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         ci = ceph_inode(inode);
 155
 156         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 157                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 158                      inode, page, page->index, offset, length);
 159                 return;
 160         }
 161
 162         /*
 163          * We can get non-dirty pages here due to races between
 164          * set_page_dirty and truncate_complete_page; just spit out a
 165          * warning, in case we end up with accounting problems later.
 166          */
 167         if (!PageDirty(page))
 168                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 169
 170         ClearPageChecked(page);
 171
 172         dout("%p invalidatepage %p idx %lu full dirty page\n",
 173              inode, page, page->index);
 174
 175         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 176         ceph_put_snap_context(snapc);
 177         page->private = 0;
 178         ClearPagePrivate(page);
 179 }
 180
 181 /* just a sanity check */
 182 static int ceph_releasepage(struct page *page, gfp_t g)
 183 {
 184         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 185         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 186         WARN_ON(PageDirty(page));
 187         WARN_ON(PagePrivate(page));
 188         return 0;
 189 }
 190
 191 /*
 192  * read a single page, without unlocking it.
 193  */
 194 static int readpage_nounlock(struct file *filp, struct page *page)
 195 {
 196         struct inode *inode = file_inode(filp);
 197         struct ceph_inode_info *ci = ceph_inode(inode);
 198         struct ceph_osd_client *osdc =
 199                 &ceph_inode_to_client(inode)->client->osdc;
 200         int err = 0;
 201         u64 len = PAGE_CACHE_SIZE;
 202
 203         dout("readpage inode %p file %p page %p index %lu\n",
 204              inode, filp, page, page->index);
 205         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 206                                   (u64) page_offset(page), &len,
 207                                   ci->i_truncate_seq, ci->i_truncate_size,
 208                                   &page, 1, 0);
 209         if (err == -ENOENT)
 210                 err = 0;
 211         if (err < 0) {
 212                 SetPageError(page);
 213                 goto out;
 214         } else if (err < PAGE_CACHE_SIZE) {
 215                 /* zero fill remainder of page */
 216                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 217         }
 218         SetPageUptodate(page);
 219
 220 out:
 221         return err < 0 ? err : 0;
 222 }
 223
 224 static int ceph_readpage(struct file *filp, struct page *page)
 225 {
 226         int r = readpage_nounlock(filp, page);
 227         unlock_page(page);
 228         return r;
 229 }
 230
 231 /*
 232  * Finish an async read(ahead) op.
 233  */
 234 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 235 {
 236         struct inode *inode = req->r_inode;
 237         struct ceph_osd_data *osd_data;
 238         int rc = req->r_result;
 239         int bytes = le32_to_cpu(msg->hdr.data_len);
 240         int num_pages;
 241         int i;
 242
 243         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 244
 245         /* unlock all pages, zeroing any data we didn't read */
 246         osd_data = osd_req_op_extent_osd_data(req, 0);
 247         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 248         num_pages = calc_pages_for((u64)osd_data->alignment,
 249                                         (u64)osd_data->length);
 250         for (i = 0; i < num_pages; i++) {
 251                 struct page *page = osd_data->pages[i];
 252
 253                 if (bytes < (int)PAGE_CACHE_SIZE) {
 254                         /* zero (remainder of) page */
 255                         int s = bytes < 0 ? 0 : bytes;
 256                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 257                 }
 258                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 259                      page->index);
 260                 flush_dcache_page(page);
 261                 SetPageUptodate(page);
 262                 unlock_page(page);
 263                 page_cache_release(page);
 264                 bytes -= PAGE_CACHE_SIZE;
 265         }
 266         kfree(osd_data->pages);
 267 }
 268
 269 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 270 {
 271         int i;
 272
 273         for (i = 0; i < num_pages; i++)
 274                 unlock_page(pages[i]);
 275 }
 276
 277 /*
 278  * start an async read(ahead) operation.  return nr_pages we submitted
 279  * a read for on success, or negative error code.
