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 <linux/ceph/osd_client.h>
  15
  16 /*
  17  * Ceph address space ops.
  18  *
  19  * There are a few funny things going on here.
  20  *
  21  * The page->private field is used to reference a struct
  22  * ceph_snap_context for _every_ dirty page.  This indicates which
  23  * snapshot the page was logically dirtied in, and thus which snap
  24  * context needs to be associated with the osd write during writeback.
  25  *
  26  * Similarly, struct ceph_inode_info maintains a set of counters to
  27  * count dirty pages on the inode.  In the absence of snapshots,
  28  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  29  *
  30  * When a snapshot is taken (that is, when the client receives
  31  * notification that a snapshot was taken), each inode with caps and
  32  * with dirty pages (dirty pages implies there is a cap) gets a new
  33  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  34  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  35  * moved to capsnap->dirty. (Unless a sync write is currently in
  36  * progress.  In that case, the capsnap is said to be "pending", new
  37  * writes cannot start, and the capsnap isn't "finalized" until the
  38  * write completes (or fails) and a final size/mtime for the inode for
  39  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  40  *
  41  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  42  * we look for the first capsnap in i_cap_snaps and write out pages in
  43  * that snap context _only_.  Then we move on to the next capsnap,
  44  * eventually reaching the "live" or "head" context (i.e., pages that
  45  * are not yet snapped) and are writing the most recently dirtied
  46  * pages.
  47  *
  48  * Invalidate and so forth must take care to ensure the dirty page
  49  * accounting is preserved.
  50  */
  51
  52 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  53 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  54         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  55          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  56
  57 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  58 {
  59         if (PagePrivate(page))
  60                 return (void *)page->private;
  61         return NULL;
  62 }
  63
  64 /*
  65  * Dirty a page.  Optimistically adjust accounting, on the assumption
  66  * that we won't race with invalidate.  If we do, readjust.
  67  */
  68 static int ceph_set_page_dirty(struct page *page)
  69 {
  70         struct address_space *mapping = page->mapping;
  71         struct inode *inode;
  72         struct ceph_inode_info *ci;
  73         int undo = 0;
  74         struct ceph_snap_context *snapc;
  75
  76         if (unlikely(!mapping))
  77                 return !TestSetPageDirty(page);
  78
  79         if (TestSetPageDirty(page)) {
  80                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  81                      mapping->host, page, page->index);
  82                 return 0;
  83         }
  84
  85         inode = mapping->host;
  86         ci = ceph_inode(inode);
  87
  88         /*
  89          * Note that we're grabbing a snapc ref here without holding
  90          * any locks!
  91          */
  92         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  93
  94         /* dirty the head */
  95         spin_lock(&ci->i_ceph_lock);
  96         if (ci->i_head_snapc == NULL)
  97                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  98         ++ci->i_wrbuffer_ref_head;
  99         if (ci->i_wrbuffer_ref == 0)
 100                 ihold(inode);
 101         ++ci->i_wrbuffer_ref;
 102         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 103              "snapc %p seq %lld (%d snaps)\n",
 104              mapping->host, page, page->index,
 105              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 106              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 107              snapc, snapc->seq, snapc->num_snaps);
 108         spin_unlock(&ci->i_ceph_lock);
 109
 110         /* now adjust page */
 111         spin_lock_irq(&mapping->tree_lock);
 112         if (page->mapping) {    /* Race with truncate? */
 113                 WARN_ON_ONCE(!PageUptodate(page));
 114                 account_page_dirtied(page, page->mapping);
 115                 radix_tree_tag_set(&mapping->page_tree,
 116                                 page_index(page), PAGECACHE_TAG_DIRTY);
 117
 118                 /*
 119                  * Reference snap context in page->private.  Also set
 120                  * PagePrivate so that we get invalidatepage callback.
 121                  */
 122                 page->private = (unsigned long)snapc;
 123                 SetPagePrivate(page);
 124         } else {
 125                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 126                 undo = 1;
 127         }
 128
 129         spin_unlock_irq(&mapping->tree_lock);
 130
 131         if (undo)
 132                 /* whoops, we failed to dirty the page */
 133                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 134
 135         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 136
 137         BUG_ON(!PageDirty(page));
 138         return 1;
 139 }
 140
 141 /*
 142  * If we are truncating the full page (i.e. offset == 0), adjust the
 143  * dirty page counters appropriately.  Only called if there is private
 144  * data on the page.
 145  */
 146 static void ceph_invalidatepage(struct page *page, unsigned long offset)
 147 {
 148         struct inode *inode;
 149         struct ceph_inode_info *ci;
 150         struct ceph_snap_context *snapc = page_snap_context(page);
 151
 152         BUG_ON(!PageLocked(page));
 153         BUG_ON(!PagePrivate(page));
 154         BUG_ON(!page->mapping);
 155
 156         inode = page->mapping->host;
 157
 158         /*
 159          * We can get non-dirty pages here due to races between
 160          * set_page_dirty and truncate_complete_page; just spit out a
 161          * warning, in case we end up with accounting problems later.
