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
 366         if (!pages) {
 367                 ret = -ENOMEM;
 368                 goto out_put;
 369         }
 370         for (i = 0; i < nr_pages; ++i) {
 371                 page = list_entry(page_list->prev, struct page, lru);
 372                 BUG_ON(PageLocked(page));
 373                 list_del(&page->lru);
 374
 375                 dout("start_read %p adding %p idx %lu\n", inode, page,
 376                      page->index);
 377                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 378                                           GFP_KERNEL)) {
 379                         ceph_fscache_uncache_page(inode, page);
 380                         put_page(page);
 381                         dout("start_read %p add_to_page_cache failed %p\n",
 382                              inode, page);
 383                         nr_pages = i;
 384                         if (nr_pages > 0) {
 385                                 len = nr_pages << PAGE_SHIFT;
 386                                 osd_req_op_extent_update(req, 0, len);
 387                                 break;
 388                         }
 389                         goto out_pages;
 390                 }
 391                 pages[i] = page;
 392         }
 393         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 394         req->r_callback = finish_read;
 395         req->r_inode = inode;
 396
 397         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 398         ret = ceph_osdc_start_request(osdc, req, false);
 399         if (ret < 0)
 400                 goto out_pages;
 401         ceph_osdc_put_request(req);
 402
 403         /* After adding locked pages to page cache, the inode holds cache cap.
 404          * So we can drop our cap refs. */
 405         if (got)
 406                 ceph_put_cap_refs(ci, got);
 407
 408         return nr_pages;
 409
 410 out_pages:
 411         for (i = 0; i < nr_pages; ++i) {
 412                 ceph_fscache_readpage_cancel(inode, pages[i]);
 413                 unlock_page(pages[i]);
 414         }
 415         ceph_put_page_vector(pages, nr_pages, false);
 416 out_put:
 417         ceph_osdc_put_request(req);
 418 out:
 419         if (got)
 420                 ceph_put_cap_refs(ci, got);
 421         return ret;
 422 }
 423
 424
 425 /*
 426  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 427  * the caller (VM) cleans them up.
 428  */
 429 static int ceph_readpages(struct file *file, struct address_space *mapping,
 430                           struct list_head *page_list, unsigned nr_pages)
 431 {
 432         struct inode *inode = file_inode(file);
 433         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 434         struct ceph_file_info *fi = file->private_data;
 435         struct ceph_rw_context *rw_ctx;
 436         int rc = 0;
 437         int max = 0;
 438
 439         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 440                 return -EINVAL;
 441
 442         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 443                                          &nr_pages);
 444
 445         if (rc == 0)
 446                 goto out;
 447
 448         rw_ctx = ceph_find_rw_context(fi);
 449         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 450         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 451              inode, file, rw_ctx, nr_pages, max);
 452         while (!list_empty(page_list)) {
 453                 rc = start_read(inode, rw_ctx, page_list, max);
 454                 if (rc < 0)
 455                         goto out;
 456         }
 457 out:
 458         ceph_fscache_readpages_cancel(inode, page_list);
 459
 460         dout("readpages %p file %p ret %d\n", inode, file, rc);
 461         return rc;
 462 }
 463
 464 struct ceph_writeback_ctl
 465 {
 466         loff_t i_size;
 467         u64 truncate_size;
 468         u32 truncate_seq;
 469         bool size_stable;
 470         bool head_snapc;
 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 *
 478 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 479                    struct ceph_snap_context *page_snapc)
 480 {
 481         struct ceph_inode_info *ci = ceph_inode(inode);
 482         struct ceph_snap_context *snapc = NULL;
 483         struct ceph_cap_snap *capsnap = NULL;
 484
 485         spin_lock(&ci->i_ceph_lock);
 486         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 487                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 488                      capsnap->context, capsnap->dirty_pages);
 489                 if (!capsnap->dirty_pages)
 490                         continue;
 491
 492                 /* get i_size, truncate_{seq,size} for page_snapc? */
 493                 if (snapc && capsnap->context != page_snapc)
 494                         continue;
 495
 496                 if (ctl) {
 497                         if (capsnap->writing) {
 498                                 ctl->i_size = i_size_read(inode);
 499                                 ctl->size_stable = false;
 500                         } else {
 501                                 ctl->i_size = capsnap->size;
 502                                 ctl->size_stable = true;
 503                         }
 504                         ctl->truncate_size = capsnap->truncate_size;
 505                         ctl->truncate_seq = capsnap->truncate_seq;
 506                         ctl->head_snapc = false;
 507                 }
 508
 509                 if (snapc)
 510                         break;
 511
 512                 snapc = ceph_get_snap_context(capsnap->context);
 513                 if (!page_snapc ||
 514                     page_snapc == snapc ||
 515                     page_snapc->seq > snapc->seq)
 516                         break;
 517         }
 518         if (!snapc && ci->i_wrbuffer_ref_head) {
 519                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 520                 dout(" head snapc %p has %d dirty pages\n",
 521                      snapc, ci->i_wrbuffer_ref_head);
 522                 if (ctl) {
 523                         ctl->i_size = i_size_read(inode);
 524                         ctl->truncate_size = ci->i_truncate_size;
 525                         ctl->truncate_seq = ci->i_truncate_seq;
 526                         ctl->size_stable = false;
 527                         ctl->head_snapc = true;
 528                 }
 529         }
 530         spin_unlock(&ci->i_ceph_lock);
 531         return snapc;
 532 }
 533
 534 static u64 get_writepages_data_length(struct inode *inode,
 535                                       struct page *page, u64 start)
 536 {
 537         struct ceph_inode_info *ci = ceph_inode(inode);
 538         struct ceph_snap_context *snapc = page_snap_context(page);
 539         struct ceph_cap_snap *capsnap = NULL;
 540         u64 end = i_size_read(inode);
 541
 542         if (snapc != ci->i_head_snapc) {
 543                 bool found = false;
 544                 spin_lock(&ci->i_ceph_lock);
 545                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 546                         if (capsnap->context == snapc) {
 547                                 if (!capsnap->writing)
 548                                         end = capsnap->size;
 549                                 found = true;
 550                                 break;
 551                         }
 552                 }
 553                 spin_unlock(&ci->i_ceph_lock);
 554                 WARN_ON(!found);
 555         }
 556         if (end > page_offset(page) + PAGE_SIZE)
 557                 end = page_offset(page) + PAGE_SIZE;
 558         return end > start ? end - start : 0;
 559 }
 560
 561 /*
 562  * Write a single page, but leave the page locked.
 563  *
 564  * If we get a write error, set the page error bit, but still adjust the
 565  * dirty page accounting (i.e., page is no longer dirty).