 280  */
 281 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 282 {
 283         struct ceph_osd_client *osdc =
 284                 &ceph_inode_to_client(inode)->client->osdc;
 285         struct ceph_inode_info *ci = ceph_inode(inode);
 286         struct page *page = list_entry(page_list->prev, struct page, lru);
 287         struct ceph_vino vino;
 288         struct ceph_osd_request *req;
 289         u64 off;
 290         u64 len;
 291         int i;
 292         struct page **pages;
 293         pgoff_t next_index;
 294         int nr_pages = 0;
 295         int ret;
 296
 297         off = (u64) page_offset(page);
 298
 299         /* count pages */
 300         next_index = page->index;
 301         list_for_each_entry_reverse(page, page_list, lru) {
 302                 if (page->index != next_index)
 303                         break;
 304                 nr_pages++;
 305                 next_index++;
 306                 if (max && nr_pages == max)
 307                         break;
 308         }
 309         len = nr_pages << PAGE_CACHE_SHIFT;
 310         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 311              off, len);
 312         vino = ceph_vino(inode);
 313         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 314                                     1, CEPH_OSD_OP_READ,
 315                                     CEPH_OSD_FLAG_READ, NULL,
 316                                     ci->i_truncate_seq, ci->i_truncate_size,
 317                                     false);
 318         if (IS_ERR(req))
 319                 return PTR_ERR(req);
 320
 321         /* build page vector */
 322         nr_pages = calc_pages_for(0, len);
 323         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 324         ret = -ENOMEM;
 325         if (!pages)
 326                 goto out;
 327         for (i = 0; i < nr_pages; ++i) {
 328                 page = list_entry(page_list->prev, struct page, lru);
 329                 BUG_ON(PageLocked(page));
 330                 list_del(&page->lru);
 331
 332                 dout("start_read %p adding %p idx %lu\n", inode, page,
 333                      page->index);
 334                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 335                                           GFP_NOFS)) {
 336                         page_cache_release(page);
 337                         dout("start_read %p add_to_page_cache failed %p\n",
 338                              inode, page);
 339                         nr_pages = i;
 340                         goto out_pages;
 341                 }
 342                 pages[i] = page;
 343         }
 344         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 345         req->r_callback = finish_read;
 346         req->r_inode = inode;
 347
 348         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 349
 350         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 351         ret = ceph_osdc_start_request(osdc, req, false);
 352         if (ret < 0)
 353                 goto out_pages;
 354         ceph_osdc_put_request(req);
 355         return nr_pages;
 356
 357 out_pages:
 358         ceph_unlock_page_vector(pages, nr_pages);
 359         ceph_release_page_vector(pages, nr_pages);
 360 out:
 361         ceph_osdc_put_request(req);
 362         return ret;
 363 }
 364
 365
 366 /*
 367  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 368  * the caller (VM) cleans them up.
 369  */
 370 static int ceph_readpages(struct file *file, struct address_space *mapping,
 371                           struct list_head *page_list, unsigned nr_pages)
 372 {
 373         struct inode *inode = file_inode(file);
 374         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 375         int rc = 0;
 376         int max = 0;
 377
 378         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 379                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 380                         >> PAGE_SHIFT;
 381
 382         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 383                 file, nr_pages,
 384              max);
 385         while (!list_empty(page_list)) {
 386                 rc = start_read(inode, page_list, max);
 387                 if (rc < 0)
 388                         goto out;
 389                 BUG_ON(rc == 0);
 390         }
 391 out:
 392         dout("readpages %p file %p ret %d\n", inode, file, rc);
 393         return rc;
 394 }
 395
 396 /*
 397  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 398  * only snap context we are allowed to write back.
 399  */
 400 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 401                                                     u64 *snap_size)
 402 {
 403         struct ceph_inode_info *ci = ceph_inode(inode);
 404         struct ceph_snap_context *snapc = NULL;
 405         struct ceph_cap_snap *capsnap = NULL;
 406
 407         spin_lock(&ci->i_ceph_lock);
 408         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 409                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 410                      capsnap->context, capsnap->dirty_pages);
 411                 if (capsnap->dirty_pages) {
 412                         snapc = ceph_get_snap_context(capsnap->context);
 413                         if (snap_size)
 414                                 *snap_size = capsnap->size;
 415                         break;
 416                 }
 417         }
 418         if (!snapc && ci->i_wrbuffer_ref_head) {
 419                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 420                 dout(" head snapc %p has %d dirty pages\n",
 421                      snapc, ci->i_wrbuffer_ref_head);
 422         }
 423         spin_unlock(&ci->i_ceph_lock);
 424         return snapc;
 425 }
 426
 427 /*
 428  * Write a single page, but leave the page locked.
 429  *
 430  * If we get a write error, set the page error bit, but still adjust the
 431  * dirty page accounting (i.e., page is no longer dirty).
