fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include <linux/ceph/osd_client.h>
  15
  16 /*
  17  * Ceph address space ops.
  18  *
  19  * There are a few funny things going on here.
  20  *
  21  * The page->private field is used to reference a struct
  22  * ceph_snap_context for _every_ dirty page.  This indicates which
  23  * snapshot the page was logically dirtied in, and thus which snap
  24  * context needs to be associated with the osd write during writeback.
  25  *
  26  * Similarly, struct ceph_inode_info maintains a set of counters to
  27  * count dirty pages on the inode.  In the absence of snapshots,
  28  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  29  *
  30  * When a snapshot is taken (that is, when the client receives
  31  * notification that a snapshot was taken), each inode with caps and
  32  * with dirty pages (dirty pages implies there is a cap) gets a new
  33  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  34  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  35  * moved to capsnap->dirty. (Unless a sync write is currently in
  36  * progress.  In that case, the capsnap is said to be "pending", new
  37  * writes cannot start, and the capsnap isn't "finalized" until the
  38  * write completes (or fails) and a final size/mtime for the inode for
  39  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  40  *
  41  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  42  * we look for the first capsnap in i_cap_snaps and write out pages in
  43  * that snap context _only_.  Then we move on to the next capsnap,
  44  * eventually reaching the "live" or "head" context (i.e., pages that
  45  * are not yet snapped) and are writing the most recently dirtied
  46  * pages.
  47  *
  48  * Invalidate and so forth must take care to ensure the dirty page
  49  * accounting is preserved.
  50  */
  51
  52 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  53 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  54         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  55          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  56
  57 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  58 {
  59         if (PagePrivate(page))
  60                 return (void *)page->private;
  61         return NULL;
  62 }
  63
  64 /*
  65  * Dirty a page.  Optimistically adjust accounting, on the assumption
  66  * that we won't race with invalidate.  If we do, readjust.
  67  */
  68 static int ceph_set_page_dirty(struct page *page)
  69 {
  70         struct address_space *mapping = page->mapping;
  71         struct inode *inode;
  72         struct ceph_inode_info *ci;
  73         int undo = 0;
  74         struct ceph_snap_context *snapc;
  75
  76         if (unlikely(!mapping))
  77                 return !TestSetPageDirty(page);
  78
  79         if (TestSetPageDirty(page)) {
  80                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  81                      mapping->host, page, page->index);
  82                 return 0;
  83         }
  84
  85         inode = mapping->host;
  86         ci = ceph_inode(inode);
  87
  88         /*
  89          * Note that we're grabbing a snapc ref here without holding
  90          * any locks!
  91          */
  92         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  93
  94         /* dirty the head */
  95         spin_lock(&ci->i_ceph_lock);
  96         if (ci->i_head_snapc == NULL)
  97                 ci->i_head_snapc = ceph_get_snap_context(snapc);
  98         ++ci->i_wrbuffer_ref_head;
  99         if (ci->i_wrbuffer_ref == 0)
 100                 ihold(inode);
 101         ++ci->i_wrbuffer_ref;
 102         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 103              "snapc %p seq %lld (%d snaps)\n",
 104              mapping->host, page, page->index,
 105              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 106              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 107              snapc, snapc->seq, snapc->num_snaps);
 108         spin_unlock(&ci->i_ceph_lock);
 109
 110         /* now adjust page */
 111         spin_lock_irq(&mapping->tree_lock);
 112         if (page->mapping) {    /* Race with truncate? */
 113                 WARN_ON_ONCE(!PageUptodate(page));
 114                 account_page_dirtied(page, page->mapping);
 115                 radix_tree_tag_set(&mapping->page_tree,
 116                                 page_index(page), PAGECACHE_TAG_DIRTY);
 117
 118                 /*
 119                  * Reference snap context in page->private.  Also set
 120                  * PagePrivate so that we get invalidatepage callback.
 121                  */
 122                 page->private = (unsigned long)snapc;
 123                 SetPagePrivate(page);
 124         } else {
 125                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 126                 undo = 1;
 127         }
 128
 129         spin_unlock_irq(&mapping->tree_lock);
 130
 131         if (undo)
 132                 /* whoops, we failed to dirty the page */
 133                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 134
 135         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 136
 137         BUG_ON(!PageDirty(page));
 138         return 1;
 139 }
 140
 141 /*
 142  * If we are truncating the full page (i.e. offset == 0), adjust the
 143  * dirty page counters appropriately.  Only called if there is private
 144  * data on the page.
 145  */
 146 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 147                                 unsigned int length)
 148 {
 149         struct inode *inode;
 150         struct ceph_inode_info *ci;
 151         struct ceph_snap_context *snapc = page_snap_context(page);
 152
 153         BUG_ON(!PageLocked(page));
 154         BUG_ON(!PagePrivate(page));
 155         BUG_ON(!page->mapping);
 156
 157         inode = page->mapping->host;
 158
 159         /*
 160          * We can get non-dirty pages here due to races between
 161          * set_page_dirty and truncate_complete_page; just spit out a
 162          * warning, in case we end up with accounting problems later.
