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         inode = page->mapping->host;
 154
 155         /*
 156          * We can get non-dirty pages here due to races between
 157          * set_page_dirty and truncate_complete_page; just spit out a
 158          * warning, in case we end up with accounting problems later.
 159          */
 160         if (!PageDirty(page))
 161                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 162
 163         if (offset == 0 && length == PAGE_CACHE_SIZE)
 164                 ClearPageChecked(page);
 165
 166         ci = ceph_inode(inode);
 167         if (offset == 0 && length == PAGE_CACHE_SIZE) {
 168                 dout("%p invalidatepage %p idx %lu full dirty page\n",
 169                      inode, page, page->index);
 170                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 171                 ceph_put_snap_context(snapc);
 172                 page->private = 0;
 173                 ClearPagePrivate(page);
 174         } else {
 175                 dout("%p invalidatepage %p idx %lu partial dirty page %u(%u)\n",
 176                      inode, page, page->index, offset, length);
 177         }
 178 }
 179
 180 /* just a sanity check */
 181 static int ceph_releasepage(struct page *page, gfp_t g)
 182 {
 183         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 184         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 185         WARN_ON(PageDirty(page));
 186         WARN_ON(PagePrivate(page));
 187         return 0;
 188 }
 189
 190 /*
 191  * read a single page, without unlocking it.
 192  */
 193 static int readpage_nounlock(struct file *filp, struct page *page)
 194 {
 195         struct inode *inode = file_inode(filp);
 196         struct ceph_inode_info *ci = ceph_inode(inode);
 197         struct ceph_osd_client *osdc =
 198                 &ceph_inode_to_client(inode)->client->osdc;
 199         int err = 0;
 200         u64 len = PAGE_CACHE_SIZE;
 201
 202         dout("readpage inode %p file %p page %p index %lu\n",
 203              inode, filp, page, page->index);
 204         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 205                                   (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 if (err < PAGE_CACHE_SIZE) {
 214                 /* zero fill remainder of page */
 215                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 216         }
 217         SetPageUptodate(page);
 218
 219 out:
 220         return err < 0 ? err : 0;
 221 }
 222
 223 static int ceph_readpage(struct file *filp, struct page *page)
 224 {
 225         int r = readpage_nounlock(filp, page);
 226         unlock_page(page);
 227         return r;
 228 }
 229
 230 /*
 231  * Finish an async read(ahead) op.
 232  */
 233 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 234 {
 235         struct inode *inode = req->r_inode;
 236         struct ceph_osd_data *osd_data;
 237         int rc = req->r_result;
 238         int bytes = le32_to_cpu(msg->hdr.data_len);
 239         int num_pages;
 240         int i;
 241
 242         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 243
 244         /* unlock all pages, zeroing any data we didn't read */
 245         osd_data = osd_req_op_extent_osd_data(req, 0);
 246         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 247         num_pages = calc_pages_for((u64)osd_data->alignment,
 248                                         (u64)osd_data->length);
 249         for (i = 0; i < num_pages; i++) {
 250                 struct page *page = osd_data->pages[i];
 251
 252                 if (bytes < (int)PAGE_CACHE_SIZE) {
 253                         /* zero (remainder of) page */
 254                         int s = bytes < 0 ? 0 : bytes;
 255                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 256                 }
 257                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 258                      page->index);
 259                 flush_dcache_page(page);
 260                 SetPageUptodate(page);
 261                 unlock_page(page);
 262                 page_cache_release(page);
 263                 bytes -= PAGE_CACHE_SIZE;
 264         }
 265         kfree(osd_data->pages);
 266 }
 267
 268 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 269 {
 270         int i;
 271
 272         for (i = 0; i < num_pages; i++)
 273                 unlock_page(pages[i]);
 274 }
 275
 276 /*
 277  * start an async read(ahead) operation.  return nr_pages we submitted
 278  * a read for on success, or negative error code.