 432  */
 433 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 434 {
 435         struct inode *inode;
 436         struct ceph_inode_info *ci;
 437         struct ceph_fs_client *fsc;
 438         struct ceph_osd_client *osdc;
 439         struct ceph_snap_context *snapc, *oldest;
 440         loff_t page_off = page_offset(page);
 441         long writeback_stat;
 442         u64 truncate_size, snap_size = 0;
 443         u32 truncate_seq;
 444         int err = 0, len = PAGE_CACHE_SIZE;
 445
 446         dout("writepage %p idx %lu\n", page, page->index);
 447
 448         if (!page->mapping || !page->mapping->host) {
 449                 dout("writepage %p - no mapping\n", page);
 450                 return -EFAULT;
 451         }
 452         inode = page->mapping->host;
 453         ci = ceph_inode(inode);
 454         fsc = ceph_inode_to_client(inode);
 455         osdc = &fsc->client->osdc;
 456
 457         /* verify this is a writeable snap context */
 458         snapc = page_snap_context(page);
 459         if (snapc == NULL) {
 460                 dout("writepage %p page %p not dirty?\n", inode, page);
 461                 goto out;
 462         }
 463         oldest = get_oldest_context(inode, &snap_size);
 464         if (snapc->seq > oldest->seq) {
 465                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 466                      inode, page, snapc);
 467                 /* we should only noop if called by kswapd */
 468                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 469                 ceph_put_snap_context(oldest);
 470                 goto out;
 471         }
 472         ceph_put_snap_context(oldest);
 473
 474         spin_lock(&ci->i_ceph_lock);
 475         truncate_seq = ci->i_truncate_seq;
 476         truncate_size = ci->i_truncate_size;
 477         if (!snap_size)
 478                 snap_size = i_size_read(inode);
 479         spin_unlock(&ci->i_ceph_lock);
 480
 481         /* is this a partial page at end of file? */
 482         if (page_off >= snap_size) {
 483                 dout("%p page eof %llu\n", page, snap_size);
 484                 goto out;
 485         }
 486         if (snap_size < page_off + len)
 487                 len = snap_size - page_off;
 488
 489         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 490              inode, page, page->index, page_off, len, snapc);
 491
 492         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 493         if (writeback_stat >
 494             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 495                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 496
 497         set_page_writeback(page);
 498         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 499                                    &ci->i_layout, snapc,
 500                                    page_off, len,
 501                                    truncate_seq, truncate_size,
 502                                    &inode->i_mtime, &page, 1);
 503         if (err < 0) {
 504                 dout("writepage setting page/mapping error %d %p\n", err, page);
 505                 SetPageError(page);
 506                 mapping_set_error(&inode->i_data, err);
 507                 if (wbc)
 508                         wbc->pages_skipped++;
 509         } else {
 510                 dout("writepage cleaned page %p\n", page);
 511                 err = 0;  /* vfs expects us to return 0 */
 512         }
 513         page->private = 0;
 514         ClearPagePrivate(page);
 515         end_page_writeback(page);
 516         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 517         ceph_put_snap_context(snapc);  /* page's reference */
 518 out:
 519         return err;
 520 }
 521
 522 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 523 {
 524         int err;
 525         struct inode *inode = page->mapping->host;
 526         BUG_ON(!inode);
 527         ihold(inode);
 528         err = writepage_nounlock(page, wbc);
 529         unlock_page(page);
 530         iput(inode);
 531         return err;
 532 }
 533
 534
 535 /*
 536  * lame release_pages helper.  release_pages() isn't exported to
 537  * modules.
 538  */
 539 static void ceph_release_pages(struct page **pages, int num)
 540 {
 541         struct pagevec pvec;
 542         int i;
 543
 544         pagevec_init(&pvec, 0);
 545         for (i = 0; i < num; i++) {
 546                 if (pagevec_add(&pvec, pages[i]) == 0)
 547                         pagevec_release(&pvec);
 548         }
 549         pagevec_release(&pvec);
 550 }
 551
 552
 553 /*
 554  * async writeback completion handler.
 555  *
 556  * If we get an error, set the mapping error bit, but not the individual
 557  * page error bits.
 558  */
 559 static void writepages_finish(struct ceph_osd_request *req,
 560                               struct ceph_msg *msg)
 561 {
 562         struct inode *inode = req->r_inode;
 563         struct ceph_inode_info *ci = ceph_inode(inode);
 564         struct ceph_osd_data *osd_data;
 565         unsigned wrote;
 566         struct page *page;
 567         int num_pages;
 568         int i;
 569         struct ceph_snap_context *snapc = req->r_snapc;
 570         struct address_space *mapping = inode->i_mapping;
 571         int rc = req->r_result;
 572         u64 bytes = req->r_ops[0].extent.length;
 573         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 574         long writeback_stat;
 575         unsigned issued = ceph_caps_issued(ci);
 576
 577         osd_data = osd_req_op_extent_osd_data(req, 0);
 578         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 579         num_pages = calc_pages_for((u64)osd_data->alignment,
 580                                         (u64)osd_data->length);
 581         if (rc >= 0) {
 582                 /*
 583                  * Assume we wrote the pages we originally sent.  The
 584                  * osd might reply with fewer pages if our writeback
 585                  * raced with a truncation and was adjusted at the osd,
 586                  * so don't believe the reply.