 162          */
 163         if (!PageDirty(page))
 164                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 165
 166         if (offset == 0)
 167                 ClearPageChecked(page);
 168
 169         ci = ceph_inode(inode);
 170         if (offset == 0) {
 171                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 172                      inode, page, page->index, offset);
 173                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 174                 ceph_put_snap_context(snapc);
 175                 page->private = 0;
 176                 ClearPagePrivate(page);
 177         } else {
 178                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
 179                      inode, page, page->index);
 180         }
 181 }
 182
 183 /* just a sanity check */
 184 static int ceph_releasepage(struct page *page, gfp_t g)
 185 {
 186         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 187         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 188         WARN_ON(PageDirty(page));
 189         WARN_ON(PagePrivate(page));
 190         return 0;
 191 }
 192
 193 /*
 194  * read a single page, without unlocking it.
 195  */
 196 static int readpage_nounlock(struct file *filp, struct page *page)
 197 {
 198         struct inode *inode = filp->f_dentry->d_inode;
 199         struct ceph_inode_info *ci = ceph_inode(inode);
 200         struct ceph_osd_client *osdc =
 201                 &ceph_inode_to_client(inode)->client->osdc;
 202         int err = 0;
 203         u64 len = PAGE_CACHE_SIZE;
 204
 205         dout("readpage inode %p file %p page %p index %lu\n",
 206              inode, filp, page, page->index);
 207         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 208                                   page->index << PAGE_CACHE_SHIFT, &len,
 209                                   ci->i_truncate_seq, ci->i_truncate_size,
 210                                   &page, 1, 0);
 211         if (err == -ENOENT)
 212                 err = 0;
 213         if (err < 0) {
 214                 SetPageError(page);
 215                 goto out;
 216         } else if (err < PAGE_CACHE_SIZE) {
 217                 /* zero fill remainder of page */
 218                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 219         }
 220         SetPageUptodate(page);
 221
 222 out:
 223         return err < 0 ? err : 0;
 224 }
 225
 226 static int ceph_readpage(struct file *filp, struct page *page)
 227 {
 228         int r = readpage_nounlock(filp, page);
 229         unlock_page(page);
 230         return r;
 231 }
 232
 233 /*
 234  * Finish an async read(ahead) op.
 235  */
 236 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 237 {
 238         struct inode *inode = req->r_inode;
 239         struct ceph_osd_reply_head *replyhead;
 240         int rc, bytes;
 241         int i;
 242
 243         /* parse reply */
 244         replyhead = msg->front.iov_base;
 245         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 246         rc = le32_to_cpu(replyhead->result);
 247         bytes = le32_to_cpu(msg->hdr.data_len);
 248
 249         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 250
 251         /* unlock all pages, zeroing any data we didn't read */
 252         for (i = 0; i < req->r_num_pages; i++, bytes -= PAGE_CACHE_SIZE) {
 253                 struct page *page = req->r_pages[i];
 254
 255                 if (bytes < (int)PAGE_CACHE_SIZE) {
 256                         /* zero (remainder of) page */
 257                         int s = bytes < 0 ? 0 : bytes;
 258                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 259                 }
 260                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 261                      page->index);
 262                 flush_dcache_page(page);
 263                 SetPageUptodate(page);
 264                 unlock_page(page);
 265                 page_cache_release(page);
 266         }
 267         kfree(req->r_pages);
 268 }
 269
 270 /*
 271  * start an async read(ahead) operation.  return nr_pages we submitted
 272  * a read for on success, or negative error code.
 273  */
 274 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 275 {
 276         struct ceph_osd_client *osdc =
 277                 &ceph_inode_to_client(inode)->client->osdc;
 278         struct ceph_inode_info *ci = ceph_inode(inode);
 279         struct page *page = list_entry(page_list->prev, struct page, lru);
 280         struct ceph_osd_request *req;
 281         u64 off;
 282         u64 len;
 283         int i;
 284         struct page **pages;
 285         pgoff_t next_index;
 286         int nr_pages = 0;
 287         int ret;
 288
 289         off = page->index << PAGE_CACHE_SHIFT;
 290
 291         /* count pages */
 292         next_index = page->index;
 293         list_for_each_entry_reverse(page, page_list, lru) {
 294                 if (page->index != next_index)
 295                         break;
 296                 nr_pages++;
 297                 next_index++;
 298                 if (max && nr_pages == max)
 299                         break;
 300         }
 301         len = nr_pages << PAGE_CACHE_SHIFT;
 302         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 303              off, len);
 304
 305         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
 306                                     off, &len,
 307                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 308                                     NULL, 0,
 309                                     ci->i_truncate_seq, ci->i_truncate_size,
 310                                     NULL, false, 1, 0);
 311         if (!req)
 312                 return -ENOMEM;
 313
 314         /* build page vector */
 315         nr_pages = len >> PAGE_CACHE_SHIFT;
 316         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 317         ret = -ENOMEM;
 318         if (!pages)
 319                 goto out;
 320         for (i = 0; i < nr_pages; ++i) {
 321                 page = list_entry(page_list->prev, struct page, lru);
 322                 BUG_ON(PageLocked(page));
 323                 list_del(&page->lru);
 324
 325                 dout("start_read %p adding %p idx %lu\n", inode, page,
 326                      page->index);
 327                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 328                                           GFP_NOFS)) {
 329                         page_cache_release(page);
 330                         dout("start_read %p add_to_page_cache failed %p\n",
 331                              inode, page);
 332                         nr_pages = i;
 333                         goto out_pages;
 334                 }
 335                 pages[i] = page;
 336         }
 337         req->r_pages = pages;
 338         req->r_num_pages = nr_pages;
 339         req->r_callback = finish_read;
 340         req->r_inode = inode;
 341
 342         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 343         ret = ceph_osdc_start_request(osdc, req, false);
 344         if (ret < 0)
 345                 goto out_pages;
 346         ceph_osdc_put_request(req);
 347         return nr_pages;
 348
 349 out_pages:
 350         ceph_release_page_vector(pages, nr_pages);
 351 out:
 352         ceph_osdc_put_request(req);
 353         return ret;
 354 }
 355
 356
 357 /*
 358  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 359  * the caller (VM) cleans them up.
