fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /* dirty the head */
  91         spin_lock(&ci->i_ceph_lock);
  92         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  93         if (__ceph_have_pending_cap_snap(ci)) {
  94                 struct ceph_cap_snap *capsnap =
  95                                 list_last_entry(&ci->i_cap_snaps,
  96                                                 struct ceph_cap_snap,
  97                                                 ci_item);
  98                 snapc = ceph_get_snap_context(capsnap->context);
  99                 capsnap->dirty_pages++;
 100         } else {
 101                 BUG_ON(!ci->i_head_snapc);
 102                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 103                 ++ci->i_wrbuffer_ref_head;
 104         }
 105         if (ci->i_wrbuffer_ref == 0)
 106                 ihold(inode);
 107         ++ci->i_wrbuffer_ref;
 108         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 109              "snapc %p seq %lld (%d snaps)\n",
 110              mapping->host, page, page->index,
 111              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 112              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 113              snapc, snapc->seq, snapc->num_snaps);
 114         spin_unlock(&ci->i_ceph_lock);
 115
 116         /*
 117          * Reference snap context in page->private.  Also set
 118          * PagePrivate so that we get invalidatepage callback.
 119          */
 120         BUG_ON(PagePrivate(page));
 121         page->private = (unsigned long)snapc;
 122         SetPagePrivate(page);
 123
 124         ret = __set_page_dirty_nobuffers(page);
 125         WARN_ON(!PageLocked(page));
 126         WARN_ON(!page->mapping);
 127
 128         return ret;
 129 }
 130
 131 /*
 132  * If we are truncating the full page (i.e. offset == 0), adjust the
 133  * dirty page counters appropriately.  Only called if there is private
 134  * data on the page.
 135  */
 136 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 137                                 unsigned int length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc = page_snap_context(page);
 142
 143         inode = page->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != PAGE_SIZE) {
 147                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 148                      inode, page, page->index, offset, length);
 149                 return;
 150         }
 151
 152         ceph_invalidate_fscache_page(inode, page);
 153
 154         if (!PagePrivate(page))
 155                 return;
 156
 157         /*
 158          * We can get non-dirty pages here due to races between
 159          * set_page_dirty and truncate_complete_page; just spit out a
 160          * warning, in case we end up with accounting problems later.
 161          */
 162         if (!PageDirty(page))
 163                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 164
 165         ClearPageChecked(page);
 166
 167         dout("%p invalidatepage %p idx %lu full dirty page\n",
 168              inode, page, page->index);
 169
 170         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 171         ceph_put_snap_context(snapc);
 172         page->private = 0;
 173         ClearPagePrivate(page);
 174 }
 175
 176 static int ceph_releasepage(struct page *page, gfp_t g)
 177 {
 178         dout("%p releasepage %p idx %lu\n", page->mapping->host,
 179              page, page->index);
 180         WARN_ON(PageDirty(page));
 181
 182         /* Can we release the page from the cache? */
 183         if (!ceph_release_fscache_page(page, g))
 184                 return 0;
 185
 186         return !PagePrivate(page);
 187 }
 188
 189 /*
 190  * read a single page, without unlocking it.
 191  */
 192 static int readpage_nounlock(struct file *filp, struct page *page)
 193 {
 194         struct inode *inode = file_inode(filp);
 195         struct ceph_inode_info *ci = ceph_inode(inode);
 196         struct ceph_osd_client *osdc =
 197                 &ceph_inode_to_client(inode)->client->osdc;
 198         int err = 0;
 199         u64 off = page_offset(page);
 200         u64 len = PAGE_SIZE;
 201
 202         if (off >= i_size_read(inode)) {
 203                 zero_user_segment(page, 0, PAGE_SIZE);
 204                 SetPageUptodate(page);
 205                 return 0;
 206         }
 207
 208         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 209                 /*
 210                  * Uptodate inline data should have been added
 211                  * into page cache while getting Fcr caps.
 212                  */
 213                 if (off == 0)
 214                         return -EINVAL;
 215                 zero_user_segment(page, 0, PAGE_SIZE);
 216                 SetPageUptodate(page);
 217                 return 0;
 218         }
 219
 220         err = ceph_readpage_from_fscache(inode, page);
 221         if (err == 0)
 222                 goto out;
 223
 224         dout("readpage inode %p file %p page %p index %lu\n",
 225              inode, filp, page, page->index);
 226         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 227                                   off, &len,
 228                                   ci->i_truncate_seq, ci->i_truncate_size,
 229                                   &page, 1, 0);
 230         if (err == -ENOENT)
 231                 err = 0;
 232         if (err < 0) {
 233                 SetPageError(page);
 234                 ceph_fscache_readpage_cancel(inode, page);
 235                 goto out;
 236         }
 237         if (err < PAGE_SIZE)
 238                 /* zero fill remainder of page */
 239                 zero_user_segment(page, err, PAGE_SIZE);
 240         else
 241                 flush_dcache_page(page);
 242
 243         SetPageUptodate(page);
 244         ceph_readpage_to_fscache(inode, page);
 245
 246 out:
 247         return err < 0 ? err : 0;
 248 }
 249
 250 static int ceph_readpage(struct file *filp, struct page *page)
 251 {
 252         int r = readpage_nounlock(filp, page);
 253         unlock_page(page);
 254         return r;
 255 }
 256
 257 /*
 258  * Finish an async read(ahead) op.
 259  */
 260 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 261 {
 262         struct inode *inode = req->r_inode;
 263         struct ceph_osd_data *osd_data;
 264         int rc = req->r_result;
 265         int bytes = le32_to_cpu(msg->hdr.data_len);
 266         int num_pages;
 267         int i;
 268
 269         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 270
 271         /* unlock all pages, zeroing any data we didn't read */
 272         osd_data = osd_req_op_extent_osd_data(req, 0);
 273         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 274         num_pages = calc_pages_for((u64)osd_data->alignment,
 275                                         (u64)osd_data->length);
 276         for (i = 0; i < num_pages; i++) {
 277                 struct page *page = osd_data->pages[i];
 278
 279                 if (rc < 0 && rc != -ENOENT)
 280                         goto unlock;
 281                 if (bytes < (int)PAGE_SIZE) {
 282                         /* zero (remainder of) page */
 283                         int s = bytes < 0 ? 0 : bytes;
 284                         zero_user_segment(page, s, PAGE_SIZE);
 285                 }
 286                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 287                      page->index);
 288                 flush_dcache_page(page);
 289                 SetPageUptodate(page);
 290                 ceph_readpage_to_fscache(inode, page);
 291 unlock:
 292                 unlock_page(page);
 293                 put_page(page);
 294                 bytes -= PAGE_SIZE;
 295         }
 296         kfree(osd_data->pages);
 297 }
 298
 299 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 300 {
 301         int i;
 302
 303         for (i = 0; i < num_pages; i++)
 304                 unlock_page(pages[i]);
 305 }
 306
 307 /*
 308  * start an async read(ahead) operation.  return nr_pages we submitted
 309  * a read for on success, or negative error code.