 587                  */
 588                 wrote = num_pages;
 589         } else {
 590                 wrote = 0;
 591                 mapping_set_error(mapping, rc);
 592         }
 593         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 594              inode, rc, bytes, wrote);
 595
 596         /* clean all pages */
 597         for (i = 0; i < num_pages; i++) {
 598                 page = osd_data->pages[i];
 599                 BUG_ON(!page);
 600                 WARN_ON(!PageUptodate(page));
 601
 602                 writeback_stat =
 603                         atomic_long_dec_return(&fsc->writeback_count);
 604                 if (writeback_stat <
 605                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 606                         clear_bdi_congested(&fsc->backing_dev_info,
 607                                             BLK_RW_ASYNC);
 608
 609                 ceph_put_snap_context(page_snap_context(page));
 610                 page->private = 0;
 611                 ClearPagePrivate(page);
 612                 dout("unlocking %d %p\n", i, page);
 613                 end_page_writeback(page);
 614
 615                 /*
 616                  * We lost the cache cap, need to truncate the page before
 617                  * it is unlocked, otherwise we'd truncate it later in the
 618                  * page truncation thread, possibly losing some data that
 619                  * raced its way in
 620                  */
 621                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 622                         generic_error_remove_page(inode->i_mapping, page);
 623
 624                 unlock_page(page);
 625         }
 626         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 627         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 628
 629         ceph_release_pages(osd_data->pages, num_pages);
 630         if (osd_data->pages_from_pool)
 631                 mempool_free(osd_data->pages,
 632                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 633         else
 634                 kfree(osd_data->pages);
 635         ceph_osdc_put_request(req);
 636 }
 637
 638 /*
 639  * initiate async writeback
 640  */
 641 static int ceph_writepages_start(struct address_space *mapping,
 642                                  struct writeback_control *wbc)
 643 {
 644         struct inode *inode = mapping->host;
 645         struct ceph_inode_info *ci = ceph_inode(inode);
 646         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 647         struct ceph_vino vino = ceph_vino(inode);
 648         pgoff_t index, start, end;
 649         int range_whole = 0;
 650         int should_loop = 1;
 651         pgoff_t max_pages = 0, max_pages_ever = 0;
 652         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 653         struct pagevec pvec;
 654         int done = 0;
 655         int rc = 0;
 656         unsigned wsize = 1 << inode->i_blkbits;
 657         struct ceph_osd_request *req = NULL;
 658         int do_sync;
 659         u64 truncate_size, snap_size;
 660         u32 truncate_seq;
 661
 662         /*
 663          * Include a 'sync' in the OSD request if this is a data
 664          * integrity write (e.g., O_SYNC write or fsync()), or if our
 665          * cap is being revoked.
 666          */
 667         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 668                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 669                 do_sync = 1;
 670         dout("writepages_start %p dosync=%d (mode=%s)\n",
 671              inode, do_sync,
 672              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 673              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 674
 675         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 676                 pr_warning("writepage_start %p on forced umount\n", inode);
 677                 return -EIO; /* we're in a forced umount, don't write! */
 678         }
 679         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 680                 wsize = fsc->mount_options->wsize;
 681         if (wsize < PAGE_CACHE_SIZE)
 682                 wsize = PAGE_CACHE_SIZE;
 683         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 684
 685         pagevec_init(&pvec, 0);
 686
 687         /* where to start/end? */
 688         if (wbc->range_cyclic) {
 689                 start = mapping->writeback_index; /* Start from prev offset */
 690                 end = -1;
 691                 dout(" cyclic, start at %lu\n", start);
 692         } else {
 693                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 694                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 695                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 696                         range_whole = 1;
 697                 should_loop = 0;
 698                 dout(" not cyclic, %lu to %lu\n", start, end);
 699         }
 700         index = start;
 701
 702 retry:
 703         /* find oldest snap context with dirty data */
 704         ceph_put_snap_context(snapc);
 705         snap_size = 0;
 706         snapc = get_oldest_context(inode, &snap_size);
 707         if (!snapc) {
 708                 /* hmm, why does writepages get called when there
 709                    is no dirty data? */
 710                 dout(" no snap context with dirty data?\n");
 711                 goto out;
 712         }
 713         if (snap_size == 0)
 714                 snap_size = i_size_read(inode);
