fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75
  76         if (PageDirty(page)) {
  77                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  78                      mapping->host, page, page->index);
  79                 BUG_ON(!PagePrivate(page));
  80                 return 0;
  81         }
  82
  83         inode = mapping->host;
  84         ci = ceph_inode(inode);
  85
  86         /* dirty the head */
  87         spin_lock(&ci->i_ceph_lock);
  88         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  89         if (__ceph_have_pending_cap_snap(ci)) {
  90                 struct ceph_cap_snap *capsnap =
  91                                 list_last_entry(&ci->i_cap_snaps,
  92                                                 struct ceph_cap_snap,
  93                                                 ci_item);
  94                 snapc = ceph_get_snap_context(capsnap->context);
  95                 capsnap->dirty_pages++;
  96         } else {
  97                 BUG_ON(!ci->i_head_snapc);
  98                 snapc = ceph_get_snap_context(ci->i_head_snapc);
  99                 ++ci->i_wrbuffer_ref_head;
 100         }
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         return __set_page_dirty_nobuffers(page);
 121 }
 122
 123 /*
 124  * If we are truncating the full page (i.e. offset == 0), adjust the
 125  * dirty page counters appropriately.  Only called if there is private
 126  * data on the page.
 127  */
 128 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 129                                 unsigned int length)
 130 {
 131         struct inode *inode;
 132         struct ceph_inode_info *ci;
 133         struct ceph_snap_context *snapc = page_snap_context(page);
 134
 135         inode = page->mapping->host;
 136         ci = ceph_inode(inode);
 137
 138         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 139                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 140                      inode, page, page->index, offset, length);
 141                 return;
 142         }
 143
 144         ceph_invalidate_fscache_page(inode, page);
 145
 146         if (!PagePrivate(page))
 147                 return;
 148
 149         /*
 150          * We can get non-dirty pages here due to races between
 151          * set_page_dirty and truncate_complete_page; just spit out a
 152          * warning, in case we end up with accounting problems later.
 153          */
 154         if (!PageDirty(page))
 155                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 156
 157         ClearPageChecked(page);
 158
 159         dout("%p invalidatepage %p idx %lu full dirty page\n",
 160              inode, page, page->index);
 161
 162         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 163         ceph_put_snap_context(snapc);
 164         page->private = 0;
 165         ClearPagePrivate(page);
 166 }
 167
 168 static int ceph_releasepage(struct page *page, gfp_t g)
 169 {
 170         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 171         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 172         WARN_ON(PageDirty(page));
 173
 174         /* Can we release the page from the cache? */
 175         if (!ceph_release_fscache_page(page, g))
 176                 return 0;
 177
 178         return !PagePrivate(page);
 179 }
 180
 181 /*
 182  * read a single page, without unlocking it.
 183  */
 184 static int ceph_do_readpage(struct file *filp, struct page *page)
 185 {
 186         struct inode *inode = file_inode(filp);
 187         struct ceph_inode_info *ci = ceph_inode(inode);
 188         struct ceph_osd_client *osdc =
 189                 &ceph_inode_to_client(inode)->client->osdc;
 190         int err = 0;
 191         u64 off = page_offset(page);
 192         u64 len = PAGE_CACHE_SIZE;
 193
 194         if (off >= i_size_read(inode)) {
 195                 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 196                 SetPageUptodate(page);
 197                 return 0;
 198         }
 199
 200         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 201                 /*
 202                  * Uptodate inline data should have been added
 203                  * into page cache while getting Fcr caps.
 204                  */
 205                 if (off == 0)
 206                         return -EINVAL;
 207                 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 208                 SetPageUptodate(page);
 209                 return 0;
 210         }
 211
 212         err = ceph_readpage_from_fscache(inode, page);
 213         if (err == 0)
 214                 return -EINPROGRESS;
 215
 216         dout("readpage inode %p file %p page %p index %lu\n",
 217              inode, filp, page, page->index);
 218         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 219                                   off, &len,
 220                                   ci->i_truncate_seq, ci->i_truncate_size,
 221                                   &page, 1, 0);
 222         if (err == -ENOENT)
 223                 err = 0;
 224         if (err < 0) {
 225                 SetPageError(page);
 226                 ceph_fscache_readpage_cancel(inode, page);
 227                 goto out;
 228         }
 229         if (err < PAGE_CACHE_SIZE)
 230                 /* zero fill remainder of page */
 231                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 232         else
 233                 flush_dcache_page(page);
 234
 235         SetPageUptodate(page);
 236         ceph_readpage_to_fscache(inode, page);
 237
 238 out:
 239         return err < 0 ? err : 0;
 240 }
 241
 242 static int ceph_readpage(struct file *filp, struct page *page)
 243 {
 244         int r = ceph_do_readpage(filp, page);
 245         if (r != -EINPROGRESS)
 246                 unlock_page(page);
 247         else
 248                 r = 0;
 249         return r;
 250 }
 251
 252 /*
 253  * Finish an async read(ahead) op.
 254  */
 255 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 256 {
 257         struct inode *inode = req->r_inode;
 258         struct ceph_osd_data *osd_data;
 259         int rc = req->r_result;
 260         int bytes = le32_to_cpu(msg->hdr.data_len);
 261         int num_pages;
 262         int i;
 263
 264         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 265
 266         /* unlock all pages, zeroing any data we didn't read */
 267         osd_data = osd_req_op_extent_osd_data(req, 0);
 268         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 269         num_pages = calc_pages_for((u64)osd_data->alignment,
 270                                         (u64)osd_data->length);
 271         for (i = 0; i < num_pages; i++) {
 272                 struct page *page = osd_data->pages[i];
 273
 274                 if (rc < 0 && rc != ENOENT)
 275                         goto unlock;
 276                 if (bytes < (int)PAGE_CACHE_SIZE) {
 277                         /* zero (remainder of) page */
 278                         int s = bytes < 0 ? 0 : bytes;
 279                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 280                 }
 281                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 282                      page->index);
 283                 flush_dcache_page(page);
 284                 SetPageUptodate(page);
 285                 ceph_readpage_to_fscache(inode, page);
 286 unlock:
 287                 unlock_page(page);
 288                 page_cache_release(page);
 289                 bytes -= PAGE_CACHE_SIZE;
 290         }
 291         kfree(osd_data->pages);
 292 }
 293
 294 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 295 {
 296         int i;
 297
 298         for (i = 0; i < num_pages; i++)
 299                 unlock_page(pages[i]);
 300 }
 301
 302 /*
 303  * start an async read(ahead) operation.  return nr_pages we submitted
 304  * a read for on success, or negative error code.