 310  */
 311 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 312 {
 313         struct ceph_osd_client *osdc =
 314                 &ceph_inode_to_client(inode)->client->osdc;
 315         struct ceph_inode_info *ci = ceph_inode(inode);
 316         struct page *page = list_entry(page_list->prev, struct page, lru);
 317         struct ceph_vino vino;
 318         struct ceph_osd_request *req;
 319         u64 off;
 320         u64 len;
 321         int i;
 322         struct page **pages;
 323         pgoff_t next_index;
 324         int nr_pages = 0;
 325         int ret;
 326
 327         off = (u64) page_offset(page);
 328
 329         /* count pages */
 330         next_index = page->index;
 331         list_for_each_entry_reverse(page, page_list, lru) {
 332                 if (page->index != next_index)
 333                         break;
 334                 nr_pages++;
 335                 next_index++;
 336                 if (max && nr_pages == max)
 337                         break;
 338         }
 339         len = nr_pages << PAGE_SHIFT;
 340         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 341              off, len);
 342         vino = ceph_vino(inode);
 343         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 344                                     0, 1, CEPH_OSD_OP_READ,
 345                                     CEPH_OSD_FLAG_READ, NULL,
 346                                     ci->i_truncate_seq, ci->i_truncate_size,
 347                                     false);
 348         if (IS_ERR(req))
 349                 return PTR_ERR(req);
 350
 351         /* build page vector */
 352         nr_pages = calc_pages_for(0, len);
 353         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 354         ret = -ENOMEM;
 355         if (!pages)
 356                 goto out;
 357         for (i = 0; i < nr_pages; ++i) {
 358                 page = list_entry(page_list->prev, struct page, lru);
 359                 BUG_ON(PageLocked(page));
 360                 list_del(&page->lru);
 361
 362                 dout("start_read %p adding %p idx %lu\n", inode, page,
 363                      page->index);
 364                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 365                                           GFP_KERNEL)) {
 366                         ceph_fscache_uncache_page(inode, page);
 367                         put_page(page);
 368                         dout("start_read %p add_to_page_cache failed %p\n",
 369                              inode, page);
 370                         nr_pages = i;
 371                         goto out_pages;
 372                 }
 373                 pages[i] = page;
 374         }
 375         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 376         req->r_callback = finish_read;
 377         req->r_inode = inode;
 378
 379         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 380
 381         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 382         ret = ceph_osdc_start_request(osdc, req, false);
 383         if (ret < 0)
 384                 goto out_pages;
 385         ceph_osdc_put_request(req);
 386         return nr_pages;
 387
 388 out_pages:
 389         ceph_unlock_page_vector(pages, nr_pages);
 390         ceph_release_page_vector(pages, nr_pages);
 391 out:
 392         ceph_osdc_put_request(req);
 393         return ret;
 394 }
 395
 396
 397 /*
 398  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 399  * the caller (VM) cleans them up.
 400  */
 401 static int ceph_readpages(struct file *file, struct address_space *mapping,
 402                           struct list_head *page_list, unsigned nr_pages)
 403 {
 404         struct inode *inode = file_inode(file);
 405         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 406         int rc = 0;
 407         int max = 0;
 408
 409         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 410                 return -EINVAL;
 411
 412         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 413                                          &nr_pages);
 414
 415         if (rc == 0)
 416                 goto out;
 417
 418         if (fsc->mount_options->rsize >= PAGE_SIZE)
 419                 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
 420                         >> PAGE_SHIFT;
 421
 422         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 423                 file, nr_pages,
 424              max);
 425         while (!list_empty(page_list)) {
 426                 rc = start_read(inode, page_list, max);
 427                 if (rc < 0)
 428                         goto out;
 429                 BUG_ON(rc == 0);
 430         }
 431 out:
 432         ceph_fscache_readpages_cancel(inode, page_list);
 433
 434         dout("readpages %p file %p ret %d\n", inode, file, rc);
 435         return rc;
 436 }
 437
 438 /*
 439  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 440  * only snap context we are allowed to write back.
 441  */
 442 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 443                                                     loff_t *snap_size)
 444 {
 445         struct ceph_inode_info *ci = ceph_inode(inode);
 446         struct ceph_snap_context *snapc = NULL;
 447         struct ceph_cap_snap *capsnap = NULL;
 448
 449         spin_lock(&ci->i_ceph_lock);
 450         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 451                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 452                      capsnap->context, capsnap->dirty_pages);
 453                 if (capsnap->dirty_pages) {
 454                         snapc = ceph_get_snap_context(capsnap->context);
 455                         if (snap_size)
 456                                 *snap_size = capsnap->size;
 457                         break;
 458                 }
 459         }
 460         if (!snapc && ci->i_wrbuffer_ref_head) {
 461                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 462                 dout(" head snapc %p has %d dirty pages\n",
 463                      snapc, ci->i_wrbuffer_ref_head);
 464         }
 465         spin_unlock(&ci->i_ceph_lock);
 466         return snapc;
 467 }
 468
 469 /*
 470  * Write a single page, but leave the page locked.
 471  *
 472  * If we get a write error, set the page error bit, but still adjust the
 473  * dirty page accounting (i.e., page is no longer dirty).
 474  */
 475 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 476 {
 477         struct inode *inode;
 478         struct ceph_inode_info *ci;
 479         struct ceph_fs_client *fsc;
 480         struct ceph_osd_client *osdc;
 481         struct ceph_snap_context *snapc, *oldest;
 482         loff_t page_off = page_offset(page);
 483         loff_t snap_size = -1;
 484         long writeback_stat;
 485         u64 truncate_size;
 486         u32 truncate_seq;
 487         int err = 0, len = PAGE_SIZE;
 488
 489         dout("writepage %p idx %lu\n", page, page->index);
 490
 491         if (!page->mapping || !page->mapping->host) {
 492                 dout("writepage %p - no mapping\n", page);
 493                 return -EFAULT;
 494         }
 495         inode = page->mapping->host;
 496         ci = ceph_inode(inode);
 497         fsc = ceph_inode_to_client(inode);
 498         osdc = &fsc->client->osdc;
 499
 500         /* verify this is a writeable snap context */
 501         snapc = page_snap_context(page);
 502         if (snapc == NULL) {
 503                 dout("writepage %p page %p not dirty?\n", inode, page);
 504                 goto out;
 505         }
 506         oldest = get_oldest_context(inode, &snap_size);
 507         if (snapc->seq > oldest->seq) {
 508                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 509                      inode, page, snapc);
 510                 /* we should only noop if called by kswapd */
 511                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 512                 ceph_put_snap_context(oldest);
 513                 goto out;
 514         }
 515         ceph_put_snap_context(oldest);
 516
 517         spin_lock(&ci->i_ceph_lock);
 518         truncate_seq = ci->i_truncate_seq;
 519         truncate_size = ci->i_truncate_size;
 520         if (snap_size == -1)
 521                 snap_size = i_size_read(inode);
 522         spin_unlock(&ci->i_ceph_lock);
 523
 524         /* is this a partial page at end of file? */
 525         if (page_off >= snap_size) {
 526                 dout("%p page eof %llu\n", page, snap_size);
 527                 goto out;
 528         }
 529         if (snap_size < page_off + len)
 530                 len = snap_size - page_off;
 531
 532         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 533              inode, page, page->index, page_off, len, snapc);
 534
 535         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 536         if (writeback_stat >
 537             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 538                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 539
 540         ceph_readpage_to_fscache(inode, page);
 541
 542         set_page_writeback(page);
 543         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 544                                    &ci->i_layout, snapc,
 545                                    page_off, len,
 546                                    truncate_seq, truncate_size,
 547                                    &inode->i_mtime, &page, 1);
 548         if (err < 0) {
 549                 dout("writepage setting page/mapping error %d %p\n", err, page);
 550                 SetPageError(page);
 551                 mapping_set_error(&inode->i_data, err);
 552                 if (wbc)
 553                         wbc->pages_skipped++;
 554         } else {
 555                 dout("writepage cleaned page %p\n", page);
 556                 err = 0;  /* vfs expects us to return 0 */
 557         }
 558         page->private = 0;
 559         ClearPagePrivate(page);
 560         end_page_writeback(page);
 561         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 562         ceph_put_snap_context(snapc);  /* page's reference */
 563 out:
 564         return err;
 565 }
 566
 567 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 568 {
 569         int err;
 570         struct inode *inode = page->mapping->host;
 571         BUG_ON(!inode);
 572         ihold(inode);
 573         err = writepage_nounlock(page, wbc);
 574         unlock_page(page);
 575         iput(inode);
 576         return err;
 577 }
 578
 579
 580 /*
 581  * lame release_pages helper.  release_pages() isn't exported to
 582  * modules.