 715         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 716              snapc, snapc->seq, snapc->num_snaps);
 717
 718         spin_lock(&ci->i_ceph_lock);
 719         truncate_seq = ci->i_truncate_seq;
 720         truncate_size = ci->i_truncate_size;
 721         if (!snap_size)
 722                 snap_size = i_size_read(inode);
 723         spin_unlock(&ci->i_ceph_lock);
 724
 725         if (last_snapc && snapc != last_snapc) {
 726                 /* if we switched to a newer snapc, restart our scan at the
 727                  * start of the original file range. */
 728                 dout("  snapc differs from last pass, restarting at %lu\n",
 729                      index);
 730                 index = start;
 731         }
 732         last_snapc = snapc;
 733
 734         while (!done && index <= end) {
 735                 int num_ops = do_sync ? 2 : 1;
 736                 unsigned i;
 737                 int first;
 738                 pgoff_t next;
 739                 int pvec_pages, locked_pages;
 740                 struct page **pages = NULL;
 741                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 742                 struct page *page;
 743                 int want;
 744                 u64 offset, len;
 745                 long writeback_stat;
 746
 747                 next = 0;
 748                 locked_pages = 0;
 749                 max_pages = max_pages_ever;
 750
 751 get_more_pages:
 752                 first = -1;
 753                 want = min(end - index,
 754                            min((pgoff_t)PAGEVEC_SIZE,
 755                                max_pages - (pgoff_t)locked_pages) - 1)
 756                         + 1;
 757                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 758                                                 PAGECACHE_TAG_DIRTY,
 759                                                 want);
 760                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 761                 if (!pvec_pages && !locked_pages)
 762                         break;
 763                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 764                         page = pvec.pages[i];
 765                         dout("? %p idx %lu\n", page, page->index);
 766                         if (locked_pages == 0)
 767                                 lock_page(page);  /* first page */
 768                         else if (!trylock_page(page))
 769                                 break;
 770
 771                         /* only dirty pages, or our accounting breaks */
 772                         if (unlikely(!PageDirty(page)) ||
 773                             unlikely(page->mapping != mapping)) {
 774                                 dout("!dirty or !mapping %p\n", page);
 775                                 unlock_page(page);
 776                                 break;
 777                         }
 778                         if (!wbc->range_cyclic && page->index > end) {
 779                                 dout("end of range %p\n", page);
 780                                 done = 1;
 781                                 unlock_page(page);
 782                                 break;
 783                         }
 784                         if (next && (page->index != next)) {
 785                                 dout("not consecutive %p\n", page);
 786                                 unlock_page(page);
 787                                 break;
 788                         }
 789                         if (wbc->sync_mode != WB_SYNC_NONE) {
 790                                 dout("waiting on writeback %p\n", page);
 791                                 wait_on_page_writeback(page);
 792                         }
 793                         if (page_offset(page) >= snap_size) {
 794                                 dout("%p page eof %llu\n", page, snap_size);
 795                                 done = 1;
 796                                 unlock_page(page);
 797                                 break;
 798                         }
 799                         if (PageWriteback(page)) {
 800                                 dout("%p under writeback\n", page);
 801                                 unlock_page(page);
 802                                 break;
 803                         }
 804
 805                         /* only if matching snap context */
 806                         pgsnapc = page_snap_context(page);
 807                         if (pgsnapc->seq > snapc->seq) {
 808                                 dout("page snapc %p %lld > oldest %p %lld\n",
 809                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 810                                 unlock_page(page);
 811                                 if (!locked_pages)
 812                                         continue; /* keep looking for snap */
 813                                 break;
 814                         }
 815
 816                         if (!clear_page_dirty_for_io(page)) {
 817                                 dout("%p !clear_page_dirty_for_io\n", page);
 818                                 unlock_page(page);
 819                                 break;
 820                         }
 821
 822                         /*
 823                          * We have something to write.  If this is
 824                          * the first locked page this time through,
 825                          * allocate an osd request and a page array
 826                          * that it will use.