 305  */
 306 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 307 {
 308         struct ceph_osd_client *osdc =
 309                 &ceph_inode_to_client(inode)->client->osdc;
 310         struct ceph_inode_info *ci = ceph_inode(inode);
 311         struct page *page = list_entry(page_list->prev, struct page, lru);
 312         struct ceph_vino vino;
 313         struct ceph_osd_request *req;
 314         u64 off;
 315         u64 len;
 316         int i;
 317         struct page **pages;
 318         pgoff_t next_index;
 319         int nr_pages = 0;
 320         int ret;
 321
 322         off = (u64) page_offset(page);
 323
 324         /* count pages */
 325         next_index = page->index;
 326         list_for_each_entry_reverse(page, page_list, lru) {
 327                 if (page->index != next_index)
 328                         break;
 329                 nr_pages++;
 330                 next_index++;
 331                 if (max && nr_pages == max)
 332                         break;
 333         }
 334         len = nr_pages << PAGE_CACHE_SHIFT;
 335         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 336              off, len);
 337         vino = ceph_vino(inode);
 338         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 339                                     0, 1, CEPH_OSD_OP_READ,
 340                                     CEPH_OSD_FLAG_READ, NULL,
 341                                     ci->i_truncate_seq, ci->i_truncate_size,
 342                                     false);
 343         if (IS_ERR(req))
 344                 return PTR_ERR(req);
 345
 346         /* build page vector */
 347         nr_pages = calc_pages_for(0, len);
 348         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 349         ret = -ENOMEM;
 350         if (!pages)
 351                 goto out;
 352         for (i = 0; i < nr_pages; ++i) {
 353                 page = list_entry(page_list->prev, struct page, lru);
 354                 BUG_ON(PageLocked(page));
 355                 list_del(&page->lru);
 356
 357                 dout("start_read %p adding %p idx %lu\n", inode, page,
 358                      page->index);
 359                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 360                                           GFP_KERNEL)) {
 361                         ceph_fscache_uncache_page(inode, page);
 362                         page_cache_release(page);
 363                         dout("start_read %p add_to_page_cache failed %p\n",
 364                              inode, page);
 365                         nr_pages = i;
 366                         goto out_pages;
 367                 }
 368                 pages[i] = page;
 369         }
 370         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 371         req->r_callback = finish_read;
 372         req->r_inode = inode;
 373
 374         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 375
 376         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 377         ret = ceph_osdc_start_request(osdc, req, false);
 378         if (ret < 0)
 379                 goto out_pages;
 380         ceph_osdc_put_request(req);
 381         return nr_pages;
 382
 383 out_pages:
 384         ceph_unlock_page_vector(pages, nr_pages);
 385         ceph_release_page_vector(pages, nr_pages);
 386 out:
 387         ceph_osdc_put_request(req);
 388         return ret;
 389 }
 390
 391
 392 /*
 393  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 394  * the caller (VM) cleans them up.
 395  */
 396 static int ceph_readpages(struct file *file, struct address_space *mapping,
 397                           struct list_head *page_list, unsigned nr_pages)
 398 {
 399         struct inode *inode = file_inode(file);
 400         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 401         int rc = 0;
 402         int max = 0;
 403
 404         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 405                 return -EINVAL;
 406
 407         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 408                                          &nr_pages);
 409
 410         if (rc == 0)
 411                 goto out;
 412
 413         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 414                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 415                         >> PAGE_SHIFT;
 416
 417         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 418                 file, nr_pages,
 419              max);
 420         while (!list_empty(page_list)) {
 421                 rc = start_read(inode, page_list, max);
 422                 if (rc < 0)
 423                         goto out;
 424                 BUG_ON(rc == 0);
 425         }
 426 out:
 427         ceph_fscache_readpages_cancel(inode, page_list);
 428
 429         dout("readpages %p file %p ret %d\n", inode, file, rc);
 430         return rc;
 431 }
 432
 433 /*
 434  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 435  * only snap context we are allowed to write back.
 436  */
 437 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 438                                                     loff_t *snap_size)
 439 {
 440         struct ceph_inode_info *ci = ceph_inode(inode);
 441         struct ceph_snap_context *snapc = NULL;
 442         struct ceph_cap_snap *capsnap = NULL;
 443
 444         spin_lock(&ci->i_ceph_lock);
 445         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 446                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 447                      capsnap->context, capsnap->dirty_pages);
 448                 if (capsnap->dirty_pages) {
 449                         snapc = ceph_get_snap_context(capsnap->context);
 450                         if (snap_size)
 451                                 *snap_size = capsnap->size;
 452                         break;
 453                 }
 454         }
 455         if (!snapc && ci->i_wrbuffer_ref_head) {
 456                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 457                 dout(" head snapc %p has %d dirty pages\n",
 458                      snapc, ci->i_wrbuffer_ref_head);
 459         }
 460         spin_unlock(&ci->i_ceph_lock);
 461         return snapc;
 462 }
 463
 464 /*
 465  * Write a single page, but leave the page locked.
 466  *
 467  * If we get a write error, set the page error bit, but still adjust the
 468  * dirty page accounting (i.e., page is no longer dirty).
 469  */
 470 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 471 {
 472         struct inode *inode;
 473         struct ceph_inode_info *ci;
 474         struct ceph_fs_client *fsc;
 475         struct ceph_osd_client *osdc;
 476         struct ceph_snap_context *snapc, *oldest;
 477         loff_t page_off = page_offset(page);
 478         loff_t snap_size = -1;
 479         long writeback_stat;
 480         u64 truncate_size;
 481         u32 truncate_seq;
 482         int err = 0, len = PAGE_CACHE_SIZE;
 483
 484         dout("writepage %p idx %lu\n", page, page->index);
 485
 486         if (!page->mapping || !page->mapping->host) {
 487                 dout("writepage %p - no mapping\n", page);
 488                 return -EFAULT;
 489         }
 490         inode = page->mapping->host;
 491         ci = ceph_inode(inode);
 492         fsc = ceph_inode_to_client(inode);
 493         osdc = &fsc->client->osdc;
 494
 495         /* verify this is a writeable snap context */
 496         snapc = page_snap_context(page);
 497         if (snapc == NULL) {
 498                 dout("writepage %p page %p not dirty?\n", inode, page);
 499                 goto out;
 500         }
 501         oldest = get_oldest_context(inode, &snap_size);
 502         if (snapc->seq > oldest->seq) {
 503                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 504                      inode, page, snapc);
 505                 /* we should only noop if called by kswapd */
 506                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 507                 ceph_put_snap_context(oldest);
 508                 goto out;
 509         }
 510         ceph_put_snap_context(oldest);
 511
 512         spin_lock(&ci->i_ceph_lock);
 513         truncate_seq = ci->i_truncate_seq;
 514         truncate_size = ci->i_truncate_size;
 515         if (snap_size == -1)
 516                 snap_size = i_size_read(inode);
 517         spin_unlock(&ci->i_ceph_lock);
 518
 519         /* is this a partial page at end of file? */
 520         if (page_off >= snap_size) {
 521                 dout("%p page eof %llu\n", page, snap_size);
 522                 goto out;
 523         }
 524         if (snap_size < page_off + len)
 525                 len = snap_size - page_off;
 526
 527         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 528              inode, page, page->index, page_off, len, snapc);
 529
 530         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 531         if (writeback_stat >
 532             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 533                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 534
 535         ceph_readpage_to_fscache(inode, page);
 536
 537         set_page_writeback(page);
 538         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 539                                    &ci->i_layout, snapc,
 540                                    page_off, len,
 541                                    truncate_seq, truncate_size,
 542                                    &inode->i_mtime, &page, 1);
 543         if (err < 0) {
 544                 dout("writepage setting page/mapping error %d %p\n", err, page);
 545                 SetPageError(page);
 546                 mapping_set_error(&inode->i_data, err);
 547                 if (wbc)
 548                         wbc->pages_skipped++;
 549         } else {
 550                 dout("writepage cleaned page %p\n", page);
 551                 err = 0;  /* vfs expects us to return 0 */
 552         }
 553         page->private = 0;
 554         ClearPagePrivate(page);
 555         end_page_writeback(page);
 556         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 557         ceph_put_snap_context(snapc);  /* page's reference */
 558 out:
 559         return err;
 560 }
 561
 562 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 563 {
 564         int err;
 565         struct inode *inode = page->mapping->host;
 566         BUG_ON(!inode);
 567         ihold(inode);
 568         err = writepage_nounlock(page, wbc);
 569         unlock_page(page);
 570         iput(inode);
 571         return err;
 572 }
 573
 574
 575 /*
 576  * lame release_pages helper.  release_pages() isn't exported to
 577  * modules.