 566  */
 567 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 568 {
 569         struct inode *inode;
 570         struct ceph_inode_info *ci;
 571         struct ceph_fs_client *fsc;
 572         struct ceph_snap_context *snapc, *oldest;
 573         loff_t page_off = page_offset(page);
 574         int err, len = PAGE_SIZE;
 575         struct ceph_writeback_ctl ceph_wbc;
 576
 577         dout("writepage %p idx %lu\n", page, page->index);
 578
 579         inode = page->mapping->host;
 580         ci = ceph_inode(inode);
 581         fsc = ceph_inode_to_client(inode);
 582
 583         /* verify this is a writeable snap context */
 584         snapc = page_snap_context(page);
 585         if (!snapc) {
 586                 dout("writepage %p page %p not dirty?\n", inode, page);
 587                 return 0;
 588         }
 589         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 590         if (snapc->seq > oldest->seq) {
 591                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 592                      inode, page, snapc);
 593                 /* we should only noop if called by kswapd */
 594                 WARN_ON(!(current->flags & PF_MEMALLOC));
 595                 ceph_put_snap_context(oldest);
 596                 redirty_page_for_writepage(wbc, page);
 597                 return 0;
 598         }
 599         ceph_put_snap_context(oldest);
 600
 601         /* is this a partial page at end of file? */
 602         if (page_off >= ceph_wbc.i_size) {
 603                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 604                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 605                 return 0;
 606         }
 607
 608         if (ceph_wbc.i_size < page_off + len)
 609                 len = ceph_wbc.i_size - page_off;
 610
 611         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 612              inode, page, page->index, page_off, len, snapc, snapc->seq);
 613
 614         if (atomic_long_inc_return(&fsc->writeback_count) >
 615             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 616                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 617
 618         set_page_writeback(page);
 619         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 620                                    &ci->i_layout, snapc, page_off, len,
 621                                    ceph_wbc.truncate_seq,
 622                                    ceph_wbc.truncate_size,
 623                                    &inode->i_mtime, &page, 1);
 624         if (err < 0) {
 625                 struct writeback_control tmp_wbc;
 626                 if (!wbc)
 627                         wbc = &tmp_wbc;
 628                 if (err == -ERESTARTSYS) {
 629                         /* killed by SIGKILL */
 630                         dout("writepage interrupted page %p\n", page);
 631                         redirty_page_for_writepage(wbc, page);
 632                         end_page_writeback(page);
 633                         return err;
 634                 }
 635                 dout("writepage setting page/mapping error %d %p\n",
 636                      err, page);
 637                 SetPageError(page);
 638                 mapping_set_error(&inode->i_data, err);
 639                 wbc->pages_skipped++;
 640         } else {
 641                 dout("writepage cleaned page %p\n", page);
 642                 err = 0;  /* vfs expects us to return 0 */
 643         }
 644         page->private = 0;
 645         ClearPagePrivate(page);
 646         end_page_writeback(page);
 647         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 648         ceph_put_snap_context(snapc);  /* page's reference */
 649
 650         if (atomic_long_dec_return(&fsc->writeback_count) <
 651             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 652                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 653
 654         return err;
 655 }
 656
 657 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 658 {
 659         int err;
 660         struct inode *inode = page->mapping->host;
 661         BUG_ON(!inode);
 662         ihold(inode);
 663         err = writepage_nounlock(page, wbc);
 664         if (err == -ERESTARTSYS) {
 665                 /* direct memory reclaimer was killed by SIGKILL. return 0
 666                  * to prevent caller from setting mapping/page error */
 667                 err = 0;
 668         }
 669         unlock_page(page);
 670         iput(inode);
 671         return err;
 672 }
 673
 674 /*
 675  * lame release_pages helper.  release_pages() isn't exported to
 676  * modules.
 677  */
 678 static void ceph_release_pages(struct page **pages, int num)
 679 {
 680         struct pagevec pvec;
 681         int i;
 682
 683         pagevec_init(&pvec);
 684         for (i = 0; i < num; i++) {
 685                 if (pagevec_add(&pvec, pages[i]) == 0)
 686                         pagevec_release(&pvec);
 687         }
 688         pagevec_release(&pvec);
 689 }
 690
 691 /*
 692  * async writeback completion handler.
 693  *
 694  * If we get an error, set the mapping error bit, but not the individual
 695  * page error bits.
 696  */
 697 static void writepages_finish(struct ceph_osd_request *req)
 698 {
 699         struct inode *inode = req->r_inode;
 700         struct ceph_inode_info *ci = ceph_inode(inode);
 701         struct ceph_osd_data *osd_data;
 702         struct page *page;
 703         int num_pages, total_pages = 0;
 704         int i, j;
 705         int rc = req->r_result;
 706         struct ceph_snap_context *snapc = req->r_snapc;
 707         struct address_space *mapping = inode->i_mapping;
 708         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 709         bool remove_page;
 710
 711         dout("writepages_finish %p rc %d\n", inode, rc);
 712         if (rc < 0) {
 713                 mapping_set_error(mapping, rc);
 714                 ceph_set_error_write(ci);
 715         } else {
 716                 ceph_clear_error_write(ci);
 717         }
 718
 719         /*
 720          * We lost the cache cap, need to truncate the page before
 721          * it is unlocked, otherwise we'd truncate it later in the
 722          * page truncation thread, possibly losing some data that
 723          * raced its way in
 724          */
 725         remove_page = !(ceph_caps_issued(ci) &
 726                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 727
 728         /* clean all pages */
 729         for (i = 0; i < req->r_num_ops; i++) {
 730                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 731                         break;
 732
 733                 osd_data = osd_req_op_extent_osd_data(req, i);
 734                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 735                 num_pages = calc_pages_for((u64)osd_data->alignment,
 736                                            (u64)osd_data->length);
 737                 total_pages += num_pages;
 738                 for (j = 0; j < num_pages; j++) {
 739                         page = osd_data->pages[j];
 740                         BUG_ON(!page);
 741                         WARN_ON(!PageUptodate(page));
 742
 743                         if (atomic_long_dec_return(&fsc->writeback_count) <
 744                              CONGESTION_OFF_THRESH(
 745                                         fsc->mount_options->congestion_kb))
 746                                 clear_bdi_congested(inode_to_bdi(inode),
 747                                                     BLK_RW_ASYNC);
 748
 749                         ceph_put_snap_context(page_snap_context(page));
 750                         page->private = 0;
 751                         ClearPagePrivate(page);
 752                         dout("unlocking %p\n", page);
 753                         end_page_writeback(page);
 754
 755                         if (remove_page)
 756                                 generic_error_remove_page(inode->i_mapping,
 757                                                           page);
 758
 759                         unlock_page(page);
 760                 }
 761                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 762                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 763
 764                 ceph_release_pages(osd_data->pages, num_pages);
 765         }
 766
 767         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 768
 769         osd_data = osd_req_op_extent_osd_data(req, 0);
 770         if (osd_data->pages_from_pool)
 771                 mempool_free(osd_data->pages,
 772                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 773         else
 774                 kfree(osd_data->pages);
 775         ceph_osdc_put_request(req);
 776 }
 777
 778 /*
 779  * initiate async writeback
 780  */
 781 static int ceph_writepages_start(struct address_space *mapping,
 782                                  struct writeback_control *wbc)
 783 {
 784         struct inode *inode = mapping->host;
 785         struct ceph_inode_info *ci = ceph_inode(inode);