 827                          */
 828                         if (locked_pages == 0) {
 829                                 BUG_ON(pages);
 830                                 /* prepare async write request */
 831                                 offset = (u64)page_offset(page);
 832                                 len = wsize;
 833                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 834                                                         &ci->i_layout, vino,
 835                                                         offset, &len, num_ops,
 836                                                         CEPH_OSD_OP_WRITE,
 837                                                         CEPH_OSD_FLAG_WRITE |
 838                                                         CEPH_OSD_FLAG_ONDISK,
 839                                                         snapc, truncate_seq,
 840                                                         truncate_size, true);
 841                                 if (IS_ERR(req)) {
 842                                         rc = PTR_ERR(req);
 843                                         unlock_page(page);
 844                                         break;
 845                                 }
 846
 847                                 req->r_callback = writepages_finish;
 848                                 req->r_inode = inode;
 849
 850                                 max_pages = calc_pages_for(0, (u64)len);
 851                                 pages = kmalloc(max_pages * sizeof (*pages),
 852                                                 GFP_NOFS);
 853                                 if (!pages) {
 854                                         pool = fsc->wb_pagevec_pool;
 855                                         pages = mempool_alloc(pool, GFP_NOFS);
 856                                         BUG_ON(!pages);
 857                                 }
 858                         }
 859
 860                         /* note position of first page in pvec */
 861                         if (first < 0)
 862                                 first = i;
 863                         dout("%p will write page %p idx %lu\n",
 864                              inode, page, page->index);
 865
 866                         writeback_stat =
 867                                atomic_long_inc_return(&fsc->writeback_count);
 868                         if (writeback_stat > CONGESTION_ON_THRESH(
 869                                     fsc->mount_options->congestion_kb)) {
 870                                 set_bdi_congested(&fsc->backing_dev_info,
 871                                                   BLK_RW_ASYNC);
 872                         }
 873
 874                         set_page_writeback(page);
 875                         pages[locked_pages] = page;
 876                         locked_pages++;
 877                         next = page->index + 1;
 878                 }
 879
 880                 /* did we get anything? */
 881                 if (!locked_pages)
 882                         goto release_pvec_pages;
 883                 if (i) {
 884                         int j;
 885                         BUG_ON(!locked_pages || first < 0);
 886
 887                         if (pvec_pages && i == pvec_pages &&
 888                             locked_pages < max_pages) {
 889                                 dout("reached end pvec, trying for more\n");
 890                                 pagevec_reinit(&pvec);
 891                                 goto get_more_pages;
 892                         }
 893
 894                         /* shift unused pages over in the pvec...  we
 895                          * will need to release them below. */
 896                         for (j = i; j < pvec_pages; j++) {
 897                                 dout(" pvec leftover page %p\n",
 898                                      pvec.pages[j]);
 899                                 pvec.pages[j-i+first] = pvec.pages[j];
 900                         }
 901                         pvec.nr -= i-first;
 902                 }
 903
 904                 /* Format the osd request message and submit the write */
 905
 906                 offset = page_offset(pages[0]);
 907                 len = min(snap_size - offset,
 908                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 909                 dout("writepages got %d pages at %llu~%llu\n",
 910                      locked_pages, offset, len);
 911
 912                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 913                                                         !!pool, false);
 914
 915                 pages = NULL;   /* request message now owns the pages array */
 916                 pool = NULL;
 917
 918                 /* Update the write op length in case we changed it */
 919
 920                 osd_req_op_extent_update(req, 0, len);
 921
 922                 vino = ceph_vino(inode);
 923                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 924                                         &inode->i_mtime);
 925
 926                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 927                 BUG_ON(rc);
 928                 req = NULL;
 929
 930                 /* continue? */
 931                 index = next;
 932                 wbc->nr_to_write -= locked_pages;
 933                 if (wbc->nr_to_write <= 0)
 934                         done = 1;
 935
 936 release_pvec_pages:
 937                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 938                      pvec.nr ? pvec.pages[0] : NULL);
 939                 pagevec_release(&pvec);
 940
 941                 if (locked_pages && !done)
 942                         goto retry;
 943         }
 944
 945         if (should_loop && !done) {
 946                 /* more to do; loop back to beginning of file */
 947                 dout("writepages looping back to beginning of file\n");
 948                 should_loop = 0;
 949                 index = 0;
 950                 goto retry;
 951         }
 952
 953         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 954                 mapping->writeback_index = index;
 955
 956 out:
 957         if (req)
 958                 ceph_osdc_put_request(req);
 959         ceph_put_snap_context(snapc);
 960         dout("writepages done, rc = %d\n", rc);
 961         return rc;
 962 }
 963
 964
 965
 966 /*
 967  * See if a given @snapc is either writeable, or already written.
 968  */
 969 static int context_is_writeable_or_written(struct inode *inode,
 970                                            struct ceph_snap_context *snapc)
 971 {
 972         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 973         int ret = !oldest || snapc->seq <= oldest->seq;
 974
 975         ceph_put_snap_context(oldest);
 976         return ret;
 977 }
 978
 979 /*
 980  * We are only allowed to write into/dirty the page if the page is
 981  * clean, or already dirty within the same snap context.
 982  *
 983  * called with page locked.