 279  */
 280 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 281 {
 282         struct ceph_osd_client *osdc =
 283                 &ceph_inode_to_client(inode)->client->osdc;
 284         struct ceph_inode_info *ci = ceph_inode(inode);
 285         struct page *page = list_entry(page_list->prev, struct page, lru);
 286         struct ceph_vino vino;
 287         struct ceph_osd_request *req;
 288         u64 off;
 289         u64 len;
 290         int i;
 291         struct page **pages;
 292         pgoff_t next_index;
 293         int nr_pages = 0;
 294         int ret;
 295
 296         off = (u64) page_offset(page);
 297
 298         /* count pages */
 299         next_index = page->index;
 300         list_for_each_entry_reverse(page, page_list, lru) {
 301                 if (page->index != next_index)
 302                         break;
 303                 nr_pages++;
 304                 next_index++;
 305                 if (max && nr_pages == max)
 306                         break;
 307         }
 308         len = nr_pages << PAGE_CACHE_SHIFT;
 309         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 310              off, len);
 311         vino = ceph_vino(inode);
 312         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 313                                     1, CEPH_OSD_OP_READ,
 314                                     CEPH_OSD_FLAG_READ, NULL,
 315                                     ci->i_truncate_seq, ci->i_truncate_size,
 316                                     false);
 317         if (IS_ERR(req))
 318                 return PTR_ERR(req);
 319
 320         /* build page vector */
 321         nr_pages = calc_pages_for(0, len);
 322         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 323         ret = -ENOMEM;
 324         if (!pages)
 325                 goto out;
 326         for (i = 0; i < nr_pages; ++i) {
 327                 page = list_entry(page_list->prev, struct page, lru);
 328                 BUG_ON(PageLocked(page));
 329                 list_del(&page->lru);
 330
 331                 dout("start_read %p adding %p idx %lu\n", inode, page,
 332                      page->index);
 333                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 334                                           GFP_NOFS)) {
 335                         page_cache_release(page);
 336                         dout("start_read %p add_to_page_cache failed %p\n",
 337                              inode, page);
 338                         nr_pages = i;
 339                         goto out_pages;
 340                 }
 341                 pages[i] = page;
 342         }
 343         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 344         req->r_callback = finish_read;
 345         req->r_inode = inode;
 346
 347         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 348
 349         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 350         ret = ceph_osdc_start_request(osdc, req, false);
 351         if (ret < 0)
 352                 goto out_pages;
 353         ceph_osdc_put_request(req);
 354         return nr_pages;
 355
 356 out_pages:
 357         ceph_unlock_page_vector(pages, nr_pages);
 358         ceph_release_page_vector(pages, nr_pages);
 359 out:
 360         ceph_osdc_put_request(req);
 361         return ret;
 362 }
 363
 364
 365 /*
 366  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 367  * the caller (VM) cleans them up.
 368  */
 369 static int ceph_readpages(struct file *file, struct address_space *mapping,
 370                           struct list_head *page_list, unsigned nr_pages)
 371 {
 372         struct inode *inode = file_inode(file);
 373         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 374         int rc = 0;
 375         int max = 0;
 376
 377         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 378                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 379                         >> PAGE_SHIFT;
 380
 381         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 382                 file, nr_pages,
 383              max);
 384         while (!list_empty(page_list)) {
 385                 rc = start_read(inode, page_list, max);
 386                 if (rc < 0)
 387                         goto out;
 388                 BUG_ON(rc == 0);
 389         }
 390 out:
 391         dout("readpages %p file %p ret %d\n", inode, file, rc);
 392         return rc;
 393 }
 394
 395 /*
 396  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 397  * only snap context we are allowed to write back.
 398  */
 399 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 400                                                     u64 *snap_size)
 401 {
 402         struct ceph_inode_info *ci = ceph_inode(inode);
 403         struct ceph_snap_context *snapc = NULL;
 404         struct ceph_cap_snap *capsnap = NULL;
 405
 406         spin_lock(&ci->i_ceph_lock);
 407         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 408                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 409                      capsnap->context, capsnap->dirty_pages);
 410                 if (capsnap->dirty_pages) {
 411                         snapc = ceph_get_snap_context(capsnap->context);
 412                         if (snap_size)
 413                                 *snap_size = capsnap->size;
 414                         break;
 415                 }
 416         }
 417         if (!snapc && ci->i_wrbuffer_ref_head) {
 418                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 419                 dout(" head snapc %p has %d dirty pages\n",
 420                      snapc, ci->i_wrbuffer_ref_head);
 421         }
 422         spin_unlock(&ci->i_ceph_lock);
 423         return snapc;
 424 }
 425
 426 /*
 427  * Write a single page, but leave the page locked.
 428  *
 429  * If we get a write error, set the page error bit, but still adjust the
 430  * dirty page accounting (i.e., page is no longer dirty).
