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
  19 /*
  20  * Ceph address space ops.
  21  *
  22  * There are a few funny things going on here.
  23  *
  24  * The page->private field is used to reference a struct
  25  * ceph_snap_context for _every_ dirty page.  This indicates which
  26  * snapshot the page was logically dirtied in, and thus which snap
  27  * context needs to be associated with the osd write during writeback.
  28  *
  29  * Similarly, struct ceph_inode_info maintains a set of counters to
  30  * count dirty pages on the inode.  In the absence of snapshots,
  31  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  32  *
  33  * When a snapshot is taken (that is, when the client receives
  34  * notification that a snapshot was taken), each inode with caps and
  35  * with dirty pages (dirty pages implies there is a cap) gets a new
  36  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  37  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  38  * moved to capsnap->dirty. (Unless a sync write is currently in
  39  * progress.  In that case, the capsnap is said to be "pending", new
  40  * writes cannot start, and the capsnap isn't "finalized" until the
  41  * write completes (or fails) and a final size/mtime for the inode for
  42  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  43  *
  44  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  45  * we look for the first capsnap in i_cap_snaps and write out pages in
  46  * that snap context _only_.  Then we move on to the next capsnap,
  47  * eventually reaching the "live" or "head" context (i.e., pages that
  48  * are not yet snapped) and are writing the most recently dirtied
  49  * pages.
  50  *
  51  * Invalidate and so forth must take care to ensure the dirty page
  52  * accounting is preserved.
  53  */
  54
  55 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  56 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  57         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  58          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  59
  60 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  61 {
  62         if (PagePrivate(page))
  63                 return (void *)page->private;
  64         return NULL;
  65 }
  66
  67 /*
  68  * Dirty a page.  Optimistically adjust accounting, on the assumption
  69  * that we won't race with invalidate.  If we do, readjust.
  70  */
  71 static int ceph_set_page_dirty(struct page *page)
  72 {
  73         struct address_space *mapping = page->mapping;
  74         struct inode *inode;
  75         struct ceph_inode_info *ci;
  76         struct ceph_snap_context *snapc;
  77         int ret;
  78
  79         if (unlikely(!mapping))
  80                 return !TestSetPageDirty(page);
  81
  82         if (PageDirty(page)) {
  83                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  84                      mapping->host, page, page->index);
  85                 BUG_ON(!PagePrivate(page));
  86                 return 0;
  87         }
  88
  89         inode = mapping->host;
  90         ci = ceph_inode(inode);
  91
  92         /* dirty the head */
  93         spin_lock(&ci->i_ceph_lock);
  94         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  95         if (__ceph_have_pending_cap_snap(ci)) {
  96                 struct ceph_cap_snap *capsnap =
  97                                 list_last_entry(&ci->i_cap_snaps,
  98                                                 struct ceph_cap_snap,
  99                                                 ci_item);
 100                 snapc = ceph_get_snap_context(capsnap->context);
 101                 capsnap->dirty_pages++;
 102         } else {
 103                 BUG_ON(!ci->i_head_snapc);
 104                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 105                 ++ci->i_wrbuffer_ref_head;
 106         }
 107         if (ci->i_wrbuffer_ref == 0)
 108                 ihold(inode);
 109         ++ci->i_wrbuffer_ref;
 110         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 111              "snapc %p seq %lld (%d snaps)\n",
 112              mapping->host, page, page->index,
 113              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 114              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 115              snapc, snapc->seq, snapc->num_snaps);
 116         spin_unlock(&ci->i_ceph_lock);
 117
 118         /*
 119          * Reference snap context in page->private.  Also set
 120          * PagePrivate so that we get invalidatepage callback.
 121          */
 122         BUG_ON(PagePrivate(page));
 123         page->private = (unsigned long)snapc;
 124         SetPagePrivate(page);
 125
 126         ret = __set_page_dirty_nobuffers(page);
 127         WARN_ON(!PageLocked(page));
 128         WARN_ON(!page->mapping);
 129
 130         return ret;
 131 }
 132
 133 /*
 134  * If we are truncating the full page (i.e. offset == 0), adjust the
 135  * dirty page counters appropriately.  Only called if there is private
 136  * data on the page.
 137  */
 138 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 139                                 unsigned int length)
 140 {
 141         struct inode *inode;
 142         struct ceph_inode_info *ci;
 143         struct ceph_snap_context *snapc = page_snap_context(page);
 144
 145         inode = page->mapping->host;
 146         ci = ceph_inode(inode);
 147
 148         if (offset != 0 || length != PAGE_SIZE) {
 149                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 150                      inode, page, page->index, offset, length);
 151                 return;
 152         }
 153
 154         ceph_invalidate_fscache_page(inode, page);
 155
 156         WARN_ON(!PageLocked(page));
 157         if (!PagePrivate(page))
 158                 return;
 159
 160         ClearPageChecked(page);
 161
 162         dout("%p invalidatepage %p idx %lu full dirty page\n",
 163              inode, page, page->index);
 164
 165         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 166         ceph_put_snap_context(snapc);
 167         page->private = 0;
 168         ClearPagePrivate(page);
 169 }
 170
 171 static int ceph_releasepage(struct page *page, gfp_t g)
 172 {
 173         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 174              page, page->index, PageDirty(page) ? "" : "not ");
 175
 176         /* Can we release the page from the cache? */
 177         if (!ceph_release_fscache_page(page, g))
 178                 return 0;
 179
 180         return !PagePrivate(page);
 181 }
 182
 183 /*
 184  * read a single page, without unlocking it.
 185  */
 186 static int ceph_do_readpage(struct file *filp, struct page *page)
 187 {
 188         struct inode *inode = file_inode(filp);
 189         struct ceph_inode_info *ci = ceph_inode(inode);
 190         struct ceph_osd_client *osdc =
 191                 &ceph_inode_to_client(inode)->client->osdc;
 192         int err = 0;
 193         u64 off = page_offset(page);
 194         u64 len = PAGE_SIZE;
 195
 196         if (off >= i_size_read(inode)) {
 197                 zero_user_segment(page, 0, PAGE_SIZE);
 198                 SetPageUptodate(page);
 199                 return 0;
 200         }
 201
 202         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 203                 /*
 204                  * Uptodate inline data should have been added
 205                  * into page cache while getting Fcr caps.