 583  */
 584 static void ceph_release_pages(struct page **pages, int num)
 585 {
 586         struct pagevec pvec;
 587         int i;
 588
 589         pagevec_init(&pvec, 0);
 590         for (i = 0; i < num; i++) {
 591                 if (pagevec_add(&pvec, pages[i]) == 0)
 592                         pagevec_release(&pvec);
 593         }
 594         pagevec_release(&pvec);
 595 }
 596
 597 /*
 598  * async writeback completion handler.
 599  *
 600  * If we get an error, set the mapping error bit, but not the individual
 601  * page error bits.
 602  */
 603 static void writepages_finish(struct ceph_osd_request *req,
 604                               struct ceph_msg *msg)
 605 {
 606         struct inode *inode = req->r_inode;
 607         struct ceph_inode_info *ci = ceph_inode(inode);
 608         struct ceph_osd_data *osd_data;
 609         struct page *page;
 610         int num_pages, total_pages = 0;
 611         int i, j;
 612         int rc = req->r_result;
 613         struct ceph_snap_context *snapc = req->r_snapc;
 614         struct address_space *mapping = inode->i_mapping;
 615         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 616         bool remove_page;
 617
 618
 619         dout("writepages_finish %p rc %d\n", inode, rc);
 620         if (rc < 0)
 621                 mapping_set_error(mapping, rc);
 622
 623         /*
 624          * We lost the cache cap, need to truncate the page before
 625          * it is unlocked, otherwise we'd truncate it later in the
 626          * page truncation thread, possibly losing some data that
 627          * raced its way in
 628          */
 629         remove_page = !(ceph_caps_issued(ci) &
 630                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 631
 632         /* clean all pages */
 633         for (i = 0; i < req->r_num_ops; i++) {
 634                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 635                         break;
 636
 637                 osd_data = osd_req_op_extent_osd_data(req, i);
 638                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 639                 num_pages = calc_pages_for((u64)osd_data->alignment,
 640                                            (u64)osd_data->length);
 641                 total_pages += num_pages;
 642                 for (j = 0; j < num_pages; j++) {
 643                         page = osd_data->pages[j];
 644                         BUG_ON(!page);
 645                         WARN_ON(!PageUptodate(page));
 646
 647                         if (atomic_long_dec_return(&fsc->writeback_count) <
 648                              CONGESTION_OFF_THRESH(
 649                                         fsc->mount_options->congestion_kb))
 650                                 clear_bdi_congested(&fsc->backing_dev_info,
 651                                                     BLK_RW_ASYNC);
 652
 653                         ceph_put_snap_context(page_snap_context(page));
 654                         page->private = 0;
 655                         ClearPagePrivate(page);
 656                         dout("unlocking %p\n", page);
 657                         end_page_writeback(page);
 658
 659                         if (remove_page)
 660                                 generic_error_remove_page(inode->i_mapping,
 661                                                           page);
 662
 663                         unlock_page(page);
 664                 }
 665                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 666                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 667
 668                 ceph_release_pages(osd_data->pages, num_pages);
 669         }
 670
 671         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 672
 673         osd_data = osd_req_op_extent_osd_data(req, 0);
 674         if (osd_data->pages_from_pool)
 675                 mempool_free(osd_data->pages,
 676                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 677         else
 678                 kfree(osd_data->pages);
 679         ceph_osdc_put_request(req);
 680 }
 681
 682 /*
 683  * initiate async writeback
 684  */
 685 static int ceph_writepages_start(struct address_space *mapping,
 686                                  struct writeback_control *wbc)
 687 {
 688         struct inode *inode = mapping->host;
 689         struct ceph_inode_info *ci = ceph_inode(inode);
 690         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 691         struct ceph_vino vino = ceph_vino(inode);
 692         pgoff_t index, start, end;
 693         int range_whole = 0;
 694         int should_loop = 1;
 695         pgoff_t max_pages = 0, max_pages_ever = 0;
 696         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 697         struct pagevec pvec;
 698         int done = 0;
 699         int rc = 0;
 700         unsigned wsize = 1 << inode->i_blkbits;
 701         struct ceph_osd_request *req = NULL;
 702         int do_sync = 0;
 703         loff_t snap_size, i_size;
 704         u64 truncate_size;
 705         u32 truncate_seq;
 706
 707         /*
 708          * Include a 'sync' in the OSD request if this is a data
 709          * integrity write (e.g., O_SYNC write or fsync()), or if our
 710          * cap is being revoked.