 431  */
 432 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 433 {
 434         struct inode *inode;
 435         struct ceph_inode_info *ci;
 436         struct ceph_fs_client *fsc;
 437         struct ceph_osd_client *osdc;
 438         struct ceph_snap_context *snapc, *oldest;
 439         loff_t page_off = page_offset(page);
 440         long writeback_stat;
 441         u64 truncate_size, snap_size = 0;
 442         u32 truncate_seq;
 443         int err = 0, len = PAGE_CACHE_SIZE;
 444
 445         dout("writepage %p idx %lu\n", page, page->index);
 446
 447         if (!page->mapping || !page->mapping->host) {
 448                 dout("writepage %p - no mapping\n", page);
 449                 return -EFAULT;
 450         }
 451         inode = page->mapping->host;
 452         ci = ceph_inode(inode);
 453         fsc = ceph_inode_to_client(inode);
 454         osdc = &fsc->client->osdc;
 455
 456         /* verify this is a writeable snap context */
 457         snapc = page_snap_context(page);
 458         if (snapc == NULL) {
 459                 dout("writepage %p page %p not dirty?\n", inode, page);
 460                 goto out;
 461         }
 462         oldest = get_oldest_context(inode, &snap_size);
 463         if (snapc->seq > oldest->seq) {
 464                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 465                      inode, page, snapc);
 466                 /* we should only noop if called by kswapd */
 467                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 468                 ceph_put_snap_context(oldest);
 469                 goto out;
 470         }
 471         ceph_put_snap_context(oldest);
 472
 473         spin_lock(&ci->i_ceph_lock);
 474         truncate_seq = ci->i_truncate_seq;
 475         truncate_size = ci->i_truncate_size;
 476         if (!snap_size)
 477                 snap_size = i_size_read(inode);
 478         spin_unlock(&ci->i_ceph_lock);
 479
 480         /* is this a partial page at end of file? */
 481         if (page_off >= snap_size) {
 482                 dout("%p page eof %llu\n", page, snap_size);
 483                 goto out;
 484         }
 485         if (snap_size < page_off + len)
 486                 len = snap_size - page_off;
 487
 488         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 489              inode, page, page->index, page_off, len, snapc);
 490
 491         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 492         if (writeback_stat >
 493             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 494                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 495
 496         set_page_writeback(page);
 497         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 498                                    &ci->i_layout, snapc,
 499                                    page_off, len,
 500                                    truncate_seq, truncate_size,
 501                                    &inode->i_mtime, &page, 1);
 502         if (err < 0) {
 503                 dout("writepage setting page/mapping error %d %p\n", err, page);
 504                 SetPageError(page);
 505                 mapping_set_error(&inode->i_data, err);
 506                 if (wbc)
 507                         wbc->pages_skipped++;
 508         } else {
 509                 dout("writepage cleaned page %p\n", page);
 510                 err = 0;  /* vfs expects us to return 0 */
 511         }
 512         page->private = 0;
 513         ClearPagePrivate(page);
 514         end_page_writeback(page);
 515         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 516         ceph_put_snap_context(snapc);  /* page's reference */
 517 out:
 518         return err;
 519 }
 520
 521 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 522 {
 523         int err;
 524         struct inode *inode = page->mapping->host;
 525         BUG_ON(!inode);
 526         ihold(inode);
 527         err = writepage_nounlock(page, wbc);
 528         unlock_page(page);
 529         iput(inode);
 530         return err;
 531 }
 532
 533
 534 /*
 535  * lame release_pages helper.  release_pages() isn't exported to
 536  * modules.
 537  */
 538 static void ceph_release_pages(struct page **pages, int num)
 539 {
 540         struct pagevec pvec;
 541         int i;
 542
 543         pagevec_init(&pvec, 0);
 544         for (i = 0; i < num; i++) {
 545                 if (pagevec_add(&pvec, pages[i]) == 0)
 546                         pagevec_release(&pvec);
 547         }
 548         pagevec_release(&pvec);
 549 }
 550
 551
 552 /*
 553  * async writeback completion handler.
 554  *
 555  * If we get an error, set the mapping error bit, but not the individual
 556  * page error bits.
 557  */
 558 static void writepages_finish(struct ceph_osd_request *req,
 559                               struct ceph_msg *msg)
 560 {
 561         struct inode *inode = req->r_inode;
 562         struct ceph_inode_info *ci = ceph_inode(inode);
 563         struct ceph_osd_data *osd_data;
 564         unsigned wrote;
 565         struct page *page;
 566         int num_pages;
 567         int i;
 568         struct ceph_snap_context *snapc = req->r_snapc;
 569         struct address_space *mapping = inode->i_mapping;
 570         int rc = req->r_result;
 571         u64 bytes = req->r_ops[0].extent.length;
 572         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 573         long writeback_stat;
 574         unsigned issued = ceph_caps_issued(ci);
 575
 576         osd_data = osd_req_op_extent_osd_data(req, 0);
 577         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 578         num_pages = calc_pages_for((u64)osd_data->alignment,
 579                                         (u64)osd_data->length);
 580         if (rc >= 0) {
 581                 /*
 582                  * Assume we wrote the pages we originally sent.  The
 583                  * osd might reply with fewer pages if our writeback
 584                  * raced with a truncation and was adjusted at the osd,
 585                  * so don't believe the reply.