 206                  */
 207                 if (off == 0)
 208                         return -EINVAL;
 209                 zero_user_segment(page, 0, PAGE_SIZE);
 210                 SetPageUptodate(page);
 211                 return 0;
 212         }
 213
 214         err = ceph_readpage_from_fscache(inode, page);
 215         if (err == 0)
 216                 return -EINPROGRESS;
 217
 218         dout("readpage inode %p file %p page %p index %lu\n",
 219              inode, filp, page, page->index);
 220         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 221                                   off, &len,
 222                                   ci->i_truncate_seq, ci->i_truncate_size,
 223                                   &page, 1, 0);
 224         if (err == -ENOENT)
 225                 err = 0;
 226         if (err < 0) {
 227                 SetPageError(page);
 228                 ceph_fscache_readpage_cancel(inode, page);
 229                 goto out;
 230         }
 231         if (err < PAGE_SIZE)
 232                 /* zero fill remainder of page */
 233                 zero_user_segment(page, err, PAGE_SIZE);
 234         else
 235                 flush_dcache_page(page);
 236
 237         SetPageUptodate(page);
 238         ceph_readpage_to_fscache(inode, page);
 239
 240 out:
 241         return err < 0 ? err : 0;
 242 }
 243
 244 static int ceph_readpage(struct file *filp, struct page *page)
 245 {
 246         int r = ceph_do_readpage(filp, page);
 247         if (r != -EINPROGRESS)
 248                 unlock_page(page);
 249         else
 250                 r = 0;
 251         return r;
 252 }
 253
 254 /*
 255  * Finish an async read(ahead) op.
 256  */
 257 static void finish_read(struct ceph_osd_request *req)
 258 {
 259         struct inode *inode = req->r_inode;
 260         struct ceph_osd_data *osd_data;
 261         int rc = req->r_result <= 0 ? req->r_result : 0;
 262         int bytes = req->r_result >= 0 ? req->r_result : 0;
 263         int num_pages;
 264         int i;
 265
 266         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 267
 268         /* unlock all pages, zeroing any data we didn't read */
 269         osd_data = osd_req_op_extent_osd_data(req, 0);
 270         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 271         num_pages = calc_pages_for((u64)osd_data->alignment,
 272                                         (u64)osd_data->length);
 273         for (i = 0; i < num_pages; i++) {
 274                 struct page *page = osd_data->pages[i];
 275
 276                 if (rc < 0 && rc != -ENOENT) {
 277                         ceph_fscache_readpage_cancel(inode, page);
 278                         goto unlock;
 279                 }
 280                 if (bytes < (int)PAGE_SIZE) {
 281                         /* zero (remainder of) page */
 282                         int s = bytes < 0 ? 0 : bytes;
 283                         zero_user_segment(page, s, PAGE_SIZE);
 284                 }
 285                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 286                      page->index);
 287                 flush_dcache_page(page);
 288                 SetPageUptodate(page);
 289                 ceph_readpage_to_fscache(inode, page);
 290 unlock:
 291                 unlock_page(page);
 292                 put_page(page);
 293                 bytes -= PAGE_SIZE;
 294         }
 295         kfree(osd_data->pages);
 296 }
 297
 298 /*
 299  * start an async read(ahead) operation.  return nr_pages we submitted
 300  * a read for on success, or negative error code.
 301  */
 302 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 303 {
 304         struct ceph_osd_client *osdc =
 305                 &ceph_inode_to_client(inode)->client->osdc;
 306         struct ceph_inode_info *ci = ceph_inode(inode);
 307         struct page *page = list_entry(page_list->prev, struct page, lru);
 308         struct ceph_vino vino;
 309         struct ceph_osd_request *req;
 310         u64 off;
 311         u64 len;
 312         int i;
 313         struct page **pages;
 314         pgoff_t next_index;
 315         int nr_pages = 0;
 316         int got = 0;
 317         int ret = 0;
 318
 319         if (!current->journal_info) {
 320                 /* caller of readpages does not hold buffer and read caps
 321                  * (fadvise, madvise and readahead cases) */
 322                 int want = CEPH_CAP_FILE_CACHE;
 323                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
 324                 if (ret < 0) {
 325                         dout("start_read %p, error getting cap\n", inode);
 326                 } else if (!(got & want)) {
 327                         dout("start_read %p, no cache cap\n", inode);
 328                         ret = 0;
 329                 }
 330                 if (ret <= 0) {
 331                         if (got)
 332                                 ceph_put_cap_refs(ci, got);
 333                         while (!list_empty(page_list)) {
 334                                 page = list_entry(page_list->prev,
 335                                                   struct page, lru);
 336                                 list_del(&page->lru);
 337                                 put_page(page);
 338                         }
 339                         return ret;
 340                 }
 341         }
 342
 343         off = (u64) page_offset(page);
 344
 345         /* count pages */
 346         next_index = page->index;
 347         list_for_each_entry_reverse(page, page_list, lru) {
 348                 if (page->index != next_index)
 349                         break;
 350                 nr_pages++;
 351                 next_index++;
 352                 if (max && nr_pages == max)
 353                         break;
 354         }
 355         len = nr_pages << PAGE_SHIFT;
 356         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 357              off, len);
 358         vino = ceph_vino(inode);
 359         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 360                                     0, 1, CEPH_OSD_OP_READ,
 361                                     CEPH_OSD_FLAG_READ, NULL,
 362                                     ci->i_truncate_seq, ci->i_truncate_size,
 363                                     false);
 364         if (IS_ERR(req)) {
 365                 ret = PTR_ERR(req);
 366                 goto out;
 367         }
 368
 369         /* build page vector */
 370         nr_pages = calc_pages_for(0, len);
 371         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 372         if (!pages) {
 373                 ret = -ENOMEM;
 374                 goto out_put;
 375         }
 376         for (i = 0; i < nr_pages; ++i) {
 377                 page = list_entry(page_list->prev, struct page, lru);
 378                 BUG_ON(PageLocked(page));
 379                 list_del(&page->lru);
 380
 381                 dout("start_read %p adding %p idx %lu\n", inode, page,
 382                      page->index);
 383                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 384                                           GFP_KERNEL)) {
 385                         ceph_fscache_uncache_page(inode, page);
 386                         put_page(page);
 387                         dout("start_read %p add_to_page_cache failed %p\n",
 388                              inode, page);
 389                         nr_pages = i;
 390                         if (nr_pages > 0) {
 391                                 len = nr_pages << PAGE_SHIFT;
 392                                 osd_req_op_extent_update(req, 0, len);
 393                                 break;
 394                         }
 395                         goto out_pages;
 396                 }
 397                 pages[i] = page;
 398         }
 399         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 400         req->r_callback = finish_read;
 401         req->r_inode = inode;
 402
 403         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 404         ret = ceph_osdc_start_request(osdc, req, false);
 405         if (ret < 0)
 406                 goto out_pages;
 407         ceph_osdc_put_request(req);
 408
 409         /* After adding locked pages to page cache, the inode holds cache cap.
