fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /*
  91          * Note that we're grabbing a snapc ref here without holding
  92          * any locks!
  93          */
  94         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         if (ci->i_head_snapc == NULL)
  99                 ci->i_head_snapc = ceph_get_snap_context(snapc);
 100         ++ci->i_wrbuffer_ref_head;
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         ret = __set_page_dirty_nobuffers(page);
 121         WARN_ON(!PageLocked(page));
 122         WARN_ON(!page->mapping);
 123
 124         return ret;
 125 }
 126
 127 /*
 128  * If we are truncating the full page (i.e. offset == 0), adjust the
 129  * dirty page counters appropriately.  Only called if there is private
 130  * data on the page.
 131  */
 132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 133                                 unsigned int length)
 134 {
 135         struct inode *inode;
 136         struct ceph_inode_info *ci;
 137         struct ceph_snap_context *snapc = page_snap_context(page);
 138
 139         inode = page->mapping->host;
 140         ci = ceph_inode(inode);
 141
 142         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 143                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 144                      inode, page, page->index, offset, length);
 145                 return;
 146         }
 147
 148         ceph_invalidate_fscache_page(inode, page);
 149
 150         if (!PagePrivate(page))
 151                 return;
 152
 153         /*
 154          * We can get non-dirty pages here due to races between
 155          * set_page_dirty and truncate_complete_page; just spit out a
 156          * warning, in case we end up with accounting problems later.
 157          */
 158         if (!PageDirty(page))
 159                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 160
 161         ClearPageChecked(page);
 162
 163         dout("%p invalidatepage %p idx %lu full dirty page\n",
 164              inode, page, page->index);
 165
 166         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 167         ceph_put_snap_context(snapc);
 168         page->private = 0;
 169         ClearPagePrivate(page);
 170 }
 171
 172 static int ceph_releasepage(struct page *page, gfp_t g)
 173 {
 174         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 175         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 176         WARN_ON(PageDirty(page));
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * read a single page, without unlocking it.
 187  */
 188 static int readpage_nounlock(struct file *filp, struct page *page)
 189 {
 190         struct inode *inode = file_inode(filp);
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_osd_client *osdc =
 193                 &ceph_inode_to_client(inode)->client->osdc;
 194         int err = 0;
 195         u64 len = PAGE_CACHE_SIZE;
 196
 197         err = ceph_readpage_from_fscache(inode, page);
 198
 199         if (err == 0)
 200                 goto out;
 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                                   (u64) page_offset(page), &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 {
 214                 if (err < PAGE_CACHE_SIZE) {
 215                 /* zero fill remainder of page */
 216                         zero_user_segment(page, err, PAGE_CACHE_SIZE);
 217                 } else {
 218                         flush_dcache_page(page);
 219                 }
 220         }
 221         SetPageUptodate(page);
 222
 223         if (err >= 0)
 224                 ceph_readpage_to_fscache(inode, page);
 225
 226 out:
 227         return err < 0 ? err : 0;
 228 }
 229
 230 static int ceph_readpage(struct file *filp, struct page *page)
 231 {
 232         int r = readpage_nounlock(filp, page);
 233         unlock_page(page);
 234         return r;
 235 }
 236
 237 /*
 238  * Finish an async read(ahead) op.
 239  */
 240 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 241 {
 242         struct inode *inode = req->r_inode;
 243         struct ceph_osd_data *osd_data;
 244         int rc = req->r_result;
 245         int bytes = le32_to_cpu(msg->hdr.data_len);
 246         int num_pages;
 247         int i;
 248
 249         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 250
 251         /* unlock all pages, zeroing any data we didn't read */
 252         osd_data = osd_req_op_extent_osd_data(req, 0);
 253         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 254         num_pages = calc_pages_for((u64)osd_data->alignment,
 255                                         (u64)osd_data->length);
 256         for (i = 0; i < num_pages; i++) {
 257                 struct page *page = osd_data->pages[i];
 258
 259                 if (rc < 0)
 260                         goto unlock;
 261                 if (bytes < (int)PAGE_CACHE_SIZE) {
 262                         /* zero (remainder of) page */
 263                         int s = bytes < 0 ? 0 : bytes;
 264                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 265                 }
 266                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 267                      page->index);
 268                 flush_dcache_page(page);
 269                 SetPageUptodate(page);
 270                 ceph_readpage_to_fscache(inode, page);
 271 unlock:
 272                 unlock_page(page);
 273                 page_cache_release(page);
 274                 bytes -= PAGE_CACHE_SIZE;
 275         }
 276         kfree(osd_data->pages);
 277 }
 278
 279 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 280 {
 281         int i;
 282
 283         for (i = 0; i < num_pages; i++)
 284                 unlock_page(pages[i]);
 285 }
 286
 287 /*
 288  * start an async read(ahead) operation.  return nr_pages we submitted
 289  * a read for on success, or negative error code.