 163          */
 164         if (!PageDirty(page))
 165                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 166
 167         if (offset == 0 && length == PAGE_CACHE_SIZE)
 168                 ClearPageChecked(page);
 169
 170         ci = ceph_inode(inode);
 171         if (offset == 0 && length == PAGE_CACHE_SIZE) {
 172                 dout("%p invalidatepage %p idx %lu full dirty page\n",
 173                      inode, page, page->index);
 174                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 175                 ceph_put_snap_context(snapc);
 176                 page->private = 0;
 177                 ClearPagePrivate(page);
 178         } else {
 179                 dout("%p invalidatepage %p idx %lu partial dirty page %u(%u)\n",
 180                      inode, page, page->index, offset, length);
 181         }
 182 }
 183
 184 /* just a sanity check */
 185 static int ceph_releasepage(struct page *page, gfp_t g)
 186 {
 187         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 188         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 189         WARN_ON(PageDirty(page));
 190         WARN_ON(PagePrivate(page));
 191         return 0;
 192 }
 193
 194 /*
 195  * read a single page, without unlocking it.
 196  */
 197 static int readpage_nounlock(struct file *filp, struct page *page)
 198 {
 199         struct inode *inode = file_inode(filp);
 200         struct ceph_inode_info *ci = ceph_inode(inode);
 201         struct ceph_osd_client *osdc =
 202                 &ceph_inode_to_client(inode)->client->osdc;
 203         int err = 0;
 204         u64 len = PAGE_CACHE_SIZE;
 205
 206         dout("readpage inode %p file %p page %p index %lu\n",
 207              inode, filp, page, page->index);
 208         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 209                                   (u64) page_offset(page), &len,
 210                                   ci->i_truncate_seq, ci->i_truncate_size,
 211                                   &page, 1, 0);
 212         if (err == -ENOENT)
 213                 err = 0;
 214         if (err < 0) {
 215                 SetPageError(page);
 216                 goto out;
 217         } else if (err < PAGE_CACHE_SIZE) {
 218                 /* zero fill remainder of page */
 219                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 220         }
 221         SetPageUptodate(page);
 222
 223 out:
 224         return err < 0 ? err : 0;
 225 }
 226
 227 static int ceph_readpage(struct file *filp, struct page *page)
 228 {
 229         int r = readpage_nounlock(filp, page);
 230         unlock_page(page);
 231         return r;
 232 }
 233
 234 /*
 235  * Finish an async read(ahead) op.
 236  */
 237 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 238 {
 239         struct inode *inode = req->r_inode;
 240         struct ceph_osd_data *osd_data;
 241         int rc = req->r_result;
 242         int bytes = le32_to_cpu(msg->hdr.data_len);
 243         int num_pages;
 244         int i;
 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         osd_data = osd_req_op_extent_osd_data(req, 0);
 250         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 251         num_pages = calc_pages_for((u64)osd_data->alignment,
 252                                         (u64)osd_data->length);
 253         for (i = 0; i < num_pages; i++) {
 254                 struct page *page = osd_data->pages[i];
 255
 256                 if (bytes < (int)PAGE_CACHE_SIZE) {
 257                         /* zero (remainder of) page */
 258                         int s = bytes < 0 ? 0 : bytes;
 259                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 260                 }
 261                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 262                      page->index);
 263                 flush_dcache_page(page);
 264                 SetPageUptodate(page);
 265                 unlock_page(page);
 266                 page_cache_release(page);
 267                 bytes -= PAGE_CACHE_SIZE;
 268         }
 269         kfree(osd_data->pages);
 270 }
 271
 272 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 273 {
 274         int i;
 275
 276         for (i = 0; i < num_pages; i++)
 277                 unlock_page(pages[i]);
 278 }
 279
 280 /*
 281  * start an async read(ahead) operation.  return nr_pages we submitted
 282  * a read for on success, or negative error code.
 283  */
 284 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 285 {
 286         struct ceph_osd_client *osdc =
 287                 &ceph_inode_to_client(inode)->client->osdc;
 288         struct ceph_inode_info *ci = ceph_inode(inode);
 289         struct page *page = list_entry(page_list->prev, struct page, lru);
 290         struct ceph_vino vino;
 291         struct ceph_osd_request *req;
 292         u64 off;
 293         u64 len;
 294         int i;
 295         struct page **pages;
 296         pgoff_t next_index;
 297         int nr_pages = 0;
 298         int ret;
 299
 300         off = (u64) page_offset(page);
 301
 302         /* count pages */
 303         next_index = page->index;
 304         list_for_each_entry_reverse(page, page_list, lru) {
 305                 if (page->index != next_index)
 306                         break;
 307                 nr_pages++;
 308                 next_index++;
 309                 if (max && nr_pages == max)
 310                         break;
 311         }
 312         len = nr_pages << PAGE_CACHE_SHIFT;
 313         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 314              off, len);
 315         vino = ceph_vino(inode);
 316         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 317                                     1, CEPH_OSD_OP_READ,
 318                                     CEPH_OSD_FLAG_READ, NULL,
 319                                     ci->i_truncate_seq, ci->i_truncate_size,
 320                                     false);
 321         if (IS_ERR(req))
 322                 return PTR_ERR(req);
 323
 324         /* build page vector */
 325         nr_pages = calc_pages_for(0, len);
 326         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 327         ret = -ENOMEM;
 328         if (!pages)
 329                 goto out;
 330         for (i = 0; i < nr_pages; ++i) {
 331                 page = list_entry(page_list->prev, struct page, lru);
 332                 BUG_ON(PageLocked(page));
 333                 list_del(&page->lru);
 334
 335                 dout("start_read %p adding %p idx %lu\n", inode, page,
 336                      page->index);
 337                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 338                                           GFP_NOFS)) {
 339                         page_cache_release(page);
 340                         dout("start_read %p add_to_page_cache failed %p\n",
 341                              inode, page);
 342                         nr_pages = i;
 343                         goto out_pages;
 344                 }
 345                 pages[i] = page;
 346         }
 347         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 348         req->r_callback = finish_read;
 349         req->r_inode = inode;
 350
 351         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 352
 353         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 354         ret = ceph_osdc_start_request(osdc, req, false);
 355         if (ret < 0)
 356                 goto out_pages;
 357         ceph_osdc_put_request(req);
 358         return nr_pages;
 359
 360 out_pages:
 361         ceph_unlock_page_vector(pages, nr_pages);
 362         ceph_release_page_vector(pages, nr_pages);
 363 out:
 364         ceph_osdc_put_request(req);
 365         return ret;
 366 }
 367
 368
 369 /*
 370  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 371  * the caller (VM) cleans them up.