 410          * So we can drop our cap refs. */
 411         if (got)
 412                 ceph_put_cap_refs(ci, got);
 413
 414         return nr_pages;
 415
 416 out_pages:
 417         for (i = 0; i < nr_pages; ++i) {
 418                 ceph_fscache_readpage_cancel(inode, pages[i]);
 419                 unlock_page(pages[i]);
 420         }
 421         ceph_put_page_vector(pages, nr_pages, false);
 422 out_put:
 423         ceph_osdc_put_request(req);
 424 out:
 425         if (got)
 426                 ceph_put_cap_refs(ci, got);
 427         return ret;
 428 }
 429
 430
 431 /*
 432  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 433  * the caller (VM) cleans them up.
 434  */
 435 static int ceph_readpages(struct file *file, struct address_space *mapping,
 436                           struct list_head *page_list, unsigned nr_pages)
 437 {
 438         struct inode *inode = file_inode(file);
 439         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 440         int rc = 0;
 441         int max = 0;
 442
 443         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 444                 return -EINVAL;
 445
 446         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 447                                          &nr_pages);
 448
 449         if (rc == 0)
 450                 goto out;
 451
 452         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 453         dout("readpages %p file %p nr_pages %d max %d\n",
 454              inode, file, nr_pages, max);
 455         while (!list_empty(page_list)) {
 456                 rc = start_read(inode, page_list, max);
 457                 if (rc < 0)
 458                         goto out;
 459         }
 460 out:
 461         ceph_fscache_readpages_cancel(inode, page_list);
 462
 463         dout("readpages %p file %p ret %d\n", inode, file, rc);
 464         return rc;
 465 }
 466
 467 struct ceph_writeback_ctl
 468 {
 469         loff_t i_size;
 470         u64 truncate_size;
 471         u32 truncate_seq;
 472         bool size_stable;
 473         bool head_snapc;
 474 };
 475
 476 /*
 477  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 478  * only snap context we are allowed to write back.
 479  */
 480 static struct ceph_snap_context *
 481 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 482                    struct ceph_snap_context *page_snapc)
 483 {
 484         struct ceph_inode_info *ci = ceph_inode(inode);
 485         struct ceph_snap_context *snapc = NULL;
 486         struct ceph_cap_snap *capsnap = NULL;
 487
 488         spin_lock(&ci->i_ceph_lock);
 489         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 490                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 491                      capsnap->context, capsnap->dirty_pages);
 492                 if (!capsnap->dirty_pages)
 493                         continue;
 494
 495                 /* get i_size, truncate_{seq,size} for page_snapc? */
 496                 if (snapc && capsnap->context != page_snapc)
 497                         continue;
 498
 499                 if (ctl) {
 500                         if (capsnap->writing) {
 501                                 ctl->i_size = i_size_read(inode);
 502                                 ctl->size_stable = false;
 503                         } else {
 504                                 ctl->i_size = capsnap->size;
 505                                 ctl->size_stable = true;
 506                         }
 507                         ctl->truncate_size = capsnap->truncate_size;
 508                         ctl->truncate_seq = capsnap->truncate_seq;
 509                         ctl->head_snapc = false;
 510                 }
 511
 512                 if (snapc)
 513                         break;
 514
 515                 snapc = ceph_get_snap_context(capsnap->context);
 516                 if (!page_snapc ||
 517                     page_snapc == snapc ||
 518                     page_snapc->seq > snapc->seq)
 519                         break;
 520         }
 521         if (!snapc && ci->i_wrbuffer_ref_head) {
 522                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 523                 dout(" head snapc %p has %d dirty pages\n",
 524                      snapc, ci->i_wrbuffer_ref_head);
 525                 if (ctl) {
 526                         ctl->i_size = i_size_read(inode);
 527                         ctl->truncate_size = ci->i_truncate_size;
 528                         ctl->truncate_seq = ci->i_truncate_seq;
 529                         ctl->size_stable = false;
 530                         ctl->head_snapc = true;
 531                 }
 532         }
 533         spin_unlock(&ci->i_ceph_lock);
 534         return snapc;
 535 }
 536
 537 static u64 get_writepages_data_length(struct inode *inode,
 538                                       struct page *page, u64 start)
 539 {
 540         struct ceph_inode_info *ci = ceph_inode(inode);
 541         struct ceph_snap_context *snapc = page_snap_context(page);
 542         struct ceph_cap_snap *capsnap = NULL;
 543         u64 end = i_size_read(inode);
 544
 545         if (snapc != ci->i_head_snapc) {
 546                 bool found = false;
 547                 spin_lock(&ci->i_ceph_lock);
 548                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 549                         if (capsnap->context == snapc) {
 550                                 if (!capsnap->writing)
 551                                         end = capsnap->size;
 552                                 found = true;
 553                                 break;
 554                         }
 555                 }
 556                 spin_unlock(&ci->i_ceph_lock);
 557                 WARN_ON(!found);
 558         }
 559         if (end > page_offset(page) + PAGE_SIZE)
 560                 end = page_offset(page) + PAGE_SIZE;
 561         return end > start ? end - start : 0;
 562 }
 563
 564 /*
 565  * Write a single page, but leave the page locked.
 566  *
 567  * If we get a write error, set the page error bit, but still adjust the
 568  * dirty page accounting (i.e., page is no longer dirty).