 360  */
 361 static int ceph_readpages(struct file *file, struct address_space *mapping,
 362                           struct list_head *page_list, unsigned nr_pages)
 363 {
 364         struct inode *inode = file->f_dentry->d_inode;
 365         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 366         int rc = 0;
 367         int max = 0;
 368
 369         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 370                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 371                         >> PAGE_SHIFT;
 372
 373         dout("readpages %p file %p nr_pages %d max %d\n", inode, file, nr_pages,
 374              max);
 375         while (!list_empty(page_list)) {
 376                 rc = start_read(inode, page_list, max);
 377                 if (rc < 0)
 378                         goto out;
 379                 BUG_ON(rc == 0);
 380         }
 381 out:
 382         dout("readpages %p file %p ret %d\n", inode, file, rc);
 383         return rc;
 384 }
 385
 386 /*
 387  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 388  * only snap context we are allowed to write back.
 389  */
 390 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 391                                                     u64 *snap_size)
 392 {
 393         struct ceph_inode_info *ci = ceph_inode(inode);
 394         struct ceph_snap_context *snapc = NULL;
 395         struct ceph_cap_snap *capsnap = NULL;
 396
 397         spin_lock(&ci->i_ceph_lock);
 398         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 399                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 400                      capsnap->context, capsnap->dirty_pages);
 401                 if (capsnap->dirty_pages) {
 402                         snapc = ceph_get_snap_context(capsnap->context);
 403                         if (snap_size)
 404                                 *snap_size = capsnap->size;
 405                         break;
 406                 }
 407         }
 408         if (!snapc && ci->i_wrbuffer_ref_head) {
 409                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 410                 dout(" head snapc %p has %d dirty pages\n",
 411                      snapc, ci->i_wrbuffer_ref_head);
 412         }
 413         spin_unlock(&ci->i_ceph_lock);
 414         return snapc;
 415 }
 416
 417 /*
 418  * Write a single page, but leave the page locked.
 419  *
 420  * If we get a write error, set the page error bit, but still adjust the
 421  * dirty page accounting (i.e., page is no longer dirty).
 422  */
 423 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 424 {
 425         struct inode *inode;
 426         struct ceph_inode_info *ci;
 427         struct ceph_fs_client *fsc;
 428         struct ceph_osd_client *osdc;
 429         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 430         int len = PAGE_CACHE_SIZE;
 431         loff_t i_size;
 432         int err = 0;
 433         struct ceph_snap_context *snapc, *oldest;
 434         u64 snap_size = 0;
 435         long writeback_stat;
 436
 437         dout("writepage %p idx %lu\n", page, page->index);
 438
 439         if (!page->mapping || !page->mapping->host) {
 440                 dout("writepage %p - no mapping\n", page);
 441                 return -EFAULT;
 442         }
 443         inode = page->mapping->host;
 444         ci = ceph_inode(inode);
 445         fsc = ceph_inode_to_client(inode);
 446         osdc = &fsc->client->osdc;
 447
 448         /* verify this is a writeable snap context */
 449         snapc = page_snap_context(page);
 450         if (snapc == NULL) {
 451                 dout("writepage %p page %p not dirty?\n", inode, page);
 452                 goto out;
 453         }
 454         oldest = get_oldest_context(inode, &snap_size);
 455         if (snapc->seq > oldest->seq) {
 456                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 457                      inode, page, snapc);
 458                 /* we should only noop if called by kswapd */
 459                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 460                 ceph_put_snap_context(oldest);
 461                 goto out;
 462         }
 463         ceph_put_snap_context(oldest);
 464
 465         /* is this a partial page at end of file? */
 466         if (snap_size)
 467                 i_size = snap_size;
 468         else
 469                 i_size = i_size_read(inode);
 470         if (i_size < page_off + len)
 471                 len = i_size - page_off;
 472
 473         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 474              inode, page, page->index, page_off, len, snapc);
 475
 476         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 477         if (writeback_stat >
 478             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 479                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 480
 481         set_page_writeback(page);
 482         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 483                                    &ci->i_layout, snapc,
 484                                    page_off, len,
 485                                    ci->i_truncate_seq, ci->i_truncate_size,
 486                                    &inode->i_mtime,
 487                                    &page, 1, 0, 0, true);
 488         if (err < 0) {
 489                 dout("writepage setting page/mapping error %d %p\n", err, page);
 490                 SetPageError(page);
 491                 mapping_set_error(&inode->i_data, err);
 492                 if (wbc)
 493                         wbc->pages_skipped++;
 494         } else {
 495                 dout("writepage cleaned page %p\n", page);
 496                 err = 0;  /* vfs expects us to return 0 */
 497         }
 498         page->private = 0;
 499         ClearPagePrivate(page);
 500         end_page_writeback(page);
 501         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 502         ceph_put_snap_context(snapc);  /* page's reference */
 503 out:
 504         return err;
 505 }
 506
 507 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 508 {
 509         int err;
 510         struct inode *inode = page->mapping->host;
 511         BUG_ON(!inode);
 512         ihold(inode);
 513         err = writepage_nounlock(page, wbc);
 514         unlock_page(page);
 515         iput(inode);
 516         return err;
 517 }
 518
 519
 520 /*
 521  * lame release_pages helper.  release_pages() isn't exported to
 522  * modules.