 586                  */
 587                 wrote = num_pages;
 588         } else {
 589                 wrote = 0;
 590                 mapping_set_error(mapping, rc);
 591         }
 592         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 593              inode, rc, bytes, wrote);
 594
 595         /* clean all pages */
 596         for (i = 0; i < num_pages; i++) {
 597                 page = osd_data->pages[i];
 598                 BUG_ON(!page);
 599                 WARN_ON(!PageUptodate(page));
 600
 601                 writeback_stat =
 602                         atomic_long_dec_return(&fsc->writeback_count);
 603                 if (writeback_stat <
 604                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 605                         clear_bdi_congested(&fsc->backing_dev_info,
 606                                             BLK_RW_ASYNC);
 607
 608                 ceph_put_snap_context(page_snap_context(page));
 609                 page->private = 0;
 610                 ClearPagePrivate(page);
 611                 dout("unlocking %d %p\n", i, page);
 612                 end_page_writeback(page);
 613
 614                 /*
 615                  * We lost the cache cap, need to truncate the page before
 616                  * it is unlocked, otherwise we'd truncate it later in the
 617                  * page truncation thread, possibly losing some data that
 618                  * raced its way in
 619                  */
 620                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 621                         generic_error_remove_page(inode->i_mapping, page);
 622
 623                 unlock_page(page);
 624         }
 625         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 626         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 627
 628         ceph_release_pages(osd_data->pages, num_pages);
 629         if (osd_data->pages_from_pool)
 630                 mempool_free(osd_data->pages,
 631                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 632         else
 633                 kfree(osd_data->pages);
 634         ceph_osdc_put_request(req);
 635 }
 636
 637 /*
 638  * initiate async writeback
 639  */
 640 static int ceph_writepages_start(struct address_space *mapping,
 641                                  struct writeback_control *wbc)
 642 {
 643         struct inode *inode = mapping->host;
 644         struct ceph_inode_info *ci = ceph_inode(inode);
 645         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 646         struct ceph_vino vino = ceph_vino(inode);
 647         pgoff_t index, start, end;
 648         int range_whole = 0;
 649         int should_loop = 1;
 650         pgoff_t max_pages = 0, max_pages_ever = 0;
 651         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 652         struct pagevec pvec;
 653         int done = 0;
 654         int rc = 0;
 655         unsigned wsize = 1 << inode->i_blkbits;
 656         struct ceph_osd_request *req = NULL;
 657         int do_sync;
 658         u64 truncate_size, snap_size;
 659         u32 truncate_seq;
 660
 661         /*
 662          * Include a 'sync' in the OSD request if this is a data
 663          * integrity write (e.g., O_SYNC write or fsync()), or if our
 664          * cap is being revoked.
 665          */
 666         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 667                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 668                 do_sync = 1;
 669         dout("writepages_start %p dosync=%d (mode=%s)\n",
 670              inode, do_sync,
 671              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 672              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 673
 674         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 675                 pr_warning("writepage_start %p on forced umount\n", inode);
 676                 return -EIO; /* we're in a forced umount, don't write! */
 677         }
 678         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 679                 wsize = fsc->mount_options->wsize;
 680         if (wsize < PAGE_CACHE_SIZE)
 681                 wsize = PAGE_CACHE_SIZE;
 682         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 683
 684         pagevec_init(&pvec, 0);
 685
 686         /* where to start/end? */
 687         if (wbc->range_cyclic) {
 688                 start = mapping->writeback_index; /* Start from prev offset */
 689                 end = -1;
 690                 dout(" cyclic, start at %lu\n", start);
 691         } else {
 692                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 693                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 694                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 695                         range_whole = 1;
 696                 should_loop = 0;
 697                 dout(" not cyclic, %lu to %lu\n", start, end);
 698         }
 699         index = start;
 700
 701 retry:
 702         /* find oldest snap context with dirty data */
 703         ceph_put_snap_context(snapc);
 704         snap_size = 0;
 705         snapc = get_oldest_context(inode, &snap_size);
 706         if (!snapc) {
 707                 /* hmm, why does writepages get called when there
 708                    is no dirty data? */
 709                 dout(" no snap context with dirty data?\n");
 710                 goto out;
 711         }
 712         if (snap_size == 0)
 713                 snap_size = i_size_read(inode);
