fs/ceph/addr.c

   1 #include "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/pagevec.h>
   9 #include <linux/task_io_accounting_ops.h>
  10
  11 #include "super.h"
  12 #include "osd_client.h"
  13
  14 /*
  15  * Ceph address space ops.
  16  *
  17  * There are a few funny things going on here.
  18  *
  19  * The page->private field is used to reference a struct
  20  * ceph_snap_context for _every_ dirty page.  This indicates which
  21  * snapshot the page was logically dirtied in, and thus which snap
  22  * context needs to be associated with the osd write during writeback.
  23  *
  24  * Similarly, struct ceph_inode_info maintains a set of counters to
  25  * count dirty pages on the inode.  In the absense of snapshots,
  26  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  27  *
  28  * When a snapshot is taken (that is, when the client receives
  29  * notification that a snapshot was taken), each inode with caps and
  30  * with dirty pages (dirty pages implies there is a cap) gets a new
  31  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  32  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  33  * moved to capsnap->dirty. (Unless a sync write is currently in
  34  * progress.  In that case, the capsnap is said to be "pending", new
  35  * writes cannot start, and the capsnap isn't "finalized" until the
  36  * write completes (or fails) and a final size/mtime for the inode for
  37  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  38  *
  39  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  40  * we look for the first capsnap in i_cap_snaps and write out pages in
  41  * that snap context _only_.  Then we move on to the next capsnap,
  42  * eventually reaching the "live" or "head" context (i.e., pages that
  43  * are not yet snapped) and are writing the most recently dirtied
  44  * pages.
  45  *
  46  * Invalidate and so forth must take care to ensure the dirty page
  47  * accounting is preserved.
  48  */
  49
  50 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  51 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  52         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  53          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  54
  55
  56
  57 /*
  58  * Dirty a page.  Optimistically adjust accounting, on the assumption
  59  * that we won't race with invalidate.  If we do, readjust.
  60  */
  61 static int ceph_set_page_dirty(struct page *page)
  62 {
  63         struct address_space *mapping = page->mapping;
  64         struct inode *inode;
  65         struct ceph_inode_info *ci;
  66         int undo = 0;
  67         struct ceph_snap_context *snapc;
  68
  69         if (unlikely(!mapping))
  70                 return !TestSetPageDirty(page);
  71
  72         if (TestSetPageDirty(page)) {
  73                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  74                      mapping->host, page, page->index);
  75                 return 0;
  76         }
  77
  78         inode = mapping->host;
  79         ci = ceph_inode(inode);
  80
  81         /*
  82          * Note that we're grabbing a snapc ref here without holding
  83          * any locks!
  84          */
  85         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  86
  87         /* dirty the head */
  88         spin_lock(&inode->i_lock);
  89         if (ci->i_wrbuffer_ref_head == 0)
  90                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  91         ++ci->i_wrbuffer_ref_head;
  92         if (ci->i_wrbuffer_ref == 0)
  93                 igrab(inode);
  94         ++ci->i_wrbuffer_ref;
  95         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
  96              "snapc %p seq %lld (%d snaps)\n",
  97              mapping->host, page, page->index,
  98              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
  99              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 100              snapc, snapc->seq, snapc->num_snaps);
 101         spin_unlock(&inode->i_lock);
 102
 103         /* now adjust page */
 104         spin_lock_irq(&mapping->tree_lock);
 105         if (page->mapping) {    /* Race with truncate? */
 106                 WARN_ON_ONCE(!PageUptodate(page));
 107
 108                 if (mapping_cap_account_dirty(mapping)) {
 109                         __inc_zone_page_state(page, NR_FILE_DIRTY);
 110                         __inc_bdi_stat(mapping->backing_dev_info,
 111                                         BDI_RECLAIMABLE);
 112                         task_io_account_write(PAGE_CACHE_SIZE);
 113                 }
 114                 radix_tree_tag_set(&mapping->page_tree,
 115                                 page_index(page), PAGECACHE_TAG_DIRTY);
 116
 117                 /*
 118                  * Reference snap context in page->private.  Also set
 119                  * PagePrivate so that we get invalidatepage callback.
 120                  */
 121                 page->private = (unsigned long)snapc;
 122                 SetPagePrivate(page);
 123         } else {
 124                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 125                 undo = 1;
 126         }
 127
 128         spin_unlock_irq(&mapping->tree_lock);
 129
 130         if (undo)
 131                 /* whoops, we failed to dirty the page */
 132                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 133
 134         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 135
 136         BUG_ON(!PageDirty(page));
 137         return 1;
 138 }
 139
 140 /*
 141  * If we are truncating the full page (i.e. offset == 0), adjust the
 142  * dirty page counters appropriately.  Only called if there is private
 143  * data on the page.