 372  */
 373 static int ceph_readpages(struct file *file, struct address_space *mapping,
 374                           struct list_head *page_list, unsigned nr_pages)
 375 {
 376         struct inode *inode = file_inode(file);
 377         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 378         int rc = 0;
 379         int max = 0;
 380
 381         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 382                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 383                         >> PAGE_SHIFT;
 384
 385         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 386                 file, nr_pages,
 387              max);
 388         while (!list_empty(page_list)) {
 389                 rc = start_read(inode, page_list, max);
 390                 if (rc < 0)
 391                         goto out;
 392                 BUG_ON(rc == 0);
 393         }
 394 out:
 395         dout("readpages %p file %p ret %d\n", inode, file, rc);
 396         return rc;
 397 }
 398
 399 /*
 400  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 401  * only snap context we are allowed to write back.
 402  */
 403 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 404                                                     u64 *snap_size)
 405 {
 406         struct ceph_inode_info *ci = ceph_inode(inode);
 407         struct ceph_snap_context *snapc = NULL;
 408         struct ceph_cap_snap *capsnap = NULL;
 409
 410         spin_lock(&ci->i_ceph_lock);
 411         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 412                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 413                      capsnap->context, capsnap->dirty_pages);
 414                 if (capsnap->dirty_pages) {
 415                         snapc = ceph_get_snap_context(capsnap->context);
 416                         if (snap_size)
 417                                 *snap_size = capsnap->size;
 418                         break;
 419                 }
 420         }
 421         if (!snapc && ci->i_wrbuffer_ref_head) {
 422                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 423                 dout(" head snapc %p has %d dirty pages\n",
 424                      snapc, ci->i_wrbuffer_ref_head);
 425         }
 426         spin_unlock(&ci->i_ceph_lock);
 427         return snapc;
 428 }
 429
 430 /*
 431  * Write a single page, but leave the page locked.
 432  *
 433  * If we get a write error, set the page error bit, but still adjust the
 434  * dirty page accounting (i.e., page is no longer dirty).
 435  */
 436 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 437 {
 438         struct inode *inode;
 439         struct ceph_inode_info *ci;
 440         struct ceph_fs_client *fsc;
 441         struct ceph_osd_client *osdc;
 442         loff_t page_off = page_offset(page);
 443         int len = PAGE_CACHE_SIZE;
 444         loff_t i_size;
 445         int err = 0;
 446         struct ceph_snap_context *snapc, *oldest;
 447         u64 snap_size = 0;
 448         long writeback_stat;
 449
 450         dout("writepage %p idx %lu\n", page, page->index);
 451
 452         if (!page->mapping || !page->mapping->host) {
 453                 dout("writepage %p - no mapping\n", page);
 454                 return -EFAULT;
 455         }
 456         inode = page->mapping->host;
 457         ci = ceph_inode(inode);
 458         fsc = ceph_inode_to_client(inode);
 459         osdc = &fsc->client->osdc;
 460
 461         /* verify this is a writeable snap context */
 462         snapc = page_snap_context(page);
 463         if (snapc == NULL) {
 464                 dout("writepage %p page %p not dirty?\n", inode, page);
 465                 goto out;
 466         }
 467         oldest = get_oldest_context(inode, &snap_size);
 468         if (snapc->seq > oldest->seq) {
 469                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 470                      inode, page, snapc);
 471                 /* we should only noop if called by kswapd */
 472                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 473                 ceph_put_snap_context(oldest);
 474                 goto out;
 475         }
 476         ceph_put_snap_context(oldest);
 477
 478         /* is this a partial page at end of file? */
 479         if (snap_size)
 480                 i_size = snap_size;
 481         else
 482                 i_size = i_size_read(inode);
 483         if (i_size < page_off + len)
 484                 len = i_size - page_off;
 485
 486         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 487              inode, page, page->index, page_off, len, snapc);
 488
 489         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 490         if (writeback_stat >
 491             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 492                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 493
 494         set_page_writeback(page);
 495         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 496                                    &ci->i_layout, snapc,
 497                                    page_off, len,
 498                                    ci->i_truncate_seq, ci->i_truncate_size,
 499                                    &inode->i_mtime, &page, 1);
 500         if (err < 0) {
 501                 dout("writepage setting page/mapping error %d %p\n", err, page);
 502                 SetPageError(page);
 503                 mapping_set_error(&inode->i_data, err);
 504                 if (wbc)
 505                         wbc->pages_skipped++;
 506         } else {
 507                 dout("writepage cleaned page %p\n", page);
 508                 err = 0;  /* vfs expects us to return 0 */
 509         }
 510         page->private = 0;
 511         ClearPagePrivate(page);
 512         end_page_writeback(page);
 513         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 514         ceph_put_snap_context(snapc);  /* page's reference */
 515 out:
 516         return err;
 517 }
 518
 519 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 520 {
 521         int err;
 522         struct inode *inode = page->mapping->host;
 523         BUG_ON(!inode);
 524         ihold(inode);
 525         err = writepage_nounlock(page, wbc);
 526         unlock_page(page);
 527         iput(inode);
 528         return err;
 529 }
 530
 531
 532 /*
 533  * lame release_pages helper.  release_pages() isn't exported to
 534  * modules.