 290  */
 291 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 292 {
 293         struct ceph_osd_client *osdc =
 294                 &ceph_inode_to_client(inode)->client->osdc;
 295         struct ceph_inode_info *ci = ceph_inode(inode);
 296         struct page *page = list_entry(page_list->prev, struct page, lru);
 297         struct ceph_vino vino;
 298         struct ceph_osd_request *req;
 299         u64 off;
 300         u64 len;
 301         int i;
 302         struct page **pages;
 303         pgoff_t next_index;
 304         int nr_pages = 0;
 305         int ret;
 306
 307         off = (u64) page_offset(page);
 308
 309         /* count pages */
 310         next_index = page->index;
 311         list_for_each_entry_reverse(page, page_list, lru) {
 312                 if (page->index != next_index)
 313                         break;
 314                 nr_pages++;
 315                 next_index++;
 316                 if (max && nr_pages == max)
 317                         break;
 318         }
 319         len = nr_pages << PAGE_CACHE_SHIFT;
 320         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 321              off, len);
 322         vino = ceph_vino(inode);
 323         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 324                                     1, CEPH_OSD_OP_READ,
 325                                     CEPH_OSD_FLAG_READ, NULL,
 326                                     ci->i_truncate_seq, ci->i_truncate_size,
 327                                     false);
 328         if (IS_ERR(req))
 329                 return PTR_ERR(req);
 330
 331         /* build page vector */
 332         nr_pages = calc_pages_for(0, len);
 333         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 334         ret = -ENOMEM;
 335         if (!pages)
 336                 goto out;
 337         for (i = 0; i < nr_pages; ++i) {
 338                 page = list_entry(page_list->prev, struct page, lru);
 339                 BUG_ON(PageLocked(page));
 340                 list_del(&page->lru);
 341
 342                 dout("start_read %p adding %p idx %lu\n", inode, page,
 343                      page->index);
 344                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 345                                           GFP_NOFS)) {
 346                         ceph_fscache_uncache_page(inode, page);
 347                         page_cache_release(page);
 348                         dout("start_read %p add_to_page_cache failed %p\n",
 349                              inode, page);
 350                         nr_pages = i;
 351                         goto out_pages;
 352                 }
 353                 pages[i] = page;
 354         }
 355         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 356         req->r_callback = finish_read;
 357         req->r_inode = inode;
 358
 359         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 360
 361         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 362         ret = ceph_osdc_start_request(osdc, req, false);
 363         if (ret < 0)
 364                 goto out_pages;
 365         ceph_osdc_put_request(req);
 366         return nr_pages;
 367
 368 out_pages:
 369         ceph_unlock_page_vector(pages, nr_pages);
 370         ceph_release_page_vector(pages, nr_pages);
 371 out:
 372         ceph_osdc_put_request(req);
 373         return ret;
 374 }
 375
 376
 377 /*
 378  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 379  * the caller (VM) cleans them up.
 380  */
 381 static int ceph_readpages(struct file *file, struct address_space *mapping,
 382                           struct list_head *page_list, unsigned nr_pages)
 383 {
 384         struct inode *inode = file_inode(file);
 385         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 386         int rc = 0;
 387         int max = 0;
 388
 389         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 390                                          &nr_pages);
 391
 392         if (rc == 0)
 393                 goto out;
 394
 395         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 396                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 397                         >> PAGE_SHIFT;
 398
 399         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 400                 file, nr_pages,
 401              max);
 402         while (!list_empty(page_list)) {
 403                 rc = start_read(inode, page_list, max);
 404                 if (rc < 0)
 405                         goto out;
 406                 BUG_ON(rc == 0);
 407         }
 408 out:
 409         ceph_fscache_readpages_cancel(inode, page_list);
 410
 411         dout("readpages %p file %p ret %d\n", inode, file, rc);
 412         return rc;
 413 }
 414
 415 /*
 416  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 417  * only snap context we are allowed to write back.
 418  */
 419 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 420                                                     u64 *snap_size)
 421 {
 422         struct ceph_inode_info *ci = ceph_inode(inode);
 423         struct ceph_snap_context *snapc = NULL;
 424         struct ceph_cap_snap *capsnap = NULL;
 425
 426         spin_lock(&ci->i_ceph_lock);
 427         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 428                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 429                      capsnap->context, capsnap->dirty_pages);
 430                 if (capsnap->dirty_pages) {
 431                         snapc = ceph_get_snap_context(capsnap->context);
 432                         if (snap_size)
 433                                 *snap_size = capsnap->size;
 434                         break;
 435                 }
 436         }
 437         if (!snapc && ci->i_wrbuffer_ref_head) {
 438                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 439                 dout(" head snapc %p has %d dirty pages\n",
 440                      snapc, ci->i_wrbuffer_ref_head);
 441         }
 442         spin_unlock(&ci->i_ceph_lock);
 443         return snapc;
 444 }
 445
 446 /*
 447  * Write a single page, but leave the page locked.