 144  */
 145 static void ceph_invalidatepage(struct page *page, unsigned long offset)
 146 {
 147         struct inode *inode;
 148         struct ceph_inode_info *ci;
 149         struct ceph_snap_context *snapc = (void *)page->private;
 150
 151         BUG_ON(!PageLocked(page));
 152         BUG_ON(!page->private);
 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(page->private);
 190         WARN_ON(PagePrivate(page));
 191         return 0;
 192 }
 193
 194 /*
 195  * read a single page, without unlocking it.
 196  */
 197 static int readpage_nounlock(struct file *filp, struct page *page)
 198 {
 199         struct inode *inode = filp->f_dentry->d_inode;
 200         struct ceph_inode_info *ci = ceph_inode(inode);
 201         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->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);
 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  * Build a vector of contiguous pages from the provided page list.
 235  */
 236 static struct page **page_vector_from_list(struct list_head *page_list,
 237                                            unsigned *nr_pages)
 238 {
 239         struct page **pages;
 240         struct page *page;
 241         int next_index, contig_pages = 0;
 242
 243         /* build page vector */
 244         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
 245         if (!pages)
 246                 return ERR_PTR(-ENOMEM);
 247
 248         BUG_ON(list_empty(page_list));
 249         next_index = list_entry(page_list->prev, struct page, lru)->index;
 250         list_for_each_entry_reverse(page, page_list, lru) {
 251                 if (page->index == next_index) {
 252                         dout("readpages page %d %p\n", contig_pages, page);
 253                         pages[contig_pages] = page;
 254                         contig_pages++;
 255                         next_index++;
 256                 } else {
 257                         break;
 258                 }
 259         }
 260         *nr_pages = contig_pages;
 261         return pages;
 262 }
 263
 264 /*
 265  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 266  * the caller (VM) cleans them up.
 267  */
 268 static int ceph_readpages(struct file *file, struct address_space *mapping,
 269                           struct list_head *page_list, unsigned nr_pages)
 270 {
 271         struct inode *inode = file->f_dentry->d_inode;
 272         struct ceph_inode_info *ci = ceph_inode(inode);
 273         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
 274         int rc = 0;
 275         struct page **pages;
 276         struct pagevec pvec;
 277         loff_t offset;
 278         u64 len;
 279
 280         dout("readpages %p file %p nr_pages %d\n",
 281              inode, file, nr_pages);
 282
 283         pages = page_vector_from_list(page_list, &nr_pages);
 284         if (IS_ERR(pages))
 285                 return PTR_ERR(pages);
 286
 287         /* guess read extent */
 288         offset = pages[0]->index << PAGE_CACHE_SHIFT;
 289         len = nr_pages << PAGE_CACHE_SHIFT;
 290         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 291                                  offset, &len,
 292                                  ci->i_truncate_seq, ci->i_truncate_size,
 293                                  pages, nr_pages);
 294         if (rc == -ENOENT)
 295                 rc = 0;
 296         if (rc < 0)
 297                 goto out;
 298
 299         /* set uptodate and add to lru in pagevec-sized chunks */
 300         pagevec_init(&pvec, 0);
 301         for (; !list_empty(page_list) && len > 0;
 302              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
 303                 struct page *page =
 304                         list_entry(page_list->prev, struct page, lru);
 305
 306                 list_del(&page->lru);
 307
 308                 if (rc < (int)PAGE_CACHE_SIZE) {
 309                         /* zero (remainder of) page */
 310                         int s = rc < 0 ? 0 : rc;
 311                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 312                 }
 313
 314                 if (add_to_page_cache(page, mapping, page->index, GFP_NOFS)) {
 315                         page_cache_release(page);
 316                         dout("readpages %p add_to_page_cache failed %p\n",
 317                              inode, page);
 318                         continue;
 319                 }
 320                 dout("readpages %p adding %p idx %lu\n", inode, page,
 321                      page->index);
 322                 flush_dcache_page(page);
 323                 SetPageUptodate(page);
 324                 unlock_page(page);
 325                 if (pagevec_add(&pvec, page) == 0)
 326                         pagevec_lru_add_file(&pvec);   /* add to lru */
 327         }
 328         pagevec_lru_add_file(&pvec);
 329         rc = 0;
 330
 331 out:
 332         kfree(pages);
 333         return rc;
 334 }
 335
 336 /*
 337  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 338  * only snap context we are allowed to write back.
 339  *
 340  * Caller holds i_lock.