 786         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 787         struct ceph_vino vino = ceph_vino(inode);
 788         pgoff_t index, start_index, end = -1;
 789         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 790         struct pagevec pvec;
 791         int rc = 0;
 792         unsigned int wsize = i_blocksize(inode);
 793         struct ceph_osd_request *req = NULL;
 794         struct ceph_writeback_ctl ceph_wbc;
 795         bool should_loop, range_whole = false;
 796         bool done = false;
 797
 798         dout("writepages_start %p (mode=%s)\n", inode,
 799              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 800              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 801
 802         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 803                 if (ci->i_wrbuffer_ref > 0) {
 804                         pr_warn_ratelimited(
 805                                 "writepage_start %p %lld forced umount\n",
 806                                 inode, ceph_ino(inode));
 807                 }
 808                 mapping_set_error(mapping, -EIO);
 809                 return -EIO; /* we're in a forced umount, don't write! */
 810         }
 811         if (fsc->mount_options->wsize < wsize)
 812                 wsize = fsc->mount_options->wsize;
 813
 814         pagevec_init(&pvec);
 815
 816         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 817         index = start_index;
 818
 819 retry:
 820         /* find oldest snap context with dirty data */
 821         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 822         if (!snapc) {
 823                 /* hmm, why does writepages get called when there
 824                    is no dirty data? */
 825                 dout(" no snap context with dirty data?\n");
 826                 goto out;
 827         }
 828         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 829              snapc, snapc->seq, snapc->num_snaps);
 830
 831         should_loop = false;
 832         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 833                 /* where to start/end? */
 834                 if (wbc->range_cyclic) {
 835                         index = start_index;
 836                         end = -1;
 837                         if (index > 0)
 838                                 should_loop = true;
 839                         dout(" cyclic, start at %lu\n", index);
 840                 } else {
 841                         index = wbc->range_start >> PAGE_SHIFT;
 842                         end = wbc->range_end >> PAGE_SHIFT;
 843                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 844                                 range_whole = true;
 845                         dout(" not cyclic, %lu to %lu\n", index, end);
 846                 }
 847         } else if (!ceph_wbc.head_snapc) {
 848                 /* Do not respect wbc->range_{start,end}. Dirty pages
 849                  * in that range can be associated with newer snapc.
 850                  * They are not writeable until we write all dirty pages
 851                  * associated with 'snapc' get written */
 852                 if (index > 0)
 853                         should_loop = true;
 854                 dout(" non-head snapc, range whole\n");
 855         }
 856
 857         ceph_put_snap_context(last_snapc);
 858         last_snapc = snapc;
 859
 860         while (!done && index <= end) {
 861                 int num_ops = 0, op_idx;
 862                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 863                 struct page **pages = NULL, **data_pages;
 864                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 865                 struct page *page;
 866                 pgoff_t strip_unit_end = 0;
 867                 u64 offset = 0, len = 0;
 868
 869                 max_pages = wsize >> PAGE_SHIFT;
 870
 871 get_more_pages:
 872                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 873                                                 end, PAGECACHE_TAG_DIRTY,
 874                                                 max_pages - locked_pages);
 875                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 876                 if (!pvec_pages && !locked_pages)
 877                         break;
 878                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 879                         page = pvec.pages[i];
 880                         dout("? %p idx %lu\n", page, page->index);
 881                         if (locked_pages == 0)
 882                                 lock_page(page);  /* first page */
 883                         else if (!trylock_page(page))
 884                                 break;
 885
 886                         /* only dirty pages, or our accounting breaks */
 887                         if (unlikely(!PageDirty(page)) ||
 888                             unlikely(page->mapping != mapping)) {
 889                                 dout("!dirty or !mapping %p\n", page);
 890                                 unlock_page(page);
 891                                 continue;
 892                         }
 893                         /* only if matching snap context */
 894                         pgsnapc = page_snap_context(page);
 895                         if (pgsnapc != snapc) {
 896                                 dout("page snapc %p %lld != oldest %p %lld\n",
 897                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 898                                 if (!should_loop &&
 899                                     !ceph_wbc.head_snapc &&
 900                                     wbc->sync_mode != WB_SYNC_NONE)
 901                                         should_loop = true;
 902                                 unlock_page(page);
 903                                 continue;
 904                         }
 905                         if (page_offset(page) >= ceph_wbc.i_size) {
 906                                 dout("%p page eof %llu\n",
 907                                      page, ceph_wbc.i_size);
 908                                 if ((ceph_wbc.size_stable ||
 909                                     page_offset(page) >= i_size_read(inode)) &&
 910                                     clear_page_dirty_for_io(page))
 911                                         mapping->a_ops->invalidatepage(page,
 912                                                                 0, PAGE_SIZE);
 913                                 unlock_page(page);
 914                                 continue;
 915                         }
 916                         if (strip_unit_end && (page->index > strip_unit_end)) {
 917                                 dout("end of strip unit %p\n", page);
 918                                 unlock_page(page);
 919                                 break;
 920                         }
 921                         if (PageWriteback(page)) {
 922                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 923                                         dout("%p under writeback\n", page);
 924                                         unlock_page(page);
 925                                         continue;
 926                                 }
 927                                 dout("waiting on writeback %p\n", page);
 928                                 wait_on_page_writeback(page);
 929                         }
 930
 931                         if (!clear_page_dirty_for_io(page)) {
 932                                 dout("%p !clear_page_dirty_for_io\n", page);
 933                                 unlock_page(page);
 934                                 continue;
 935                         }
 936
 937                         /*
 938                          * We have something to write.  If this is
 939                          * the first locked page this time through,
 940                          * calculate max possinle write size and
 941                          * allocate a page array
 942                          */
 943                         if (locked_pages == 0) {
 944                                 u64 objnum;
 945                                 u64 objoff;
 946                                 u32 xlen;
 947
 948                                 /* prepare async write request */
 949                                 offset = (u64)page_offset(page);
 950                                 ceph_calc_file_object_mapping(&ci->i_layout,
 951                                                               offset, wsize,
 952                                                               &objnum, &objoff,
 953                                                               &xlen);
 954                                 len = xlen;
 955
 956                                 num_ops = 1;
 957                                 strip_unit_end = page->index +
 958                                         ((len - 1) >> PAGE_SHIFT);
 959
 960                                 BUG_ON(pages);
 961                                 max_pages = calc_pages_for(0, (u64)len);
 962                                 pages = kmalloc_array(max_pages,