 714         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 715              snapc, snapc->seq, snapc->num_snaps);
 716
 717         spin_lock(&ci->i_ceph_lock);
 718         truncate_seq = ci->i_truncate_seq;
 719         truncate_size = ci->i_truncate_size;
 720         if (!snap_size)
 721                 snap_size = i_size_read(inode);
 722         spin_unlock(&ci->i_ceph_lock);
 723
 724         if (last_snapc && snapc != last_snapc) {
 725                 /* if we switched to a newer snapc, restart our scan at the
 726                  * start of the original file range. */
 727                 dout("  snapc differs from last pass, restarting at %lu\n",
 728                      index);
 729                 index = start;
 730         }
 731         last_snapc = snapc;
 732
 733         while (!done && index <= end) {
 734                 int num_ops = do_sync ? 2 : 1;
 735                 unsigned i;
 736                 int first;
 737                 pgoff_t next;
 738                 int pvec_pages, locked_pages;
 739                 struct page **pages = NULL;
 740                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 741                 struct page *page;
 742                 int want;
 743                 u64 offset, len;
 744                 long writeback_stat;
 745
 746                 next = 0;
 747                 locked_pages = 0;
 748                 max_pages = max_pages_ever;
 749
 750 get_more_pages:
 751                 first = -1;
 752                 want = min(end - index,
 753                            min((pgoff_t)PAGEVEC_SIZE,
 754                                max_pages - (pgoff_t)locked_pages) - 1)
 755                         + 1;
 756                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 757                                                 PAGECACHE_TAG_DIRTY,
 758                                                 want);
 759                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 760                 if (!pvec_pages && !locked_pages)
 761                         break;
 762                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 763                         page = pvec.pages[i];
 764                         dout("? %p idx %lu\n", page, page->index);
 765                         if (locked_pages == 0)
 766                                 lock_page(page);  /* first page */
 767                         else if (!trylock_page(page))
 768                                 break;
 769
 770                         /* only dirty pages, or our accounting breaks */
 771                         if (unlikely(!PageDirty(page)) ||
 772                             unlikely(page->mapping != mapping)) {
 773                                 dout("!dirty or !mapping %p\n", page);
 774                                 unlock_page(page);
 775                                 break;
 776                         }
 777                         if (!wbc->range_cyclic && page->index > end) {
 778                                 dout("end of range %p\n", page);
 779                                 done = 1;
 780                                 unlock_page(page);
 781                                 break;
 782                         }
 783                         if (next && (page->index != next)) {
 784                                 dout("not consecutive %p\n", page);
 785                                 unlock_page(page);
 786                                 break;
 787                         }
 788                         if (wbc->sync_mode != WB_SYNC_NONE) {
 789                                 dout("waiting on writeback %p\n", page);
 790                                 wait_on_page_writeback(page);
 791                         }
 792                         if (page_offset(page) >= snap_size) {
 793                                 dout("%p page eof %llu\n", page, snap_size);
 794                                 done = 1;
 795                                 unlock_page(page);
 796                                 break;
 797                         }
 798                         if (PageWriteback(page)) {
 799                                 dout("%p under writeback\n", page);
 800                                 unlock_page(page);
 801                                 break;
 802                         }
 803
 804                         /* only if matching snap context */
 805                         pgsnapc = page_snap_context(page);
 806                         if (pgsnapc->seq > snapc->seq) {
 807                                 dout("page snapc %p %lld > oldest %p %lld\n",
 808                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 809                                 unlock_page(page);
 810                                 if (!locked_pages)
 811                                         continue; /* keep looking for snap */
 812                                 break;
 813                         }
 814
 815                         if (!clear_page_dirty_for_io(page)) {
 816                                 dout("%p !clear_page_dirty_for_io\n", page);
 817                                 unlock_page(page);
 818                                 break;
 819                         }
 820
 821                         /*
 822                          * We have something to write.  If this is
 823                          * the first locked page this time through,
 824                          * allocate an osd request and a page array
 825                          * that it will use.