 711          */
 712         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 713                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 714                 do_sync = 1;
 715         dout("writepages_start %p dosync=%d (mode=%s)\n",
 716              inode, do_sync,
 717              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 718              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 719
 720         if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 721                 pr_warn("writepage_start %p on forced umount\n", inode);
 722                 truncate_pagecache(inode, 0);
 723                 mapping_set_error(mapping, -EIO);
 724                 return -EIO; /* we're in a forced umount, don't write! */
 725         }
 726         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 727                 wsize = fsc->mount_options->wsize;
 728         if (wsize < PAGE_SIZE)
 729                 wsize = PAGE_SIZE;
 730         max_pages_ever = wsize >> PAGE_SHIFT;
 731
 732         pagevec_init(&pvec, 0);
 733
 734         /* where to start/end? */
 735         if (wbc->range_cyclic) {
 736                 start = mapping->writeback_index; /* Start from prev offset */
 737                 end = -1;
 738                 dout(" cyclic, start at %lu\n", start);
 739         } else {
 740                 start = wbc->range_start >> PAGE_SHIFT;
 741                 end = wbc->range_end >> PAGE_SHIFT;
 742                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 743                         range_whole = 1;
 744                 should_loop = 0;
 745                 dout(" not cyclic, %lu to %lu\n", start, end);
 746         }
 747         index = start;
 748
 749 retry:
 750         /* find oldest snap context with dirty data */
 751         ceph_put_snap_context(snapc);
 752         snap_size = -1;
 753         snapc = get_oldest_context(inode, &snap_size);
 754         if (!snapc) {
 755                 /* hmm, why does writepages get called when there
 756                    is no dirty data? */
 757                 dout(" no snap context with dirty data?\n");
 758                 goto out;
 759         }
 760         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 761              snapc, snapc->seq, snapc->num_snaps);
 762
 763         spin_lock(&ci->i_ceph_lock);
 764         truncate_seq = ci->i_truncate_seq;
 765         truncate_size = ci->i_truncate_size;
 766         i_size = i_size_read(inode);
 767         spin_unlock(&ci->i_ceph_lock);
 768
 769         if (last_snapc && snapc != last_snapc) {
 770                 /* if we switched to a newer snapc, restart our scan at the
 771                  * start of the original file range. */
 772                 dout("  snapc differs from last pass, restarting at %lu\n",
 773                      index);
 774                 index = start;
 775         }
 776         last_snapc = snapc;
 777
 778         while (!done && index <= end) {
 779                 unsigned i;
 780                 int first;
 781                 pgoff_t strip_unit_end = 0;
 782                 int num_ops = 0, op_idx;
 783                 int pvec_pages, locked_pages = 0;
 784                 struct page **pages = NULL, **data_pages;
 785                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 786                 struct page *page;
 787                 int want;
 788                 u64 offset = 0, len = 0;
 789
 790                 max_pages = max_pages_ever;
 791
 792 get_more_pages:
 793                 first = -1;
 794                 want = min(end - index,
 795                            min((pgoff_t)PAGEVEC_SIZE,
 796                                max_pages - (pgoff_t)locked_pages) - 1)
 797                         + 1;
 798                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 799                                                 PAGECACHE_TAG_DIRTY,
 800                                                 want);
 801                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 802                 if (!pvec_pages && !locked_pages)
 803                         break;
 804                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 805                         page = pvec.pages[i];
 806                         dout("? %p idx %lu\n", page, page->index);
 807                         if (locked_pages == 0)
 808                                 lock_page(page);  /* first page */
 809                         else if (!trylock_page(page))
 810                                 break;
 811
 812                         /* only dirty pages, or our accounting breaks */
 813                         if (unlikely(!PageDirty(page)) ||
 814                             unlikely(page->mapping != mapping)) {
 815                                 dout("!dirty or !mapping %p\n", page);
 816                                 unlock_page(page);
 817                                 break;
 818                         }
 819                         if (!wbc->range_cyclic && page->index > end) {
 820                                 dout("end of range %p\n", page);
 821                                 done = 1;
 822                                 unlock_page(page);
 823                                 break;
 824                         }
 825                         if (strip_unit_end && (page->index > strip_unit_end)) {
 826                                 dout("end of strip unit %p\n", page);
 827                                 unlock_page(page);
 828                                 break;
 829                         }
 830                         if (wbc->sync_mode != WB_SYNC_NONE) {
 831                                 dout("waiting on writeback %p\n", page);
 832                                 wait_on_page_writeback(page);
 833                         }
 834                         if (page_offset(page) >=
 835                             (snap_size == -1 ? i_size : snap_size)) {
 836                                 dout("%p page eof %llu\n", page,
 837                                      (snap_size == -1 ? i_size : snap_size));
 838                                 done = 1;
 839                                 unlock_page(page);
 840                                 break;
 841                         }
 842                         if (PageWriteback(page)) {
 843                                 dout("%p under writeback\n", page);
 844                                 unlock_page(page);
 845                                 break;
 846                         }
 847
 848                         /* only if matching snap context */
 849                         pgsnapc = page_snap_context(page);
 850                         if (pgsnapc->seq > snapc->seq) {
 851                                 dout("page snapc %p %lld > oldest %p %lld\n",
 852                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 853                                 unlock_page(page);
 854                                 if (!locked_pages)
 855                                         continue; /* keep looking for snap */
 856                                 break;
 857                         }
 858
 859                         if (!clear_page_dirty_for_io(page)) {
 860                                 dout("%p !clear_page_dirty_for_io\n", page);
 861                                 unlock_page(page);
 862                                 break;
 863                         }
 864
 865                         /*
 866                          * We have something to write.  If this is
 867                          * the first locked page this time through,
 868                          * calculate max possinle write size and
 869                          * allocate a page array
 870                          */
 871                         if (locked_pages == 0) {
 872                                 u64 objnum;
 873                                 u64 objoff;
 874
 875                                 /* prepare async write request */
 876                                 offset = (u64)page_offset(page);
 877                                 len = wsize;
 878
 879                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
 880                                                                 offset, len,
 881                                                                 &objnum, &objoff,
 882                                                                 &len);
 883                                 if (rc < 0) {
 884                                         unlock_page(page);
 885                                         break;
 886                                 }
 887
 888                                 num_ops = 1 + do_sync;
 889                                 strip_unit_end = page->index +
 890                                         ((len - 1) >> PAGE_SHIFT);
 891
 892                                 BUG_ON(pages);
 893                                 max_pages = calc_pages_for(0, (u64)len);
 894                                 pages = kmalloc(max_pages * sizeof (*pages),
 895                                                 GFP_NOFS);
 896                                 if (!pages) {
 897                                         pool = fsc->wb_pagevec_pool;
 898                                         pages = mempool_alloc(pool, GFP_NOFS);
 899                                         BUG_ON(!pages);
 900                                 }
 901
 902                                 len = 0;
 903                         } else if (page->index !=
 904                                    (offset + len) >> PAGE_SHIFT) {
 905                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 906                                                         CEPH_OSD_MAX_OPS)) {
 907                                         redirty_page_for_writepage(wbc, page);
 908                                         unlock_page(page);
 909                                         break;
 910                                 }
 911
 912                                 num_ops++;
 913                                 offset = (u64)page_offset(page);
 914                                 len = 0;
 915                         }
 916
 917                         /* note position of first page in pvec */
 918                         if (first < 0)
 919                                 first = i;
 920                         dout("%p will write page %p idx %lu\n",
 921                              inode, page, page->index);
 922
 923                         if (atomic_long_inc_return(&fsc->writeback_count) >
 924                             CONGESTION_ON_THRESH(
 925                                     fsc->mount_options->congestion_kb)) {
 926                                 set_bdi_congested(&fsc->backing_dev_info,
 927                                                   BLK_RW_ASYNC);
 928                         }
 929
 930                         pages[locked_pages] = page;
 931                         locked_pages++;
 932                         len += PAGE_SIZE;
 933                 }
 934
 935                 /* did we get anything? */
 936                 if (!locked_pages)
 937                         goto release_pvec_pages;
 938                 if (i) {
 939                         int j;
 940                         BUG_ON(!locked_pages || first < 0);
 941
 942                         if (pvec_pages && i == pvec_pages &&
 943                             locked_pages < max_pages) {
 944                                 dout("reached end pvec, trying for more\n");
 945                                 pagevec_reinit(&pvec);
 946                                 goto get_more_pages;
 947                         }
 948
 949                         /* shift unused pages over in the pvec...  we
 950                          * will need to release them below. */
 951                         for (j = i; j < pvec_pages; j++) {
 952                                 dout(" pvec leftover page %p\n", pvec.pages[j]);