 984  * return success with page locked,
 985  * or any failure (incl -EAGAIN) with page unlocked.
 986  */
 987 static int ceph_update_writeable_page(struct file *file,
 988                             loff_t pos, unsigned len,
 989                             struct page *page)
 990 {
 991         struct inode *inode = file_inode(file);
 992         struct ceph_inode_info *ci = ceph_inode(inode);
 993         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 994         loff_t page_off = pos & PAGE_CACHE_MASK;
 995         int pos_in_page = pos & ~PAGE_CACHE_MASK;
 996         int end_in_page = pos_in_page + len;
 997         loff_t i_size;
 998         int r;
 999         struct ceph_snap_context *snapc, *oldest;
1000
1001 retry_locked:
1002         /* writepages currently holds page lock, but if we change that later, */
1003         wait_on_page_writeback(page);
1004
1005         /* check snap context */
1006         BUG_ON(!ci->i_snap_realm);
1007         down_read(&mdsc->snap_rwsem);
1008         BUG_ON(!ci->i_snap_realm->cached_context);
1009         snapc = page_snap_context(page);
1010         if (snapc && snapc != ci->i_head_snapc) {
1011                 /*
1012                  * this page is already dirty in another (older) snap
1013                  * context!  is it writeable now?
1014                  */
1015                 oldest = get_oldest_context(inode, NULL);
1016                 up_read(&mdsc->snap_rwsem);
1017
1018                 if (snapc->seq > oldest->seq) {
1019                         ceph_put_snap_context(oldest);
1020                         dout(" page %p snapc %p not current or oldest\n",
1021                              page, snapc);
1022                         /*
1023                          * queue for writeback, and wait for snapc to
1024                          * be writeable or written
1025                          */
1026                         snapc = ceph_get_snap_context(snapc);
1027                         unlock_page(page);
1028                         ceph_queue_writeback(inode);
1029                         r = wait_event_interruptible(ci->i_cap_wq,
1030                                context_is_writeable_or_written(inode, snapc));
1031                         ceph_put_snap_context(snapc);
1032                         if (r == -ERESTARTSYS)
1033                                 return r;
1034                         return -EAGAIN;
1035                 }
1036                 ceph_put_snap_context(oldest);
1037
1038                 /* yay, writeable, do it now (without dropping page lock) */
1039                 dout(" page %p snapc %p not current, but oldest\n",
1040                      page, snapc);
1041                 if (!clear_page_dirty_for_io(page))
1042                         goto retry_locked;
1043                 r = writepage_nounlock(page, NULL);
1044                 if (r < 0)
1045                         goto fail_nosnap;
1046                 goto retry_locked;
1047         }
1048
1049         if (PageUptodate(page)) {
1050                 dout(" page %p already uptodate\n", page);
1051                 return 0;
1052         }
1053
1054         /* full page? */
1055         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1056                 return 0;
1057
1058         /* past end of file? */
1059         i_size = inode->i_size;   /* caller holds i_mutex */
1060
1061         if (i_size + len > inode->i_sb->s_maxbytes) {
1062                 /* file is too big */
1063                 r = -EINVAL;
1064                 goto fail;
1065         }
1066
1067         if (page_off >= i_size ||
1068             (pos_in_page == 0 && (pos+len) >= i_size &&
1069              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1070                 dout(" zeroing %p 0 - %d and %d - %d\n",
1071                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1072                 zero_user_segments(page,
1073                                    0, pos_in_page,
1074                                    end_in_page, PAGE_CACHE_SIZE);
1075                 return 0;
1076         }
1077
1078         /* we need to read it. */
1079         up_read(&mdsc->snap_rwsem);
1080         r = readpage_nounlock(file, page);
1081         if (r < 0)
1082                 goto fail_nosnap;
1083         goto retry_locked;
1084
1085 fail:
1086         up_read(&mdsc->snap_rwsem);
1087 fail_nosnap:
1088         unlock_page(page);
1089         return r;
1090 }
1091
1092 /*
1093  * We are only allowed to write into/dirty the page if the page is
1094  * clean, or already dirty within the same snap context.
1095  */
1096 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1097                             loff_t pos, unsigned len, unsigned flags,
1098                             struct page **pagep, void **fsdata)
1099 {
1100         struct inode *inode = file_inode(file);
1101         struct page *page;
1102         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1103         int r;
1104
1105         do {
1106                 /* get a page */
1107                 page = grab_cache_page_write_begin(mapping, index, 0);
1108                 if (!page)
1109                         return -ENOMEM;
1110                 *pagep = page;
1111
1112                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1113                      inode, page, (int)pos, (int)len);
1114
1115                 r = ceph_update_writeable_page(file, pos, len, page);
1116         } while (r == -EAGAIN);
1117
1118         return r;
1119 }
1120
1121 /*
1122  * we don't do anything in here that simple_write_end doesn't do
1123  * except adjust dirty page accounting and drop read lock on
1124  * mdsc->snap_rwsem.