 341  */
 342 static struct ceph_snap_context *__get_oldest_context(struct inode *inode,
 343                                                       u64 *snap_size)
 344 {
 345         struct ceph_inode_info *ci = ceph_inode(inode);
 346         struct ceph_snap_context *snapc = NULL;
 347         struct ceph_cap_snap *capsnap = NULL;
 348
 349         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 350                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 351                      capsnap->context, capsnap->dirty_pages);
 352                 if (capsnap->dirty_pages) {
 353                         snapc = ceph_get_snap_context(capsnap->context);
 354                         if (snap_size)
 355                                 *snap_size = capsnap->size;
 356                         break;
 357                 }
 358         }
 359         if (!snapc && ci->i_snap_realm) {
 360                 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
 361                 dout(" head snapc %p has %d dirty pages\n",
 362                      snapc, ci->i_wrbuffer_ref_head);
 363         }
 364         return snapc;
 365 }
 366
 367 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 368                                                     u64 *snap_size)
 369 {
 370         struct ceph_snap_context *snapc = NULL;
 371
 372         spin_lock(&inode->i_lock);
 373         snapc = __get_oldest_context(inode, snap_size);
 374         spin_unlock(&inode->i_lock);
 375         return snapc;
 376 }
 377
 378 /*
 379  * Write a single page, but leave the page locked.
 380  *
 381  * If we get a write error, set the page error bit, but still adjust the
 382  * dirty page accounting (i.e., page is no longer dirty).
 383  */
 384 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 385 {
 386         struct inode *inode;
 387         struct ceph_inode_info *ci;
 388         struct ceph_client *client;
 389         struct ceph_osd_client *osdc;
 390         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
 391         int len = PAGE_CACHE_SIZE;
 392         loff_t i_size;
 393         int err = 0;
 394         struct ceph_snap_context *snapc;
 395         u64 snap_size = 0;
 396         long writeback_stat;
 397
 398         dout("writepage %p idx %lu\n", page, page->index);
 399
 400         if (!page->mapping || !page->mapping->host) {
 401                 dout("writepage %p - no mapping\n", page);
 402                 return -EFAULT;
 403         }
 404         inode = page->mapping->host;
 405         ci = ceph_inode(inode);
 406         client = ceph_inode_to_client(inode);
 407         osdc = &client->osdc;
 408
 409         /* verify this is a writeable snap context */
 410         snapc = (void *)page->private;
 411         if (snapc == NULL) {
 412                 dout("writepage %p page %p not dirty?\n", inode, page);
 413                 goto out;
 414         }
 415         if (snapc != get_oldest_context(inode, &snap_size)) {
 416                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 417                      inode, page, (void *)page->private);
 418                 /* we should only noop if called by kswapd */
 419                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 420                 goto out;
 421         }
 422
 423         /* is this a partial page at end of file? */
 424         if (snap_size)
 425                 i_size = snap_size;
 426         else
 427                 i_size = i_size_read(inode);
 428         if (i_size < page_off + len)
 429                 len = i_size - page_off;
 430
 431         dout("writepage %p page %p index %lu on %llu~%u\n",
 432              inode, page, page->index, page_off, len);
 433
 434         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 435         if (writeback_stat >
 436             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
 437                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 438
 439         set_page_writeback(page);
 440         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 441                                    &ci->i_layout, snapc,
 442                                    page_off, len,
 443                                    ci->i_truncate_seq, ci->i_truncate_size,
 444                                    &inode->i_mtime,
 445                                    &page, 1, 0, 0, true);
 446         if (err < 0) {
 447                 dout("writepage setting page/mapping error %d %p\n", err, page);
 448                 SetPageError(page);
 449                 mapping_set_error(&inode->i_data, err);
 450                 if (wbc)
 451                         wbc->pages_skipped++;
 452         } else {
 453                 dout("writepage cleaned page %p\n", page);
 454                 err = 0;  /* vfs expects us to return 0 */
 455         }
 456         page->private = 0;
 457         ClearPagePrivate(page);
 458         end_page_writeback(page);
 459         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 460         ceph_put_snap_context(snapc);
 461 out:
 462         return err;
 463 }
 464
 465 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 466 {
 467         int err;
 468         struct inode *inode = page->mapping->host;
 469         BUG_ON(!inode);
 470         igrab(inode);
 471         err = writepage_nounlock(page, wbc);
 472         unlock_page(page);
 473         iput(inode);
 474         return err;
 475 }
 476
 477
 478 /*
 479  * lame release_pages helper.  release_pages() isn't exported to
 480  * modules.
 481  */
 482 static void ceph_release_pages(struct page **pages, int num)
 483 {
 484         struct pagevec pvec;
 485         int i;
 486
 487         pagevec_init(&pvec, 0);
 488         for (i = 0; i < num; i++) {
 489                 if (pagevec_add(&pvec, pages[i]) == 0)
 490                         pagevec_release(&pvec);
 491         }
 492         pagevec_release(&pvec);
 493 }
 494
 495
 496 /*
 497  * async writeback completion handler.
 498  *
 499  * If we get an error, set the mapping error bit, but not the individual
 500  * page error bits.