 535  */
 536 static void ceph_release_pages(struct page **pages, int num)
 537 {
 538         struct pagevec pvec;
 539         int i;
 540
 541         pagevec_init(&pvec, 0);
 542         for (i = 0; i < num; i++) {
 543                 if (pagevec_add(&pvec, pages[i]) == 0)
 544                         pagevec_release(&pvec);
 545         }
 546         pagevec_release(&pvec);
 547 }
 548
 549
 550 /*
 551  * async writeback completion handler.
 552  *
 553  * If we get an error, set the mapping error bit, but not the individual
 554  * page error bits.
 555  */
 556 static void writepages_finish(struct ceph_osd_request *req,
 557                               struct ceph_msg *msg)
 558 {
 559         struct inode *inode = req->r_inode;
 560         struct ceph_inode_info *ci = ceph_inode(inode);
 561         struct ceph_osd_data *osd_data;
 562         unsigned wrote;
 563         struct page *page;
 564         int num_pages;
 565         int i;
 566         struct ceph_snap_context *snapc = req->r_snapc;
 567         struct address_space *mapping = inode->i_mapping;
 568         int rc = req->r_result;
 569         u64 bytes = req->r_ops[0].extent.length;
 570         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 571         long writeback_stat;
 572         unsigned issued = ceph_caps_issued(ci);
 573
 574         osd_data = osd_req_op_extent_osd_data(req, 0);
 575         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 576         num_pages = calc_pages_for((u64)osd_data->alignment,
 577                                         (u64)osd_data->length);
 578         if (rc >= 0) {
 579                 /*
 580                  * Assume we wrote the pages we originally sent.  The
 581                  * osd might reply with fewer pages if our writeback
 582                  * raced with a truncation and was adjusted at the osd,
 583                  * so don't believe the reply.
 584                  */
 585                 wrote = num_pages;
 586         } else {
 587                 wrote = 0;
 588                 mapping_set_error(mapping, rc);
 589         }
 590         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 591              inode, rc, bytes, wrote);
 592
 593         /* clean all pages */
 594         for (i = 0; i < num_pages; i++) {
 595                 page = osd_data->pages[i];
 596                 BUG_ON(!page);
 597                 WARN_ON(!PageUptodate(page));
 598
 599                 writeback_stat =
 600                         atomic_long_dec_return(&fsc->writeback_count);
 601                 if (writeback_stat <
 602                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 603                         clear_bdi_congested(&fsc->backing_dev_info,
 604                                             BLK_RW_ASYNC);
 605
 606                 ceph_put_snap_context(page_snap_context(page));
 607                 page->private = 0;
 608                 ClearPagePrivate(page);
 609                 dout("unlocking %d %p\n", i, page);
 610                 end_page_writeback(page);
 611
 612                 /*
 613                  * We lost the cache cap, need to truncate the page before
 614                  * it is unlocked, otherwise we'd truncate it later in the
 615                  * page truncation thread, possibly losing some data that
 616                  * raced its way in
 617                  */
 618                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 619                         generic_error_remove_page(inode->i_mapping, page);
 620
 621                 unlock_page(page);
 622         }
 623         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 624         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 625
 626         ceph_release_pages(osd_data->pages, num_pages);
 627         if (osd_data->pages_from_pool)
 628                 mempool_free(osd_data->pages,
 629                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 630         else
 631                 kfree(osd_data->pages);
 632         ceph_osdc_put_request(req);
 633 }
 634
 635 static struct ceph_osd_request *
 636 ceph_writepages_osd_request(struct inode *inode, u64 offset, u64 *len,
 637                                 struct ceph_snap_context *snapc, int num_ops)
 638 {
 639         struct ceph_fs_client *fsc;
 640         struct ceph_inode_info *ci;
 641         struct ceph_vino vino;
 642
 643         fsc = ceph_inode_to_client(inode);
 644         ci = ceph_inode(inode);
 645         vino = ceph_vino(inode);
 646         /* BUG_ON(vino.snap != CEPH_NOSNAP); */
 647
 648         return ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 649                         vino, offset, len, num_ops, CEPH_OSD_OP_WRITE,
 650                         CEPH_OSD_FLAG_WRITE|CEPH_OSD_FLAG_ONDISK,
 651                         snapc, ci->i_truncate_seq, ci->i_truncate_size, true);
 652 }
 653
 654 /*
 655  * initiate async writeback
 656  */
 657 static int ceph_writepages_start(struct address_space *mapping,
 658                                  struct writeback_control *wbc)
 659 {
 660         struct inode *inode = mapping->host;
 661         struct ceph_inode_info *ci = ceph_inode(inode);
 662         struct ceph_fs_client *fsc;
 663         pgoff_t index, start, end;
 664         int range_whole = 0;
 665         int should_loop = 1;
 666         pgoff_t max_pages = 0, max_pages_ever = 0;
 667         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 668         struct pagevec pvec;
 669         int done = 0;
 670         int rc = 0;
 671         unsigned wsize = 1 << inode->i_blkbits;
 672         struct ceph_osd_request *req = NULL;
 673         int do_sync;
 674         u64 snap_size;
 675
 676         /*
 677          * Include a 'sync' in the OSD request if this is a data
 678          * integrity write (e.g., O_SYNC write or fsync()), or if our
 679          * cap is being revoked.