1125  */
1126 static int ceph_write_end(struct file *file, struct address_space *mapping,
1127                           loff_t pos, unsigned len, unsigned copied,
1128                           struct page *page, void *fsdata)
1129 {
1130         struct inode *inode = file_inode(file);
1131         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1132         struct ceph_mds_client *mdsc = fsc->mdsc;
1133         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1134         int check_cap = 0;
1135
1136         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1137              inode, page, (int)pos, (int)copied, (int)len);
1138
1139         /* zero the stale part of the page if we did a short copy */
1140         if (copied < len)
1141                 zero_user_segment(page, from+copied, len);
1142
1143         /* did file size increase? */
1144         /* (no need for i_size_read(); we caller holds i_mutex */
1145         if (pos+copied > inode->i_size)
1146                 check_cap = ceph_inode_set_size(inode, pos+copied);
1147
1148         if (!PageUptodate(page))
1149                 SetPageUptodate(page);
1150
1151         set_page_dirty(page);
1152
1153         unlock_page(page);
1154         up_read(&mdsc->snap_rwsem);
1155         page_cache_release(page);
1156
1157         if (check_cap)
1158                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1159
1160         return copied;
1161 }
1162
1163 /*
1164  * we set .direct_IO to indicate direct io is supported, but since we
1165  * intercept O_DIRECT reads and writes early, this function should
1166  * never get called.
1167  */
1168 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1169                               const struct iovec *iov,
1170                               loff_t pos, unsigned long nr_segs)
1171 {
1172         WARN_ON(1);
1173         return -EINVAL;
1174 }
1175
1176 const struct address_space_operations ceph_aops = {
1177         .readpage = ceph_readpage,
1178         .readpages = ceph_readpages,
1179         .writepage = ceph_writepage,
1180         .writepages = ceph_writepages_start,
1181         .write_begin = ceph_write_begin,
1182         .write_end = ceph_write_end,
1183         .set_page_dirty = ceph_set_page_dirty,
1184         .invalidatepage = ceph_invalidatepage,
1185         .releasepage = ceph_releasepage,
1186         .direct_IO = ceph_direct_io,
1187 };
1188
1189
1190 /*
1191  * vm ops
1192  */
1193
1194 /*
1195  * Reuse write_begin here for simplicity.
1196  */
1197 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1198 {
1199         struct inode *inode = file_inode(vma->vm_file);
1200         struct page *page = vmf->page;
1201         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1202         loff_t off = page_offset(page);
1203         loff_t size, len;
1204         int ret;
1205
1206         /* Update time before taking page lock */
1207         file_update_time(vma->vm_file);
1208
1209         size = i_size_read(inode);
1210         if (off + PAGE_CACHE_SIZE <= size)
1211                 len = PAGE_CACHE_SIZE;
1212         else
1213                 len = size & ~PAGE_CACHE_MASK;
1214
1215         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1216              off, len, page, page->index);
1217
1218         lock_page(page);
1219
1220         ret = VM_FAULT_NOPAGE;
1221         if ((off > size) ||
1222             (page->mapping != inode->i_mapping))
1223                 goto out;
1224
1225         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1226         if (ret == 0) {
1227                 /* success.  we'll keep the page locked. */
1228                 set_page_dirty(page);
1229                 up_read(&mdsc->snap_rwsem);
1230                 ret = VM_FAULT_LOCKED;
1231         } else {
1232                 if (ret == -ENOMEM)
1233                         ret = VM_FAULT_OOM;
1234                 else
1235                         ret = VM_FAULT_SIGBUS;
1236         }
1237 out:
1238         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1239         if (ret != VM_FAULT_LOCKED)
1240                 unlock_page(page);
1241         return ret;
1242 }
1243
1244 static struct vm_operations_struct ceph_vmops = {
1245         .fault          = filemap_fault,
1246         .page_mkwrite   = ceph_page_mkwrite,
1247         .remap_pages    = generic_file_remap_pages,
1248 };
1249
1250 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1251 {
1252         struct address_space *mapping = file->f_mapping;
1253
1254         if (!mapping->a_ops->readpage)
1255                 return -ENOEXEC;
1256         file_accessed(file);
1257         vma->vm_ops = &ceph_vmops;
1258         return 0;
1259 }