 501  */
 502 static void writepages_finish(struct ceph_osd_request *req,
 503                               struct ceph_msg *msg)
 504 {
 505         struct inode *inode = req->r_inode;
 506         struct ceph_osd_reply_head *replyhead;
 507         struct ceph_osd_op *op;
 508         struct ceph_inode_info *ci = ceph_inode(inode);
 509         unsigned wrote;
 510         struct page *page;
 511         int i;
 512         struct ceph_snap_context *snapc = req->r_snapc;
 513         struct address_space *mapping = inode->i_mapping;
 514         struct writeback_control *wbc = req->r_wbc;
 515         __s32 rc = -EIO;
 516         u64 bytes = 0;
 517         struct ceph_client *client = ceph_inode_to_client(inode);
 518         long writeback_stat;
 519         unsigned issued = __ceph_caps_issued(ci, NULL);
 520
 521         /* parse reply */
 522         replyhead = msg->front.iov_base;
 523         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 524         op = (void *)(replyhead + 1);
 525         rc = le32_to_cpu(replyhead->result);
 526         bytes = le64_to_cpu(op->extent.length);
 527
 528         if (rc >= 0) {
 529                 /*
 530                  * Assume we wrote the pages we originally sent.  The
 531                  * osd might reply with fewer pages if our writeback
 532                  * raced with a truncation and was adjusted at the osd,
 533                  * so don't believe the reply.
 534                  */
 535                 wrote = req->r_num_pages;
 536         } else {
 537                 wrote = 0;
 538                 mapping_set_error(mapping, rc);
 539         }
 540         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 541              inode, rc, bytes, wrote);
 542
 543         /* clean all pages */
 544         for (i = 0; i < req->r_num_pages; i++) {
 545                 page = req->r_pages[i];
 546                 BUG_ON(!page);
 547                 WARN_ON(!PageUptodate(page));
 548
 549                 writeback_stat =
 550                         atomic_long_dec_return(&client->writeback_count);
 551                 if (writeback_stat <
 552                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
 553                         clear_bdi_congested(&client->backing_dev_info,
 554                                             BLK_RW_ASYNC);
 555
 556                 if (i >= wrote) {
 557                         dout("inode %p skipping page %p\n", inode, page);
 558                         wbc->pages_skipped++;
 559                 }
 560                 page->private = 0;
 561                 ClearPagePrivate(page);
 562                 ceph_put_snap_context(snapc);
 563                 dout("unlocking %d %p\n", i, page);
 564                 end_page_writeback(page);
 565
 566                 /*
 567                  * We lost the cache cap, need to truncate the page before
 568                  * it is unlocked, otherwise we'd truncate it later in the
 569                  * page truncation thread, possibly losing some data that
 570                  * raced its way in
 571                  */
 572                 if ((issued & CEPH_CAP_FILE_CACHE) == 0)
 573                         generic_error_remove_page(inode->i_mapping, page);
 574
 575                 unlock_page(page);
 576         }
 577         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 578         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 579
 580         ceph_release_pages(req->r_pages, req->r_num_pages);
 581         if (req->r_pages_from_pool)
 582                 mempool_free(req->r_pages,
 583                              ceph_client(inode->i_sb)->wb_pagevec_pool);
 584         else
 585                 kfree(req->r_pages);
 586         ceph_osdc_put_request(req);
 587 }
 588
 589 /*
 590  * allocate a page vec, either directly, or if necessary, via a the
 591  * mempool.  we avoid the mempool if we can because req->r_num_pages
 592  * may be less than the maximum write size.
 593  */
 594 static void alloc_page_vec(struct ceph_client *client,
 595                            struct ceph_osd_request *req)
 596 {
 597         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 598                                GFP_NOFS);
 599         if (!req->r_pages) {
 600                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
 601                 req->r_pages_from_pool = 1;
 602                 WARN_ON(!req->r_pages);
 603         }
 604 }
 605
 606 /*
 607  * initiate async writeback
 608  */
 609 static int ceph_writepages_start(struct address_space *mapping,
 610                                  struct writeback_control *wbc)
 611 {
 612         struct inode *inode = mapping->host;
 613         struct backing_dev_info *bdi = mapping->backing_dev_info;
 614         struct ceph_inode_info *ci = ceph_inode(inode);
 615         struct ceph_client *client;
 616         pgoff_t index, start, end;
 617         int range_whole = 0;
 618         int should_loop = 1;
 619         pgoff_t max_pages = 0, max_pages_ever = 0;
 620         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL;
 621         struct pagevec pvec;
 622         int done = 0;
 623         int rc = 0;
 624         unsigned wsize = 1 << inode->i_blkbits;
 625         struct ceph_osd_request *req = NULL;
 626         int do_sync;
 627         u64 snap_size = 0;
 628
 629         /*
 630          * Include a 'sync' in the OSD request if this is a data
 631          * integrity write (e.g., O_SYNC write or fsync()), or if our
 632          * cap is being revoked.