 569  */
 570 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 571 {
 572         struct inode *inode;
 573         struct ceph_inode_info *ci;
 574         struct ceph_fs_client *fsc;
 575         struct ceph_snap_context *snapc, *oldest;
 576         loff_t page_off = page_offset(page);
 577         long writeback_stat;
 578         int err, len = PAGE_SIZE;
 579         struct ceph_writeback_ctl ceph_wbc;
 580
 581         dout("writepage %p idx %lu\n", page, page->index);
 582
 583         inode = page->mapping->host;
 584         ci = ceph_inode(inode);
 585         fsc = ceph_inode_to_client(inode);
 586
 587         /* verify this is a writeable snap context */
 588         snapc = page_snap_context(page);
 589         if (!snapc) {
 590                 dout("writepage %p page %p not dirty?\n", inode, page);
 591                 return 0;
 592         }
 593         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 594         if (snapc->seq > oldest->seq) {
 595                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 596                      inode, page, snapc);
 597                 /* we should only noop if called by kswapd */
 598                 WARN_ON(!(current->flags & PF_MEMALLOC));
 599                 ceph_put_snap_context(oldest);
 600                 redirty_page_for_writepage(wbc, page);
 601                 return 0;
 602         }
 603         ceph_put_snap_context(oldest);
 604
 605         /* is this a partial page at end of file? */
 606         if (page_off >= ceph_wbc.i_size) {
 607                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 608                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 609                 return 0;
 610         }
 611
 612         if (ceph_wbc.i_size < page_off + len)
 613                 len = ceph_wbc.i_size - page_off;
 614
 615         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 616              inode, page, page->index, page_off, len, snapc, snapc->seq);
 617
 618         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 619         if (writeback_stat >
 620             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 621                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 622
 623         set_page_writeback(page);
 624         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 625                                    &ci->i_layout, snapc, page_off, len,
 626                                    ceph_wbc.truncate_seq,
 627                                    ceph_wbc.truncate_size,
 628                                    &inode->i_mtime, &page, 1);
 629         if (err < 0) {
 630                 struct writeback_control tmp_wbc;
 631                 if (!wbc)
 632                         wbc = &tmp_wbc;
 633                 if (err == -ERESTARTSYS) {
 634                         /* killed by SIGKILL */
 635                         dout("writepage interrupted page %p\n", page);
 636                         redirty_page_for_writepage(wbc, page);
 637                         end_page_writeback(page);
 638                         return err;
 639                 }
 640                 dout("writepage setting page/mapping error %d %p\n",
 641                      err, page);
 642                 SetPageError(page);
 643                 mapping_set_error(&inode->i_data, err);
 644                 wbc->pages_skipped++;
 645         } else {
 646                 dout("writepage cleaned page %p\n", page);
 647                 err = 0;  /* vfs expects us to return 0 */
 648         }
 649         page->private = 0;
 650         ClearPagePrivate(page);
 651         end_page_writeback(page);
 652         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 653         ceph_put_snap_context(snapc);  /* page's reference */
 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, 0);
 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 stop, 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, 0);
 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 || wbc->sync_mode != WB_SYNC_NONE)
 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         stop = false;
 861         while (!stop && index <= end) {
 862                 int num_ops = 0, op_idx;
 863                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 864                 struct page **pages = NULL, **data_pages;
 865                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 866                 struct page *page;
 867                 pgoff_t strip_unit_end = 0;
 868                 u64 offset = 0, len = 0;
 869
 870                 max_pages = wsize >> PAGE_SHIFT;
 871
 872 get_more_pages:
 873                 pvec_pages = min_t(unsigned, PAGEVEC_SIZE,
 874                                    max_pages - locked_pages);
 875                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 876                                                 end, PAGECACHE_TAG_DIRTY,
 877                                                 pvec_pages);
 878                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 879                 if (!pvec_pages && !locked_pages)
 880                         break;
 881                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 882                         page = pvec.pages[i];
 883                         dout("? %p idx %lu\n", page, page->index);
 884                         if (locked_pages == 0)
 885                                 lock_page(page);  /* first page */
 886                         else if (!trylock_page(page))
 887                                 break;
 888
 889                         /* only dirty pages, or our accounting breaks */
 890                         if (unlikely(!PageDirty(page)) ||
 891                             unlikely(page->mapping != mapping)) {
 892                                 dout("!dirty or !mapping %p\n", page);
 893                                 unlock_page(page);
 894                                 continue;
 895                         }
 896                         if (strip_unit_end && (page->index > strip_unit_end)) {
 897                                 dout("end of strip unit %p\n", page);
 898                                 unlock_page(page);
 899                                 break;
 900                         }
 901                         if (page_offset(page) >= ceph_wbc.i_size) {
 902                                 dout("%p page eof %llu\n",
 903                                      page, ceph_wbc.i_size);
 904                                 /* not done if range_cyclic */
 905                                 stop = true;
 906                                 unlock_page(page);
 907                                 break;
 908                         }
 909                         if (PageWriteback(page)) {
 910                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 911                                         dout("%p under writeback\n", page);
 912                                         unlock_page(page);
 913                                         continue;
 914                                 }
 915                                 dout("waiting on writeback %p\n", page);
 916                                 wait_on_page_writeback(page);
 917                         }
 918
 919                         /* only if matching snap context */
 920                         pgsnapc = page_snap_context(page);
 921                         if (pgsnapc != snapc) {
 922                                 dout("page snapc %p %lld != oldest %p %lld\n",
 923                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 924                                 unlock_page(page);
 925                                 continue;
 926                         }
 927
 928                         if (!clear_page_dirty_for_io(page)) {
 929                                 dout("%p !clear_page_dirty_for_io\n", page);
 930                                 unlock_page(page);
 931                                 continue;
 932                         }
 933
 934                         /*
 935                          * We have something to write.  If this is
 936                          * the first locked page this time through,
 937                          * calculate max possinle write size and
 938                          * allocate a page array
 939                          */
 940                         if (locked_pages == 0) {
 941                                 u64 objnum;
 942                                 u64 objoff;
 943
 944                                 /* prepare async write request */
 945                                 offset = (u64)page_offset(page);
 946                                 len = wsize;
 947
 948                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 949                                                                 offset, len,
 950                                                                 &objnum, &objoff,
 951                                                                 &len);
 952                                 if (rc < 0) {
 953                                         unlock_page(page);
 954                                         break;
 955                                 }
 956
 957                                 num_ops = 1;
 958                                 strip_unit_end = page->index +
 959                                         ((len - 1) >> PAGE_SHIFT);
 960
 961                                 BUG_ON(pages);
 962                                 max_pages = calc_pages_for(0, (u64)len);
 963                                 pages = kmalloc(max_pages * 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(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 = stop = 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                                                         PAGEVEC_SIZE))) {
1169                                 for (i = 0; i < nr; i++) {
1170                                         page = pvec.pages[i];
1171                                         if (page_snap_context(page) != snapc)
1172                                                 continue;
1173                                         wait_on_page_writeback(page);
1174                                 }
1175                                 pagevec_release(&pvec);
1176                                 cond_resched();
1177                         }
1178                 }
1179
1180                 start_index = 0;
1181                 index = 0;
1182                 goto retry;
1183         }
1184
1185         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1186                 mapping->writeback_index = index;
1187
1188 out:
1189         ceph_osdc_put_request(req);
1190         ceph_put_snap_context(last_snapc);
1191         dout("writepages dend - startone, rc = %d\n", rc);
1192         return rc;
1193 }
1194
1195
1196
1197 /*
1198  * See if a given @snapc is either writeable, or already written.