 963                                                       sizeof(*pages),
 964                                                       GFP_NOFS);
 965                                 if (!pages) {
 966                                         pool = fsc->wb_pagevec_pool;
 967                                         pages = mempool_alloc(pool, GFP_NOFS);
 968                                         BUG_ON(!pages);
 969                                 }
 970
 971                                 len = 0;
 972                         } else if (page->index !=
 973                                    (offset + len) >> PAGE_SHIFT) {
 974                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 975                                                         CEPH_OSD_MAX_OPS)) {
 976                                         redirty_page_for_writepage(wbc, page);
 977                                         unlock_page(page);
 978                                         break;
 979                                 }
 980
 981                                 num_ops++;
 982                                 offset = (u64)page_offset(page);
 983                                 len = 0;
 984                         }
 985
 986                         /* note position of first page in pvec */
 987                         dout("%p will write page %p idx %lu\n",
 988                              inode, page, page->index);
 989
 990                         if (atomic_long_inc_return(&fsc->writeback_count) >
 991                             CONGESTION_ON_THRESH(
 992                                     fsc->mount_options->congestion_kb)) {
 993                                 set_bdi_congested(inode_to_bdi(inode),
 994                                                   BLK_RW_ASYNC);
 995                         }
 996
 997
 998                         pages[locked_pages++] = page;
 999                         pvec.pages[i] = NULL;
1000
1001                         len += PAGE_SIZE;
1002                 }
1003
1004                 /* did we get anything? */
1005                 if (!locked_pages)
1006                         goto release_pvec_pages;
1007                 if (i) {
1008                         unsigned j, n = 0;
1009                         /* shift unused page to beginning of pvec */
1010                         for (j = 0; j < pvec_pages; j++) {
1011                                 if (!pvec.pages[j])
1012                                         continue;
1013                                 if (n < j)
1014                                         pvec.pages[n] = pvec.pages[j];
1015                                 n++;
1016                         }
1017                         pvec.nr = n;
1018
1019                         if (pvec_pages && i == pvec_pages &&
1020                             locked_pages < max_pages) {
1021                                 dout("reached end pvec, trying for more\n");
1022                                 pagevec_release(&pvec);
1023                                 goto get_more_pages;
1024                         }
1025                 }
1026
1027 new_request:
1028                 offset = page_offset(pages[0]);
1029                 len = wsize;
1030
1031                 req = ceph_osdc_new_request(&fsc->client->osdc,
1032                                         &ci->i_layout, vino,
1033                                         offset, &len, 0, num_ops,
1034                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1035                                         snapc, ceph_wbc.truncate_seq,
1036                                         ceph_wbc.truncate_size, false);
1037                 if (IS_ERR(req)) {
1038                         req = ceph_osdc_new_request(&fsc->client->osdc,
1039                                                 &ci->i_layout, vino,
1040                                                 offset, &len, 0,
1041                                                 min(num_ops,
1042                                                     CEPH_OSD_SLAB_OPS),
1043                                                 CEPH_OSD_OP_WRITE,
1044                                                 CEPH_OSD_FLAG_WRITE,
1045                                                 snapc, ceph_wbc.truncate_seq,
1046                                                 ceph_wbc.truncate_size, true);
1047                         BUG_ON(IS_ERR(req));
1048                 }
1049                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1050                              PAGE_SIZE - offset);
1051
1052                 req->r_callback = writepages_finish;
1053                 req->r_inode = inode;
1054
1055                 /* Format the osd request message and submit the write */
1056                 len = 0;
1057                 data_pages = pages;
1058                 op_idx = 0;
1059                 for (i = 0; i < locked_pages; i++) {
1060                         u64 cur_offset = page_offset(pages[i]);
1061                         if (offset + len != cur_offset) {
1062                                 if (op_idx + 1 == req->r_num_ops)
1063                                         break;
1064                                 osd_req_op_extent_dup_last(req, op_idx,
1065                                                            cur_offset - offset);
1066                                 dout("writepages got pages at %llu~%llu\n",
1067                                      offset, len);
1068                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1069                                                         data_pages, len, 0,
1070                                                         !!pool, false);
1071                                 osd_req_op_extent_update(req, op_idx, len);
1072
1073                                 len = 0;
1074                                 offset = cur_offset;
1075                                 data_pages = pages + i;
1076                                 op_idx++;
1077                         }
1078
1079                         set_page_writeback(pages[i]);
1080                         len += PAGE_SIZE;
1081                 }
1082
1083                 if (ceph_wbc.size_stable) {
1084                         len = min(len, ceph_wbc.i_size - offset);
1085                 } else if (i == locked_pages) {
1086                         /* writepages_finish() clears writeback pages
1087                          * according to the data length, so make sure
1088                          * data length covers all locked pages */
1089                         u64 min_len = len + 1 - PAGE_SIZE;
1090                         len = get_writepages_data_length(inode, pages[i - 1],
1091                                                          offset);
1092                         len = max(len, min_len);
1093                 }
1094                 dout("writepages got pages at %llu~%llu\n", offset, len);
1095
1096                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1097                                                  0, !!pool, false);
1098                 osd_req_op_extent_update(req, op_idx, len);
1099
1100                 BUG_ON(op_idx + 1 != req->r_num_ops);
1101
1102                 pool = NULL;
1103                 if (i < locked_pages) {
1104                         BUG_ON(num_ops <= req->r_num_ops);
1105                         num_ops -= req->r_num_ops;
1106                         locked_pages -= i;
1107
1108                         /* allocate new pages array for next request */
1109                         data_pages = pages;
1110                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1111                                               GFP_NOFS);
1112                         if (!pages) {
1113                                 pool = fsc->wb_pagevec_pool;
1114                                 pages = mempool_alloc(pool, GFP_NOFS);
1115                                 BUG_ON(!pages);
1116                         }
1117                         memcpy(pages, data_pages + i,
1118                                locked_pages * sizeof(*pages));
1119                         memset(data_pages + i, 0,
1120                                locked_pages * sizeof(*pages));
1121                 } else {
1122                         BUG_ON(num_ops != req->r_num_ops);
1123                         index = pages[i - 1]->index + 1;
1124                         /* request message now owns the pages array */
1125                         pages = NULL;
1126                 }
1127
1128                 req->r_mtime = inode->i_mtime;
1129                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1130                 BUG_ON(rc);
1131                 req = NULL;
1132
1133                 wbc->nr_to_write -= i;
1134                 if (pages)
1135                         goto new_request;
1136
1137                 /*
1138                  * We stop writing back only if we are not doing
1139                  * integrity sync. In case of integrity sync we have to
1140                  * keep going until we have written all the pages
1141                  * we tagged for writeback prior to entering this loop.