 953                                 pvec.pages[j-i+first] = pvec.pages[j];
 954                         }
 955                         pvec.nr -= i-first;
 956                 }
 957
 958 new_request:
 959                 offset = page_offset(pages[0]);
 960                 len = wsize;
 961
 962                 req = ceph_osdc_new_request(&fsc->client->osdc,
 963                                         &ci->i_layout, vino,
 964                                         offset, &len, 0, num_ops,
 965                                         CEPH_OSD_OP_WRITE,
 966                                         CEPH_OSD_FLAG_WRITE |
 967                                         CEPH_OSD_FLAG_ONDISK,
 968                                         snapc, truncate_seq,
 969                                         truncate_size, false);
 970                 if (IS_ERR(req)) {
 971                         req = ceph_osdc_new_request(&fsc->client->osdc,
 972                                                 &ci->i_layout, vino,
 973                                                 offset, &len, 0,
 974                                                 min(num_ops,
 975                                                     CEPH_OSD_SLAB_OPS),
 976                                                 CEPH_OSD_OP_WRITE,
 977                                                 CEPH_OSD_FLAG_WRITE |
 978                                                 CEPH_OSD_FLAG_ONDISK,
 979                                                 snapc, truncate_seq,
 980                                                 truncate_size, true);
 981                         BUG_ON(IS_ERR(req));
 982                 }
 983                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
 984                              PAGE_SIZE - offset);
 985
 986                 req->r_callback = writepages_finish;
 987                 req->r_inode = inode;
 988
 989                 /* Format the osd request message and submit the write */
 990                 len = 0;
 991                 data_pages = pages;
 992                 op_idx = 0;
 993                 for (i = 0; i < locked_pages; i++) {
 994                         u64 cur_offset = page_offset(pages[i]);
 995                         if (offset + len != cur_offset) {
 996                                 if (op_idx + do_sync + 1 == req->r_num_ops)
 997                                         break;
 998                                 osd_req_op_extent_dup_last(req, op_idx,
 999                                                            cur_offset - offset);
1000                                 dout("writepages got pages at %llu~%llu\n",
1001                                      offset, len);
1002                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1003                                                         data_pages, len, 0,
1004                                                         !!pool, false);
1005                                 osd_req_op_extent_update(req, op_idx, len);
1006
1007                                 len = 0;
1008                                 offset = cur_offset;
1009                                 data_pages = pages + i;
1010                                 op_idx++;
1011                         }
1012
1013                         set_page_writeback(pages[i]);
1014                         len += PAGE_SIZE;
1015                 }
1016
1017                 if (snap_size != -1) {
1018                         len = min(len, snap_size - offset);
1019                 } else if (i == locked_pages) {
1020                         /* writepages_finish() clears writeback pages
1021                          * according to the data length, so make sure
1022                          * data length covers all locked pages */
1023                         u64 min_len = len + 1 - PAGE_SIZE;
1024                         len = min(len, (u64)i_size_read(inode) - offset);
1025                         len = max(len, min_len);
1026                 }
1027                 dout("writepages got pages at %llu~%llu\n", offset, len);
1028
1029                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1030                                                  0, !!pool, false);
1031                 osd_req_op_extent_update(req, op_idx, len);
1032
1033                 if (do_sync) {
1034                         op_idx++;
1035                         osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1036                 }
1037                 BUG_ON(op_idx + 1 != req->r_num_ops);
1038
1039                 pool = NULL;
1040                 if (i < locked_pages) {
1041                         BUG_ON(num_ops <= req->r_num_ops);
1042                         num_ops -= req->r_num_ops;
1043                         num_ops += do_sync;
1044                         locked_pages -= i;
1045
1046                         /* allocate new pages array for next request */
1047                         data_pages = pages;
1048                         pages = kmalloc(locked_pages * sizeof (*pages),
1049                                         GFP_NOFS);
1050                         if (!pages) {
1051                                 pool = fsc->wb_pagevec_pool;
1052                                 pages = mempool_alloc(pool, GFP_NOFS);
1053                                 BUG_ON(!pages);
1054                         }
1055                         memcpy(pages, data_pages + i,
1056                                locked_pages * sizeof(*pages));
1057                         memset(data_pages + i, 0,
1058                                locked_pages * sizeof(*pages));
1059                 } else {
1060                         BUG_ON(num_ops != req->r_num_ops);
1061                         index = pages[i - 1]->index + 1;
1062                         /* request message now owns the pages array */
1063                         pages = NULL;
1064                 }
1065
1066                 vino = ceph_vino(inode);
1067                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
1068                                         &inode->i_mtime);
1069
1070                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1071                 BUG_ON(rc);
1072                 req = NULL;
1073
1074                 wbc->nr_to_write -= i;
1075                 if (pages)
1076                         goto new_request;
1077
1078                 if (wbc->nr_to_write <= 0)
1079                         done = 1;
1080
1081 release_pvec_pages:
1082                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1083                      pvec.nr ? pvec.pages[0] : NULL);
1084                 pagevec_release(&pvec);
1085
1086                 if (locked_pages && !done)
1087                         goto retry;
1088         }
1089
1090         if (should_loop && !done) {
1091                 /* more to do; loop back to beginning of file */
1092                 dout("writepages looping back to beginning of file\n");
1093                 should_loop = 0;
1094                 index = 0;
1095                 goto retry;
1096         }
1097
1098         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1099                 mapping->writeback_index = index;
1100
1101 out:
1102         if (req)
1103                 ceph_osdc_put_request(req);
1104         ceph_put_snap_context(snapc);
1105         dout("writepages done, rc = %d\n", rc);
1106         return rc;
1107 }
1108
1109
1110
1111 /*
1112  * See if a given @snapc is either writeable, or already written.
1113  */
1114 static int context_is_writeable_or_written(struct inode *inode,
1115                                            struct ceph_snap_context *snapc)
1116 {
1117         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1118         int ret = !oldest || snapc->seq <= oldest->seq;
1119
1120         ceph_put_snap_context(oldest);
1121         return ret;
1122 }
1123
1124 /*
1125  * We are only allowed to write into/dirty the page if the page is
1126  * clean, or already dirty within the same snap context.
1127  *
1128  * called with page locked.
1129  * return success with page locked,
1130  * or any failure (incl -EAGAIN) with page unlocked.
1131  */
1132 static int ceph_update_writeable_page(struct file *file,
1133                             loff_t pos, unsigned len,
1134                             struct page *page)
1135 {
1136         struct inode *inode = file_inode(file);
1137         struct ceph_inode_info *ci = ceph_inode(inode);
1138         loff_t page_off = pos & PAGE_MASK;
1139         int pos_in_page = pos & ~PAGE_MASK;
1140         int end_in_page = pos_in_page + len;
1141         loff_t i_size;
1142         int r;
1143         struct ceph_snap_context *snapc, *oldest;
1144
1145 retry_locked:
1146         /* writepages currently holds page lock, but if we change that later, */
1147         wait_on_page_writeback(page);
1148
1149         snapc = page_snap_context(page);
1150         if (snapc && snapc != ci->i_head_snapc) {
1151                 /*
1152                  * this page is already dirty in another (older) snap
1153                  * context!  is it writeable now?
1154                  */
1155                 oldest = get_oldest_context(inode, NULL);
1156
1157                 if (snapc->seq > oldest->seq) {
1158                         ceph_put_snap_context(oldest);
1159                         dout(" page %p snapc %p not current or oldest\n",
1160                              page, snapc);
1161                         /*
1162                          * queue for writeback, and wait for snapc to
1163                          * be writeable or written
1164                          */
1165                         snapc = ceph_get_snap_context(snapc);
1166                         unlock_page(page);
1167                         ceph_queue_writeback(inode);
1168                         r = wait_event_interruptible(ci->i_cap_wq,
1169                                context_is_writeable_or_written(inode, snapc));
1170                         ceph_put_snap_context(snapc);
1171                         if (r == -ERESTARTSYS)
1172                                 return r;
1173                         return -EAGAIN;
1174                 }
1175                 ceph_put_snap_context(oldest);
1176
1177                 /* yay, writeable, do it now (without dropping page lock) */
1178                 dout(" page %p snapc %p not current, but oldest\n",
1179                      page, snapc);
1180                 if (!clear_page_dirty_for_io(page))
1181                         goto retry_locked;
1182                 r = writepage_nounlock(page, NULL);
1183                 if (r < 0)
1184                         goto fail_nosnap;
1185                 goto retry_locked;
1186         }
1187
1188         if (PageUptodate(page)) {
1189                 dout(" page %p already uptodate\n", page);
1190                 return 0;
1191         }
1192
1193         /* full page? */
1194         if (pos_in_page == 0 && len == PAGE_SIZE)
1195                 return 0;
1196
1197         /* past end of file? */
1198         i_size = i_size_read(inode);
1199
1200         if (page_off >= i_size ||
1201             (pos_in_page == 0 && (pos+len) >= i_size &&
1202              end_in_page - pos_in_page != PAGE_SIZE)) {
1203                 dout(" zeroing %p 0 - %d and %d - %d\n",
1204                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1205                 zero_user_segments(page,
1206                                    0, pos_in_page,
1207                                    end_in_page, PAGE_SIZE);
1208                 return 0;
1209         }
1210
1211         /* we need to read it. */
1212         r = readpage_nounlock(file, page);
1213         if (r < 0)
1214                 goto fail_nosnap;
1215         goto retry_locked;
1216 fail_nosnap:
1217         unlock_page(page);
1218         return r;
1219 }
1220
1221 /*
1222  * We are only allowed to write into/dirty the page if the page is
1223  * clean, or already dirty within the same snap context.