 826                          */
 827                         if (locked_pages == 0) {
 828                                 BUG_ON(pages);
 829                                 /* prepare async write request */
 830                                 offset = (u64)page_offset(page);
 831                                 len = wsize;
 832                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 833                                                         &ci->i_layout, vino,
 834                                                         offset, &len, num_ops,
 835                                                         CEPH_OSD_OP_WRITE,
 836                                                         CEPH_OSD_FLAG_WRITE |
 837                                                         CEPH_OSD_FLAG_ONDISK,
 838                                                         snapc, truncate_seq,
 839                                                         truncate_size, true);
 840                                 if (IS_ERR(req)) {
 841                                         rc = PTR_ERR(req);
 842                                         unlock_page(page);
 843                                         break;
 844                                 }
 845
 846                                 req->r_callback = writepages_finish;
 847                                 req->r_inode = inode;
 848
 849                                 max_pages = calc_pages_for(0, (u64)len);
 850                                 pages = kmalloc(max_pages * sizeof (*pages),
 851                                                 GFP_NOFS);
 852                                 if (!pages) {
 853                                         pool = fsc->wb_pagevec_pool;
 854                                         pages = mempool_alloc(pool, GFP_NOFS);
 855                                         BUG_ON(!pages);
 856                                 }
 857                         }
 858
 859                         /* note position of first page in pvec */
 860                         if (first < 0)
 861                                 first = i;
 862                         dout("%p will write page %p idx %lu\n",
 863                              inode, page, page->index);
 864
 865                         writeback_stat =
 866                                atomic_long_inc_return(&fsc->writeback_count);
 867                         if (writeback_stat > CONGESTION_ON_THRESH(
 868                                     fsc->mount_options->congestion_kb)) {
 869                                 set_bdi_congested(&fsc->backing_dev_info,
 870                                                   BLK_RW_ASYNC);
 871                         }
 872
 873                         set_page_writeback(page);
 874                         pages[locked_pages] = page;
 875                         locked_pages++;
 876                         next = page->index + 1;
 877                 }
 878
 879                 /* did we get anything? */
 880                 if (!locked_pages)
 881                         goto release_pvec_pages;
 882                 if (i) {
 883                         int j;
 884                         BUG_ON(!locked_pages || first < 0);
 885
 886                         if (pvec_pages && i == pvec_pages &&
 887                             locked_pages < max_pages) {
 888                                 dout("reached end pvec, trying for more\n");
 889                                 pagevec_reinit(&pvec);
 890                                 goto get_more_pages;
 891                         }
 892
 893                         /* shift unused pages over in the pvec...  we
 894                          * will need to release them below. */
 895                         for (j = i; j < pvec_pages; j++) {
 896                                 dout(" pvec leftover page %p\n",
 897                                      pvec.pages[j]);
 898                                 pvec.pages[j-i+first] = pvec.pages[j];
 899                         }
 900                         pvec.nr -= i-first;
 901                 }
 902
 903                 /* Format the osd request message and submit the write */
 904
 905                 offset = page_offset(pages[0]);
 906                 len = min(snap_size - offset,
 907                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 908                 dout("writepages got %d pages at %llu~%llu\n",
 909                      locked_pages, offset, len);
 910
 911                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 912                                                         !!pool, false);
 913
 914                 pages = NULL;   /* request message now owns the pages array */
 915                 pool = NULL;
 916
 917                 /* Update the write op length in case we changed it */
 918
 919                 osd_req_op_extent_update(req, 0, len);
 920
 921                 vino = ceph_vino(inode);
 922                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 923                                         &inode->i_mtime);
 924
 925                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 926                 BUG_ON(rc);
 927                 req = NULL;
 928
 929                 /* continue? */
 930                 index = next;
 931                 wbc->nr_to_write -= locked_pages;
 932                 if (wbc->nr_to_write <= 0)
 933                         done = 1;
 934
 935 release_pvec_pages:
 936                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 937                      pvec.nr ? pvec.pages[0] : NULL);
 938                 pagevec_release(&pvec);
 939
 940                 if (locked_pages && !done)
 941                         goto retry;
 942         }
 943
 944         if (should_loop && !done) {
 945                 /* more to do; loop back to beginning of file */
 946                 dout("writepages looping back to beginning of file\n");
 947                 should_loop = 0;
 948                 index = 0;
 949                 goto retry;
 950         }
 951
 952         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 953                 mapping->writeback_index = index;
 954
 955 out:
 956         if (req)
 957                 ceph_osdc_put_request(req);
 958         ceph_put_snap_context(snapc);
 959         dout("writepages done, rc = %d\n", rc);
 960         return rc;
 961 }
 962
 963
 964
 965 /*
 966  * See if a given @snapc is either writeable, or already written.
 967  */
 968 static int context_is_writeable_or_written(struct inode *inode,
 969                                            struct ceph_snap_context *snapc)
 970 {
 971         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 972         int ret = !oldest || snapc->seq <= oldest->seq;
 973
 974         ceph_put_snap_context(oldest);
 975         return ret;
 976 }
 977
 978 /*
 979  * We are only allowed to write into/dirty the page if the page is
 980  * clean, or already dirty within the same snap context.