 523  */
 524 static void ceph_release_pages(struct page **pages, int num)
 525 {
 526         struct pagevec pvec;
 527         int i;
 528
 529         pagevec_init(&pvec, 0);
 530         for (i = 0; i < num; i++) {
 531                 if (pagevec_add(&pvec, pages[i]) == 0)
 532                         pagevec_release(&pvec);
 533         }
 534         pagevec_release(&pvec);
 535 }
 536
 537
 538 /*
 539  * async writeback completion handler.
 540  *
 541  * If we get an error, set the mapping error bit, but not the individual
 542  * page error bits.
 543  */
 544 static void writepages_finish(struct ceph_osd_request *req,
 545                               struct ceph_msg *msg)
 546 {
 547         struct inode *inode = req->r_inode;
 548         struct ceph_osd_reply_head *replyhead;
 549         struct ceph_osd_op *op;
 550         struct ceph_inode_info *ci = ceph_inode(inode);
 551         unsigned wrote;
 552         struct page *page;
 553         int i;
 554         struct ceph_snap_context *snapc = req->r_snapc;
 555         struct address_space *mapping = inode->i_mapping;
 556         __s32 rc = -EIO;
 557         u64 bytes = 0;
 558         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 559         long writeback_stat;
 560         unsigned issued = ceph_caps_issued(ci);
 561
 562         /* parse reply */
 563         replyhead = msg->front.iov_base;
 564         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 565         op = (void *)(replyhead + 1);
 566         rc = le32_to_cpu(replyhead->result);
 567         bytes = le64_to_cpu(op->extent.length);
 568
 569         if (rc >= 0) {
 570                 /*
 571                  * Assume we wrote the pages we originally sent.  The
 572                  * osd might reply with fewer pages if our writeback
 573                  * raced with a truncation and was adjusted at the osd,
 574                  * so don't believe the reply.
 575                  */
 576                 wrote = req->r_num_pages;
 577         } else {
 578                 wrote = 0;
 579                 mapping_set_error(mapping, rc);
 580         }
 581         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 582              inode, rc, bytes, wrote);
 583
 584         /* clean all pages */
 585         for (i = 0; i < req->r_num_pages; i++) {
 586                 page = req->r_pages[i];
 587                 BUG_ON(!page);
 588                 WARN_ON(!PageUptodate(page));
 589
 590                 writeback_stat =
 591                         atomic_long_dec_return(&fsc->writeback_count);
 592                 if (writeback_stat <
 593                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 594                         clear_bdi_congested(&fsc->backing_dev_info,
 595                                             BLK_RW_ASYNC);
 596
 597                 ceph_put_snap_context(page_snap_context(page));
 598                 page->private = 0;
 599                 ClearPagePrivate(page);
 600                 dout("unlocking %d %p\n", i, page);
 601                 end_page_writeback(page);
 602
 603                 /*
 604                  * We lost the cache cap, need to truncate the page before
 605                  * it is unlocked, otherwise we'd truncate it later in the
 606                  * page truncation thread, possibly losing some data that
 607                  * raced its way in
 608                  */
 609                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 610                         generic_error_remove_page(inode->i_mapping, page);
 611
 612                 unlock_page(page);
 613         }
 614         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 615         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 616
 617         ceph_release_pages(req->r_pages, req->r_num_pages);
 618         if (req->r_pages_from_pool)
 619                 mempool_free(req->r_pages,
 620                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 621         else
 622                 kfree(req->r_pages);
 623         ceph_osdc_put_request(req);
 624 }
 625
 626 /*
 627  * allocate a page vec, either directly, or if necessary, via a the
 628  * mempool.  we avoid the mempool if we can because req->r_num_pages
 629  * may be less than the maximum write size.