1224  */
1225 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1226                             loff_t pos, unsigned len, unsigned flags,
1227                             struct page **pagep, void **fsdata)
1228 {
1229         struct inode *inode = file_inode(file);
1230         struct page *page;
1231         pgoff_t index = pos >> PAGE_SHIFT;
1232         int r;
1233
1234         do {
1235                 /* get a page */
1236                 page = grab_cache_page_write_begin(mapping, index, 0);
1237                 if (!page)
1238                         return -ENOMEM;
1239
1240                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1241                      inode, page, (int)pos, (int)len);
1242
1243                 r = ceph_update_writeable_page(file, pos, len, page);
1244                 if (r < 0)
1245                         put_page(page);
1246                 else
1247                         *pagep = page;
1248         } while (r == -EAGAIN);
1249
1250         return r;
1251 }
1252
1253 /*
1254  * we don't do anything in here that simple_write_end doesn't do
1255  * except adjust dirty page accounting
1256  */
1257 static int ceph_write_end(struct file *file, struct address_space *mapping,
1258                           loff_t pos, unsigned len, unsigned copied,
1259                           struct page *page, void *fsdata)
1260 {
1261         struct inode *inode = file_inode(file);
1262         unsigned from = pos & (PAGE_SIZE - 1);
1263         int check_cap = 0;
1264
1265         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1266              inode, page, (int)pos, (int)copied, (int)len);
1267
1268         /* zero the stale part of the page if we did a short copy */
1269         if (copied < len)
1270                 zero_user_segment(page, from+copied, len);
1271
1272         /* did file size increase? */
1273         if (pos+copied > i_size_read(inode))
1274                 check_cap = ceph_inode_set_size(inode, pos+copied);
1275
1276         if (!PageUptodate(page))
1277                 SetPageUptodate(page);
1278
1279         set_page_dirty(page);
1280
1281         unlock_page(page);
1282         put_page(page);
1283
1284         if (check_cap)
1285                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1286
1287         return copied;
1288 }
1289
1290 /*
1291  * we set .direct_IO to indicate direct io is supported, but since we
1292  * intercept O_DIRECT reads and writes early, this function should
1293  * never get called.
1294  */
1295 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter,
1296                               loff_t pos)
1297 {
1298         WARN_ON(1);
1299         return -EINVAL;
1300 }
1301
1302 const struct address_space_operations ceph_aops = {
1303         .readpage = ceph_readpage,
1304         .readpages = ceph_readpages,
1305         .writepage = ceph_writepage,
1306         .writepages = ceph_writepages_start,
1307         .write_begin = ceph_write_begin,
1308         .write_end = ceph_write_end,
1309         .set_page_dirty = ceph_set_page_dirty,
1310         .invalidatepage = ceph_invalidatepage,
1311         .releasepage = ceph_releasepage,
1312         .direct_IO = ceph_direct_io,
1313 };
1314
1315
1316 /*
1317  * vm ops
1318  */
1319 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1320 {
1321         struct inode *inode = file_inode(vma->vm_file);
1322         struct ceph_inode_info *ci = ceph_inode(inode);
1323         struct ceph_file_info *fi = vma->vm_file->private_data;
1324         struct page *pinned_page = NULL;
1325         loff_t off = vmf->pgoff << PAGE_SHIFT;
1326         int want, got, ret;
1327
1328         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1329              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1330         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1331                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1332         else
1333                 want = CEPH_CAP_FILE_CACHE;
1334         while (1) {
1335                 got = 0;
1336                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1337                                     -1, &got, &pinned_page);
1338                 if (ret == 0)
1339                         break;
1340                 if (ret != -ERESTARTSYS) {
1341                         WARN_ON(1);
1342                         return VM_FAULT_SIGBUS;
1343                 }
1344         }
1345         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1346              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1347
1348         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1349             ci->i_inline_version == CEPH_INLINE_NONE)
1350                 ret = filemap_fault(vma, vmf);
1351         else
1352                 ret = -EAGAIN;
1353
1354         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1355              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1356         if (pinned_page)
1357                 put_page(pinned_page);
1358         ceph_put_cap_refs(ci, got);
1359
1360         if (ret != -EAGAIN)
1361                 return ret;
1362
1363         /* read inline data */
1364         if (off >= PAGE_SIZE) {
1365                 /* does not support inline data > PAGE_SIZE */
1366                 ret = VM_FAULT_SIGBUS;
1367         } else {
1368                 int ret1;
1369                 struct address_space *mapping = inode->i_mapping;
1370                 struct page *page = find_or_create_page(mapping, 0,
1371                                                 mapping_gfp_constraint(mapping,
1372                                                 ~__GFP_FS));
1373                 if (!page) {
1374                         ret = VM_FAULT_OOM;
1375                         goto out;
1376                 }
1377                 ret1 = __ceph_do_getattr(inode, page,
1378                                          CEPH_STAT_CAP_INLINE_DATA, true);
1379                 if (ret1 < 0 || off >= i_size_read(inode)) {
1380                         unlock_page(page);
1381                         put_page(page);
1382                         ret = VM_FAULT_SIGBUS;
1383                         goto out;
1384                 }
1385                 if (ret1 < PAGE_SIZE)
1386                         zero_user_segment(page, ret1, PAGE_SIZE);
1387                 else
1388                         flush_dcache_page(page);
1389                 SetPageUptodate(page);
1390                 vmf->page = page;
1391                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1392         }
1393 out:
1394         dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1395              inode, off, (size_t)PAGE_SIZE, ret);
1396         return ret;
1397 }
1398
1399 /*
1400  * Reuse write_begin here for simplicity.