 981  *
 982  * called with page locked.
 983  * return success with page locked,
 984  * or any failure (incl -EAGAIN) with page unlocked.
 985  */
 986 static int ceph_update_writeable_page(struct file *file,
 987                             loff_t pos, unsigned len,
 988                             struct page *page)
 989 {
 990         struct inode *inode = file_inode(file);
 991         struct ceph_inode_info *ci = ceph_inode(inode);
 992         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 993         loff_t page_off = pos & PAGE_CACHE_MASK;
 994         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 995         int end_in_page = pos_in_page + len;
 996         loff_t i_size;
 997         int r;
 998         struct ceph_snap_context *snapc, *oldest;
 999
1000 retry_locked:
1001         /* writepages currently holds page lock, but if we change that later, */
1002         wait_on_page_writeback(page);
1003
1004         /* check snap context */
1005         BUG_ON(!ci->i_snap_realm);
1006         down_read(&mdsc->snap_rwsem);
1007         BUG_ON(!ci->i_snap_realm->cached_context);
1008         snapc = page_snap_context(page);
1009         if (snapc && snapc != ci->i_head_snapc) {
1010                 /*
1011                  * this page is already dirty in another (older) snap
1012                  * context!  is it writeable now?
1013                  */
1014                 oldest = get_oldest_context(inode, NULL);
1015                 up_read(&mdsc->snap_rwsem);
1016
1017                 if (snapc->seq > oldest->seq) {
1018                         ceph_put_snap_context(oldest);
1019                         dout(" page %p snapc %p not current or oldest\n",
1020                              page, snapc);
1021                         /*
1022                          * queue for writeback, and wait for snapc to
1023                          * be writeable or written
1024                          */
1025                         snapc = ceph_get_snap_context(snapc);
1026                         unlock_page(page);
1027                         ceph_queue_writeback(inode);
1028                         r = wait_event_interruptible(ci->i_cap_wq,
1029                                context_is_writeable_or_written(inode, snapc));
1030                         ceph_put_snap_context(snapc);
1031                         if (r == -ERESTARTSYS)
1032                                 return r;
1033                         return -EAGAIN;
1034                 }
1035                 ceph_put_snap_context(oldest);
1036
1037                 /* yay, writeable, do it now (without dropping page lock) */
1038                 dout(" page %p snapc %p not current, but oldest\n",
1039                      page, snapc);
1040                 if (!clear_page_dirty_for_io(page))
1041                         goto retry_locked;
1042                 r = writepage_nounlock(page, NULL);
1043                 if (r < 0)
1044                         goto fail_nosnap;
1045                 goto retry_locked;
1046         }
1047
1048         if (PageUptodate(page)) {
1049                 dout(" page %p already uptodate\n", page);
1050                 return 0;
1051         }
1052
1053         /* full page? */
1054         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1055                 return 0;
1056
1057         /* past end of file? */
1058         i_size = inode->i_size;   /* caller holds i_mutex */
1059
1060         if (i_size + len > inode->i_sb->s_maxbytes) {
1061                 /* file is too big */
1062                 r = -EINVAL;
1063                 goto fail;
1064         }
1065
1066         if (page_off >= i_size ||
1067             (pos_in_page == 0 && (pos+len) >= i_size &&
1068              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1069                 dout(" zeroing %p 0 - %d and %d - %d\n",
1070                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1071                 zero_user_segments(page,
1072                                    0, pos_in_page,
1073                                    end_in_page, PAGE_CACHE_SIZE);
1074                 return 0;
1075         }
1076
1077         /* we need to read it. */
1078         up_read(&mdsc->snap_rwsem);
1079         r = readpage_nounlock(file, page);
1080         if (r < 0)
1081                 goto fail_nosnap;
1082         goto retry_locked;
1083
1084 fail:
1085         up_read(&mdsc->snap_rwsem);
1086 fail_nosnap:
1087         unlock_page(page);
1088         return r;
1089 }
1090
1091 /*
1092  * We are only allowed to write into/dirty the page if the page is
1093  * clean, or already dirty within the same snap context.
1094  */
1095 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1096                             loff_t pos, unsigned len, unsigned flags,
1097                             struct page **pagep, void **fsdata)
1098 {
1099         struct inode *inode = file_inode(file);
1100         struct page *page;
1101         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1102         int r;
1103
1104         do {
1105                 /* get a page */
1106                 page = grab_cache_page_write_begin(mapping, index, 0);
1107                 if (!page)
1108                         return -ENOMEM;
1109                 *pagep = page;
1110
1111                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1112                      inode, page, (int)pos, (int)len);
1113
1114                 r = ceph_update_writeable_page(file, pos, len, page);
1115         } while (r == -EAGAIN);
1116
1117         return r;
1118 }
1119
1120 /*
1121  * we don't do anything in here that simple_write_end doesn't do
1122  * except adjust dirty page accounting and drop read lock on
1123  * mdsc->snap_rwsem.