1142                  */
1143                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1144                         done = true;
1145
1146 release_pvec_pages:
1147                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1148                      pvec.nr ? pvec.pages[0] : NULL);
1149                 pagevec_release(&pvec);
1150         }
1151
1152         if (should_loop && !done) {
1153                 /* more to do; loop back to beginning of file */
1154                 dout("writepages looping back to beginning of file\n");
1155                 end = start_index - 1; /* OK even when start_index == 0 */
1156
1157                 /* to write dirty pages associated with next snapc,
1158                  * we need to wait until current writes complete */
1159                 if (wbc->sync_mode != WB_SYNC_NONE &&
1160                     start_index == 0 && /* all dirty pages were checked */
1161                     !ceph_wbc.head_snapc) {
1162                         struct page *page;
1163                         unsigned i, nr;
1164                         index = 0;
1165                         while ((index <= end) &&
1166                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1167                                                 PAGECACHE_TAG_WRITEBACK))) {
1168                                 for (i = 0; i < nr; i++) {
1169                                         page = pvec.pages[i];
1170                                         if (page_snap_context(page) != snapc)
1171                                                 continue;
1172                                         wait_on_page_writeback(page);
1173                                 }
1174                                 pagevec_release(&pvec);
1175                                 cond_resched();
1176                         }
1177                 }
1178
1179                 start_index = 0;
1180                 index = 0;
1181                 goto retry;
1182         }
1183
1184         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1185                 mapping->writeback_index = index;
1186
1187 out:
1188         ceph_osdc_put_request(req);
1189         ceph_put_snap_context(last_snapc);
1190         dout("writepages dend - startone, rc = %d\n", rc);
1191         return rc;
1192 }
1193
1194
1195
1196 /*
1197  * See if a given @snapc is either writeable, or already written.
1198  */
1199 static int context_is_writeable_or_written(struct inode *inode,
1200                                            struct ceph_snap_context *snapc)
1201 {
1202         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1203         int ret = !oldest || snapc->seq <= oldest->seq;
1204
1205         ceph_put_snap_context(oldest);
1206         return ret;
1207 }
1208
1209 /*
1210  * We are only allowed to write into/dirty the page if the page is
1211  * clean, or already dirty within the same snap context.
1212  *
1213  * called with page locked.
1214  * return success with page locked,
1215  * or any failure (incl -EAGAIN) with page unlocked.
1216  */
1217 static int ceph_update_writeable_page(struct file *file,
1218                             loff_t pos, unsigned len,
1219                             struct page *page)
1220 {
1221         struct inode *inode = file_inode(file);
1222         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1223         struct ceph_inode_info *ci = ceph_inode(inode);
1224         loff_t page_off = pos & PAGE_MASK;
1225         int pos_in_page = pos & ~PAGE_MASK;
1226         int end_in_page = pos_in_page + len;
1227         loff_t i_size;
1228         int r;
1229         struct ceph_snap_context *snapc, *oldest;
1230
1231         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1232                 dout(" page %p forced umount\n", page);
1233                 unlock_page(page);
1234                 return -EIO;
1235         }
1236
1237 retry_locked:
1238         /* writepages currently holds page lock, but if we change that later, */
1239         wait_on_page_writeback(page);
1240
1241         snapc = page_snap_context(page);
1242         if (snapc && snapc != ci->i_head_snapc) {
1243                 /*
1244                  * this page is already dirty in another (older) snap
1245                  * context!  is it writeable now?
1246                  */
1247                 oldest = get_oldest_context(inode, NULL, NULL);
1248                 if (snapc->seq > oldest->seq) {
1249                         ceph_put_snap_context(oldest);
1250                         dout(" page %p snapc %p not current or oldest\n",
1251                              page, snapc);
1252                         /*
1253                          * queue for writeback, and wait for snapc to
1254                          * be writeable or written
1255                          */
1256                         snapc = ceph_get_snap_context(snapc);
1257                         unlock_page(page);
1258                         ceph_queue_writeback(inode);
1259                         r = wait_event_killable(ci->i_cap_wq,
1260                                context_is_writeable_or_written(inode, snapc));
1261                         ceph_put_snap_context(snapc);
1262                         if (r == -ERESTARTSYS)
1263                                 return r;
1264                         return -EAGAIN;
1265                 }
1266                 ceph_put_snap_context(oldest);
1267
1268                 /* yay, writeable, do it now (without dropping page lock) */
1269                 dout(" page %p snapc %p not current, but oldest\n",
1270                      page, snapc);
1271                 if (!clear_page_dirty_for_io(page))
1272                         goto retry_locked;
1273                 r = writepage_nounlock(page, NULL);
1274                 if (r < 0)
1275                         goto fail_unlock;
1276                 goto retry_locked;
1277         }
1278
1279         if (PageUptodate(page)) {
1280                 dout(" page %p already uptodate\n", page);
1281                 return 0;
1282         }
1283
1284         /* full page? */
1285         if (pos_in_page == 0 && len == PAGE_SIZE)
1286                 return 0;
1287
1288         /* past end of file? */
1289         i_size = i_size_read(inode);
1290
1291         if (page_off >= i_size ||
1292             (pos_in_page == 0 && (pos+len) >= i_size &&
1293              end_in_page - pos_in_page != PAGE_SIZE)) {
1294                 dout(" zeroing %p 0 - %d and %d - %d\n",
1295                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1296                 zero_user_segments(page,
1297                                    0, pos_in_page,
1298                                    end_in_page, PAGE_SIZE);
1299                 return 0;
1300         }
1301
1302         /* we need to read it. */
1303         r = ceph_do_readpage(file, page);
1304         if (r < 0) {
1305                 if (r == -EINPROGRESS)
1306                         return -EAGAIN;
1307                 goto fail_unlock;
1308         }
1309         goto retry_locked;
1310 fail_unlock:
1311         unlock_page(page);
1312         return r;
1313 }
1314
1315 /*
1316  * We are only allowed to write into/dirty the page if the page is
1317  * clean, or already dirty within the same snap context.
1318  */
1319 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1320                             loff_t pos, unsigned len, unsigned flags,
1321                             struct page **pagep, void **fsdata)
1322 {
1323         struct inode *inode = file_inode(file);
1324         struct page *page;
1325         pgoff_t index = pos >> PAGE_SHIFT;
1326         int r;
1327
1328         do {
1329                 /* get a page */
1330                 page = grab_cache_page_write_begin(mapping, index, 0);
1331                 if (!page)
1332                         return -ENOMEM;
1333
1334                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1335                      inode, page, (int)pos, (int)len);
1336
1337                 r = ceph_update_writeable_page(file, pos, len, page);
1338                 if (r < 0)
1339                         put_page(page);
1340                 else
1341                         *pagep = page;
1342         } while (r == -EAGAIN);
1343
1344         return r;
1345 }
1346
1347 /*
1348  * we don't do anything in here that simple_write_end doesn't do
1349  * except adjust dirty page accounting
1350  */
1351 static int ceph_write_end(struct file *file, struct address_space *mapping,
1352                           loff_t pos, unsigned len, unsigned copied,
1353                           struct page *page, void *fsdata)
1354 {
1355         struct inode *inode = file_inode(file);
1356         bool check_cap = false;
1357
1358         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1359              inode, page, (int)pos, (int)copied, (int)len);
1360
1361         /* zero the stale part of the page if we did a short copy */
1362         if (!PageUptodate(page)) {
1363                 if (copied < len) {
1364                         copied = 0;
1365                         goto out;
1366                 }
1367                 SetPageUptodate(page);
1368         }
1369
1370         /* did file size increase? */
1371         if (pos+copied > i_size_read(inode))
1372                 check_cap = ceph_inode_set_size(inode, pos+copied);
1373
1374         set_page_dirty(page);
1375
1376 out:
1377         unlock_page(page);
1378         put_page(page);
1379
1380         if (check_cap)
1381                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1382
1383         return copied;
1384 }
1385
1386 /*
1387  * we set .direct_IO to indicate direct io is supported, but since we
1388  * intercept O_DIRECT reads and writes early, this function should
1389  * never get called.