 578  */
 579 static void ceph_release_pages(struct page **pages, int num)
 580 {
 581         struct pagevec pvec;
 582         int i;
 583
 584         pagevec_init(&pvec, 0);
 585         for (i = 0; i < num; i++) {
 586                 if (pagevec_add(&pvec, pages[i]) == 0)
 587                         pagevec_release(&pvec);
 588         }
 589         pagevec_release(&pvec);
 590 }
 591
 592 /*
 593  * async writeback completion handler.
 594  *
 595  * If we get an error, set the mapping error bit, but not the individual
 596  * page error bits.
 597  */
 598 static void writepages_finish(struct ceph_osd_request *req,
 599                               struct ceph_msg *msg)
 600 {
 601         struct inode *inode = req->r_inode;
 602         struct ceph_inode_info *ci = ceph_inode(inode);
 603         struct ceph_osd_data *osd_data;
 604         unsigned wrote;
 605         struct page *page;
 606         int num_pages;
 607         int i;
 608         struct ceph_snap_context *snapc = req->r_snapc;
 609         struct address_space *mapping = inode->i_mapping;
 610         int rc = req->r_result;
 611         u64 bytes = req->r_ops[0].extent.length;
 612         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 613         long writeback_stat;
 614         unsigned issued = ceph_caps_issued(ci);
 615
 616         osd_data = osd_req_op_extent_osd_data(req, 0);
 617         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 618         num_pages = calc_pages_for((u64)osd_data->alignment,
 619                                         (u64)osd_data->length);
 620         if (rc >= 0) {
 621                 /*
 622                  * Assume we wrote the pages we originally sent.  The
 623                  * osd might reply with fewer pages if our writeback
 624                  * raced with a truncation and was adjusted at the osd,
 625                  * so don't believe the reply.
 626                  */
 627                 wrote = num_pages;
 628         } else {
 629                 wrote = 0;
 630                 mapping_set_error(mapping, rc);
 631         }
 632         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 633              inode, rc, bytes, wrote);
 634
 635         /* clean all pages */
 636         for (i = 0; i < num_pages; i++) {
 637                 page = osd_data->pages[i];
 638                 BUG_ON(!page);
 639                 WARN_ON(!PageUptodate(page));
 640
 641                 writeback_stat =
 642                         atomic_long_dec_return(&fsc->writeback_count);
 643                 if (writeback_stat <
 644                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 645                         clear_bdi_congested(&fsc->backing_dev_info,
 646                                             BLK_RW_ASYNC);
 647
 648                 ceph_put_snap_context(page_snap_context(page));
 649                 page->private = 0;
 650                 ClearPagePrivate(page);
 651                 dout("unlocking %d %p\n", i, page);
 652                 end_page_writeback(page);
 653
 654                 /*
 655                  * We lost the cache cap, need to truncate the page before
 656                  * it is unlocked, otherwise we'd truncate it later in the
 657                  * page truncation thread, possibly losing some data that
 658                  * raced its way in
 659                  */
 660                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 661                         generic_error_remove_page(inode->i_mapping, page);
 662
 663                 unlock_page(page);
 664         }
 665         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 666         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 667
 668         ceph_release_pages(osd_data->pages, num_pages);
 669         if (osd_data->pages_from_pool)
 670                 mempool_free(osd_data->pages,
 671                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 672         else
 673                 kfree(osd_data->pages);
 674         ceph_osdc_put_request(req);
 675 }
 676
 677 /*
 678  * initiate async writeback
 679  */
 680 static int ceph_writepages_start(struct address_space *mapping,
 681                                  struct writeback_control *wbc)
 682 {
 683         struct inode *inode = mapping->host;
 684         struct ceph_inode_info *ci = ceph_inode(inode);
 685         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 686         struct ceph_vino vino = ceph_vino(inode);
 687         pgoff_t index, start, end;
 688         int range_whole = 0;
 689         int should_loop = 1;
 690         pgoff_t max_pages = 0, max_pages_ever = 0;
 691         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 692         struct pagevec pvec;
 693         int done = 0;
 694         int rc = 0;
 695         unsigned int wsize = i_blocksize(inode);
 696         struct ceph_osd_request *req = NULL;
 697         int do_sync = 0;
 698         loff_t snap_size, i_size;
 699         u64 truncate_size;
 700         u32 truncate_seq;
 701
 702         /*
 703          * Include a 'sync' in the OSD request if this is a data
 704          * integrity write (e.g., O_SYNC write or fsync()), or if our
 705          * cap is being revoked.