1199  */
1200 static int context_is_writeable_or_written(struct inode *inode,
1201                                            struct ceph_snap_context *snapc)
1202 {
1203         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1204         int ret = !oldest || snapc->seq <= oldest->seq;
1205
1206         ceph_put_snap_context(oldest);
1207         return ret;
1208 }
1209
1210 /*
1211  * We are only allowed to write into/dirty the page if the page is
1212  * clean, or already dirty within the same snap context.
1213  *
1214  * called with page locked.
1215  * return success with page locked,
1216  * or any failure (incl -EAGAIN) with page unlocked.
1217  */
1218 static int ceph_update_writeable_page(struct file *file,
1219                             loff_t pos, unsigned len,
1220                             struct page *page)
1221 {
1222         struct inode *inode = file_inode(file);
1223         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1224         struct ceph_inode_info *ci = ceph_inode(inode);
1225         loff_t page_off = pos & PAGE_MASK;
1226         int pos_in_page = pos & ~PAGE_MASK;
1227         int end_in_page = pos_in_page + len;
1228         loff_t i_size;
1229         int r;
1230         struct ceph_snap_context *snapc, *oldest;
1231
1232         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1233                 dout(" page %p forced umount\n", page);
1234                 unlock_page(page);
1235                 return -EIO;
1236         }
1237
1238 retry_locked:
1239         /* writepages currently holds page lock, but if we change that later, */
1240         wait_on_page_writeback(page);
1241
1242         snapc = page_snap_context(page);
1243         if (snapc && snapc != ci->i_head_snapc) {
1244                 /*
1245                  * this page is already dirty in another (older) snap
1246                  * context!  is it writeable now?
1247                  */
1248                 oldest = get_oldest_context(inode, NULL, NULL);
1249                 if (snapc->seq > oldest->seq) {
1250                         ceph_put_snap_context(oldest);
1251                         dout(" page %p snapc %p not current or oldest\n",
1252                              page, snapc);
1253                         /*
1254                          * queue for writeback, and wait for snapc to
1255                          * be writeable or written
1256                          */
1257                         snapc = ceph_get_snap_context(snapc);
1258                         unlock_page(page);
1259                         ceph_queue_writeback(inode);
1260                         r = wait_event_killable(ci->i_cap_wq,
1261                                context_is_writeable_or_written(inode, snapc));
1262                         ceph_put_snap_context(snapc);
1263                         if (r == -ERESTARTSYS)
1264                                 return r;
1265                         return -EAGAIN;
1266                 }
1267                 ceph_put_snap_context(oldest);
1268
1269                 /* yay, writeable, do it now (without dropping page lock) */
1270                 dout(" page %p snapc %p not current, but oldest\n",
1271                      page, snapc);
1272                 if (!clear_page_dirty_for_io(page))
1273                         goto retry_locked;
1274                 r = writepage_nounlock(page, NULL);
1275                 if (r < 0)
1276                         goto fail_unlock;
1277                 goto retry_locked;
1278         }
1279
1280         if (PageUptodate(page)) {
1281                 dout(" page %p already uptodate\n", page);
1282                 return 0;
1283         }
1284
1285         /* full page? */
1286         if (pos_in_page == 0 && len == PAGE_SIZE)
1287                 return 0;
1288
1289         /* past end of file? */
1290         i_size = i_size_read(inode);
1291
1292         if (page_off >= i_size ||
1293             (pos_in_page == 0 && (pos+len) >= i_size &&
1294              end_in_page - pos_in_page != PAGE_SIZE)) {
1295                 dout(" zeroing %p 0 - %d and %d - %d\n",
1296                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1297                 zero_user_segments(page,
1298                                    0, pos_in_page,
1299                                    end_in_page, PAGE_SIZE);
1300                 return 0;
1301         }
1302
1303         /* we need to read it. */
1304         r = ceph_do_readpage(file, page);
1305         if (r < 0) {
1306                 if (r == -EINPROGRESS)
1307                         return -EAGAIN;
1308                 goto fail_unlock;
1309         }
1310         goto retry_locked;
1311 fail_unlock:
1312         unlock_page(page);
1313         return r;
1314 }
1315
1316 /*
1317  * We are only allowed to write into/dirty the page if the page is
1318  * clean, or already dirty within the same snap context.
1319  */
1320 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1321                             loff_t pos, unsigned len, unsigned flags,
1322                             struct page **pagep, void **fsdata)
1323 {
1324         struct inode *inode = file_inode(file);
1325         struct page *page;
1326         pgoff_t index = pos >> PAGE_SHIFT;
1327         int r;
1328
1329         do {
1330                 /* get a page */
1331                 page = grab_cache_page_write_begin(mapping, index, 0);
1332                 if (!page)
1333                         return -ENOMEM;
1334
1335                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1336                      inode, page, (int)pos, (int)len);
1337
1338                 r = ceph_update_writeable_page(file, pos, len, page);
1339                 if (r < 0)
1340                         put_page(page);
1341                 else
1342                         *pagep = page;
1343         } while (r == -EAGAIN);
1344
1345         return r;
1346 }
1347
1348 /*
1349  * we don't do anything in here that simple_write_end doesn't do
1350  * except adjust dirty page accounting
1351  */
1352 static int ceph_write_end(struct file *file, struct address_space *mapping,
1353                           loff_t pos, unsigned len, unsigned copied,
1354                           struct page *page, void *fsdata)
1355 {
1356         struct inode *inode = file_inode(file);
1357         bool check_cap = false;
1358
1359         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1360              inode, page, (int)pos, (int)copied, (int)len);
1361
1362         /* zero the stale part of the page if we did a short copy */
1363         if (!PageUptodate(page)) {
1364                 if (copied < len) {
1365                         copied = 0;
1366                         goto out;
1367                 }
1368                 SetPageUptodate(page);
1369         }
1370
1371         /* did file size increase? */
1372         if (pos+copied > i_size_read(inode))
1373                 check_cap = ceph_inode_set_size(inode, pos+copied);
1374
1375         set_page_dirty(page);
1376
1377 out:
1378         unlock_page(page);
1379         put_page(page);
1380
1381         if (check_cap)
1382                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1383
1384         return copied;
1385 }
1386
1387 /*
1388  * we set .direct_IO to indicate direct io is supported, but since we
1389  * intercept O_DIRECT reads and writes early, this function should
1390  * never get called.