1390  */
1391 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1392 {
1393         WARN_ON(1);
1394         return -EINVAL;
1395 }
1396
1397 const struct address_space_operations ceph_aops = {
1398         .readpage = ceph_readpage,
1399         .readpages = ceph_readpages,
1400         .writepage = ceph_writepage,
1401         .writepages = ceph_writepages_start,
1402         .write_begin = ceph_write_begin,
1403         .write_end = ceph_write_end,
1404         .set_page_dirty = ceph_set_page_dirty,
1405         .invalidatepage = ceph_invalidatepage,
1406         .releasepage = ceph_releasepage,
1407         .direct_IO = ceph_direct_io,
1408 };
1409
1410 static void ceph_block_sigs(sigset_t *oldset)
1411 {
1412         sigset_t mask;
1413         siginitsetinv(&mask, sigmask(SIGKILL));
1414         sigprocmask(SIG_BLOCK, &mask, oldset);
1415 }
1416
1417 static void ceph_restore_sigs(sigset_t *oldset)
1418 {
1419         sigprocmask(SIG_SETMASK, oldset, NULL);
1420 }
1421
1422 /*
1423  * vm ops
1424  */
1425 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1426 {
1427         struct vm_area_struct *vma = vmf->vma;
1428         struct inode *inode = file_inode(vma->vm_file);
1429         struct ceph_inode_info *ci = ceph_inode(inode);
1430         struct ceph_file_info *fi = vma->vm_file->private_data;
1431         struct page *pinned_page = NULL;
1432         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1433         int want, got, err;
1434         sigset_t oldset;
1435         vm_fault_t ret = VM_FAULT_SIGBUS;
1436
1437         ceph_block_sigs(&oldset);
1438
1439         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1440              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1441         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1442                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1443         else
1444                 want = CEPH_CAP_FILE_CACHE;
1445
1446         got = 0;
1447         err = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1448         if (err < 0)
1449                 goto out_restore;
1450
1451         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1452              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1453
1454         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1455             ci->i_inline_version == CEPH_INLINE_NONE) {
1456                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1457                 ceph_add_rw_context(fi, &rw_ctx);
1458                 ret = filemap_fault(vmf);
1459                 ceph_del_rw_context(fi, &rw_ctx);
1460                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1461                         inode, off, (size_t)PAGE_SIZE,
1462                                 ceph_cap_string(got), ret);
1463         } else
1464                 err = -EAGAIN;
1465
1466         if (pinned_page)
1467                 put_page(pinned_page);
1468         ceph_put_cap_refs(ci, got);
1469
1470         if (err != -EAGAIN)
1471                 goto out_restore;
1472
1473         /* read inline data */
1474         if (off >= PAGE_SIZE) {
1475                 /* does not support inline data > PAGE_SIZE */
1476                 ret = VM_FAULT_SIGBUS;
1477         } else {
1478                 struct address_space *mapping = inode->i_mapping;
1479                 struct page *page = find_or_create_page(mapping, 0,
1480                                                 mapping_gfp_constraint(mapping,
1481                                                 ~__GFP_FS));
1482                 if (!page) {
1483                         ret = VM_FAULT_OOM;
1484                         goto out_inline;
1485                 }
1486                 err = __ceph_do_getattr(inode, page,
1487                                          CEPH_STAT_CAP_INLINE_DATA, true);
1488                 if (err < 0 || off >= i_size_read(inode)) {
1489                         unlock_page(page);
1490                         put_page(page);
1491                         if (err == -ENOMEM)
1492                                 ret = VM_FAULT_OOM;
1493                         else
1494                                 ret = VM_FAULT_SIGBUS;
1495                         goto out_inline;
1496                 }
1497                 if (err < PAGE_SIZE)
1498                         zero_user_segment(page, err, PAGE_SIZE);
1499                 else
1500                         flush_dcache_page(page);
1501                 SetPageUptodate(page);
1502                 vmf->page = page;
1503                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1504 out_inline:
1505                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1506                      inode, off, (size_t)PAGE_SIZE, ret);
1507         }
1508 out_restore:
1509         ceph_restore_sigs(&oldset);
1510         if (err < 0)
1511                 ret = vmf_error(err);
1512
1513         return ret;
1514 }
1515
1516 /*
1517  * Reuse write_begin here for simplicity.
1518  */
1519 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1520 {
1521         struct vm_area_struct *vma = vmf->vma;
1522         struct inode *inode = file_inode(vma->vm_file);
1523         struct ceph_inode_info *ci = ceph_inode(inode);
1524         struct ceph_file_info *fi = vma->vm_file->private_data;
1525         struct ceph_cap_flush *prealloc_cf;
1526         struct page *page = vmf->page;
1527         loff_t off = page_offset(page);
1528         loff_t size = i_size_read(inode);
1529         size_t len;
1530         int want, got, err;
1531         sigset_t oldset;
1532         vm_fault_t ret = VM_FAULT_SIGBUS;
1533
1534         prealloc_cf = ceph_alloc_cap_flush();
1535         if (!prealloc_cf)
1536                 return VM_FAULT_OOM;
1537
1538         ceph_block_sigs(&oldset);
1539
1540         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1541                 struct page *locked_page = NULL;
1542                 if (off == 0) {
1543                         lock_page(page);
1544                         locked_page = page;
1545                 }
1546                 err = ceph_uninline_data(vma->vm_file, locked_page);
1547                 if (locked_page)
1548                         unlock_page(locked_page);
1549                 if (err < 0)
1550                         goto out_free;
1551         }
1552
1553         if (off + PAGE_SIZE <= size)
1554                 len = PAGE_SIZE;
1555         else
1556                 len = size & ~PAGE_MASK;
1557
1558         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1559              inode, ceph_vinop(inode), off, len, size);
1560         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1561                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1562         else
1563                 want = CEPH_CAP_FILE_BUFFER;
1564
1565         got = 0;
1566         err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1567                             &got, NULL);
1568         if (err < 0)
1569                 goto out_free;
1570
1571         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1572              inode, off, len, ceph_cap_string(got));
1573
1574         /* Update time before taking page lock */
1575         file_update_time(vma->vm_file);
1576
1577         do {
1578                 lock_page(page);
1579
1580                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1581                         unlock_page(page);
1582                         ret = VM_FAULT_NOPAGE;
1583                         break;
1584                 }
1585
1586                 err = ceph_update_writeable_page(vma->vm_file, off, len, page);
1587                 if (err >= 0) {
1588                         /* success.  we'll keep the page locked. */
1589                         set_page_dirty(page);
1590                         ret = VM_FAULT_LOCKED;
1591                 }
1592         } while (err == -EAGAIN);
1593
1594         if (ret == VM_FAULT_LOCKED ||
1595             ci->i_inline_version != CEPH_INLINE_NONE) {
1596                 int dirty;
1597                 spin_lock(&ci->i_ceph_lock);
1598                 ci->i_inline_version = CEPH_INLINE_NONE;
1599                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1600                                                &prealloc_cf);
1601                 spin_unlock(&ci->i_ceph_lock);
1602                 if (dirty)
1603                         __mark_inode_dirty(inode, dirty);
1604         }
1605
1606         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1607              inode, off, len, ceph_cap_string(got), ret);
1608         ceph_put_cap_refs(ci, got);
1609 out_free:
1610         ceph_restore_sigs(&oldset);