 706          */
 707         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 708                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 709                 do_sync = 1;
 710         dout("writepages_start %p dosync=%d (mode=%s)\n",
 711              inode, do_sync,
 712              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 713              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 714
 715         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 716                 pr_warn("writepage_start %p on forced umount\n", inode);
 717                 truncate_pagecache(inode, 0);
 718                 mapping_set_error(mapping, -EIO);
 719                 return -EIO; /* we're in a forced umount, don't write! */
 720         }
 721         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 722                 wsize = fsc->mount_options->wsize;
 723         if (wsize < PAGE_CACHE_SIZE)
 724                 wsize = PAGE_CACHE_SIZE;
 725         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 726
 727         pagevec_init(&pvec, 0);
 728
 729         /* where to start/end? */
 730         if (wbc->range_cyclic) {
 731                 start = mapping->writeback_index; /* Start from prev offset */
 732                 end = -1;
 733                 dout(" cyclic, start at %lu\n", start);
 734         } else {
 735                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 736                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 737                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 738                         range_whole = 1;
 739                 should_loop = 0;
 740                 dout(" not cyclic, %lu to %lu\n", start, end);
 741         }
 742         index = start;
 743
 744 retry:
 745         /* find oldest snap context with dirty data */
 746         ceph_put_snap_context(snapc);
 747         snap_size = -1;
 748         snapc = get_oldest_context(inode, &snap_size);
 749         if (!snapc) {
 750                 /* hmm, why does writepages get called when there
 751                    is no dirty data? */
 752                 dout(" no snap context with dirty data?\n");
 753                 goto out;
 754         }
 755         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 756              snapc, snapc->seq, snapc->num_snaps);
 757
 758         spin_lock(&ci->i_ceph_lock);
 759         truncate_seq = ci->i_truncate_seq;
 760         truncate_size = ci->i_truncate_size;
 761         i_size = i_size_read(inode);
 762         spin_unlock(&ci->i_ceph_lock);
 763
 764         if (last_snapc && snapc != last_snapc) {
 765                 /* if we switched to a newer snapc, restart our scan at the
 766                  * start of the original file range. */
 767                 dout("  snapc differs from last pass, restarting at %lu\n",
 768                      index);
 769                 index = start;
 770         }
 771         last_snapc = snapc;
 772
 773         while (!done && index <= end) {
 774                 unsigned i;
 775                 int first;
 776                 pgoff_t next;
 777                 int pvec_pages, locked_pages;
 778                 struct page **pages = NULL;
 779                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 780                 struct page *page;
 781                 int want;
 782                 u64 offset, len;
 783                 long writeback_stat;
 784
 785                 next = 0;
 786                 locked_pages = 0;
 787                 max_pages = max_pages_ever;
 788
 789 get_more_pages:
 790                 first = -1;
 791                 want = min(end - index,
 792                            min((pgoff_t)PAGEVEC_SIZE,
 793                                max_pages - (pgoff_t)locked_pages) - 1)
 794                         + 1;
 795                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 796                                                 PAGECACHE_TAG_DIRTY,
 797                                                 want);
 798                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 799                 if (!pvec_pages && !locked_pages)
 800                         break;
 801                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 802                         page = pvec.pages[i];
 803                         dout("? %p idx %lu\n", page, page->index);
 804                         if (locked_pages == 0)
 805                                 lock_page(page);  /* first page */
 806                         else if (!trylock_page(page))
 807                                 break;
 808
 809                         /* only dirty pages, or our accounting breaks */
 810                         if (unlikely(!PageDirty(page)) ||
 811                             unlikely(page->mapping != mapping)) {
 812                                 dout("!dirty or !mapping %p\n", page);
 813                                 unlock_page(page);
 814                                 break;
 815                         }
 816                         if (!wbc->range_cyclic && page->index > end) {
 817                                 dout("end of range %p\n", page);
 818                                 done = 1;
 819                                 unlock_page(page);
 820                                 break;
 821                         }
 822                         if (next && (page->index != next)) {
 823                                 dout("not consecutive %p\n", page);
 824                                 unlock_page(page);
 825                                 break;
 826                         }
 827                         if (wbc->sync_mode != WB_SYNC_NONE) {
 828                                 dout("waiting on writeback %p\n", page);
 829                                 wait_on_page_writeback(page);
 830                         }
 831                         if (page_offset(page) >=
 832                             (snap_size == -1 ? i_size : snap_size)) {
 833                                 dout("%p page eof %llu\n", page,
 834                                      (snap_size == -1 ? i_size : snap_size));
 835                                 done = 1;
 836                                 unlock_page(page);
 837                                 break;
 838                         }
 839                         if (PageWriteback(page)) {
 840                                 dout("%p under writeback\n", page);
 841                                 unlock_page(page);
 842                                 break;
 843                         }
 844
 845                         /* only if matching snap context */
 846                         pgsnapc = page_snap_context(page);
 847                         if (pgsnapc->seq > snapc->seq) {
 848                                 dout("page snapc %p %lld > oldest %p %lld\n",
 849                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 850                                 unlock_page(page);
 851                                 if (!locked_pages)
 852                                         continue; /* keep looking for snap */
 853                                 break;
 854                         }
 855
 856                         if (!clear_page_dirty_for_io(page)) {
 857                                 dout("%p !clear_page_dirty_for_io\n", page);
 858                                 unlock_page(page);
 859                                 break;
 860                         }
 861
 862                         /*
 863                          * We have something to write.  If this is
 864                          * the first locked page this time through,
 865                          * allocate an osd request and a page array
 866                          * that it will use.
 867                          */
 868                         if (locked_pages == 0) {
 869                                 BUG_ON(pages);
 870                                 /* prepare async write request */
 871                                 offset = (u64)page_offset(page);
 872                                 len = wsize;
 873                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 874                                                         &ci->i_layout, vino,
 875                                                         offset, &len, 0,
 876                                                         do_sync ? 2 : 1,
 877                                                         CEPH_OSD_OP_WRITE,
 878                                                         CEPH_OSD_FLAG_WRITE |
 879                                                         CEPH_OSD_FLAG_ONDISK,
 880                                                         snapc, truncate_seq,
 881                                                         truncate_size, true);
 882                                 if (IS_ERR(req)) {
 883                                         rc = PTR_ERR(req);
 884                                         unlock_page(page);
 885                                         break;
 886                                 }
 887
 888                                 if (do_sync)
 889                                         osd_req_op_init(req, 1,
 890                                                         CEPH_OSD_OP_STARTSYNC, 0);
 891
 892                                 req->r_callback = writepages_finish;
 893                                 req->r_inode = inode;
 894
 895                                 max_pages = calc_pages_for(0, (u64)len);
 896                                 pages = kmalloc(max_pages * sizeof (*pages),
 897                                                 GFP_NOFS);
 898                                 if (!pages) {
 899                                         pool = fsc->wb_pagevec_pool;
 900                                         pages = mempool_alloc(pool, GFP_NOFS);