1124  */
1125 static int ceph_write_end(struct file *file, struct address_space *mapping,
1126                           loff_t pos, unsigned len, unsigned copied,
1127                           struct page *page, void *fsdata)
1128 {
1129         struct inode *inode = file_inode(file);
1130         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1131         struct ceph_mds_client *mdsc = fsc->mdsc;
1132         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1133         int check_cap = 0;
1134
1135         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1136              inode, page, (int)pos, (int)copied, (int)len);
1137
1138         /* zero the stale part of the page if we did a short copy */
1139         if (copied < len)
1140                 zero_user_segment(page, from+copied, len);
1141
1142         /* did file size increase? */
1143         /* (no need for i_size_read(); we caller holds i_mutex */
1144         if (pos+copied > inode->i_size)
1145                 check_cap = ceph_inode_set_size(inode, pos+copied);
1146
1147         if (!PageUptodate(page))
1148                 SetPageUptodate(page);
1149
1150         set_page_dirty(page);
1151
1152         unlock_page(page);
1153         up_read(&mdsc->snap_rwsem);
1154         page_cache_release(page);
1155
1156         if (check_cap)
1157                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1158
1159         return copied;
1160 }
1161
1162 /*
1163  * we set .direct_IO to indicate direct io is supported, but since we
1164  * intercept O_DIRECT reads and writes early, this function should
1165  * never get called.
1166  */
1167 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1168                               const struct iovec *iov,
1169                               loff_t pos, unsigned long nr_segs)
1170 {
1171         WARN_ON(1);
1172         return -EINVAL;
1173 }
1174
1175 const struct address_space_operations ceph_aops = {
1176         .readpage = ceph_readpage,
1177         .readpages = ceph_readpages,
1178         .writepage = ceph_writepage,
1179         .writepages = ceph_writepages_start,
1180         .write_begin = ceph_write_begin,
1181         .write_end = ceph_write_end,
1182         .set_page_dirty = ceph_set_page_dirty,
1183         .invalidatepage = ceph_invalidatepage,
1184         .releasepage = ceph_releasepage,
1185         .direct_IO = ceph_direct_io,
1186 };
1187
1188
1189 /*
1190  * vm ops
1191  */
1192
1193 /*
1194  * Reuse write_begin here for simplicity.
1195  */
1196 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1197 {
1198         struct inode *inode = file_inode(vma->vm_file);
1199         struct page *page = vmf->page;
1200         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1201         loff_t off = page_offset(page);
1202         loff_t size, len;
1203         int ret;
1204
1205         /* Update time before taking page lock */
1206         file_update_time(vma->vm_file);
1207
1208         size = i_size_read(inode);
1209         if (off + PAGE_CACHE_SIZE <= size)
1210                 len = PAGE_CACHE_SIZE;
1211         else
1212                 len = size & ~PAGE_CACHE_MASK;
1213
1214         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1215              off, len, page, page->index);
1216
1217         lock_page(page);
1218
1219         ret = VM_FAULT_NOPAGE;
1220         if ((off > size) ||
1221             (page->mapping != inode->i_mapping))
1222                 goto out;
1223
1224         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1225         if (ret == 0) {
1226                 /* success.  we'll keep the page locked. */
1227                 set_page_dirty(page);
1228                 up_read(&mdsc->snap_rwsem);
1229                 ret = VM_FAULT_LOCKED;
1230         } else {
1231                 if (ret == -ENOMEM)
1232                         ret = VM_FAULT_OOM;
1233                 else
1234                         ret = VM_FAULT_SIGBUS;
1235         }
1236 out:
1237         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1238         if (ret != VM_FAULT_LOCKED)
1239                 unlock_page(page);
1240         return ret;
1241 }
1242
1243 static struct vm_operations_struct ceph_vmops = {
1244         .fault          = filemap_fault,
1245         .page_mkwrite   = ceph_page_mkwrite,
1246         .remap_pages    = generic_file_remap_pages,
1247 };
1248
1249 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1250 {
1251         struct address_space *mapping = file->f_mapping;
1252
1253         if (!mapping->a_ops->readpage)
1254                 return -ENOEXEC;
1255         file_accessed(file);
1256         vma->vm_ops = &ceph_vmops;
1257         return 0;
1258 }