 630  */
 631 static void alloc_page_vec(struct ceph_fs_client *fsc,
 632                            struct ceph_osd_request *req)
 633 {
 634         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 635                                GFP_NOFS);
 636         if (!req->r_pages) {
 637                 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
 638                 req->r_pages_from_pool = 1;
 639                 WARN_ON(!req->r_pages);
 640         }
 641 }
 642
 643 /*
 644  * initiate async writeback
 645  */
 646 static int ceph_writepages_start(struct address_space *mapping,
 647                                  struct writeback_control *wbc)
 648 {
 649         struct inode *inode = mapping->host;
 650         struct ceph_inode_info *ci = ceph_inode(inode);
 651         struct ceph_fs_client *fsc;
 652         pgoff_t index, start, end;
 653         int range_whole = 0;
 654         int should_loop = 1;
 655         pgoff_t max_pages = 0, max_pages_ever = 0;
 656         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 657         struct pagevec pvec;
 658         int done = 0;
 659         int rc = 0;
 660         unsigned wsize = 1 << inode->i_blkbits;
 661         struct ceph_osd_request *req = NULL;
 662         int do_sync;
 663         u64 snap_size = 0;
 664
 665         /*
 666          * Include a 'sync' in the OSD request if this is a data
 667          * integrity write (e.g., O_SYNC write or fsync()), or if our
 668          * cap is being revoked.
 669          */
 670         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 671         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 672                 do_sync = 1;
 673         dout("writepages_start %p dosync=%d (mode=%s)\n",
 674              inode, do_sync,
 675              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 676              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 677
 678         fsc = ceph_inode_to_client(inode);
 679         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 680                 pr_warning("writepage_start %p on forced umount\n", inode);
 681                 return -EIO; /* we're in a forced umount, don't write! */
 682         }
 683         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 684                 wsize = fsc->mount_options->wsize;
 685         if (wsize < PAGE_CACHE_SIZE)
 686                 wsize = PAGE_CACHE_SIZE;
 687         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 688
 689         pagevec_init(&pvec, 0);
 690
 691         /* where to start/end? */
 692         if (wbc->range_cyclic) {
 693                 start = mapping->writeback_index; /* Start from prev offset */
 694                 end = -1;
 695                 dout(" cyclic, start at %lu\n", start);
 696         } else {
 697                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 698                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 699                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 700                         range_whole = 1;
 701                 should_loop = 0;
 702                 dout(" not cyclic, %lu to %lu\n", start, end);
 703         }
 704         index = start;
 705
 706 retry:
 707         /* find oldest snap context with dirty data */
 708         ceph_put_snap_context(snapc);
 709         snapc = get_oldest_context(inode, &snap_size);
 710         if (!snapc) {
 711                 /* hmm, why does writepages get called when there
 712                    is no dirty data? */
 713                 dout(" no snap context with dirty data?\n");
 714                 goto out;
 715         }
 716         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 717              snapc, snapc->seq, snapc->num_snaps);
 718         if (last_snapc && snapc != last_snapc) {
 719                 /* if we switched to a newer snapc, restart our scan at the
 720                  * start of the original file range. */
 721                 dout("  snapc differs from last pass, restarting at %lu\n",
 722                      index);
 723                 index = start;
 724         }
 725         last_snapc = snapc;
 726
 727         while (!done && index <= end) {
 728                 unsigned i;
 729                 int first;
 730                 pgoff_t next;
 731                 int pvec_pages, locked_pages;
 732                 struct page *page;
 733                 int want;
 734                 u64 offset, len;
 735                 struct ceph_osd_request_head *reqhead;
 736                 struct ceph_osd_op *op;
 737                 long writeback_stat;
 738
 739                 next = 0;
 740                 locked_pages = 0;
 741                 max_pages = max_pages_ever;
 742
 743 get_more_pages:
 744                 first = -1;
 745                 want = min(end - index,
 746                            min((pgoff_t)PAGEVEC_SIZE,
 747                                max_pages - (pgoff_t)locked_pages) - 1)
 748                         + 1;
 749                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 750                                                 PAGECACHE_TAG_DIRTY,
 751                                                 want);
 752                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 753                 if (!pvec_pages && !locked_pages)
 754                         break;
 755                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 756                         page = pvec.pages[i];
 757                         dout("? %p idx %lu\n", page, page->index);
 758                         if (locked_pages == 0)
 759                                 lock_page(page);  /* first page */
 760                         else if (!trylock_page(page))
 761                                 break;
 762
 763                         /* only dirty pages, or our accounting breaks */
 764                         if (unlikely(!PageDirty(page)) ||
 765                             unlikely(page->mapping != mapping)) {
 766                                 dout("!dirty or !mapping %p\n", page);
 767                                 unlock_page(page);
 768                                 break;
 769                         }
 770                         if (!wbc->range_cyclic && page->index > end) {
 771                                 dout("end of range %p\n", page);
 772                                 done = 1;
 773                                 unlock_page(page);
 774                                 break;
 775                         }
 776                         if (next && (page->index != next)) {
 777                                 dout("not consecutive %p\n", page);
 778                                 unlock_page(page);
 779                                 break;
 780                         }
 781                         if (wbc->sync_mode != WB_SYNC_NONE) {
 782                                 dout("waiting on writeback %p\n", page);
 783                                 wait_on_page_writeback(page);
 784                         }
 785                         if ((snap_size && page_offset(page) > snap_size) ||
 786                             (!snap_size &&
 787                              page_offset(page) > i_size_read(inode))) {
 788                                 dout("%p page eof %llu\n", page, snap_size ?
