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