 633          */
 634         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 635         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 636                 do_sync = 1;
 637         dout("writepages_start %p dosync=%d (mode=%s)\n",
 638              inode, do_sync,
 639              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 640              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 641
 642         client = ceph_inode_to_client(inode);
 643         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
 644                 pr_warning("writepage_start %p on forced umount\n", inode);
 645                 return -EIO; /* we're in a forced umount, don't write! */
 646         }
 647         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
 648                 wsize = client->mount_args->wsize;
 649         if (wsize < PAGE_CACHE_SIZE)
 650                 wsize = PAGE_CACHE_SIZE;
 651         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 652
 653         pagevec_init(&pvec, 0);
 654
 655         /* ?? */
 656         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 657                 dout(" writepages congested\n");
 658                 wbc->encountered_congestion = 1;
 659                 goto out_final;
 660         }
 661
 662         /* where to start/end? */
 663         if (wbc->range_cyclic) {
 664                 start = mapping->writeback_index; /* Start from prev offset */
 665                 end = -1;
 666                 dout(" cyclic, start at %lu\n", start);
 667         } else {
 668                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 669                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 670                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 671                         range_whole = 1;
 672                 should_loop = 0;
 673                 dout(" not cyclic, %lu to %lu\n", start, end);
 674         }
 675         index = start;
 676
 677 retry:
 678         /* find oldest snap context with dirty data */
 679         ceph_put_snap_context(snapc);
 680         snapc = get_oldest_context(inode, &snap_size);
 681         if (!snapc) {
 682                 /* hmm, why does writepages get called when there
 683                    is no dirty data? */
 684                 dout(" no snap context with dirty data?\n");
 685                 goto out;
 686         }
 687         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 688              snapc, snapc->seq, snapc->num_snaps);
 689         if (last_snapc && snapc != last_snapc) {
 690                 /* if we switched to a newer snapc, restart our scan at the
 691                  * start of the original file range. */
 692                 dout("  snapc differs from last pass, restarting at %lu\n",
 693                      index);
 694                 index = start;
 695         }
 696         last_snapc = snapc;
 697
 698         while (!done && index <= end) {
 699                 unsigned i;
 700                 int first;
 701                 pgoff_t next;
 702                 int pvec_pages, locked_pages;
 703                 struct page *page;
 704                 int want;
 705                 u64 offset, len;
 706                 struct ceph_osd_request_head *reqhead;
 707                 struct ceph_osd_op *op;
 708                 long writeback_stat;
 709
 710                 next = 0;
 711                 locked_pages = 0;
 712                 max_pages = max_pages_ever;
 713
 714 get_more_pages:
 715                 first = -1;
 716                 want = min(end - index,
 717                            min((pgoff_t)PAGEVEC_SIZE,
 718                                max_pages - (pgoff_t)locked_pages) - 1)
 719                         + 1;
 720                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 721                                                 PAGECACHE_TAG_DIRTY,
 722                                                 want);
 723                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 724                 if (!pvec_pages && !locked_pages)
 725                         break;
 726                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 727                         page = pvec.pages[i];
 728                         dout("? %p idx %lu\n", page, page->index);
 729                         if (locked_pages == 0)
 730                                 lock_page(page);  /* first page */
 731                         else if (!trylock_page(page))
 732                                 break;
 733
 734                         /* only dirty pages, or our accounting breaks */
 735                         if (unlikely(!PageDirty(page)) ||
 736                             unlikely(page->mapping != mapping)) {
 737                                 dout("!dirty or !mapping %p\n", page);
 738                                 unlock_page(page);
 739                                 break;
 740                         }
 741                         if (!wbc->range_cyclic && page->index > end) {
 742                                 dout("end of range %p\n", page);
 743                                 done = 1;
 744                                 unlock_page(page);
 745                                 break;
 746                         }
 747                         if (next && (page->index != next)) {
 748                                 dout("not consecutive %p\n", page);
 749                                 unlock_page(page);
 750                                 break;
 751                         }
 752                         if (wbc->sync_mode != WB_SYNC_NONE) {
 753                                 dout("waiting on writeback %p\n", page);
 754                                 wait_on_page_writeback(page);
 755                         }
 756                         if ((snap_size && page_offset(page) > snap_size) ||
 757                             (!snap_size &&
 758                              page_offset(page) > i_size_read(inode))) {
 759                                 dout("%p page eof %llu\n", page, snap_size ?