 448  *
 449  * If we get a write error, set the page error bit, but still adjust the
 450  * dirty page accounting (i.e., page is no longer dirty).
 451  */
 452 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 453 {
 454         struct inode *inode;
 455         struct ceph_inode_info *ci;
 456         struct ceph_fs_client *fsc;
 457         struct ceph_osd_client *osdc;
 458         struct ceph_snap_context *snapc, *oldest;
 459         loff_t page_off = page_offset(page);
 460         long writeback_stat;
 461         u64 truncate_size, snap_size = 0;
 462         u32 truncate_seq;
 463         int err = 0, len = PAGE_CACHE_SIZE;
 464
 465         dout("writepage %p idx %lu\n", page, page->index);
 466
 467         if (!page->mapping || !page->mapping->host) {
 468                 dout("writepage %p - no mapping\n", page);
 469                 return -EFAULT;
 470         }
 471         inode = page->mapping->host;
 472         ci = ceph_inode(inode);
 473         fsc = ceph_inode_to_client(inode);
 474         osdc = &fsc->client->osdc;
 475
 476         /* verify this is a writeable snap context */
 477         snapc = page_snap_context(page);
 478         if (snapc == NULL) {
 479                 dout("writepage %p page %p not dirty?\n", inode, page);
 480                 goto out;
 481         }
 482         oldest = get_oldest_context(inode, &snap_size);
 483         if (snapc->seq > oldest->seq) {
 484                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 485                      inode, page, snapc);
 486                 /* we should only noop if called by kswapd */
 487                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 488                 ceph_put_snap_context(oldest);
 489                 goto out;
 490         }
 491         ceph_put_snap_context(oldest);
 492
 493         spin_lock(&ci->i_ceph_lock);
 494         truncate_seq = ci->i_truncate_seq;
 495         truncate_size = ci->i_truncate_size;
 496         if (!snap_size)
 497                 snap_size = i_size_read(inode);
 498         spin_unlock(&ci->i_ceph_lock);
 499
 500         /* is this a partial page at end of file? */
 501         if (page_off >= snap_size) {
 502                 dout("%p page eof %llu\n", page, snap_size);
 503                 goto out;
 504         }
 505         if (snap_size < page_off + len)
 506                 len = snap_size - page_off;
 507
 508         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 509              inode, page, page->index, page_off, len, snapc);
 510
 511         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 512         if (writeback_stat >
 513             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 514                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 515
 516         ceph_readpage_to_fscache(inode, page);
 517
 518         set_page_writeback(page);
 519         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 520                                    &ci->i_layout, snapc,
 521                                    page_off, len,
 522                                    truncate_seq, truncate_size,
 523                                    &inode->i_mtime, &page, 1);
 524         if (err < 0) {
 525                 dout("writepage setting page/mapping error %d %p\n", err, page);
 526                 SetPageError(page);
 527                 mapping_set_error(&inode->i_data, err);
 528                 if (wbc)
 529                         wbc->pages_skipped++;
 530         } else {
 531                 dout("writepage cleaned page %p\n", page);
 532                 err = 0;  /* vfs expects us to return 0 */
 533         }
 534         page->private = 0;
 535         ClearPagePrivate(page);
 536         end_page_writeback(page);
 537         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 538         ceph_put_snap_context(snapc);  /* page's reference */
 539 out:
 540         return err;
 541 }
 542
 543 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 544 {
 545         int err;
 546         struct inode *inode = page->mapping->host;
 547         BUG_ON(!inode);
 548         ihold(inode);
 549         err = writepage_nounlock(page, wbc);
 550         unlock_page(page);
 551         iput(inode);
 552         return err;
 553 }
 554
 555
 556 /*
 557  * lame release_pages helper.  release_pages() isn't exported to
 558  * modules.
 559  */
 560 static void ceph_release_pages(struct page **pages, int num)
 561 {
 562         struct pagevec pvec;
 563         int i;
 564
 565         pagevec_init(&pvec, 0);
 566         for (i = 0; i < num; i++) {
 567                 if (pagevec_add(&pvec, pages[i]) == 0)
 568                         pagevec_release(&pvec);
 569         }
 570         pagevec_release(&pvec);
 571 }
 572
 573 /*
 574  * async writeback completion handler.
 575  *
 576  * If we get an error, set the mapping error bit, but not the individual
 577  * page error bits.