 901                                         BUG_ON(!pages);
 902                                 }
 903                         }
 904
 905                         /* note position of first page in pvec */
 906                         if (first < 0)
 907                                 first = i;
 908                         dout("%p will write page %p idx %lu\n",
 909                              inode, page, page->index);
 910
 911                         writeback_stat =
 912                                atomic_long_inc_return(&fsc->writeback_count);
 913                         if (writeback_stat > CONGESTION_ON_THRESH(
 914                                     fsc->mount_options->congestion_kb)) {
 915                                 set_bdi_congested(&fsc->backing_dev_info,
 916                                                   BLK_RW_ASYNC);
 917                         }
 918
 919                         set_page_writeback(page);
 920                         pages[locked_pages] = page;
 921                         locked_pages++;
 922                         next = page->index + 1;
 923                 }
 924
 925                 /* did we get anything? */
 926                 if (!locked_pages)
 927                         goto release_pvec_pages;
 928                 if (i) {
 929                         int j;
 930                         BUG_ON(!locked_pages || first < 0);
 931
 932                         if (pvec_pages && i == pvec_pages &&
 933                             locked_pages < max_pages) {
 934                                 dout("reached end pvec, trying for more\n");
 935                                 pagevec_reinit(&pvec);
 936                                 goto get_more_pages;
 937                         }
 938
 939                         /* shift unused pages over in the pvec...  we
 940                          * will need to release them below. */
 941                         for (j = i; j < pvec_pages; j++) {
 942                                 dout(" pvec leftover page %p\n",
 943                                      pvec.pages[j]);
 944                                 pvec.pages[j-i+first] = pvec.pages[j];
 945                         }
 946                         pvec.nr -= i-first;
 947                 }
 948
 949                 /* Format the osd request message and submit the write */
 950                 offset = page_offset(pages[0]);
 951                 len = (u64)locked_pages << PAGE_CACHE_SHIFT;
 952                 if (snap_size == -1) {
 953                         len = min(len, (u64)i_size_read(inode) - offset);
 954                          /* writepages_finish() clears writeback pages
 955                           * according to the data length, so make sure
 956                           * data length covers all locked pages */
 957                         len = max(len, 1 +
 958                                 ((u64)(locked_pages - 1) << PAGE_CACHE_SHIFT));
 959                 } else {
 960                         len = min(len, snap_size - offset);
 961                 }
 962                 dout("writepages got %d pages at %llu~%llu\n",
 963                      locked_pages, offset, len);
 964
 965                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 966                                                         !!pool, false);
 967
 968                 pages = NULL;   /* request message now owns the pages array */
 969                 pool = NULL;
 970
 971                 /* Update the write op length in case we changed it */
 972
 973                 osd_req_op_extent_update(req, 0, len);
 974
 975                 vino = ceph_vino(inode);
 976                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 977                                         &inode->i_mtime);
 978
 979                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 980                 BUG_ON(rc);
 981                 req = NULL;
 982
 983                 /* continue? */
 984                 index = next;
 985                 wbc->nr_to_write -= locked_pages;
 986                 if (wbc->nr_to_write <= 0)
 987                         done = 1;
 988
 989 release_pvec_pages:
 990                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 991                      pvec.nr ? pvec.pages[0] : NULL);
 992                 pagevec_release(&pvec);
 993
 994                 if (locked_pages && !done)
 995                         goto retry;
 996         }
 997
 998         if (should_loop && !done) {
 999                 /* more to do; loop back to beginning of file */
1000                 dout("writepages looping back to beginning of file\n");
1001                 should_loop = 0;
1002                 index = 0;
1003                 goto retry;
1004         }
1005
1006         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1007                 mapping->writeback_index = index;
1008
1009 out:
1010         if (req)
1011                 ceph_osdc_put_request(req);
1012         ceph_put_snap_context(snapc);
1013         dout("writepages done, rc = %d\n", rc);
1014         return rc;
1015 }
1016
1017
1018
1019 /*
1020  * See if a given @snapc is either writeable, or already written.
1021  */
1022 static int context_is_writeable_or_written(struct inode *inode,
1023                                            struct ceph_snap_context *snapc)
1024 {
1025         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1026         int ret = !oldest || snapc->seq <= oldest->seq;
1027
1028         ceph_put_snap_context(oldest);
1029         return ret;
1030 }
1031
1032 /*
1033  * We are only allowed to write into/dirty the page if the page is
1034  * clean, or already dirty within the same snap context.
1035  *
1036  * called with page locked.
1037  * return success with page locked,
1038  * or any failure (incl -EAGAIN) with page unlocked.
1039  */
1040 static int ceph_update_writeable_page(struct file *file,
1041                             loff_t pos, unsigned len,
1042                             struct page *page)
1043 {
1044         struct inode *inode = file_inode(file);
1045         struct ceph_inode_info *ci = ceph_inode(inode);
1046         loff_t page_off = pos & PAGE_CACHE_MASK;
1047         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1048         int end_in_page = pos_in_page + len;
1049         loff_t i_size;
1050         int r;
1051         struct ceph_snap_context *snapc, *oldest;
1052
1053 retry_locked:
1054         /* writepages currently holds page lock, but if we change that later, */
1055         wait_on_page_writeback(page);
1056
1057         snapc = page_snap_context(page);
1058         if (snapc && snapc != ci->i_head_snapc) {
1059                 /*
1060                  * this page is already dirty in another (older) snap
1061                  * context!  is it writeable now?
1062                  */
1063                 oldest = get_oldest_context(inode, NULL);
1064
1065                 if (snapc->seq > oldest->seq) {
1066                         ceph_put_snap_context(oldest);
1067                         dout(" page %p snapc %p not current or oldest\n",
1068                              page, snapc);
1069                         /*
1070                          * queue for writeback, and wait for snapc to
1071                          * be writeable or written
1072                          */
1073                         snapc = ceph_get_snap_context(snapc);
1074                         unlock_page(page);
1075                         ceph_queue_writeback(inode);
1076                         r = wait_event_interruptible(ci->i_cap_wq,
1077                                context_is_writeable_or_written(inode, snapc));
1078                         ceph_put_snap_context(snapc);
1079                         if (r == -ERESTARTSYS)
1080                                 return r;
1081                         return -EAGAIN;
1082                 }
1083                 ceph_put_snap_context(oldest);
1084
1085                 /* yay, writeable, do it now (without dropping page lock) */
1086                 dout(" page %p snapc %p not current, but oldest\n",
1087                      page, snapc);
1088                 if (!clear_page_dirty_for_io(page))
1089                         goto retry_locked;
1090                 r = writepage_nounlock(page, NULL);
1091                 if (r < 0)
1092                         goto fail_unlock;
1093                 goto retry_locked;
1094         }
1095
1096         if (PageUptodate(page)) {
1097                 dout(" page %p already uptodate\n", page);
1098                 return 0;
1099         }
1100
1101         /* full page? */
1102         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1103                 return 0;
1104
1105         /* past end of file? */
1106         i_size = inode->i_size;   /* caller holds i_mutex */
1107
1108         if (page_off >= i_size ||
1109             (pos_in_page == 0 && (pos+len) >= i_size &&
1110              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1111                 dout(" zeroing %p 0 - %d and %d - %d\n",
1112                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1113                 zero_user_segments(page,
1114                                    0, pos_in_page,
1115                                    end_in_page, PAGE_CACHE_SIZE);
1116                 return 0;
1117         }
1118
1119         /* we need to read it. */
1120         r = ceph_do_readpage(file, page);
1121         if (r < 0) {
1122                 if (r == -EINPROGRESS)
1123                         return -EAGAIN;
1124                 goto fail_unlock;
1125         }
1126         goto retry_locked;
1127 fail_unlock:
1128         unlock_page(page);
1129         return r;
1130 }
1131
1132 /*
1133  * We are only allowed to write into/dirty the page if the page is
1134  * clean, or already dirty within the same snap context.