 760                                      snap_size : i_size_read(inode));
 761                                 done = 1;
 762                                 unlock_page(page);
 763                                 break;
 764                         }
 765                         if (PageWriteback(page)) {
 766                                 dout("%p under writeback\n", page);
 767                                 unlock_page(page);
 768                                 break;
 769                         }
 770
 771                         /* only if matching snap context */
 772                         if (snapc != (void *)page->private) {
 773                                 dout("page snapc %p != oldest %p\n",
 774                                      (void *)page->private, snapc);
 775                                 unlock_page(page);
 776                                 if (!locked_pages)
 777                                         continue; /* keep looking for snap */
 778                                 break;
 779                         }
 780
 781                         if (!clear_page_dirty_for_io(page)) {
 782                                 dout("%p !clear_page_dirty_for_io\n", page);
 783                                 unlock_page(page);
 784                                 break;
 785                         }
 786
 787                         /* ok */
 788                         if (locked_pages == 0) {
 789                                 /* prepare async write request */
 790                                 offset = page->index << PAGE_CACHE_SHIFT;
 791                                 len = wsize;
 792                                 req = ceph_osdc_new_request(&client->osdc,
 793                                             &ci->i_layout,
 794                                             ceph_vino(inode),
 795                                             offset, &len,
 796                                             CEPH_OSD_OP_WRITE,
 797                                             CEPH_OSD_FLAG_WRITE |
 798                                                     CEPH_OSD_FLAG_ONDISK,
 799                                             snapc, do_sync,
 800                                             ci->i_truncate_seq,
 801                                             ci->i_truncate_size,
 802                                             &inode->i_mtime, true, 1);
 803                                 max_pages = req->r_num_pages;
 804
 805                                 alloc_page_vec(client, req);
 806                                 req->r_callback = writepages_finish;
 807                                 req->r_inode = inode;
 808                                 req->r_wbc = wbc;
 809                         }
 810
 811                         /* note position of first page in pvec */
 812                         if (first < 0)
 813                                 first = i;
 814                         dout("%p will write page %p idx %lu\n",
 815                              inode, page, page->index);
 816
 817                         writeback_stat = atomic_long_inc_return(&client->writeback_count);
 818                         if (writeback_stat > CONGESTION_ON_THRESH(client->mount_args->congestion_kb)) {
 819                                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
 820                         }
 821
 822                         set_page_writeback(page);
 823                         req->r_pages[locked_pages] = page;
 824                         locked_pages++;
 825                         next = page->index + 1;
 826                 }
 827
 828                 /* did we get anything? */
 829                 if (!locked_pages)
 830                         goto release_pvec_pages;
 831                 if (i) {
 832                         int j;
 833                         BUG_ON(!locked_pages || first < 0);
 834
 835                         if (pvec_pages && i == pvec_pages &&
 836                             locked_pages < max_pages) {
 837                                 dout("reached end pvec, trying for more\n");
 838                                 pagevec_reinit(&pvec);
 839                                 goto get_more_pages;
 840                         }
 841
 842                         /* shift unused pages over in the pvec...  we
 843                          * will need to release them below. */
 844                         for (j = i; j < pvec_pages; j++) {
 845                                 dout(" pvec leftover page %p\n",
 846                                      pvec.pages[j]);
 847                                 pvec.pages[j-i+first] = pvec.pages[j];
 848                         }
 849                         pvec.nr -= i-first;
 850                 }
 851
 852                 /* submit the write */
 853                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 854                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 855                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 856                 dout("writepages got %d pages at %llu~%llu\n",
 857                      locked_pages, offset, len);
 858
 859                 /* revise final length, page count */
 860                 req->r_num_pages = locked_pages;
 861                 reqhead = req->r_request->front.iov_base;
 862                 op = (void *)(reqhead + 1);
 863                 op->extent.length = cpu_to_le64(len);
 864                 op->payload_len = cpu_to_le32(len);
 865                 req->r_request->hdr.data_len = cpu_to_le32(len);
 866
 867                 ceph_osdc_start_request(&client->osdc, req, true);
 868                 req = NULL;
 869
 870                 /* continue? */
 871                 index = next;
 872                 wbc->nr_to_write -= locked_pages;
 873                 if (wbc->nr_to_write <= 0)
 874                         done = 1;
 875
 876 release_pvec_pages:
 877                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 878                      pvec.nr ? pvec.pages[0] : NULL);
 879                 pagevec_release(&pvec);
 880
 881                 if (locked_pages && !done)
 882                         goto retry;
 883         }
 884
 885         if (should_loop && !done) {
 886                 /* more to do; loop back to beginning of file */
 887                 dout("writepages looping back to beginning of file\n");
 888                 should_loop = 0;
 889                 index = 0;
 890                 goto retry;
 891         }
 892
 893         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 894                 mapping->writeback_index = index;
 895
 896 out:
 897         if (req)
 898                 ceph_osdc_put_request(req);
 899         if (rc > 0)
 900                 rc = 0;  /* vfs expects us to return 0 */
 901         ceph_put_snap_context(snapc);
 902         dout("writepages done, rc = %d\n", rc);
 903 out_final:
 904         return rc;
 905 }
 906
 907
 908
 909 /*
 910  * See if a given @snapc is either writeable, or already written.