 578  */
 579 static void writepages_finish(struct ceph_osd_request *req,
 580                               struct ceph_msg *msg)
 581 {
 582         struct inode *inode = req->r_inode;
 583         struct ceph_inode_info *ci = ceph_inode(inode);
 584         struct ceph_osd_data *osd_data;
 585         unsigned wrote;
 586         struct page *page;
 587         int num_pages;
 588         int i;
 589         struct ceph_snap_context *snapc = req->r_snapc;
 590         struct address_space *mapping = inode->i_mapping;
 591         int rc = req->r_result;
 592         u64 bytes = req->r_ops[0].extent.length;
 593         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 594         long writeback_stat;
 595         unsigned issued = ceph_caps_issued(ci);
 596
 597         osd_data = osd_req_op_extent_osd_data(req, 0);
 598         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 599         num_pages = calc_pages_for((u64)osd_data->alignment,
 600                                         (u64)osd_data->length);
 601         if (rc >= 0) {
 602                 /*
 603                  * Assume we wrote the pages we originally sent.  The
 604                  * osd might reply with fewer pages if our writeback
 605                  * raced with a truncation and was adjusted at the osd,
 606                  * so don't believe the reply.
 607                  */
 608                 wrote = num_pages;
 609         } else {
 610                 wrote = 0;
 611                 mapping_set_error(mapping, rc);
 612         }
 613         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 614              inode, rc, bytes, wrote);
 615
 616         /* clean all pages */
 617         for (i = 0; i < num_pages; i++) {
 618                 page = osd_data->pages[i];
 619                 BUG_ON(!page);
 620                 WARN_ON(!PageUptodate(page));
 621
 622                 writeback_stat =
 623                         atomic_long_dec_return(&fsc->writeback_count);
 624                 if (writeback_stat <
 625                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 626                         clear_bdi_congested(&fsc->backing_dev_info,
 627                                             BLK_RW_ASYNC);
 628
 629                 ceph_put_snap_context(page_snap_context(page));
 630                 page->private = 0;
 631                 ClearPagePrivate(page);
 632                 dout("unlocking %d %p\n", i, page);
 633                 end_page_writeback(page);
 634
 635                 /*
 636                  * We lost the cache cap, need to truncate the page before
 637                  * it is unlocked, otherwise we'd truncate it later in the
 638                  * page truncation thread, possibly losing some data that
 639                  * raced its way in
 640                  */
 641                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 642                         generic_error_remove_page(inode->i_mapping, page);
 643
 644                 unlock_page(page);
 645         }
 646         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 647         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 648
 649         ceph_release_pages(osd_data->pages, num_pages);
 650         if (osd_data->pages_from_pool)
 651                 mempool_free(osd_data->pages,
 652                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 653         else
 654                 kfree(osd_data->pages);
 655         ceph_osdc_put_request(req);
 656 }
 657
 658 /*
 659  * initiate async writeback
 660  */
 661 static int ceph_writepages_start(struct address_space *mapping,
 662                                  struct writeback_control *wbc)
 663 {
 664         struct inode *inode = mapping->host;
 665         struct ceph_inode_info *ci = ceph_inode(inode);
 666         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 667         struct ceph_vino vino = ceph_vino(inode);
 668         pgoff_t index, start, end;
 669         int range_whole = 0;
 670         int should_loop = 1;
 671         pgoff_t max_pages = 0, max_pages_ever = 0;
 672         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 673         struct pagevec pvec;
 674         int done = 0;
 675         int rc = 0;
 676         unsigned wsize = 1 << inode->i_blkbits;
 677         struct ceph_osd_request *req = NULL;
 678         int do_sync;
 679         u64 truncate_size, snap_size;
 680         u32 truncate_seq;
 681
 682         /*
 683          * Include a 'sync' in the OSD request if this is a data
 684          * integrity write (e.g., O_SYNC write or fsync()), or if our
 685          * cap is being revoked.