 789                                      snap_size : i_size_read(inode));
 790                                 done = 1;
 791                                 unlock_page(page);
 792                                 break;
 793                         }
 794                         if (PageWriteback(page)) {
 795                                 dout("%p under writeback\n", page);
 796                                 unlock_page(page);
 797                                 break;
 798                         }
 799
 800                         /* only if matching snap context */
 801                         pgsnapc = page_snap_context(page);
 802                         if (pgsnapc->seq > snapc->seq) {
 803                                 dout("page snapc %p %lld > oldest %p %lld\n",
 804                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 805                                 unlock_page(page);
 806                                 if (!locked_pages)
 807                                         continue; /* keep looking for snap */
 808                                 break;
 809                         }
 810
 811                         if (!clear_page_dirty_for_io(page)) {
 812                                 dout("%p !clear_page_dirty_for_io\n", page);
 813                                 unlock_page(page);
 814                                 break;
 815                         }
 816
 817                         /* ok */
 818                         if (locked_pages == 0) {
 819                                 /* prepare async write request */
 820                                 offset = (unsigned long long)page->index
 821                                         << PAGE_CACHE_SHIFT;
 822                                 len = wsize;
 823                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 824                                             &ci->i_layout,
 825                                             ceph_vino(inode),
 826                                             offset, &len,
 827                                             CEPH_OSD_OP_WRITE,
 828                                             CEPH_OSD_FLAG_WRITE |
 829                                                     CEPH_OSD_FLAG_ONDISK,
 830                                             snapc, do_sync,
 831                                             ci->i_truncate_seq,
 832                                             ci->i_truncate_size,
 833                                             &inode->i_mtime, true, 1, 0);
 834
 835                                 if (!req) {
 836                                         rc = -ENOMEM;
 837                                         unlock_page(page);
 838                                         break;
 839                                 }
 840
 841                                 max_pages = req->r_num_pages;
 842
 843                                 alloc_page_vec(fsc, req);
 844                                 req->r_callback = writepages_finish;
 845                                 req->r_inode = inode;
 846                         }
 847
 848                         /* note position of first page in pvec */
 849                         if (first < 0)
 850                                 first = i;
 851                         dout("%p will write page %p idx %lu\n",
 852                              inode, page, page->index);
 853
 854                         writeback_stat =
 855                                atomic_long_inc_return(&fsc->writeback_count);
 856                         if (writeback_stat > CONGESTION_ON_THRESH(
 857                                     fsc->mount_options->congestion_kb)) {
 858                                 set_bdi_congested(&fsc->backing_dev_info,
 859                                                   BLK_RW_ASYNC);
 860                         }
 861
 862                         set_page_writeback(page);
 863                         req->r_pages[locked_pages] = page;
 864                         locked_pages++;
 865                         next = page->index + 1;
 866                 }
 867
 868                 /* did we get anything? */
 869                 if (!locked_pages)
 870                         goto release_pvec_pages;
 871                 if (i) {
 872                         int j;
 873                         BUG_ON(!locked_pages || first < 0);
 874
 875                         if (pvec_pages && i == pvec_pages &&
 876                             locked_pages < max_pages) {
 877                                 dout("reached end pvec, trying for more\n");
 878                                 pagevec_reinit(&pvec);
 879                                 goto get_more_pages;
 880                         }
 881
 882                         /* shift unused pages over in the pvec...  we
 883                          * will need to release them below. */
 884                         for (j = i; j < pvec_pages; j++) {
 885                                 dout(" pvec leftover page %p\n",
 886                                      pvec.pages[j]);
 887                                 pvec.pages[j-i+first] = pvec.pages[j];
 888                         }
 889                         pvec.nr -= i-first;
 890                 }
 891
 892                 /* submit the write */
 893                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 894                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 895                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 896                 dout("writepages got %d pages at %llu~%llu\n",
 897                      locked_pages, offset, len);
 898
 899                 /* revise final length, page count */
 900                 req->r_num_pages = locked_pages;
 901                 reqhead = req->r_request->front.iov_base;
 902                 op = (void *)(reqhead + 1);
 903                 op->extent.length = cpu_to_le64(len);
 904                 op->payload_len = cpu_to_le32(len);
 905                 req->r_request->hdr.data_len = cpu_to_le32(len);
 906
 907                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 908                 BUG_ON(rc);
 909                 req = NULL;
 910
 911                 /* continue? */
 912                 index = next;
 913                 wbc->nr_to_write -= locked_pages;
 914                 if (wbc->nr_to_write <= 0)
 915                         done = 1;
 916
 917 release_pvec_pages:
 918                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 919                      pvec.nr ? pvec.pages[0] : NULL);
 920                 pagevec_release(&pvec);
 921
 922                 if (locked_pages && !done)
 923                         goto retry;
 924         }
 925
 926         if (should_loop && !done) {
 927                 /* more to do; loop back to beginning of file */
 928                 dout("writepages looping back to beginning of file\n");
 929                 should_loop = 0;
 930                 index = 0;
 931                 goto retry;
 932         }
 933
 934         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 935                 mapping->writeback_index = index;
 936
 937 out:
 938         if (req)
 939                 ceph_osdc_put_request(req);
 940         ceph_put_snap_context(snapc);
 941         dout("writepages done, rc = %d\n", rc);
 942         return rc;
 943 }
 944
 945
 946
 947 /*
 948  * See if a given @snapc is either writeable, or already written.