 911  */
 912 static int context_is_writeable_or_written(struct inode *inode,
 913                                            struct ceph_snap_context *snapc)
 914 {
 915         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 916         return !oldest || snapc->seq <= oldest->seq;
 917 }
 918
 919 /*
 920  * We are only allowed to write into/dirty the page if the page is
 921  * clean, or already dirty within the same snap context.
 922  *
 923  * called with page locked.
 924  * return success with page locked,
 925  * or any failure (incl -EAGAIN) with page unlocked.
 926  */
 927 static int ceph_update_writeable_page(struct file *file,
 928                             loff_t pos, unsigned len,
 929                             struct page *page)
 930 {
 931         struct inode *inode = file->f_dentry->d_inode;
 932         struct ceph_inode_info *ci = ceph_inode(inode);
 933         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
 934         loff_t page_off = pos & PAGE_CACHE_MASK;
 935         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 936         int end_in_page = pos_in_page + len;
 937         loff_t i_size;
 938         struct ceph_snap_context *snapc;
 939         int r;
 940
 941 retry_locked:
 942         /* writepages currently holds page lock, but if we change that later, */
 943         wait_on_page_writeback(page);
 944
 945         /* check snap context */
 946         BUG_ON(!ci->i_snap_realm);
 947         down_read(&mdsc->snap_rwsem);
 948         BUG_ON(!ci->i_snap_realm->cached_context);
 949         if (page->private &&
 950             (void *)page->private != ci->i_snap_realm->cached_context) {
 951                 /*
 952                  * this page is already dirty in another (older) snap
 953                  * context!  is it writeable now?
 954                  */
 955                 snapc = get_oldest_context(inode, NULL);
 956                 up_read(&mdsc->snap_rwsem);
 957
 958                 if (snapc != (void *)page->private) {
 959                         dout(" page %p snapc %p not current or oldest\n",
 960                              page, (void *)page->private);
 961                         /*
 962                          * queue for writeback, and wait for snapc to
 963                          * be writeable or written
 964                          */
 965                         snapc = ceph_get_snap_context((void *)page->private);
 966                         unlock_page(page);
 967                         ceph_queue_writeback(inode);
 968                         r = wait_event_interruptible(ci->i_cap_wq,
 969                                context_is_writeable_or_written(inode, snapc));
 970                         ceph_put_snap_context(snapc);
 971                         if (r == -ERESTARTSYS)
 972                                 return r;
 973                         return -EAGAIN;
 974                 }
 975
 976                 /* yay, writeable, do it now (without dropping page lock) */
 977                 dout(" page %p snapc %p not current, but oldest\n",
 978                      page, snapc);
 979                 if (!clear_page_dirty_for_io(page))
 980                         goto retry_locked;
 981                 r = writepage_nounlock(page, NULL);
 982                 if (r < 0)
 983                         goto fail_nosnap;
 984                 goto retry_locked;
 985         }
 986
 987         if (PageUptodate(page)) {
 988                 dout(" page %p already uptodate\n", page);
 989                 return 0;
 990         }
 991
 992         /* full page? */
 993         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
 994                 return 0;
 995
 996         /* past end of file? */
 997         i_size = inode->i_size;   /* caller holds i_mutex */
 998
 999         if (i_size + len > inode->i_sb->s_maxbytes) {
1000                 /* file is too big */
1001                 r = -EINVAL;
1002                 goto fail;
1003         }
1004
1005         if (page_off >= i_size ||
1006             (pos_in_page == 0 && (pos+len) >= i_size &&
1007              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1008                 dout(" zeroing %p 0 - %d and %d - %d\n",
1009                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1010                 zero_user_segments(page,
1011                                    0, pos_in_page,
1012                                    end_in_page, PAGE_CACHE_SIZE);
1013                 return 0;
1014         }
1015
1016         /* we need to read it. */
1017         up_read(&mdsc->snap_rwsem);
1018         r = readpage_nounlock(file, page);
1019         if (r < 0)
1020                 goto fail_nosnap;
1021         goto retry_locked;
1022
1023 fail:
1024         up_read(&mdsc->snap_rwsem);
1025 fail_nosnap:
1026         unlock_page(page);
1027         return r;
1028 }
1029
1030 /*
1031  * We are only allowed to write into/dirty the page if the page is
1032  * clean, or already dirty within the same snap context.