1135  */
1136 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1137                             loff_t pos, unsigned len, unsigned flags,
1138                             struct page **pagep, void **fsdata)
1139 {
1140         struct inode *inode = file_inode(file);
1141         struct page *page;
1142         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1143         int r;
1144
1145         do {
1146                 /* get a page */
1147                 page = grab_cache_page_write_begin(mapping, index, 0);
1148                 if (!page)
1149                         return -ENOMEM;
1150                 *pagep = page;
1151
1152                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1153                      inode, page, (int)pos, (int)len);
1154
1155                 r = ceph_update_writeable_page(file, pos, len, page);
1156                 if (r < 0)
1157                         page_cache_release(page);
1158                 else
1159                         *pagep = page;
1160         } while (r == -EAGAIN);
1161
1162         return r;
1163 }
1164
1165 /*
1166  * we don't do anything in here that simple_write_end doesn't do
1167  * except adjust dirty page accounting
1168  */
1169 static int ceph_write_end(struct file *file, struct address_space *mapping,
1170                           loff_t pos, unsigned len, unsigned copied,
1171                           struct page *page, void *fsdata)
1172 {
1173         struct inode *inode = file_inode(file);
1174         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1175         int check_cap = 0;
1176
1177         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1178              inode, page, (int)pos, (int)copied, (int)len);
1179
1180         /* zero the stale part of the page if we did a short copy */
1181         if (copied < len)
1182                 zero_user_segment(page, from+copied, len);
1183
1184         /* did file size increase? */
1185         /* (no need for i_size_read(); we caller holds i_mutex */
1186         if (pos+copied > inode->i_size)
1187                 check_cap = ceph_inode_set_size(inode, pos+copied);
1188
1189         if (!PageUptodate(page))
1190                 SetPageUptodate(page);
1191
1192         set_page_dirty(page);
1193
1194         unlock_page(page);
1195         page_cache_release(page);
1196
1197         if (check_cap)
1198                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1199
1200         return copied;
1201 }
1202
1203 /*
1204  * we set .direct_IO to indicate direct io is supported, but since we
1205  * intercept O_DIRECT reads and writes early, this function should
1206  * never get called.
1207  */
1208 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter,
1209                               loff_t pos)
1210 {
1211         WARN_ON(1);
1212         return -EINVAL;
1213 }
1214
1215 const struct address_space_operations ceph_aops = {
1216         .readpage = ceph_readpage,
1217         .readpages = ceph_readpages,
1218         .writepage = ceph_writepage,
1219         .writepages = ceph_writepages_start,
1220         .write_begin = ceph_write_begin,
1221         .write_end = ceph_write_end,
1222         .set_page_dirty = ceph_set_page_dirty,
1223         .invalidatepage = ceph_invalidatepage,
1224         .releasepage = ceph_releasepage,
1225         .direct_IO = ceph_direct_io,
1226 };
1227
1228
1229 /*
1230  * vm ops
1231  */
1232 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1233 {
1234         struct inode *inode = file_inode(vma->vm_file);
1235         struct ceph_inode_info *ci = ceph_inode(inode);
1236         struct ceph_file_info *fi = vma->vm_file->private_data;
1237         struct page *pinned_page = NULL;
1238         loff_t off = (loff_t)vmf->pgoff << PAGE_CACHE_SHIFT;
1239         int want, got, ret;
1240
1241         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1242              inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1243         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1244                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1245         else
1246                 want = CEPH_CAP_FILE_CACHE;
1247         while (1) {
1248                 got = 0;
1249                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1250                                     -1, &got, &pinned_page);
1251                 if (ret == 0)
1252                         break;
1253                 if (ret != -ERESTARTSYS) {
1254                         WARN_ON(1);
1255                         return VM_FAULT_SIGBUS;
1256                 }
1257         }
1258         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1259              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1260
1261         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1262             ci->i_inline_version == CEPH_INLINE_NONE)
1263                 ret = filemap_fault(vma, vmf);
1264         else
1265                 ret = -EAGAIN;
1266
1267         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1268              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1269         if (pinned_page)
1270                 page_cache_release(pinned_page);
1271         ceph_put_cap_refs(ci, got);
1272
1273         if (ret != -EAGAIN)
1274                 return ret;
1275
1276         /* read inline data */
1277         if (off >= PAGE_CACHE_SIZE) {
1278                 /* does not support inline data > PAGE_SIZE */
1279                 ret = VM_FAULT_SIGBUS;
1280         } else {
1281                 int ret1;
1282                 struct address_space *mapping = inode->i_mapping;
1283                 struct page *page = find_or_create_page(mapping, 0,
1284                                                 mapping_gfp_constraint(mapping,
1285                                                 ~__GFP_FS));
1286                 if (!page) {
1287                         ret = VM_FAULT_OOM;
1288                         goto out;
1289                 }
1290                 ret1 = __ceph_do_getattr(inode, page,
1291                                          CEPH_STAT_CAP_INLINE_DATA, true);
1292                 if (ret1 < 0 || off >= i_size_read(inode)) {
1293                         unlock_page(page);
1294                         page_cache_release(page);
1295                         ret = VM_FAULT_SIGBUS;
1296                         goto out;
1297                 }
1298                 if (ret1 < PAGE_CACHE_SIZE)
1299                         zero_user_segment(page, ret1, PAGE_CACHE_SIZE);
1300                 else
1301                         flush_dcache_page(page);
1302                 SetPageUptodate(page);
1303                 vmf->page = page;
1304                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1305         }
1306 out:
1307         dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1308              inode, off, (size_t)PAGE_CACHE_SIZE, ret);
1309         return ret;
1310 }
1311
1312 /*
1313  * Reuse write_begin here for simplicity.