1611         ceph_free_cap_flush(prealloc_cf);
1612         if (err < 0)
1613                 ret = vmf_error(err);
1614         return ret;
1615 }
1616
1617 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1618                            char *data, size_t len)
1619 {
1620         struct address_space *mapping = inode->i_mapping;
1621         struct page *page;
1622
1623         if (locked_page) {
1624                 page = locked_page;
1625         } else {
1626                 if (i_size_read(inode) == 0)
1627                         return;
1628                 page = find_or_create_page(mapping, 0,
1629                                            mapping_gfp_constraint(mapping,
1630                                            ~__GFP_FS));
1631                 if (!page)
1632                         return;
1633                 if (PageUptodate(page)) {
1634                         unlock_page(page);
1635                         put_page(page);
1636                         return;
1637                 }
1638         }
1639
1640         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1641              inode, ceph_vinop(inode), len, locked_page);
1642
1643         if (len > 0) {
1644                 void *kaddr = kmap_atomic(page);
1645                 memcpy(kaddr, data, len);
1646                 kunmap_atomic(kaddr);
1647         }
1648
1649         if (page != locked_page) {
1650                 if (len < PAGE_SIZE)
1651                         zero_user_segment(page, len, PAGE_SIZE);
1652                 else
1653                         flush_dcache_page(page);
1654
1655                 SetPageUptodate(page);
1656                 unlock_page(page);
1657                 put_page(page);
1658         }
1659 }
1660
1661 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1662 {
1663         struct inode *inode = file_inode(filp);
1664         struct ceph_inode_info *ci = ceph_inode(inode);
1665         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1666         struct ceph_osd_request *req;
1667         struct page *page = NULL;
1668         u64 len, inline_version;
1669         int err = 0;
1670         bool from_pagecache = false;
1671
1672         spin_lock(&ci->i_ceph_lock);
1673         inline_version = ci->i_inline_version;
1674         spin_unlock(&ci->i_ceph_lock);
1675
1676         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1677              inode, ceph_vinop(inode), inline_version);
1678
1679         if (inline_version == 1 || /* initial version, no data */
1680             inline_version == CEPH_INLINE_NONE)
1681                 goto out;
1682
1683         if (locked_page) {
1684                 page = locked_page;
1685                 WARN_ON(!PageUptodate(page));
1686         } else if (ceph_caps_issued(ci) &
1687                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1688                 page = find_get_page(inode->i_mapping, 0);
1689                 if (page) {
1690                         if (PageUptodate(page)) {
1691                                 from_pagecache = true;
1692                                 lock_page(page);
1693                         } else {
1694                                 put_page(page);
1695                                 page = NULL;
1696                         }
1697                 }
1698         }
1699
1700         if (page) {
1701                 len = i_size_read(inode);
1702                 if (len > PAGE_SIZE)
1703                         len = PAGE_SIZE;
1704         } else {
1705                 page = __page_cache_alloc(GFP_NOFS);
1706                 if (!page) {
1707                         err = -ENOMEM;
1708                         goto out;
1709                 }
1710                 err = __ceph_do_getattr(inode, page,
1711                                         CEPH_STAT_CAP_INLINE_DATA, true);
1712                 if (err < 0) {
1713                         /* no inline data */
1714                         if (err == -ENODATA)
1715                                 err = 0;
1716                         goto out;
1717                 }
1718                 len = err;
1719         }
1720
1721         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1722                                     ceph_vino(inode), 0, &len, 0, 1,
1723                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1724                                     NULL, 0, 0, false);
1725         if (IS_ERR(req)) {
1726                 err = PTR_ERR(req);
1727                 goto out;
1728         }
1729
1730         req->r_mtime = inode->i_mtime;
1731         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1732         if (!err)
1733                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1734         ceph_osdc_put_request(req);
1735         if (err < 0)
1736                 goto out;
1737
1738         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1739                                     ceph_vino(inode), 0, &len, 1, 3,
1740                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1741                                     NULL, ci->i_truncate_seq,
1742                                     ci->i_truncate_size, false);
1743         if (IS_ERR(req)) {
1744                 err = PTR_ERR(req);
1745                 goto out;
1746         }
1747
1748         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1749
1750         {
1751                 __le64 xattr_buf = cpu_to_le64(inline_version);
1752                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1753                                             "inline_version", &xattr_buf,
1754                                             sizeof(xattr_buf),
1755                                             CEPH_OSD_CMPXATTR_OP_GT,
1756                                             CEPH_OSD_CMPXATTR_MODE_U64);
1757                 if (err)
1758                         goto out_put;
1759         }
1760
1761         {
1762                 char xattr_buf[32];
1763                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1764                                          "%llu", inline_version);
1765                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1766                                             "inline_version",
1767                                             xattr_buf, xattr_len, 0, 0);
1768                 if (err)
1769                         goto out_put;
1770         }
1771
1772         req->r_mtime = inode->i_mtime;
1773         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1774         if (!err)
1775                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1776 out_put:
1777         ceph_osdc_put_request(req);
1778         if (err == -ECANCELED)
1779                 err = 0;
1780 out:
1781         if (page && page != locked_page) {
1782                 if (from_pagecache) {
1783                         unlock_page(page);
1784                         put_page(page);
1785                 } else
1786                         __free_pages(page, 0);
1787         }
1788
1789         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1790              inode, ceph_vinop(inode), inline_version, err);
1791         return err;
1792 }
1793
1794 static const struct vm_operations_struct ceph_vmops = {
1795         .fault          = ceph_filemap_fault,
1796         .page_mkwrite   = ceph_page_mkwrite,
1797 };
1798
1799 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1800 {
1801         struct address_space *mapping = file->f_mapping;
1802
1803         if (!mapping->a_ops->readpage)
1804                 return -ENOEXEC;
1805         file_accessed(file);
1806         vma->vm_ops = &ceph_vmops;
1807         return 0;
1808 }
1809
1810 enum {
1811         POOL_READ       = 1,
1812         POOL_WRITE      = 2,
1813 };
1814
1815 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1816                                 s64 pool, struct ceph_string *pool_ns)
1817 {
1818         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1819         struct ceph_mds_client *mdsc = fsc->mdsc;
1820         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1821         struct rb_node **p, *parent;
1822         struct ceph_pool_perm *perm;
1823         struct page **pages;
1824         size_t pool_ns_len;
1825         int err = 0, err2 = 0, have = 0;
1826
1827         down_read(&mdsc->pool_perm_rwsem);
1828         p = &mdsc->pool_perm_tree.rb_node;
1829         while (*p) {
1830                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1831                 if (pool < perm->pool)
1832                         p = &(*p)->rb_left;
1833                 else if (pool > perm->pool)
1834                         p = &(*p)->rb_right;