1033  */
1034 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1035                             loff_t pos, unsigned len, unsigned flags,
1036                             struct page **pagep, void **fsdata)
1037 {
1038         struct inode *inode = file->f_dentry->d_inode;
1039         struct page *page;
1040         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1041         int r;
1042
1043         do {
1044                 /* get a page */
1045                 page = grab_cache_page_write_begin(mapping, index, 0);
1046                 if (!page)
1047                         return -ENOMEM;
1048                 *pagep = page;
1049
1050                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1051                 inode, page, (int)pos, (int)len);
1052
1053                 r = ceph_update_writeable_page(file, pos, len, page);
1054         } while (r == -EAGAIN);
1055
1056         return r;
1057 }
1058
1059 /*
1060  * we don't do anything in here that simple_write_end doesn't do
1061  * except adjust dirty page accounting and drop read lock on
1062  * mdsc->snap_rwsem.
1063  */
1064 static int ceph_write_end(struct file *file, struct address_space *mapping,
1065                           loff_t pos, unsigned len, unsigned copied,
1066                           struct page *page, void *fsdata)
1067 {
1068         struct inode *inode = file->f_dentry->d_inode;
1069         struct ceph_client *client = ceph_inode_to_client(inode);
1070         struct ceph_mds_client *mdsc = &client->mdsc;
1071         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1072         int check_cap = 0;
1073
1074         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1075              inode, page, (int)pos, (int)copied, (int)len);
1076
1077         /* zero the stale part of the page if we did a short copy */
1078         if (copied < len)
1079                 zero_user_segment(page, from+copied, len);
1080
1081         /* did file size increase? */
1082         /* (no need for i_size_read(); we caller holds i_mutex */
1083         if (pos+copied > inode->i_size)
1084                 check_cap = ceph_inode_set_size(inode, pos+copied);
1085
1086         if (!PageUptodate(page))
1087                 SetPageUptodate(page);
1088
1089         set_page_dirty(page);
1090
1091         unlock_page(page);
1092         up_read(&mdsc->snap_rwsem);
1093         page_cache_release(page);
1094
1095         if (check_cap)
1096                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1097
1098         return copied;
1099 }
1100
1101 /*
1102  * we set .direct_IO to indicate direct io is supported, but since we
1103  * intercept O_DIRECT reads and writes early, this function should
1104  * never get called.
1105  */
1106 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1107                               const struct iovec *iov,
1108                               loff_t pos, unsigned long nr_segs)
1109 {
1110         WARN_ON(1);
1111         return -EINVAL;
1112 }
1113
1114 const struct address_space_operations ceph_aops = {
1115         .readpage = ceph_readpage,
1116         .readpages = ceph_readpages,
1117         .writepage = ceph_writepage,
1118         .writepages = ceph_writepages_start,
1119         .write_begin = ceph_write_begin,
1120         .write_end = ceph_write_end,
1121         .set_page_dirty = ceph_set_page_dirty,
1122         .invalidatepage = ceph_invalidatepage,
1123         .releasepage = ceph_releasepage,
1124         .direct_IO = ceph_direct_io,
1125 };
1126
1127
1128 /*
1129  * vm ops
1130  */
1131
1132 /*
1133  * Reuse write_begin here for simplicity.
1134  */
1135 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1136 {
1137         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1138         struct page *page = vmf->page;
1139         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1140         loff_t off = page->index << PAGE_CACHE_SHIFT;
1141         loff_t size, len;
1142         int ret;
1143
1144         size = i_size_read(inode);
1145         if (off + PAGE_CACHE_SIZE <= size)
1146                 len = PAGE_CACHE_SIZE;
1147         else
1148                 len = size & ~PAGE_CACHE_MASK;
1149
1150         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1151              off, len, page, page->index);
1152
1153         lock_page(page);
1154
1155         ret = VM_FAULT_NOPAGE;
1156         if ((off > size) ||
1157             (page->mapping != inode->i_mapping))
1158                 goto out;
1159
1160         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1161         if (ret == 0) {
1162                 /* success.  we'll keep the page locked. */
1163                 set_page_dirty(page);
1164                 up_read(&mdsc->snap_rwsem);
1165                 ret = VM_FAULT_LOCKED;
1166         } else {
1167                 if (ret == -ENOMEM)
1168                         ret = VM_FAULT_OOM;
1169                 else
1170                         ret = VM_FAULT_SIGBUS;
1171         }
1172 out:
1173         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1174         if (ret != VM_FAULT_LOCKED)
1175                 unlock_page(page);
1176         return ret;
1177 }
1178
1179 static struct vm_operations_struct ceph_vmops = {
1180         .fault          = filemap_fault,
1181         .page_mkwrite   = ceph_page_mkwrite,
1182 };
1183
1184 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1185 {
1186         struct address_space *mapping = file->f_mapping;
1187
1188         if (!mapping->a_ops->readpage)
1189                 return -ENOEXEC;
1190         file_accessed(file);
1191         vma->vm_ops = &ceph_vmops;
1192         vma->vm_flags |= VM_CAN_NONLINEAR;
1193         return 0;
1194 }