 680          */
 681         do_sync = wbc->sync_mode == WB_SYNC_ALL;
 682         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 683                 do_sync = 1;
 684         dout("writepages_start %p dosync=%d (mode=%s)\n",
 685              inode, do_sync,
 686              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 687              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 688
 689         fsc = ceph_inode_to_client(inode);
 690         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 691                 pr_warning("writepage_start %p on forced umount\n", inode);
 692                 return -EIO; /* we're in a forced umount, don't write! */
 693         }
 694         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 695                 wsize = fsc->mount_options->wsize;
 696         if (wsize < PAGE_CACHE_SIZE)
 697                 wsize = PAGE_CACHE_SIZE;
 698         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 699
 700         pagevec_init(&pvec, 0);
 701
 702         /* where to start/end? */
 703         if (wbc->range_cyclic) {
 704                 start = mapping->writeback_index; /* Start from prev offset */
 705                 end = -1;
 706                 dout(" cyclic, start at %lu\n", start);
 707         } else {
 708                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 709                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 710                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 711                         range_whole = 1;
 712                 should_loop = 0;
 713                 dout(" not cyclic, %lu to %lu\n", start, end);
 714         }
 715         index = start;
 716
 717 retry:
 718         /* find oldest snap context with dirty data */
 719         ceph_put_snap_context(snapc);
 720         snap_size = 0;
 721         snapc = get_oldest_context(inode, &snap_size);
 722         if (!snapc) {
 723                 /* hmm, why does writepages get called when there
 724                    is no dirty data? */
 725                 dout(" no snap context with dirty data?\n");
 726                 goto out;
 727         }
 728         if (snap_size == 0)
 729                 snap_size = i_size_read(inode);
 730         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 731              snapc, snapc->seq, snapc->num_snaps);
 732         if (last_snapc && snapc != last_snapc) {
 733                 /* if we switched to a newer snapc, restart our scan at the
 734                  * start of the original file range. */
 735                 dout("  snapc differs from last pass, restarting at %lu\n",
 736                      index);
 737                 index = start;
 738         }
 739         last_snapc = snapc;
 740
 741         while (!done && index <= end) {
 742                 int num_ops = do_sync ? 2 : 1;
 743                 struct ceph_vino vino;
 744                 unsigned i;
 745                 int first;
 746                 pgoff_t next;
 747                 int pvec_pages, locked_pages;
 748                 struct page **pages = NULL;
 749                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 750                 struct page *page;
 751                 int want;
 752                 u64 offset, len;
 753                 long writeback_stat;
 754
 755                 next = 0;
 756                 locked_pages = 0;
 757                 max_pages = max_pages_ever;
 758
 759 get_more_pages:
 760                 first = -1;
 761                 want = min(end - index,
 762                            min((pgoff_t)PAGEVEC_SIZE,
 763                                max_pages - (pgoff_t)locked_pages) - 1)
 764                         + 1;
 765                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 766                                                 PAGECACHE_TAG_DIRTY,
 767                                                 want);
 768                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 769                 if (!pvec_pages && !locked_pages)
 770                         break;
 771                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 772                         page = pvec.pages[i];
 773                         dout("? %p idx %lu\n", page, page->index);
 774                         if (locked_pages == 0)
 775                                 lock_page(page);  /* first page */
 776                         else if (!trylock_page(page))
 777                                 break;
 778
 779                         /* only dirty pages, or our accounting breaks */
 780                         if (unlikely(!PageDirty(page)) ||
 781                             unlikely(page->mapping != mapping)) {
 782                                 dout("!dirty or !mapping %p\n", page);
 783                                 unlock_page(page);
 784                                 break;
 785                         }
 786                         if (!wbc->range_cyclic && page->index > end) {
 787                                 dout("end of range %p\n", page);
 788                                 done = 1;
 789                                 unlock_page(page);
 790                                 break;
 791                         }
 792                         if (next && (page->index != next)) {
 793                                 dout("not consecutive %p\n", page);
 794                                 unlock_page(page);
 795                                 break;
 796                         }
 797                         if (wbc->sync_mode != WB_SYNC_NONE) {
 798                                 dout("waiting on writeback %p\n", page);
 799                                 wait_on_page_writeback(page);
 800                         }
 801                         if (page_offset(page) >= snap_size) {
 802                                 dout("%p page eof %llu\n", page, snap_size);
 803                                 done = 1;
 804                                 unlock_page(page);
 805                                 break;
 806                         }
 807                         if (PageWriteback(page)) {
 808                                 dout("%p under writeback\n", page);
 809                                 unlock_page(page);
 810                                 break;
 811                         }
 812
 813                         /* only if matching snap context */
 814                         pgsnapc = page_snap_context(page);
 815                         if (pgsnapc->seq > snapc->seq) {
 816                                 dout("page snapc %p %lld > oldest %p %lld\n",
 817                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 818                                 unlock_page(page);
 819                                 if (!locked_pages)
 820                                         continue; /* keep looking for snap */
 821                                 break;
 822                         }
 823
 824                         if (!clear_page_dirty_for_io(page)) {
 825                                 dout("%p !clear_page_dirty_for_io\n", page);
 826                                 unlock_page(page);
 827                                 break;
 828                         }
 829
 830                         /*
 831                          * We have something to write.  If this is
 832                          * the first locked page this time through,
 833                          * allocate an osd request and a page array
 834                          * that it will use.