1835                 else {
1836                         int ret = ceph_compare_string(pool_ns,
1837                                                 perm->pool_ns,
1838                                                 perm->pool_ns_len);
1839                         if (ret < 0)
1840                                 p = &(*p)->rb_left;
1841                         else if (ret > 0)
1842                                 p = &(*p)->rb_right;
1843                         else {
1844                                 have = perm->perm;
1845                                 break;
1846                         }
1847                 }
1848         }
1849         up_read(&mdsc->pool_perm_rwsem);
1850         if (*p)
1851                 goto out;
1852
1853         if (pool_ns)
1854                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1855                      pool, (int)pool_ns->len, pool_ns->str);
1856         else
1857                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1858
1859         down_write(&mdsc->pool_perm_rwsem);
1860         p = &mdsc->pool_perm_tree.rb_node;
1861         parent = NULL;
1862         while (*p) {
1863                 parent = *p;
1864                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1865                 if (pool < perm->pool)
1866                         p = &(*p)->rb_left;
1867                 else if (pool > perm->pool)
1868                         p = &(*p)->rb_right;
1869                 else {
1870                         int ret = ceph_compare_string(pool_ns,
1871                                                 perm->pool_ns,
1872                                                 perm->pool_ns_len);
1873                         if (ret < 0)
1874                                 p = &(*p)->rb_left;
1875                         else if (ret > 0)
1876                                 p = &(*p)->rb_right;
1877                         else {
1878                                 have = perm->perm;
1879                                 break;
1880                         }
1881                 }
1882         }
1883         if (*p) {
1884                 up_write(&mdsc->pool_perm_rwsem);
1885                 goto out;
1886         }
1887
1888         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1889                                          1, false, GFP_NOFS);
1890         if (!rd_req) {
1891                 err = -ENOMEM;
1892                 goto out_unlock;
1893         }
1894
1895         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1896         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1897         rd_req->r_base_oloc.pool = pool;
1898         if (pool_ns)
1899                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1900         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1901
1902         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1903         if (err)
1904                 goto out_unlock;
1905
1906         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1907                                          1, false, GFP_NOFS);
1908         if (!wr_req) {
1909                 err = -ENOMEM;
1910                 goto out_unlock;
1911         }
1912
1913         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1914         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1915         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1916         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1917
1918         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1919         if (err)
1920                 goto out_unlock;
1921
1922         /* one page should be large enough for STAT data */
1923         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1924         if (IS_ERR(pages)) {
1925                 err = PTR_ERR(pages);
1926                 goto out_unlock;
1927         }
1928
1929         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1930                                      0, false, true);
1931         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1932
1933         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1934         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1935
1936         if (!err)
1937                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1938         if (!err2)
1939                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1940
1941         if (err >= 0 || err == -ENOENT)
1942                 have |= POOL_READ;
1943         else if (err != -EPERM)
1944                 goto out_unlock;
1945
1946         if (err2 == 0 || err2 == -EEXIST)
1947                 have |= POOL_WRITE;
1948         else if (err2 != -EPERM) {
1949                 err = err2;
1950                 goto out_unlock;
1951         }
1952
1953         pool_ns_len = pool_ns ? pool_ns->len : 0;
1954         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1955         if (!perm) {
1956                 err = -ENOMEM;
1957                 goto out_unlock;
1958         }
1959
1960         perm->pool = pool;
1961         perm->perm = have;
1962         perm->pool_ns_len = pool_ns_len;
1963         if (pool_ns_len > 0)
1964                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1965         perm->pool_ns[pool_ns_len] = 0;
1966
1967         rb_link_node(&perm->node, parent, p);
1968         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1969         err = 0;
1970 out_unlock:
1971         up_write(&mdsc->pool_perm_rwsem);
1972
1973         ceph_osdc_put_request(rd_req);
1974         ceph_osdc_put_request(wr_req);
1975 out:
1976         if (!err)
1977                 err = have;
1978         if (pool_ns)
1979                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1980                      pool, (int)pool_ns->len, pool_ns->str, err);
1981         else
1982                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1983         return err;
1984 }
1985
1986 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1987 {
1988         s64 pool;
1989         struct ceph_string *pool_ns;
1990         int ret, flags;
1991
1992         if (ci->i_vino.snap != CEPH_NOSNAP) {
1993                 /*
1994                  * Pool permission check needs to write to the first object.
1995                  * But for snapshot, head of the first object may have alread
1996                  * been deleted. Skip check to avoid creating orphan object.
1997                  */
1998                 return 0;
1999         }
2000
2001         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2002                                 NOPOOLPERM))
2003                 return 0;
2004
2005         spin_lock(&ci->i_ceph_lock);
2006         flags = ci->i_ceph_flags;
2007         pool = ci->i_layout.pool_id;
2008         spin_unlock(&ci->i_ceph_lock);
2009 check:
2010         if (flags & CEPH_I_POOL_PERM) {
2011                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2012                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2013                              pool);
2014                         return -EPERM;
2015                 }
2016                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2017                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2018                              pool);
2019                         return -EPERM;
2020                 }
2021                 return 0;
2022         }
2023
2024         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2025         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2026         ceph_put_string(pool_ns);
2027         if (ret < 0)
2028                 return ret;
2029
2030         flags = CEPH_I_POOL_PERM;
2031         if (ret & POOL_READ)
2032                 flags |= CEPH_I_POOL_RD;
2033         if (ret & POOL_WRITE)
2034                 flags |= CEPH_I_POOL_WR;
2035
2036         spin_lock(&ci->i_ceph_lock);
2037         if (pool == ci->i_layout.pool_id &&
2038             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2039                 ci->i_ceph_flags |= flags;
2040         } else {
2041                 pool = ci->i_layout.pool_id;
2042                 flags = ci->i_ceph_flags;
2043         }
2044         spin_unlock(&ci->i_ceph_lock);
2045         goto check;
2046 }
2047
2048 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2049 {
2050         struct ceph_pool_perm *perm;
2051         struct rb_node *n;
2052
2053         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2054                 n = rb_first(&mdsc->pool_perm_tree);
2055                 perm = rb_entry(n, struct ceph_pool_perm, node);
2056                 rb_erase(n, &mdsc->pool_perm_tree);
2057                 kfree(perm);
2058         }
2059 }