 686          */
 687         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 688                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 689                 do_sync = 1;
 690         dout("writepages_start %p dosync=%d (mode=%s)\n",
 691              inode, do_sync,
 692              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 693              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 694
 695         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 696                 pr_warning("writepage_start %p on forced umount\n", inode);
 697                 return -EIO; /* we're in a forced umount, don't write! */
 698         }
 699         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 700                 wsize = fsc->mount_options->wsize;
 701         if (wsize < PAGE_CACHE_SIZE)
 702                 wsize = PAGE_CACHE_SIZE;
 703         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 704
 705         pagevec_init(&pvec, 0);
 706
 707         /* where to start/end? */
 708         if (wbc->range_cyclic) {
 709                 start = mapping->writeback_index; /* Start from prev offset */
 710                 end = -1;
 711                 dout(" cyclic, start at %lu\n", start);
 712         } else {
 713                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 714                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 715                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 716                         range_whole = 1;
 717                 should_loop = 0;
 718                 dout(" not cyclic, %lu to %lu\n", start, end);
 719         }
 720         index = start;
 721
 722 retry:
 723         /* find oldest snap context with dirty data */
 724         ceph_put_snap_context(snapc);
 725         snap_size = 0;
 726         snapc = get_oldest_context(inode, &snap_size);
 727         if (!snapc) {
 728                 /* hmm, why does writepages get called when there
 729                    is no dirty data? */
 730                 dout(" no snap context with dirty data?\n");
 731                 goto out;
 732         }
 733         if (snap_size == 0)
 734                 snap_size = i_size_read(inode);
 735         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 736              snapc, snapc->seq, snapc->num_snaps);
 737
 738         spin_lock(&ci->i_ceph_lock);
 739         truncate_seq = ci->i_truncate_seq;
 740         truncate_size = ci->i_truncate_size;
 741         if (!snap_size)
 742                 snap_size = i_size_read(inode);
 743         spin_unlock(&ci->i_ceph_lock);
 744
 745         if (last_snapc && snapc != last_snapc) {
 746                 /* if we switched to a newer snapc, restart our scan at the
 747                  * start of the original file range. */
 748                 dout("  snapc differs from last pass, restarting at %lu\n",
 749                      index);
 750                 index = start;
 751         }
 752         last_snapc = snapc;
 753
 754         while (!done && index <= end) {
 755                 int num_ops = do_sync ? 2 : 1;
 756                 unsigned i;
 757                 int first;
 758                 pgoff_t next;
 759                 int pvec_pages, locked_pages;
 760                 struct page **pages = NULL;
 761                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 762                 struct page *page;
 763                 int want;
 764                 u64 offset, len;
 765                 long writeback_stat;
 766
 767                 next = 0;
 768                 locked_pages = 0;
 769                 max_pages = max_pages_ever;
 770
 771 get_more_pages:
 772                 first = -1;
 773                 want = min(end - index,
 774                            min((pgoff_t)PAGEVEC_SIZE,
 775                                max_pages - (pgoff_t)locked_pages) - 1)
 776                         + 1;
 777                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 778                                                 PAGECACHE_TAG_DIRTY,
 779                                                 want);
 780                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 781                 if (!pvec_pages && !locked_pages)
 782                         break;
 783                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 784                         page = pvec.pages[i];
 785                         dout("? %p idx %lu\n", page, page->index);
 786                         if (locked_pages == 0)
 787                                 lock_page(page);  /* first page */
 788                         else if (!trylock_page(page))
 789                                 break;
 790
 791                         /* only dirty pages, or our accounting breaks */
 792                         if (unlikely(!PageDirty(page)) ||
 793                             unlikely(page->mapping != mapping)) {
 794                                 dout("!dirty or !mapping %p\n", page);
 795                                 unlock_page(page);
 796                                 break;
 797                         }
 798                         if (!wbc->range_cyclic && page->index > end) {
 799                                 dout("end of range %p\n", page);
 800                                 done = 1;
 801                                 unlock_page(page);
 802                                 break;
 803                         }
 804                         if (next && (page->index != next)) {
 805                                 dout("not consecutive %p\n", page);
 806                                 unlock_page(page);
 807                                 break;
 808                         }
 809                         if (wbc->sync_mode != WB_SYNC_NONE) {
 810                                 dout("waiting on writeback %p\n", page);
 811                                 wait_on_page_writeback(page);
 812                         }
 813                         if (page_offset(page) >= snap_size) {
 814                                 dout("%p page eof %llu\n", page, snap_size);
 815                                 done = 1;
 816                                 unlock_page(page);
 817                                 break;
 818                         }
 819                         if (PageWriteback(page)) {
 820                                 dout("%p under writeback\n", page);
 821                                 unlock_page(page);
 822                                 break;
 823                         }
 824
 825                         /* only if matching snap context */
 826                         pgsnapc = page_snap_context(page);
 827                         if (pgsnapc->seq > snapc->seq) {
 828                                 dout("page snapc %p %lld > oldest %p %lld\n",
 829                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 830                                 unlock_page(page);
 831                                 if (!locked_pages)
 832                                         continue; /* keep looking for snap */
 833                                 break;
 834                         }
 835
 836                         if (!clear_page_dirty_for_io(page)) {
 837                                 dout("%p !clear_page_dirty_for_io\n", page);
 838                                 unlock_page(page);
 839                                 break;
 840                         }
 841
 842                         /*
 843                          * We have something to write.  If this is
 844                          * the first locked page this time through,
 845                          * allocate an osd request and a page array
 846                          * that it will use.