 949  */
 950 static int context_is_writeable_or_written(struct inode *inode,
 951                                            struct ceph_snap_context *snapc)
 952 {
 953         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 954         int ret = !oldest || snapc->seq <= oldest->seq;
 955
 956         ceph_put_snap_context(oldest);
 957         return ret;
 958 }
 959
 960 /*
 961  * We are only allowed to write into/dirty the page if the page is
 962  * clean, or already dirty within the same snap context.
 963  *
 964  * called with page locked.
 965  * return success with page locked,
 966  * or any failure (incl -EAGAIN) with page unlocked.
 967  */
 968 static int ceph_update_writeable_page(struct file *file,
 969                             loff_t pos, unsigned len,
 970                             struct page *page)
 971 {
 972         struct inode *inode = file->f_dentry->d_inode;
 973         struct ceph_inode_info *ci = ceph_inode(inode);
 974         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 975         loff_t page_off = pos & PAGE_CACHE_MASK;
 976         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 977         int end_in_page = pos_in_page + len;
 978         loff_t i_size;
 979         int r;
 980         struct ceph_snap_context *snapc, *oldest;
 981
 982 retry_locked:
 983         /* writepages currently holds page lock, but if we change that later, */
 984         wait_on_page_writeback(page);
 985
 986         /* check snap context */
 987         BUG_ON(!ci->i_snap_realm);
 988         down_read(&mdsc->snap_rwsem);
 989         BUG_ON(!ci->i_snap_realm->cached_context);
 990         snapc = page_snap_context(page);
 991         if (snapc && snapc != ci->i_head_snapc) {
 992                 /*
 993                  * this page is already dirty in another (older) snap
 994                  * context!  is it writeable now?
 995                  */
 996                 oldest = get_oldest_context(inode, NULL);
 997                 up_read(&mdsc->snap_rwsem);
 998
 999                 if (snapc->seq > oldest->seq) {
1000                         ceph_put_snap_context(oldest);
1001                         dout(" page %p snapc %p not current or oldest\n",
1002                              page, snapc);
1003                         /*
1004                          * queue for writeback, and wait for snapc to
1005                          * be writeable or written
1006                          */
1007                         snapc = ceph_get_snap_context(snapc);
1008                         unlock_page(page);
1009                         ceph_queue_writeback(inode);
1010                         r = wait_event_interruptible(ci->i_cap_wq,
1011                                context_is_writeable_or_written(inode, snapc));
1012                         ceph_put_snap_context(snapc);
1013                         if (r == -ERESTARTSYS)
1014                                 return r;
1015                         return -EAGAIN;
1016                 }
1017                 ceph_put_snap_context(oldest);
1018
1019                 /* yay, writeable, do it now (without dropping page lock) */
1020                 dout(" page %p snapc %p not current, but oldest\n",
1021                      page, snapc);
1022                 if (!clear_page_dirty_for_io(page))
1023                         goto retry_locked;
1024                 r = writepage_nounlock(page, NULL);
1025                 if (r < 0)
1026                         goto fail_nosnap;
1027                 goto retry_locked;
1028         }
1029
1030         if (PageUptodate(page)) {
1031                 dout(" page %p already uptodate\n", page);
1032                 return 0;
1033         }
1034
1035         /* full page? */
1036         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1037                 return 0;
1038
1039         /* past end of file? */
1040         i_size = inode->i_size;   /* caller holds i_mutex */
1041
1042         if (i_size + len > inode->i_sb->s_maxbytes) {
1043                 /* file is too big */
1044                 r = -EINVAL;
1045                 goto fail;
1046         }
1047
1048         if (page_off >= i_size ||
1049             (pos_in_page == 0 && (pos+len) >= i_size &&
1050              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1051                 dout(" zeroing %p 0 - %d and %d - %d\n",
1052                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1053                 zero_user_segments(page,
1054                                    0, pos_in_page,
1055                                    end_in_page, PAGE_CACHE_SIZE);
1056                 return 0;
1057         }
1058
1059         /* we need to read it. */
1060         up_read(&mdsc->snap_rwsem);
1061         r = readpage_nounlock(file, page);
1062         if (r < 0)
1063                 goto fail_nosnap;
1064         goto retry_locked;
1065
1066 fail:
1067         up_read(&mdsc->snap_rwsem);
1068 fail_nosnap:
1069         unlock_page(page);
1070         return r;
1071 }
1072
1073 /*
1074  * We are only allowed to write into/dirty the page if the page is
1075  * clean, or already dirty within the same snap context.