1401  */
1402 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1403 {
1404         struct inode *inode = file_inode(vma->vm_file);
1405         struct ceph_inode_info *ci = ceph_inode(inode);
1406         struct ceph_file_info *fi = vma->vm_file->private_data;
1407         struct ceph_cap_flush *prealloc_cf;
1408         struct page *page = vmf->page;
1409         loff_t off = page_offset(page);
1410         loff_t size = i_size_read(inode);
1411         size_t len;
1412         int want, got, ret;
1413
1414         prealloc_cf = ceph_alloc_cap_flush();
1415         if (!prealloc_cf)
1416                 return VM_FAULT_SIGBUS;
1417
1418         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1419                 struct page *locked_page = NULL;
1420                 if (off == 0) {
1421                         lock_page(page);
1422                         locked_page = page;
1423                 }
1424                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1425                 if (locked_page)
1426                         unlock_page(locked_page);
1427                 if (ret < 0) {
1428                         ret = VM_FAULT_SIGBUS;
1429                         goto out_free;
1430                 }
1431         }
1432
1433         if (off + PAGE_SIZE <= size)
1434                 len = PAGE_SIZE;
1435         else
1436                 len = size & ~PAGE_MASK;
1437
1438         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1439              inode, ceph_vinop(inode), off, len, size);
1440         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1441                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1442         else
1443                 want = CEPH_CAP_FILE_BUFFER;
1444         while (1) {
1445                 got = 0;
1446                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1447                                     &got, NULL);
1448                 if (ret == 0)
1449                         break;
1450                 if (ret != -ERESTARTSYS) {
1451                         WARN_ON(1);
1452                         ret = VM_FAULT_SIGBUS;
1453                         goto out_free;
1454                 }
1455         }
1456         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1457              inode, off, len, ceph_cap_string(got));
1458
1459         /* Update time before taking page lock */
1460         file_update_time(vma->vm_file);
1461
1462         lock_page(page);
1463
1464         ret = VM_FAULT_NOPAGE;
1465         if ((off > size) ||
1466             (page->mapping != inode->i_mapping)) {
1467                 unlock_page(page);
1468                 goto out;
1469         }
1470
1471         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1472         if (ret >= 0) {
1473                 /* success.  we'll keep the page locked. */
1474                 set_page_dirty(page);
1475                 ret = VM_FAULT_LOCKED;
1476         } else {
1477                 if (ret == -ENOMEM)
1478                         ret = VM_FAULT_OOM;
1479                 else
1480                         ret = VM_FAULT_SIGBUS;
1481         }
1482 out:
1483         if (ret == VM_FAULT_LOCKED ||
1484             ci->i_inline_version != CEPH_INLINE_NONE) {
1485                 int dirty;
1486                 spin_lock(&ci->i_ceph_lock);
1487                 ci->i_inline_version = CEPH_INLINE_NONE;
1488                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1489                                                &prealloc_cf);
1490                 spin_unlock(&ci->i_ceph_lock);
1491                 if (dirty)
1492                         __mark_inode_dirty(inode, dirty);
1493         }
1494
1495         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1496              inode, off, len, ceph_cap_string(got), ret);
1497         ceph_put_cap_refs(ci, got);
1498 out_free:
1499         ceph_free_cap_flush(prealloc_cf);
1500
1501         return ret;
1502 }
1503
1504 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1505                            char *data, size_t len)
1506 {
1507         struct address_space *mapping = inode->i_mapping;
1508         struct page *page;
1509
1510         if (locked_page) {
1511                 page = locked_page;
1512         } else {
1513                 if (i_size_read(inode) == 0)
1514                         return;
1515                 page = find_or_create_page(mapping, 0,
1516                                            mapping_gfp_constraint(mapping,
1517                                            ~__GFP_FS));
1518                 if (!page)
1519                         return;
1520                 if (PageUptodate(page)) {
1521                         unlock_page(page);
1522                         put_page(page);
1523                         return;
1524                 }
1525         }
1526
1527         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1528              inode, ceph_vinop(inode), len, locked_page);
1529
1530         if (len > 0) {
1531                 void *kaddr = kmap_atomic(page);
1532                 memcpy(kaddr, data, len);
1533                 kunmap_atomic(kaddr);
1534         }
1535
1536         if (page != locked_page) {
1537                 if (len < PAGE_SIZE)
1538                         zero_user_segment(page, len, PAGE_SIZE);
1539                 else
1540                         flush_dcache_page(page);
1541
1542                 SetPageUptodate(page);
1543                 unlock_page(page);
1544                 put_page(page);
1545         }
1546 }
1547
1548 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1549 {
1550         struct inode *inode = file_inode(filp);
1551         struct ceph_inode_info *ci = ceph_inode(inode);
1552         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1553         struct ceph_osd_request *req;
1554         struct page *page = NULL;
1555         u64 len, inline_version;
1556         int err = 0;
1557         bool from_pagecache = false;
1558
1559         spin_lock(&ci->i_ceph_lock);
1560         inline_version = ci->i_inline_version;
1561         spin_unlock(&ci->i_ceph_lock);
1562
1563         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1564              inode, ceph_vinop(inode), inline_version);
1565
1566         if (inline_version == 1 || /* initial version, no data */
1567             inline_version == CEPH_INLINE_NONE)
1568                 goto out;
1569
1570         if (locked_page) {
1571                 page = locked_page;
1572                 WARN_ON(!PageUptodate(page));
1573         } else if (ceph_caps_issued(ci) &
1574                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1575                 page = find_get_page(inode->i_mapping, 0);
1576                 if (page) {
1577                         if (PageUptodate(page)) {
1578                                 from_pagecache = true;
1579                                 lock_page(page);
1580                         } else {
1581                                 put_page(page);
1582                                 page = NULL;
1583                         }
1584                 }
1585         }
1586
1587         if (page) {
1588                 len = i_size_read(inode);
1589                 if (len > PAGE_SIZE)
1590                         len = PAGE_SIZE;
1591         } else {
1592                 page = __page_cache_alloc(GFP_NOFS);
1593                 if (!page) {
1594                         err = -ENOMEM;
1595                         goto out;
1596                 }
1597                 err = __ceph_do_getattr(inode, page,
1598                                         CEPH_STAT_CAP_INLINE_DATA, true);
1599                 if (err < 0) {
1600                         /* no inline data */
1601                         if (err == -ENODATA)
1602                                 err = 0;
1603                         goto out;
1604                 }
1605                 len = err;
1606         }
1607
1608         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1609                                     ceph_vino(inode), 0, &len, 0, 1,
1610                                     CEPH_OSD_OP_CREATE,
1611                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1612                                     NULL, 0, 0, false);
1613         if (IS_ERR(req)) {
1614                 err = PTR_ERR(req);
1615                 goto out;
1616         }
1617
1618         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1619         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1620         if (!err)
1621                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1622         ceph_osdc_put_request(req);
1623         if (err < 0)
1624                 goto out;
1625
1626         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1627                                     ceph_vino(inode), 0, &len, 1, 3,
1628                                     CEPH_OSD_OP_WRITE,
1629                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1630                                     NULL, ci->i_truncate_seq,
1631                                     ci->i_truncate_size, false);
1632         if (IS_ERR(req)) {
1633                 err = PTR_ERR(req);
1634                 goto out;
1635         }
1636
1637         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1638
1639         {
1640                 __le64 xattr_buf = cpu_to_le64(inline_version);
1641                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1642                                             "inline_version", &xattr_buf,
1643                                             sizeof(xattr_buf),