1314  */
1315 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1316 {
1317         struct inode *inode = file_inode(vma->vm_file);
1318         struct ceph_inode_info *ci = ceph_inode(inode);
1319         struct ceph_file_info *fi = vma->vm_file->private_data;
1320         struct ceph_cap_flush *prealloc_cf;
1321         struct page *page = vmf->page;
1322         loff_t off = page_offset(page);
1323         loff_t size = i_size_read(inode);
1324         size_t len;
1325         int want, got, ret;
1326
1327         prealloc_cf = ceph_alloc_cap_flush();
1328         if (!prealloc_cf)
1329                 return VM_FAULT_SIGBUS;
1330
1331         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1332                 struct page *locked_page = NULL;
1333                 if (off == 0) {
1334                         lock_page(page);
1335                         locked_page = page;
1336                 }
1337                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1338                 if (locked_page)
1339                         unlock_page(locked_page);
1340                 if (ret < 0) {
1341                         ret = VM_FAULT_SIGBUS;
1342                         goto out_free;
1343                 }
1344         }
1345
1346         if (off + PAGE_CACHE_SIZE <= size)
1347                 len = PAGE_CACHE_SIZE;
1348         else
1349                 len = size & ~PAGE_CACHE_MASK;
1350
1351         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1352              inode, ceph_vinop(inode), off, len, size);
1353         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1354                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1355         else
1356                 want = CEPH_CAP_FILE_BUFFER;
1357         while (1) {
1358                 got = 0;
1359                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1360                                     &got, NULL);
1361                 if (ret == 0)
1362                         break;
1363                 if (ret != -ERESTARTSYS) {
1364                         WARN_ON(1);
1365                         ret = VM_FAULT_SIGBUS;
1366                         goto out_free;
1367                 }
1368         }
1369         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1370              inode, off, len, ceph_cap_string(got));
1371
1372         /* Update time before taking page lock */
1373         file_update_time(vma->vm_file);
1374
1375         lock_page(page);
1376
1377         ret = VM_FAULT_NOPAGE;
1378         if ((off > size) ||
1379             (page->mapping != inode->i_mapping))
1380                 goto out;
1381
1382         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1383         if (ret == 0) {
1384                 /* success.  we'll keep the page locked. */
1385                 set_page_dirty(page);
1386                 ret = VM_FAULT_LOCKED;
1387         } else {
1388                 if (ret == -ENOMEM)
1389                         ret = VM_FAULT_OOM;
1390                 else
1391                         ret = VM_FAULT_SIGBUS;
1392         }
1393 out:
1394         if (ret != VM_FAULT_LOCKED)
1395                 unlock_page(page);
1396         if (ret == VM_FAULT_LOCKED ||
1397             ci->i_inline_version != CEPH_INLINE_NONE) {
1398                 int dirty;
1399                 spin_lock(&ci->i_ceph_lock);
1400                 ci->i_inline_version = CEPH_INLINE_NONE;
1401                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1402                                                &prealloc_cf);
1403                 spin_unlock(&ci->i_ceph_lock);
1404                 if (dirty)
1405                         __mark_inode_dirty(inode, dirty);
1406         }
1407
1408         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1409              inode, off, len, ceph_cap_string(got), ret);
1410         ceph_put_cap_refs(ci, got);
1411 out_free:
1412         ceph_free_cap_flush(prealloc_cf);
1413
1414         return ret;
1415 }
1416
1417 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1418                            char *data, size_t len)
1419 {
1420         struct address_space *mapping = inode->i_mapping;
1421         struct page *page;
1422
1423         if (locked_page) {
1424                 page = locked_page;
1425         } else {
1426                 if (i_size_read(inode) == 0)
1427                         return;
1428                 page = find_or_create_page(mapping, 0,
1429                                            mapping_gfp_constraint(mapping,
1430                                            ~__GFP_FS));
1431                 if (!page)
1432                         return;
1433                 if (PageUptodate(page)) {
1434                         unlock_page(page);
1435                         page_cache_release(page);
1436                         return;
1437                 }
1438         }
1439
1440         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1441              inode, ceph_vinop(inode), len, locked_page);
1442
1443         if (len > 0) {
1444                 void *kaddr = kmap_atomic(page);
1445                 memcpy(kaddr, data, len);
1446                 kunmap_atomic(kaddr);
1447         }
1448
1449         if (page != locked_page) {
1450                 if (len < PAGE_CACHE_SIZE)
1451                         zero_user_segment(page, len, PAGE_CACHE_SIZE);
1452                 else
1453                         flush_dcache_page(page);
1454
1455                 SetPageUptodate(page);
1456                 unlock_page(page);
1457                 page_cache_release(page);
1458         }
1459 }
1460
1461 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1462 {
1463         struct inode *inode = file_inode(filp);
1464         struct ceph_inode_info *ci = ceph_inode(inode);
1465         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1466         struct ceph_osd_request *req;
1467         struct page *page = NULL;
1468         u64 len, inline_version;
1469         int err = 0;
1470         bool from_pagecache = false;
1471
1472         spin_lock(&ci->i_ceph_lock);
1473         inline_version = ci->i_inline_version;
1474         spin_unlock(&ci->i_ceph_lock);
1475
1476         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1477              inode, ceph_vinop(inode), inline_version);
1478
1479         if (inline_version == 1 || /* initial version, no data */
1480             inline_version == CEPH_INLINE_NONE)
1481                 goto out;
1482
1483         if (locked_page) {
1484                 page = locked_page;
1485                 WARN_ON(!PageUptodate(page));
1486         } else if (ceph_caps_issued(ci) &
1487                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1488                 page = find_get_page(inode->i_mapping, 0);
1489                 if (page) {
1490                         if (PageUptodate(page)) {
1491                                 from_pagecache = true;
1492                                 lock_page(page);
1493                         } else {
1494                                 page_cache_release(page);
1495                                 page = NULL;
1496                         }
1497                 }
1498         }
1499
1500         if (page) {
1501                 len = i_size_read(inode);
1502                 if (len > PAGE_CACHE_SIZE)
1503                         len = PAGE_CACHE_SIZE;
1504         } else {
1505                 page = __page_cache_alloc(GFP_NOFS);
1506                 if (!page) {
1507                         err = -ENOMEM;
1508                         goto out;
1509                 }
1510                 err = __ceph_do_getattr(inode, page,
1511                                         CEPH_STAT_CAP_INLINE_DATA, true);
1512                 if (err < 0) {
1513                         /* no inline data */
1514                         if (err == -ENODATA)
1515                                 err = 0;
1516                         goto out;
1517                 }
1518                 len = err;
1519         }
1520
1521         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1522                                     ceph_vino(inode), 0, &len, 0, 1,
1523                                     CEPH_OSD_OP_CREATE,
1524                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1525                                     ceph_empty_snapc, 0, 0, false);
1526         if (IS_ERR(req)) {
1527                 err = PTR_ERR(req);
1528                 goto out;
1529         }
1530
1531         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1532         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1533         if (!err)
1534                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1535         ceph_osdc_put_request(req);
1536         if (err < 0)
1537                 goto out;
1538
1539         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1540                                     ceph_vino(inode), 0, &len, 1, 3,
1541                                     CEPH_OSD_OP_WRITE,
1542                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1543                                     ceph_empty_snapc,
1544                                     ci->i_truncate_seq, ci->i_truncate_size,
1545                                     false);
1546         if (IS_ERR(req)) {
1547                 err = PTR_ERR(req);
1548                 goto out;
1549         }
1550
1551         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1552
1553         {
1554                 __le64 xattr_buf = cpu_to_le64(inline_version);
1555                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1556                                             "inline_version", &xattr_buf,