 835                          */
 836                         if (locked_pages == 0) {
 837                                 size_t size;
 838
 839                                 BUG_ON(pages);
 840
 841                                 /* prepare async write request */
 842                                 offset = (u64)page_offset(page);
 843                                 len = wsize;
 844                                 req = ceph_writepages_osd_request(inode,
 845                                                         offset, &len, snapc,
 846                                                         num_ops);
 847
 848                                 if (IS_ERR(req)) {
 849                                         rc = PTR_ERR(req);
 850                                         unlock_page(page);
 851                                         break;
 852                                 }
 853
 854                                 req->r_callback = writepages_finish;
 855                                 req->r_inode = inode;
 856
 857                                 max_pages = calc_pages_for(0, (u64)len);
 858                                 size = max_pages * sizeof (*pages);
 859                                 pages = kmalloc(size, GFP_NOFS);
 860                                 if (!pages) {
 861                                         pool = fsc->wb_pagevec_pool;
 862                                         pages = mempool_alloc(pool, GFP_NOFS);
 863                                         BUG_ON(!pages);
 864                                 }
 865                         }
 866
 867                         /* note position of first page in pvec */
 868                         if (first < 0)
 869                                 first = i;
 870                         dout("%p will write page %p idx %lu\n",
 871                              inode, page, page->index);
 872
 873                         writeback_stat =
 874                                atomic_long_inc_return(&fsc->writeback_count);
 875                         if (writeback_stat > CONGESTION_ON_THRESH(
 876                                     fsc->mount_options->congestion_kb)) {
 877                                 set_bdi_congested(&fsc->backing_dev_info,
 878                                                   BLK_RW_ASYNC);
 879                         }
 880
 881                         set_page_writeback(page);
 882                         pages[locked_pages] = page;
 883                         locked_pages++;
 884                         next = page->index + 1;
 885                 }
 886
 887                 /* did we get anything? */
 888                 if (!locked_pages)
 889                         goto release_pvec_pages;
 890                 if (i) {
 891                         int j;
 892                         BUG_ON(!locked_pages || first < 0);
 893
 894                         if (pvec_pages && i == pvec_pages &&
 895                             locked_pages < max_pages) {
 896                                 dout("reached end pvec, trying for more\n");
 897                                 pagevec_reinit(&pvec);
 898                                 goto get_more_pages;
 899                         }
 900
 901                         /* shift unused pages over in the pvec...  we
 902                          * will need to release them below. */
 903                         for (j = i; j < pvec_pages; j++) {
 904                                 dout(" pvec leftover page %p\n",
 905                                      pvec.pages[j]);
 906                                 pvec.pages[j-i+first] = pvec.pages[j];
 907                         }
 908                         pvec.nr -= i-first;
 909                 }
 910
 911                 /* Format the osd request message and submit the write */
 912
 913                 offset = page_offset(pages[0]);
 914                 len = min(snap_size - offset,
 915                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 916                 dout("writepages got %d pages at %llu~%llu\n",
 917                      locked_pages, offset, len);
 918
 919                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 920                                                         !!pool, false);
 921
 922                 pages = NULL;   /* request message now owns the pages array */
 923                 pool = NULL;
 924
 925                 /* Update the write op length in case we changed it */
 926
 927                 osd_req_op_extent_update(req, 0, len);
 928
 929                 vino = ceph_vino(inode);
 930                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 931                                         &inode->i_mtime);
 932
 933                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 934                 BUG_ON(rc);
 935                 req = NULL;
 936
 937                 /* continue? */
 938                 index = next;
 939                 wbc->nr_to_write -= locked_pages;
 940                 if (wbc->nr_to_write <= 0)
 941                         done = 1;
 942
 943 release_pvec_pages:
 944                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 945                      pvec.nr ? pvec.pages[0] : NULL);
 946                 pagevec_release(&pvec);
 947
 948                 if (locked_pages && !done)
 949                         goto retry;
 950         }
 951
 952         if (should_loop && !done) {
 953                 /* more to do; loop back to beginning of file */
 954                 dout("writepages looping back to beginning of file\n");
 955                 should_loop = 0;
 956                 index = 0;
 957                 goto retry;
 958         }
 959
 960         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 961                 mapping->writeback_index = index;
 962
 963 out:
 964         if (req)
 965                 ceph_osdc_put_request(req);
 966         ceph_put_snap_context(snapc);
 967         dout("writepages done, rc = %d\n", rc);
 968         return rc;
 969 }
 970
 971
 972
 973 /*
 974  * See if a given @snapc is either writeable, or already written.