1644                                             CEPH_OSD_CMPXATTR_OP_GT,
1645                                             CEPH_OSD_CMPXATTR_MODE_U64);
1646                 if (err)
1647                         goto out_put;
1648         }
1649
1650         {
1651                 char xattr_buf[32];
1652                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1653                                          "%llu", inline_version);
1654                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1655                                             "inline_version",
1656                                             xattr_buf, xattr_len, 0, 0);
1657                 if (err)
1658                         goto out_put;
1659         }
1660
1661         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1662         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1663         if (!err)
1664                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1665 out_put:
1666         ceph_osdc_put_request(req);
1667         if (err == -ECANCELED)
1668                 err = 0;
1669 out:
1670         if (page && page != locked_page) {
1671                 if (from_pagecache) {
1672                         unlock_page(page);
1673                         put_page(page);
1674                 } else
1675                         __free_pages(page, 0);
1676         }
1677
1678         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1679              inode, ceph_vinop(inode), inline_version, err);
1680         return err;
1681 }
1682
1683 static const struct vm_operations_struct ceph_vmops = {
1684         .fault          = ceph_filemap_fault,
1685         .page_mkwrite   = ceph_page_mkwrite,
1686 };
1687
1688 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1689 {
1690         struct address_space *mapping = file->f_mapping;
1691
1692         if (!mapping->a_ops->readpage)
1693                 return -ENOEXEC;
1694         file_accessed(file);
1695         vma->vm_ops = &ceph_vmops;
1696         return 0;
1697 }
1698
1699 enum {
1700         POOL_READ       = 1,
1701         POOL_WRITE      = 2,
1702 };
1703
1704 static int __ceph_pool_perm_get(struct ceph_inode_info *ci, u32 pool)
1705 {
1706         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1707         struct ceph_mds_client *mdsc = fsc->mdsc;
1708         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1709         struct rb_node **p, *parent;
1710         struct ceph_pool_perm *perm;
1711         struct page **pages;
1712         int err = 0, err2 = 0, have = 0;
1713
1714         down_read(&mdsc->pool_perm_rwsem);
1715         p = &mdsc->pool_perm_tree.rb_node;
1716         while (*p) {
1717                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1718                 if (pool < perm->pool)
1719                         p = &(*p)->rb_left;
1720                 else if (pool > perm->pool)
1721                         p = &(*p)->rb_right;
1722                 else {
1723                         have = perm->perm;
1724                         break;
1725                 }
1726         }
1727         up_read(&mdsc->pool_perm_rwsem);
1728         if (*p)
1729                 goto out;
1730
1731         dout("__ceph_pool_perm_get pool %u no perm cached\n", pool);
1732
1733         down_write(&mdsc->pool_perm_rwsem);
1734         parent = NULL;
1735         while (*p) {
1736                 parent = *p;
1737                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1738                 if (pool < perm->pool)
1739                         p = &(*p)->rb_left;
1740                 else if (pool > perm->pool)
1741                         p = &(*p)->rb_right;
1742                 else {
1743                         have = perm->perm;
1744                         break;
1745                 }
1746         }
1747         if (*p) {
1748                 up_write(&mdsc->pool_perm_rwsem);
1749                 goto out;
1750         }
1751
1752         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1753                                          1, false, GFP_NOFS);
1754         if (!rd_req) {
1755                 err = -ENOMEM;
1756                 goto out_unlock;
1757         }
1758
1759         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1760         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1761         rd_req->r_base_oloc.pool = pool;
1762         snprintf(rd_req->r_base_oid.name, sizeof(rd_req->r_base_oid.name),
1763                  "%llx.00000000", ci->i_vino.ino);
1764         rd_req->r_base_oid.name_len = strlen(rd_req->r_base_oid.name);
1765
1766         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1767                                          1, false, GFP_NOFS);
1768         if (!wr_req) {
1769                 err = -ENOMEM;
1770                 goto out_unlock;
1771         }
1772
1773         wr_req->r_flags = CEPH_OSD_FLAG_WRITE |
1774                           CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK;
1775         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1776         wr_req->r_base_oloc.pool = pool;
1777         wr_req->r_base_oid = rd_req->r_base_oid;
1778
1779         /* one page should be large enough for STAT data */
1780         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1781         if (IS_ERR(pages)) {
1782                 err = PTR_ERR(pages);
1783                 goto out_unlock;
1784         }
1785
1786         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1787                                      0, false, true);
1788         ceph_osdc_build_request(rd_req, 0, NULL, CEPH_NOSNAP,
1789                                 &ci->vfs_inode.i_mtime);
1790         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1791
1792         ceph_osdc_build_request(wr_req, 0, NULL, CEPH_NOSNAP,
1793                                 &ci->vfs_inode.i_mtime);
1794         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1795
1796         if (!err)
1797                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1798         if (!err2)
1799                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1800
1801         if (err >= 0 || err == -ENOENT)
1802                 have |= POOL_READ;
1803         else if (err != -EPERM)
1804                 goto out_unlock;
1805
1806         if (err2 == 0 || err2 == -EEXIST)
1807                 have |= POOL_WRITE;
1808         else if (err2 != -EPERM) {
1809                 err = err2;
1810                 goto out_unlock;
1811         }
1812
1813         perm = kmalloc(sizeof(*perm), GFP_NOFS);
1814         if (!perm) {
1815                 err = -ENOMEM;
1816                 goto out_unlock;
1817         }
1818
1819         perm->pool = pool;
1820         perm->perm = have;
1821         rb_link_node(&perm->node, parent, p);
1822         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1823         err = 0;
1824 out_unlock:
1825         up_write(&mdsc->pool_perm_rwsem);
1826
1827         if (rd_req)
1828                 ceph_osdc_put_request(rd_req);
1829         if (wr_req)
1830                 ceph_osdc_put_request(wr_req);
1831 out:
1832         if (!err)
1833                 err = have;
1834         dout("__ceph_pool_perm_get pool %u result = %d\n", pool, err);
1835         return err;
1836 }
1837
1838 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1839 {
1840         u32 pool;
1841         int ret, flags;
1842
1843         /* does not support pool namespace yet */
1844         if (ci->i_pool_ns_len)
1845                 return -EIO;
1846
1847         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1848                                 NOPOOLPERM))
1849                 return 0;
1850
1851         spin_lock(&ci->i_ceph_lock);
1852         flags = ci->i_ceph_flags;
1853         pool = ceph_file_layout_pg_pool(ci->i_layout);
1854         spin_unlock(&ci->i_ceph_lock);
1855 check:
1856         if (flags & CEPH_I_POOL_PERM) {
1857                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1858                         dout("ceph_pool_perm_check pool %u no read perm\n",
1859                              pool);
1860                         return -EPERM;
1861                 }
1862                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1863                         dout("ceph_pool_perm_check pool %u no write perm\n",
1864                              pool);
1865                         return -EPERM;
1866                 }
1867                 return 0;
1868         }
1869
1870         ret = __ceph_pool_perm_get(ci, pool);
1871         if (ret < 0)
1872                 return ret;
1873
1874         flags = CEPH_I_POOL_PERM;
1875         if (ret & POOL_READ)
1876                 flags |= CEPH_I_POOL_RD;
1877         if (ret & POOL_WRITE)
1878                 flags |= CEPH_I_POOL_WR;
1879
1880         spin_lock(&ci->i_ceph_lock);
1881         if (pool == ceph_file_layout_pg_pool(ci->i_layout)) {
1882                 ci->i_ceph_flags = flags;
1883         } else {
1884                 pool = ceph_file_layout_pg_pool(ci->i_layout);
1885                 flags = ci->i_ceph_flags;
1886         }
1887         spin_unlock(&ci->i_ceph_lock);
1888         goto check;
1889 }
1890
1891 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1892 {
1893         struct ceph_pool_perm *perm;
1894         struct rb_node *n;
1895
1896         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1897                 n = rb_first(&mdsc->pool_perm_tree);
1898                 perm = rb_entry(n, struct ceph_pool_perm, node);
1899                 rb_erase(n, &mdsc->pool_perm_tree);
1900                 kfree(perm);
1901         }
1902 }