1076  */
1077 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1078                             loff_t pos, unsigned len, unsigned flags,
1079                             struct page **pagep, void **fsdata)
1080 {
1081         struct inode *inode = file->f_dentry->d_inode;
1082         struct page *page;
1083         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1084         int r;
1085
1086         do {
1087                 /* get a page */
1088                 page = grab_cache_page_write_begin(mapping, index, 0);
1089                 if (!page)
1090                         return -ENOMEM;
1091                 *pagep = page;
1092
1093                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1094                      inode, page, (int)pos, (int)len);
1095
1096                 r = ceph_update_writeable_page(file, pos, len, page);
1097         } while (r == -EAGAIN);
1098
1099         return r;
1100 }
1101
1102 /*
1103  * we don't do anything in here that simple_write_end doesn't do
1104  * except adjust dirty page accounting and drop read lock on
1105  * mdsc->snap_rwsem.
1106  */
1107 static int ceph_write_end(struct file *file, struct address_space *mapping,
1108                           loff_t pos, unsigned len, unsigned copied,
1109                           struct page *page, void *fsdata)
1110 {
1111         struct inode *inode = file->f_dentry->d_inode;
1112         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1113         struct ceph_mds_client *mdsc = fsc->mdsc;
1114         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1115         int check_cap = 0;
1116
1117         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1118              inode, page, (int)pos, (int)copied, (int)len);
1119
1120         /* zero the stale part of the page if we did a short copy */
1121         if (copied < len)
1122                 zero_user_segment(page, from+copied, len);
1123
1124         /* did file size increase? */
1125         /* (no need for i_size_read(); we caller holds i_mutex */
1126         if (pos+copied > inode->i_size)
1127                 check_cap = ceph_inode_set_size(inode, pos+copied);
1128
1129         if (!PageUptodate(page))
1130                 SetPageUptodate(page);
1131
1132         set_page_dirty(page);
1133
1134         unlock_page(page);
1135         up_read(&mdsc->snap_rwsem);
1136         page_cache_release(page);
1137
1138         if (check_cap)
1139                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1140
1141         return copied;
1142 }
1143
1144 /*
1145  * we set .direct_IO to indicate direct io is supported, but since we
1146  * intercept O_DIRECT reads and writes early, this function should
1147  * never get called.
1148  */
1149 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1150                               const struct iovec *iov,
1151                               loff_t pos, unsigned long nr_segs)
1152 {
1153         WARN_ON(1);
1154         return -EINVAL;
1155 }
1156
1157 const struct address_space_operations ceph_aops = {
1158         .readpage = ceph_readpage,
1159         .readpages = ceph_readpages,
1160         .writepage = ceph_writepage,
1161         .writepages = ceph_writepages_start,
1162         .write_begin = ceph_write_begin,
1163         .write_end = ceph_write_end,
1164         .set_page_dirty = ceph_set_page_dirty,
1165         .invalidatepage = ceph_invalidatepage,
1166         .releasepage = ceph_releasepage,
1167         .direct_IO = ceph_direct_io,
1168 };
1169
1170
1171 /*
1172  * vm ops
1173  */
1174
1175 /*
1176  * Reuse write_begin here for simplicity.
1177  */
1178 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1179 {
1180         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1181         struct page *page = vmf->page;
1182         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1183         loff_t off = page->index << PAGE_CACHE_SHIFT;
1184         loff_t size, len;
1185         int ret;
1186
1187         /* Update time before taking page lock */
1188         file_update_time(vma->vm_file);
1189
1190         size = i_size_read(inode);
1191         if (off + PAGE_CACHE_SIZE <= size)
1192                 len = PAGE_CACHE_SIZE;
1193         else
1194                 len = size & ~PAGE_CACHE_MASK;
1195
1196         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1197              off, len, page, page->index);
1198
1199         lock_page(page);
1200
1201         ret = VM_FAULT_NOPAGE;
1202         if ((off > size) ||
1203             (page->mapping != inode->i_mapping))
1204                 goto out;
1205
1206         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1207         if (ret == 0) {
1208                 /* success.  we'll keep the page locked. */
1209                 set_page_dirty(page);
1210                 up_read(&mdsc->snap_rwsem);
1211                 ret = VM_FAULT_LOCKED;
1212         } else {
1213                 if (ret == -ENOMEM)
1214                         ret = VM_FAULT_OOM;
1215                 else
1216                         ret = VM_FAULT_SIGBUS;
1217         }
1218 out:
1219         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1220         if (ret != VM_FAULT_LOCKED)
1221                 unlock_page(page);
1222         return ret;
1223 }
1224
1225 static struct vm_operations_struct ceph_vmops = {
1226         .fault          = filemap_fault,
1227         .page_mkwrite   = ceph_page_mkwrite,
1228 };
1229
1230 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1231 {
1232         struct address_space *mapping = file->f_mapping;
1233
1234         if (!mapping->a_ops->readpage)
1235                 return -ENOEXEC;
1236         file_accessed(file);
1237         vma->vm_ops = &ceph_vmops;
1238         vma->vm_flags |= VM_CAN_NONLINEAR;
1239         return 0;
1240 }