 975  */
 976 static int context_is_writeable_or_written(struct inode *inode,
 977                                            struct ceph_snap_context *snapc)
 978 {
 979         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 980         int ret = !oldest || snapc->seq <= oldest->seq;
 981
 982         ceph_put_snap_context(oldest);
 983         return ret;
 984 }
 985
 986 /*
 987  * We are only allowed to write into/dirty the page if the page is
 988  * clean, or already dirty within the same snap context.
 989  *
 990  * called with page locked.
 991  * return success with page locked,
 992  * or any failure (incl -EAGAIN) with page unlocked.
 993  */
 994 static int ceph_update_writeable_page(struct file *file,
 995                             loff_t pos, unsigned len,
 996                             struct page *page)
 997 {
 998         struct inode *inode = file_inode(file);
 999         struct ceph_inode_info *ci = ceph_inode(inode);
1000         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1001         loff_t page_off = pos & PAGE_CACHE_MASK;
1002         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1003         int end_in_page = pos_in_page + len;
1004         loff_t i_size;
1005         int r;
1006         struct ceph_snap_context *snapc, *oldest;
1007
1008 retry_locked:
1009         /* writepages currently holds page lock, but if we change that later, */
1010         wait_on_page_writeback(page);
1011
1012         /* check snap context */
1013         BUG_ON(!ci->i_snap_realm);
1014         down_read(&mdsc->snap_rwsem);
1015         BUG_ON(!ci->i_snap_realm->cached_context);
1016         snapc = page_snap_context(page);
1017         if (snapc && snapc != ci->i_head_snapc) {
1018                 /*
1019                  * this page is already dirty in another (older) snap
1020                  * context!  is it writeable now?
1021                  */
1022                 oldest = get_oldest_context(inode, NULL);
1023                 up_read(&mdsc->snap_rwsem);
1024
1025                 if (snapc->seq > oldest->seq) {
1026                         ceph_put_snap_context(oldest);
1027                         dout(" page %p snapc %p not current or oldest\n",
1028                              page, snapc);
1029                         /*
1030                          * queue for writeback, and wait for snapc to
1031                          * be writeable or written
1032                          */
1033                         snapc = ceph_get_snap_context(snapc);
1034                         unlock_page(page);
1035                         ceph_queue_writeback(inode);
1036                         r = wait_event_interruptible(ci->i_cap_wq,
1037                                context_is_writeable_or_written(inode, snapc));
1038                         ceph_put_snap_context(snapc);
1039                         if (r == -ERESTARTSYS)
1040                                 return r;
1041                         return -EAGAIN;
1042                 }
1043                 ceph_put_snap_context(oldest);
1044
1045                 /* yay, writeable, do it now (without dropping page lock) */
1046                 dout(" page %p snapc %p not current, but oldest\n",
1047                      page, snapc);
1048                 if (!clear_page_dirty_for_io(page))
1049                         goto retry_locked;
1050                 r = writepage_nounlock(page, NULL);
1051                 if (r < 0)
1052                         goto fail_nosnap;
1053                 goto retry_locked;
1054         }
1055
1056         if (PageUptodate(page)) {
1057                 dout(" page %p already uptodate\n", page);
1058                 return 0;
1059         }
1060
1061         /* full page? */
1062         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1063                 return 0;
1064
1065         /* past end of file? */
1066         i_size = inode->i_size;   /* caller holds i_mutex */
1067
1068         if (i_size + len > inode->i_sb->s_maxbytes) {
1069                 /* file is too big */
1070                 r = -EINVAL;
1071                 goto fail;
1072         }
1073
1074         if (page_off >= i_size ||
1075             (pos_in_page == 0 && (pos+len) >= i_size &&
1076              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1077                 dout(" zeroing %p 0 - %d and %d - %d\n",
1078                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1079                 zero_user_segments(page,
1080                                    0, pos_in_page,
1081                                    end_in_page, PAGE_CACHE_SIZE);
1082                 return 0;
1083         }
1084
1085         /* we need to read it. */
1086         up_read(&mdsc->snap_rwsem);
1087         r = readpage_nounlock(file, page);
1088         if (r < 0)
1089                 goto fail_nosnap;
1090         goto retry_locked;
1091
1092 fail:
1093         up_read(&mdsc->snap_rwsem);
1094 fail_nosnap:
1095         unlock_page(page);
1096         return r;
1097 }
1098
1099 /*
1100  * We are only allowed to write into/dirty the page if the page is
1101  * clean, or already dirty within the same snap context.