1557                                             sizeof(xattr_buf),
1558                                             CEPH_OSD_CMPXATTR_OP_GT,
1559                                             CEPH_OSD_CMPXATTR_MODE_U64);
1560                 if (err)
1561                         goto out_put;
1562         }
1563
1564         {
1565                 char xattr_buf[32];
1566                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1567                                          "%llu", inline_version);
1568                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1569                                             "inline_version",
1570                                             xattr_buf, xattr_len, 0, 0);
1571                 if (err)
1572                         goto out_put;
1573         }
1574
1575         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1576         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1577         if (!err)
1578                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1579 out_put:
1580         ceph_osdc_put_request(req);
1581         if (err == -ECANCELED)
1582                 err = 0;
1583 out:
1584         if (page && page != locked_page) {
1585                 if (from_pagecache) {
1586                         unlock_page(page);
1587                         page_cache_release(page);
1588                 } else
1589                         __free_pages(page, 0);
1590         }
1591
1592         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1593              inode, ceph_vinop(inode), inline_version, err);
1594         return err;
1595 }
1596
1597 static const struct vm_operations_struct ceph_vmops = {
1598         .fault          = ceph_filemap_fault,
1599         .page_mkwrite   = ceph_page_mkwrite,
1600 };
1601
1602 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1603 {
1604         struct address_space *mapping = file->f_mapping;
1605
1606         if (!mapping->a_ops->readpage)
1607                 return -ENOEXEC;
1608         file_accessed(file);
1609         vma->vm_ops = &ceph_vmops;
1610         return 0;
1611 }
1612
1613 enum {
1614         POOL_READ       = 1,
1615         POOL_WRITE      = 2,
1616 };
1617
1618 static int __ceph_pool_perm_get(struct ceph_inode_info *ci, u32 pool)
1619 {
1620         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1621         struct ceph_mds_client *mdsc = fsc->mdsc;
1622         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1623         struct rb_node **p, *parent;
1624         struct ceph_pool_perm *perm;
1625         struct page **pages;
1626         int err = 0, err2 = 0, have = 0;
1627
1628         down_read(&mdsc->pool_perm_rwsem);
1629         p = &mdsc->pool_perm_tree.rb_node;
1630         while (*p) {
1631                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1632                 if (pool < perm->pool)
1633                         p = &(*p)->rb_left;
1634                 else if (pool > perm->pool)
1635                         p = &(*p)->rb_right;
1636                 else {
1637                         have = perm->perm;
1638                         break;
1639                 }
1640         }
1641         up_read(&mdsc->pool_perm_rwsem);
1642         if (*p)
1643                 goto out;
1644
1645         dout("__ceph_pool_perm_get pool %u no perm cached\n", pool);
1646
1647         down_write(&mdsc->pool_perm_rwsem);
1648         parent = NULL;
1649         while (*p) {
1650                 parent = *p;
1651                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1652                 if (pool < perm->pool)
1653                         p = &(*p)->rb_left;
1654                 else if (pool > perm->pool)
1655                         p = &(*p)->rb_right;
1656                 else {
1657                         have = perm->perm;
1658                         break;
1659                 }
1660         }
1661         if (*p) {
1662                 up_write(&mdsc->pool_perm_rwsem);
1663                 goto out;
1664         }
1665
1666         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1667                                          ceph_empty_snapc,
1668                                          1, false, GFP_NOFS);
1669         if (!rd_req) {
1670                 err = -ENOMEM;
1671                 goto out_unlock;
1672         }
1673
1674         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1675         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1676         rd_req->r_base_oloc.pool = pool;
1677         snprintf(rd_req->r_base_oid.name, sizeof(rd_req->r_base_oid.name),
1678                  "%llx.00000000", ci->i_vino.ino);
1679         rd_req->r_base_oid.name_len = strlen(rd_req->r_base_oid.name);
1680
1681         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1682                                          ceph_empty_snapc,
1683                                          1, false, GFP_NOFS);
1684         if (!wr_req) {
1685                 err = -ENOMEM;
1686                 goto out_unlock;
1687         }
1688
1689         wr_req->r_flags = CEPH_OSD_FLAG_WRITE |
1690                           CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK;
1691         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1692         wr_req->r_base_oloc.pool = pool;
1693         wr_req->r_base_oid = rd_req->r_base_oid;
1694
1695         /* one page should be large enough for STAT data */
1696         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1697         if (IS_ERR(pages)) {
1698                 err = PTR_ERR(pages);
1699                 goto out_unlock;
1700         }
1701
1702         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1703                                      0, false, true);
1704         ceph_osdc_build_request(rd_req, 0, NULL, CEPH_NOSNAP,
1705                                 &ci->vfs_inode.i_mtime);
1706         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1707
1708         ceph_osdc_build_request(wr_req, 0, NULL, CEPH_NOSNAP,
1709                                 &ci->vfs_inode.i_mtime);
1710         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1711
1712         if (!err)
1713                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1714         if (!err2)
1715                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1716
1717         if (err >= 0 || err == -ENOENT)
1718                 have |= POOL_READ;
1719         else if (err != -EPERM)
1720                 goto out_unlock;
1721
1722         if (err2 == 0 || err2 == -EEXIST)
1723                 have |= POOL_WRITE;
1724         else if (err2 != -EPERM) {
1725                 err = err2;
1726                 goto out_unlock;
1727         }
1728
1729         perm = kmalloc(sizeof(*perm), GFP_NOFS);
1730         if (!perm) {
1731                 err = -ENOMEM;
1732                 goto out_unlock;
1733         }
1734
1735         perm->pool = pool;
1736         perm->perm = have;
1737         rb_link_node(&perm->node, parent, p);
1738         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1739         err = 0;
1740 out_unlock:
1741         up_write(&mdsc->pool_perm_rwsem);
1742
1743         if (rd_req)
1744                 ceph_osdc_put_request(rd_req);
1745         if (wr_req)
1746                 ceph_osdc_put_request(wr_req);
1747 out:
1748         if (!err)
1749                 err = have;
1750         dout("__ceph_pool_perm_get pool %u result = %d\n", pool, err);
1751         return err;
1752 }
1753
1754 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1755 {
1756         u32 pool;
1757         int ret, flags;
1758
1759         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1760                                 NOPOOLPERM))
1761                 return 0;
1762
1763         spin_lock(&ci->i_ceph_lock);
1764         flags = ci->i_ceph_flags;
1765         pool = ceph_file_layout_pg_pool(ci->i_layout);
1766         spin_unlock(&ci->i_ceph_lock);
1767 check:
1768         if (flags & CEPH_I_POOL_PERM) {
1769                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1770                         dout("ceph_pool_perm_check pool %u no read perm\n",
1771                              pool);
1772                         return -EPERM;
1773                 }
1774                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1775                         dout("ceph_pool_perm_check pool %u no write perm\n",
1776                              pool);
1777                         return -EPERM;
1778                 }
1779                 return 0;
1780         }
1781
1782         ret = __ceph_pool_perm_get(ci, pool);
1783         if (ret < 0)
1784                 return ret;
1785
1786         flags = CEPH_I_POOL_PERM;
1787         if (ret & POOL_READ)
1788                 flags |= CEPH_I_POOL_RD;
1789         if (ret & POOL_WRITE)
1790                 flags |= CEPH_I_POOL_WR;
1791
1792         spin_lock(&ci->i_ceph_lock);
1793         if (pool == ceph_file_layout_pg_pool(ci->i_layout)) {
1794                 ci->i_ceph_flags = flags;
1795         } else {
1796                 pool = ceph_file_layout_pg_pool(ci->i_layout);
1797                 flags = ci->i_ceph_flags;
1798         }
1799         spin_unlock(&ci->i_ceph_lock);
1800         goto check;
1801 }
1802
1803 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1804 {
1805         struct ceph_pool_perm *perm;
1806         struct rb_node *n;
1807
1808         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1809                 n = rb_first(&mdsc->pool_perm_tree);
1810                 perm = rb_entry(n, struct ceph_pool_perm, node);
1811                 rb_erase(n, &mdsc->pool_perm_tree);
1812                 kfree(perm);
1813         }
1814 }