1391  */
1392 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1393 {
1394         WARN_ON(1);
1395         return -EINVAL;
1396 }
1397
1398 const struct address_space_operations ceph_aops = {
1399         .readpage = ceph_readpage,
1400         .readpages = ceph_readpages,
1401         .writepage = ceph_writepage,
1402         .writepages = ceph_writepages_start,
1403         .write_begin = ceph_write_begin,
1404         .write_end = ceph_write_end,
1405         .set_page_dirty = ceph_set_page_dirty,
1406         .invalidatepage = ceph_invalidatepage,
1407         .releasepage = ceph_releasepage,
1408         .direct_IO = ceph_direct_io,
1409 };
1410
1411 static void ceph_block_sigs(sigset_t *oldset)
1412 {
1413         sigset_t mask;
1414         siginitsetinv(&mask, sigmask(SIGKILL));
1415         sigprocmask(SIG_BLOCK, &mask, oldset);
1416 }
1417
1418 static void ceph_restore_sigs(sigset_t *oldset)
1419 {
1420         sigprocmask(SIG_SETMASK, oldset, NULL);
1421 }
1422
1423 /*
1424  * vm ops
1425  */
1426 static int ceph_filemap_fault(struct vm_fault *vmf)
1427 {
1428         struct vm_area_struct *vma = vmf->vma;
1429         struct inode *inode = file_inode(vma->vm_file);
1430         struct ceph_inode_info *ci = ceph_inode(inode);
1431         struct ceph_file_info *fi = vma->vm_file->private_data;
1432         struct page *pinned_page = NULL;
1433         loff_t off = vmf->pgoff << PAGE_SHIFT;
1434         int want, got, ret;
1435         sigset_t oldset;
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         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1448         if (ret < 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                 current->journal_info = vma->vm_file;
1457                 ret = filemap_fault(vmf);
1458                 current->journal_info = NULL;
1459         } else
1460                 ret = -EAGAIN;
1461
1462         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1463              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1464         if (pinned_page)
1465                 put_page(pinned_page);
1466         ceph_put_cap_refs(ci, got);
1467
1468         if (ret != -EAGAIN)
1469                 goto out_restore;
1470
1471         /* read inline data */
1472         if (off >= PAGE_SIZE) {
1473                 /* does not support inline data > PAGE_SIZE */
1474                 ret = VM_FAULT_SIGBUS;
1475         } else {
1476                 int ret1;
1477                 struct address_space *mapping = inode->i_mapping;
1478                 struct page *page = find_or_create_page(mapping, 0,
1479                                                 mapping_gfp_constraint(mapping,
1480                                                 ~__GFP_FS));
1481                 if (!page) {
1482                         ret = VM_FAULT_OOM;
1483                         goto out_inline;
1484                 }
1485                 ret1 = __ceph_do_getattr(inode, page,
1486                                          CEPH_STAT_CAP_INLINE_DATA, true);
1487                 if (ret1 < 0 || off >= i_size_read(inode)) {
1488                         unlock_page(page);
1489                         put_page(page);
1490                         if (ret1 < 0)
1491                                 ret = ret1;
1492                         else
1493                                 ret = VM_FAULT_SIGBUS;
1494                         goto out_inline;
1495                 }
1496                 if (ret1 < PAGE_SIZE)
1497                         zero_user_segment(page, ret1, PAGE_SIZE);
1498                 else
1499                         flush_dcache_page(page);
1500                 SetPageUptodate(page);
1501                 vmf->page = page;
1502                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1503 out_inline:
1504                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1505                      inode, off, (size_t)PAGE_SIZE, ret);
1506         }
1507 out_restore:
1508         ceph_restore_sigs(&oldset);
1509         if (ret < 0)
1510                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1511
1512         return ret;
1513 }
1514
1515 /*
1516  * Reuse write_begin here for simplicity.
1517  */
1518 static int ceph_page_mkwrite(struct vm_fault *vmf)
1519 {
1520         struct vm_area_struct *vma = vmf->vma;
1521         struct inode *inode = file_inode(vma->vm_file);
1522         struct ceph_inode_info *ci = ceph_inode(inode);
1523         struct ceph_file_info *fi = vma->vm_file->private_data;
1524         struct ceph_cap_flush *prealloc_cf;
1525         struct page *page = vmf->page;
1526         loff_t off = page_offset(page);
1527         loff_t size = i_size_read(inode);
1528         size_t len;
1529         int want, got, ret;
1530         sigset_t oldset;
1531
1532         prealloc_cf = ceph_alloc_cap_flush();
1533         if (!prealloc_cf)
1534                 return VM_FAULT_OOM;
1535
1536         ceph_block_sigs(&oldset);
1537
1538         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1539                 struct page *locked_page = NULL;
1540                 if (off == 0) {
1541                         lock_page(page);
1542                         locked_page = page;
1543                 }
1544                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1545                 if (locked_page)
1546                         unlock_page(locked_page);
1547                 if (ret < 0)
1548                         goto out_free;
1549         }
1550
1551         if (off + PAGE_SIZE <= size)
1552                 len = PAGE_SIZE;
1553         else
1554                 len = size & ~PAGE_MASK;
1555
1556         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1557              inode, ceph_vinop(inode), off, len, size);
1558         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1559                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1560         else
1561                 want = CEPH_CAP_FILE_BUFFER;
1562
1563         got = 0;
1564         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1565                             &got, NULL);
1566         if (ret < 0)
1567                 goto out_free;
1568
1569         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1570              inode, off, len, ceph_cap_string(got));
1571
1572         /* Update time before taking page lock */
1573         file_update_time(vma->vm_file);
1574
1575         do {
1576                 lock_page(page);
1577
1578                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1579                         unlock_page(page);
1580                         ret = VM_FAULT_NOPAGE;
1581                         break;
1582                 }
1583
1584                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1585                 if (ret >= 0) {
1586                         /* success.  we'll keep the page locked. */
1587                         set_page_dirty(page);
1588                         ret = VM_FAULT_LOCKED;
1589                 }
1590         } while (ret == -EAGAIN);
1591
1592         if (ret == VM_FAULT_LOCKED ||
1593             ci->i_inline_version != CEPH_INLINE_NONE) {
1594                 int dirty;
1595                 spin_lock(&ci->i_ceph_lock);
1596                 ci->i_inline_version = CEPH_INLINE_NONE;
1597                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1598                                                &prealloc_cf);
1599                 spin_unlock(&ci->i_ceph_lock);
1600                 if (dirty)
1601                         __mark_inode_dirty(inode, dirty);
1602         }
1603
1604         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1605              inode, off, len, ceph_cap_string(got), ret);
1606         ceph_put_cap_refs(ci, got);
1607 out_free:
1608         ceph_restore_sigs(&oldset);
1609         ceph_free_cap_flush(prealloc_cf);
1610         if (ret < 0)
1611                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1612         return ret;
1613 }
1614
1615 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1616                            char *data, size_t len)
1617 {
1618         struct address_space *mapping = inode->i_mapping;
1619         struct page *page;
1620
1621         if (locked_page) {
1622                 page = locked_page;
1623         } else {
1624                 if (i_size_read(inode) == 0)
1625                         return;