 847                          */
 848                         if (locked_pages == 0) {
 849                                 BUG_ON(pages);
 850                                 /* prepare async write request */
 851                                 offset = (u64)page_offset(page);
 852                                 len = wsize;
 853                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 854                                                         &ci->i_layout, vino,
 855                                                         offset, &len, num_ops,
 856                                                         CEPH_OSD_OP_WRITE,
 857                                                         CEPH_OSD_FLAG_WRITE |
 858                                                         CEPH_OSD_FLAG_ONDISK,
 859                                                         snapc, truncate_seq,
 860                                                         truncate_size, true);
 861                                 if (IS_ERR(req)) {
 862                                         rc = PTR_ERR(req);
 863                                         unlock_page(page);
 864                                         break;
 865                                 }
 866
 867                                 req->r_callback = writepages_finish;
 868                                 req->r_inode = inode;
 869
 870                                 max_pages = calc_pages_for(0, (u64)len);
 871                                 pages = kmalloc(max_pages * sizeof (*pages),
 872                                                 GFP_NOFS);
 873                                 if (!pages) {
 874                                         pool = fsc->wb_pagevec_pool;
 875                                         pages = mempool_alloc(pool, GFP_NOFS);
 876                                         BUG_ON(!pages);
 877                                 }
 878                         }
 879
 880                         /* note position of first page in pvec */
 881                         if (first < 0)
 882                                 first = i;
 883                         dout("%p will write page %p idx %lu\n",
 884                              inode, page, page->index);
 885
 886                         writeback_stat =
 887                                atomic_long_inc_return(&fsc->writeback_count);
 888                         if (writeback_stat > CONGESTION_ON_THRESH(
 889                                     fsc->mount_options->congestion_kb)) {
 890                                 set_bdi_congested(&fsc->backing_dev_info,
 891                                                   BLK_RW_ASYNC);
 892                         }
 893
 894                         set_page_writeback(page);
 895                         pages[locked_pages] = page;
 896                         locked_pages++;
 897                         next = page->index + 1;
 898                 }
 899
 900                 /* did we get anything? */
 901                 if (!locked_pages)
 902                         goto release_pvec_pages;
 903                 if (i) {
 904                         int j;
 905                         BUG_ON(!locked_pages || first < 0);
 906
 907                         if (pvec_pages && i == pvec_pages &&
 908                             locked_pages < max_pages) {
 909                                 dout("reached end pvec, trying for more\n");
 910                                 pagevec_reinit(&pvec);
 911                                 goto get_more_pages;
 912                         }
 913
 914                         /* shift unused pages over in the pvec...  we
 915                          * will need to release them below. */
 916                         for (j = i; j < pvec_pages; j++) {
 917                                 dout(" pvec leftover page %p\n",
 918                                      pvec.pages[j]);
 919                                 pvec.pages[j-i+first] = pvec.pages[j];
 920                         }
 921                         pvec.nr -= i-first;
 922                 }
 923
 924                 /* Format the osd request message and submit the write */
 925
 926                 offset = page_offset(pages[0]);
 927                 len = min(snap_size - offset,
 928                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 929                 dout("writepages got %d pages at %llu~%llu\n",
 930                      locked_pages, offset, len);
 931
 932                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 933                                                         !!pool, false);
 934
 935                 pages = NULL;   /* request message now owns the pages array */
 936                 pool = NULL;
 937
 938                 /* Update the write op length in case we changed it */
 939
 940                 osd_req_op_extent_update(req, 0, len);
 941
 942                 vino = ceph_vino(inode);
 943                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 944                                         &inode->i_mtime);
 945
 946                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 947                 BUG_ON(rc);
 948                 req = NULL;
 949
 950                 /* continue? */
 951                 index = next;
 952                 wbc->nr_to_write -= locked_pages;
 953                 if (wbc->nr_to_write <= 0)
 954                         done = 1;
 955
 956 release_pvec_pages:
 957                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 958                      pvec.nr ? pvec.pages[0] : NULL);
 959                 pagevec_release(&pvec);
 960
 961                 if (locked_pages && !done)
 962                         goto retry;
 963         }
 964
 965         if (should_loop && !done) {
 966                 /* more to do; loop back to beginning of file */
 967                 dout("writepages looping back to beginning of file\n");
 968                 should_loop = 0;
 969                 index = 0;
 970                 goto retry;
 971         }
 972
 973         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 974                 mapping->writeback_index = index;
 975
 976 out:
 977         if (req)
 978                 ceph_osdc_put_request(req);
 979         ceph_put_snap_context(snapc);
 980         dout("writepages done, rc = %d\n", rc);
 981         return rc;
 982 }
 983
 984
 985
 986 /*
 987  * See if a given @snapc is either writeable, or already written.