1102  */
1103 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1104                             loff_t pos, unsigned len, unsigned flags,
1105                             struct page **pagep, void **fsdata)
1106 {
1107         struct inode *inode = file_inode(file);
1108         struct page *page;
1109         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1110         int r;
1111
1112         do {
1113                 /* get a page */
1114                 page = grab_cache_page_write_begin(mapping, index, 0);
1115                 if (!page)
1116                         return -ENOMEM;
1117                 *pagep = page;
1118
1119                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1120                      inode, page, (int)pos, (int)len);
1121
1122                 r = ceph_update_writeable_page(file, pos, len, page);
1123         } while (r == -EAGAIN);
1124
1125         return r;
1126 }
1127
1128 /*
1129  * we don't do anything in here that simple_write_end doesn't do
1130  * except adjust dirty page accounting and drop read lock on
1131  * mdsc->snap_rwsem.
1132  */
1133 static int ceph_write_end(struct file *file, struct address_space *mapping,
1134                           loff_t pos, unsigned len, unsigned copied,
1135                           struct page *page, void *fsdata)
1136 {
1137         struct inode *inode = file_inode(file);
1138         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1139         struct ceph_mds_client *mdsc = fsc->mdsc;
1140         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1141         int check_cap = 0;
1142
1143         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1144              inode, page, (int)pos, (int)copied, (int)len);
1145
1146         /* zero the stale part of the page if we did a short copy */
1147         if (copied < len)
1148                 zero_user_segment(page, from+copied, len);
1149
1150         /* did file size increase? */
1151         /* (no need for i_size_read(); we caller holds i_mutex */
1152         if (pos+copied > inode->i_size)
1153                 check_cap = ceph_inode_set_size(inode, pos+copied);
1154
1155         if (!PageUptodate(page))
1156                 SetPageUptodate(page);
1157
1158         set_page_dirty(page);
1159
1160         unlock_page(page);
1161         up_read(&mdsc->snap_rwsem);
1162         page_cache_release(page);
1163
1164         if (check_cap)
1165                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1166
1167         return copied;
1168 }
1169
1170 /*
1171  * we set .direct_IO to indicate direct io is supported, but since we
1172  * intercept O_DIRECT reads and writes early, this function should
1173  * never get called.
1174  */
1175 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1176                               const struct iovec *iov,
1177                               loff_t pos, unsigned long nr_segs)
1178 {
1179         WARN_ON(1);
1180         return -EINVAL;
1181 }
1182
1183 const struct address_space_operations ceph_aops = {
1184         .readpage = ceph_readpage,
1185         .readpages = ceph_readpages,
1186         .writepage = ceph_writepage,
1187         .writepages = ceph_writepages_start,
1188         .write_begin = ceph_write_begin,
1189         .write_end = ceph_write_end,
1190         .set_page_dirty = ceph_set_page_dirty,
1191         .invalidatepage = ceph_invalidatepage,
1192         .releasepage = ceph_releasepage,
1193         .direct_IO = ceph_direct_io,
1194 };
1195
1196
1197 /*
1198  * vm ops
1199  */
1200
1201 /*
1202  * Reuse write_begin here for simplicity.
1203  */
1204 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1205 {
1206         struct inode *inode = file_inode(vma->vm_file);
1207         struct page *page = vmf->page;
1208         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1209         loff_t off = page_offset(page);
1210         loff_t size, len;
1211         int ret;
1212
1213         /* Update time before taking page lock */
1214         file_update_time(vma->vm_file);
1215
1216         size = i_size_read(inode);
1217         if (off + PAGE_CACHE_SIZE <= size)
1218                 len = PAGE_CACHE_SIZE;
1219         else
1220                 len = size & ~PAGE_CACHE_MASK;
1221
1222         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1223              off, len, page, page->index);
1224
1225         lock_page(page);
1226
1227         ret = VM_FAULT_NOPAGE;
1228         if ((off > size) ||
1229             (page->mapping != inode->i_mapping))
1230                 goto out;
1231
1232         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1233         if (ret == 0) {
1234                 /* success.  we'll keep the page locked. */
1235                 set_page_dirty(page);
1236                 up_read(&mdsc->snap_rwsem);
1237                 ret = VM_FAULT_LOCKED;
1238         } else {
1239                 if (ret == -ENOMEM)
1240                         ret = VM_FAULT_OOM;
1241                 else
1242                         ret = VM_FAULT_SIGBUS;
1243         }
1244 out:
1245         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1246         if (ret != VM_FAULT_LOCKED)
1247                 unlock_page(page);
1248         return ret;
1249 }
1250
1251 static struct vm_operations_struct ceph_vmops = {
1252         .fault          = filemap_fault,
1253         .page_mkwrite   = ceph_page_mkwrite,
1254         .remap_pages    = generic_file_remap_pages,
1255 };
1256
1257 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1258 {
1259         struct address_space *mapping = file->f_mapping;
1260
1261         if (!mapping->a_ops->readpage)
1262                 return -ENOEXEC;
1263         file_accessed(file);
1264         vma->vm_ops = &ceph_vmops;
1265         return 0;
1266 }