1626                 page = find_or_create_page(mapping, 0,
1627                                            mapping_gfp_constraint(mapping,
1628                                            ~__GFP_FS));
1629                 if (!page)
1630                         return;
1631                 if (PageUptodate(page)) {
1632                         unlock_page(page);
1633                         put_page(page);
1634                         return;
1635                 }
1636         }
1637
1638         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1639              inode, ceph_vinop(inode), len, locked_page);
1640
1641         if (len > 0) {
1642                 void *kaddr = kmap_atomic(page);
1643                 memcpy(kaddr, data, len);
1644                 kunmap_atomic(kaddr);
1645         }
1646
1647         if (page != locked_page) {
1648                 if (len < PAGE_SIZE)
1649                         zero_user_segment(page, len, PAGE_SIZE);
1650                 else
1651                         flush_dcache_page(page);
1652
1653                 SetPageUptodate(page);
1654                 unlock_page(page);
1655                 put_page(page);
1656         }
1657 }
1658
1659 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1660 {
1661         struct inode *inode = file_inode(filp);
1662         struct ceph_inode_info *ci = ceph_inode(inode);
1663         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1664         struct ceph_osd_request *req;
1665         struct page *page = NULL;
1666         u64 len, inline_version;
1667         int err = 0;
1668         bool from_pagecache = false;
1669
1670         spin_lock(&ci->i_ceph_lock);
1671         inline_version = ci->i_inline_version;
1672         spin_unlock(&ci->i_ceph_lock);
1673
1674         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1675              inode, ceph_vinop(inode), inline_version);
1676
1677         if (inline_version == 1 || /* initial version, no data */
1678             inline_version == CEPH_INLINE_NONE)
1679                 goto out;
1680
1681         if (locked_page) {
1682                 page = locked_page;
1683                 WARN_ON(!PageUptodate(page));
1684         } else if (ceph_caps_issued(ci) &
1685                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1686                 page = find_get_page(inode->i_mapping, 0);
1687                 if (page) {
1688                         if (PageUptodate(page)) {
1689                                 from_pagecache = true;
1690                                 lock_page(page);
1691                         } else {
1692                                 put_page(page);
1693                                 page = NULL;
1694                         }
1695                 }
1696         }
1697
1698         if (page) {
1699                 len = i_size_read(inode);
1700                 if (len > PAGE_SIZE)
1701                         len = PAGE_SIZE;
1702         } else {
1703                 page = __page_cache_alloc(GFP_NOFS);
1704                 if (!page) {
1705                         err = -ENOMEM;
1706                         goto out;
1707                 }
1708                 err = __ceph_do_getattr(inode, page,
1709                                         CEPH_STAT_CAP_INLINE_DATA, true);
1710                 if (err < 0) {
1711                         /* no inline data */
1712                         if (err == -ENODATA)
1713                                 err = 0;
1714                         goto out;
1715                 }
1716                 len = err;
1717         }
1718
1719         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1720                                     ceph_vino(inode), 0, &len, 0, 1,
1721                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1722                                     NULL, 0, 0, false);
1723         if (IS_ERR(req)) {
1724                 err = PTR_ERR(req);
1725                 goto out;
1726         }
1727
1728         req->r_mtime = inode->i_mtime;
1729         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1730         if (!err)
1731                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1732         ceph_osdc_put_request(req);
1733         if (err < 0)
1734                 goto out;
1735
1736         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1737                                     ceph_vino(inode), 0, &len, 1, 3,
1738                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1739                                     NULL, ci->i_truncate_seq,
1740                                     ci->i_truncate_size, false);
1741         if (IS_ERR(req)) {
1742                 err = PTR_ERR(req);
1743                 goto out;
1744         }
1745
1746         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1747
1748         {
1749                 __le64 xattr_buf = cpu_to_le64(inline_version);
1750                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1751                                             "inline_version", &xattr_buf,
1752                                             sizeof(xattr_buf),
1753                                             CEPH_OSD_CMPXATTR_OP_GT,
1754                                             CEPH_OSD_CMPXATTR_MODE_U64);
1755                 if (err)
1756                         goto out_put;
1757         }
1758
1759         {
1760                 char xattr_buf[32];
1761                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1762                                          "%llu", inline_version);
1763                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1764                                             "inline_version",
1765                                             xattr_buf, xattr_len, 0, 0);
1766                 if (err)
1767                         goto out_put;
1768         }
1769
1770         req->r_mtime = inode->i_mtime;
1771         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1772         if (!err)
1773                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1774 out_put:
1775         ceph_osdc_put_request(req);
1776         if (err == -ECANCELED)
1777                 err = 0;
1778 out:
1779         if (page && page != locked_page) {
1780                 if (from_pagecache) {
1781                         unlock_page(page);
1782                         put_page(page);
1783                 } else
1784                         __free_pages(page, 0);
1785         }
1786
1787         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1788              inode, ceph_vinop(inode), inline_version, err);
1789         return err;
1790 }
1791
1792 static const struct vm_operations_struct ceph_vmops = {
1793         .fault          = ceph_filemap_fault,
1794         .page_mkwrite   = ceph_page_mkwrite,
1795 };
1796
1797 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1798 {
1799         struct address_space *mapping = file->f_mapping;
1800
1801         if (!mapping->a_ops->readpage)
1802                 return -ENOEXEC;
1803         file_accessed(file);
1804         vma->vm_ops = &ceph_vmops;
1805         return 0;
1806 }
1807
1808 enum {
1809         POOL_READ       = 1,
1810         POOL_WRITE      = 2,
1811 };
1812
1813 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1814                                 s64 pool, struct ceph_string *pool_ns)
1815 {
1816         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1817         struct ceph_mds_client *mdsc = fsc->mdsc;
1818         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1819         struct rb_node **p, *parent;
1820         struct ceph_pool_perm *perm;
1821         struct page **pages;
1822         size_t pool_ns_len;
1823         int err = 0, err2 = 0, have = 0;
1824
1825         down_read(&mdsc->pool_perm_rwsem);
1826         p = &mdsc->pool_perm_tree.rb_node;
1827         while (*p) {
1828                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1829                 if (pool < perm->pool)
1830                         p = &(*p)->rb_left;
1831                 else if (pool > perm->pool)
1832                         p = &(*p)->rb_right;
1833                 else {
1834                         int ret = ceph_compare_string(pool_ns,
1835                                                 perm->pool_ns,
1836                                                 perm->pool_ns_len);
1837                         if (ret < 0)
1838                                 p = &(*p)->rb_left;