 988  */
 989 static int context_is_writeable_or_written(struct inode *inode,
 990                                            struct ceph_snap_context *snapc)
 991 {
 992         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 993         int ret = !oldest || snapc->seq <= oldest->seq;
 994
 995         ceph_put_snap_context(oldest);
 996         return ret;
 997 }
 998
 999 /*
1000  * We are only allowed to write into/dirty the page if the page is
1001  * clean, or already dirty within the same snap context.
1002  *
1003  * called with page locked.
1004  * return success with page locked,
1005  * or any failure (incl -EAGAIN) with page unlocked.
1006  */
1007 static int ceph_update_writeable_page(struct file *file,
1008                             loff_t pos, unsigned len,
1009                             struct page *page)
1010 {
1011         struct inode *inode = file_inode(file);
1012         struct ceph_inode_info *ci = ceph_inode(inode);
1013         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1014         loff_t page_off = pos & PAGE_CACHE_MASK;
1015         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1016         int end_in_page = pos_in_page + len;
1017         loff_t i_size;
1018         int r;
1019         struct ceph_snap_context *snapc, *oldest;
1020
1021 retry_locked:
1022         /* writepages currently holds page lock, but if we change that later, */
1023         wait_on_page_writeback(page);
1024
1025         /* check snap context */
1026         BUG_ON(!ci->i_snap_realm);
1027         down_read(&mdsc->snap_rwsem);
1028         BUG_ON(!ci->i_snap_realm->cached_context);
1029         snapc = page_snap_context(page);
1030         if (snapc && snapc != ci->i_head_snapc) {
1031                 /*
1032                  * this page is already dirty in another (older) snap
1033                  * context!  is it writeable now?
1034                  */
1035                 oldest = get_oldest_context(inode, NULL);
1036                 up_read(&mdsc->snap_rwsem);
1037
1038                 if (snapc->seq > oldest->seq) {
1039                         ceph_put_snap_context(oldest);
1040                         dout(" page %p snapc %p not current or oldest\n",
1041                              page, snapc);
1042                         /*
1043                          * queue for writeback, and wait for snapc to
1044                          * be writeable or written
1045                          */
1046                         snapc = ceph_get_snap_context(snapc);
1047                         unlock_page(page);
1048                         ceph_queue_writeback(inode);
1049                         r = wait_event_interruptible(ci->i_cap_wq,
1050                                context_is_writeable_or_written(inode, snapc));
1051                         ceph_put_snap_context(snapc);
1052                         if (r == -ERESTARTSYS)
1053                                 return r;
1054                         return -EAGAIN;
1055                 }
1056                 ceph_put_snap_context(oldest);
1057
1058                 /* yay, writeable, do it now (without dropping page lock) */
1059                 dout(" page %p snapc %p not current, but oldest\n",
1060                      page, snapc);
1061                 if (!clear_page_dirty_for_io(page))
1062                         goto retry_locked;
1063                 r = writepage_nounlock(page, NULL);
1064                 if (r < 0)
1065                         goto fail_nosnap;
1066                 goto retry_locked;
1067         }
1068
1069         if (PageUptodate(page)) {
1070                 dout(" page %p already uptodate\n", page);
1071                 return 0;
1072         }
1073
1074         /* full page? */
1075         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1076                 return 0;
1077
1078         /* past end of file? */
1079         i_size = inode->i_size;   /* caller holds i_mutex */
1080
1081         if (i_size + len > inode->i_sb->s_maxbytes) {
1082                 /* file is too big */
1083                 r = -EINVAL;
1084                 goto fail;
1085         }
1086
1087         if (page_off >= i_size ||
1088             (pos_in_page == 0 && (pos+len) >= i_size &&
1089              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1090                 dout(" zeroing %p 0 - %d and %d - %d\n",
1091                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1092                 zero_user_segments(page,
1093                                    0, pos_in_page,
1094                                    end_in_page, PAGE_CACHE_SIZE);
1095                 return 0;
1096         }
1097
1098         /* we need to read it. */
1099         up_read(&mdsc->snap_rwsem);
1100         r = readpage_nounlock(file, page);
1101         if (r < 0)
1102                 goto fail_nosnap;
1103         goto retry_locked;
1104
1105 fail:
1106         up_read(&mdsc->snap_rwsem);
1107 fail_nosnap:
1108         unlock_page(page);
1109         return r;
1110 }
1111
1112 /*
1113  * We are only allowed to write into/dirty the page if the page is
1114  * clean, or already dirty within the same snap context.
1115  */
1116 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1117                             loff_t pos, unsigned len, unsigned flags,
1118                             struct page **pagep, void **fsdata)
1119 {
1120         struct inode *inode = file_inode(file);
1121         struct page *page;
1122         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1123         int r;
1124
1125         do {
1126                 /* get a page */
1127                 page = grab_cache_page_write_begin(mapping, index, 0);
1128                 if (!page)
1129                         return -ENOMEM;
1130                 *pagep = page;
1131
1132                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1133                      inode, page, (int)pos, (int)len);
1134
1135                 r = ceph_update_writeable_page(file, pos, len, page);
1136         } while (r == -EAGAIN);
1137
1138         return r;
1139 }
1140
1141 /*
1142  * we don't do anything in here that simple_write_end doesn't do
1143  * except adjust dirty page accounting and drop read lock on
1144  * mdsc->snap_rwsem.
1145  */
1146 static int ceph_write_end(struct file *file, struct address_space *mapping,
1147                           loff_t pos, unsigned len, unsigned copied,
1148                           struct page *page, void *fsdata)
1149 {
1150         struct inode *inode = file_inode(file);
1151         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1152         struct ceph_mds_client *mdsc = fsc->mdsc;
1153         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1154         int check_cap = 0;
1155
1156         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1157              inode, page, (int)pos, (int)copied, (int)len);
1158
1159         /* zero the stale part of the page if we did a short copy */
1160         if (copied < len)
1161                 zero_user_segment(page, from+copied, len);
1162
1163         /* did file size increase? */
1164         /* (no need for i_size_read(); we caller holds i_mutex */
1165         if (pos+copied > inode->i_size)
1166                 check_cap = ceph_inode_set_size(inode, pos+copied);
1167
1168         if (!PageUptodate(page))
1169                 SetPageUptodate(page);
1170
1171         set_page_dirty(page);
1172
1173         unlock_page(page);
1174         up_read(&mdsc->snap_rwsem);
1175         page_cache_release(page);
1176
1177         if (check_cap)
1178                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1179
1180         return copied;
1181 }
1182
1183 /*
1184  * we set .direct_IO to indicate direct io is supported, but since we
1185  * intercept O_DIRECT reads and writes early, this function should
1186  * never get called.
1187  */
1188 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1189                               const struct iovec *iov,
1190                               loff_t pos, unsigned long nr_segs)
1191 {
1192         WARN_ON(1);
1193         return -EINVAL;
1194 }
1195
1196 const struct address_space_operations ceph_aops = {
1197         .readpage = ceph_readpage,
1198         .readpages = ceph_readpages,
1199         .writepage = ceph_writepage,
1200         .writepages = ceph_writepages_start,
1201         .write_begin = ceph_write_begin,
1202         .write_end = ceph_write_end,
1203         .set_page_dirty = ceph_set_page_dirty,
1204         .invalidatepage = ceph_invalidatepage,
1205         .releasepage = ceph_releasepage,
1206         .direct_IO = ceph_direct_io,
1207 };
1208
1209
1210 /*
1211  * vm ops
1212  */
1213
1214 /*
1215  * Reuse write_begin here for simplicity.
1216  */
1217 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1218 {
1219         struct inode *inode = file_inode(vma->vm_file);
1220         struct page *page = vmf->page;
1221         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1222         loff_t off = page_offset(page);
1223         loff_t size, len;
1224         int ret;
1225
1226         /* Update time before taking page lock */
1227         file_update_time(vma->vm_file);
1228
1229         size = i_size_read(inode);
1230         if (off + PAGE_CACHE_SIZE <= size)
1231                 len = PAGE_CACHE_SIZE;
1232         else
1233                 len = size & ~PAGE_CACHE_MASK;
1234
1235         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1236              off, len, page, page->index);
1237
1238         lock_page(page);
1239
1240         ret = VM_FAULT_NOPAGE;
1241         if ((off > size) ||
1242             (page->mapping != inode->i_mapping))
1243                 goto out;
1244
1245         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1246         if (ret == 0) {
1247                 /* success.  we'll keep the page locked. */
1248                 set_page_dirty(page);
1249                 up_read(&mdsc->snap_rwsem);
1250                 ret = VM_FAULT_LOCKED;
1251         } else {
1252                 if (ret == -ENOMEM)
1253                         ret = VM_FAULT_OOM;
1254                 else
1255                         ret = VM_FAULT_SIGBUS;
1256         }
1257 out:
1258         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1259         if (ret != VM_FAULT_LOCKED)
1260                 unlock_page(page);
1261         return ret;
1262 }
1263
1264 static struct vm_operations_struct ceph_vmops = {
1265         .fault          = filemap_fault,
1266         .page_mkwrite   = ceph_page_mkwrite,
1267         .remap_pages    = generic_file_remap_pages,
1268 };
1269
1270 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1271 {
1272         struct address_space *mapping = file->f_mapping;
1273
1274         if (!mapping->a_ops->readpage)
1275                 return -ENOEXEC;
1276         file_accessed(file);
1277         vma->vm_ops = &ceph_vmops;
1278         return 0;
1279 }