1839                         else if (ret > 0)
1840                                 p = &(*p)->rb_right;
1841                         else {
1842                                 have = perm->perm;
1843                                 break;
1844                         }
1845                 }
1846         }
1847         up_read(&mdsc->pool_perm_rwsem);
1848         if (*p)
1849                 goto out;
1850
1851         if (pool_ns)
1852                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1853                      pool, (int)pool_ns->len, pool_ns->str);
1854         else
1855                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1856
1857         down_write(&mdsc->pool_perm_rwsem);
1858         p = &mdsc->pool_perm_tree.rb_node;
1859         parent = NULL;
1860         while (*p) {
1861                 parent = *p;
1862                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1863                 if (pool < perm->pool)
1864                         p = &(*p)->rb_left;
1865                 else if (pool > perm->pool)
1866                         p = &(*p)->rb_right;
1867                 else {
1868                         int ret = ceph_compare_string(pool_ns,
1869                                                 perm->pool_ns,
1870                                                 perm->pool_ns_len);
1871                         if (ret < 0)
1872                                 p = &(*p)->rb_left;
1873                         else if (ret > 0)
1874                                 p = &(*p)->rb_right;
1875                         else {
1876                                 have = perm->perm;
1877                                 break;
1878                         }
1879                 }
1880         }
1881         if (*p) {
1882                 up_write(&mdsc->pool_perm_rwsem);
1883                 goto out;
1884         }
1885
1886         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1887                                          1, false, GFP_NOFS);
1888         if (!rd_req) {
1889                 err = -ENOMEM;
1890                 goto out_unlock;
1891         }
1892
1893         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1894         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1895         rd_req->r_base_oloc.pool = pool;
1896         if (pool_ns)
1897                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1898         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1899
1900         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1901         if (err)
1902                 goto out_unlock;
1903
1904         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1905                                          1, false, GFP_NOFS);
1906         if (!wr_req) {
1907                 err = -ENOMEM;
1908                 goto out_unlock;
1909         }
1910
1911         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1912         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1913         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1914         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1915
1916         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1917         if (err)
1918                 goto out_unlock;
1919
1920         /* one page should be large enough for STAT data */
1921         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1922         if (IS_ERR(pages)) {
1923                 err = PTR_ERR(pages);
1924                 goto out_unlock;
1925         }
1926
1927         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1928                                      0, false, true);
1929         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1930
1931         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1932         wr_req->r_abort_on_full = true;
1933         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1934
1935         if (!err)
1936                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1937         if (!err2)
1938                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1939
1940         if (err >= 0 || err == -ENOENT)
1941                 have |= POOL_READ;
1942         else if (err != -EPERM)
1943                 goto out_unlock;
1944
1945         if (err2 == 0 || err2 == -EEXIST)
1946                 have |= POOL_WRITE;
1947         else if (err2 != -EPERM) {
1948                 err = err2;
1949                 goto out_unlock;
1950         }
1951
1952         pool_ns_len = pool_ns ? pool_ns->len : 0;
1953         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1954         if (!perm) {
1955                 err = -ENOMEM;
1956                 goto out_unlock;
1957         }
1958
1959         perm->pool = pool;
1960         perm->perm = have;
1961         perm->pool_ns_len = pool_ns_len;
1962         if (pool_ns_len > 0)
1963                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1964         perm->pool_ns[pool_ns_len] = 0;
1965
1966         rb_link_node(&perm->node, parent, p);
1967         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1968         err = 0;
1969 out_unlock:
1970         up_write(&mdsc->pool_perm_rwsem);
1971
1972         ceph_osdc_put_request(rd_req);
1973         ceph_osdc_put_request(wr_req);
1974 out:
1975         if (!err)
1976                 err = have;
1977         if (pool_ns)
1978                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1979                      pool, (int)pool_ns->len, pool_ns->str, err);
1980         else
1981                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1982         return err;
1983 }
1984
1985 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1986 {
1987         s64 pool;
1988         struct ceph_string *pool_ns;
1989         int ret, flags;
1990
1991         if (ci->i_vino.snap != CEPH_NOSNAP) {
1992                 /*
1993                  * Pool permission check needs to write to the first object.
1994                  * But for snapshot, head of the first object may have alread
1995                  * been deleted. Skip check to avoid creating orphan object.
1996                  */
1997                 return 0;
1998         }
1999
2000         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2001                                 NOPOOLPERM))
2002                 return 0;
2003
2004         spin_lock(&ci->i_ceph_lock);
2005         flags = ci->i_ceph_flags;
2006         pool = ci->i_layout.pool_id;
2007         spin_unlock(&ci->i_ceph_lock);
2008 check:
2009         if (flags & CEPH_I_POOL_PERM) {
2010                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2011                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2012                              pool);
2013                         return -EPERM;
2014                 }
2015                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2016                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2017                              pool);
2018                         return -EPERM;
2019                 }
2020                 return 0;
2021         }
2022
2023         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2024         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2025         ceph_put_string(pool_ns);
2026         if (ret < 0)
2027                 return ret;
2028
2029         flags = CEPH_I_POOL_PERM;
2030         if (ret & POOL_READ)
2031                 flags |= CEPH_I_POOL_RD;
2032         if (ret & POOL_WRITE)
2033                 flags |= CEPH_I_POOL_WR;
2034
2035         spin_lock(&ci->i_ceph_lock);
2036         if (pool == ci->i_layout.pool_id &&
2037             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2038                 ci->i_ceph_flags |= flags;
2039         } else {
2040                 pool = ci->i_layout.pool_id;
2041                 flags = ci->i_ceph_flags;
2042         }
2043         spin_unlock(&ci->i_ceph_lock);
2044         goto check;
2045 }
2046
2047 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2048 {
2049         struct ceph_pool_perm *perm;
2050         struct rb_node *n;
2051
2052         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2053                 n = rb_first(&mdsc->pool_perm_tree);
2054                 perm = rb_entry(n, struct ceph_pool_perm, node);
2055                 rb_erase(n, &mdsc->pool_perm_tree);
2056                 kfree(perm);
2057         }
2058 }