fs/ceph/addr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/backing-dev.h>
   5 #include <linux/fs.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>    /* generic_writepages */
   9 #include <linux/slab.h>
  10 #include <linux/pagevec.h>
  11 #include <linux/task_io_accounting_ops.h>
  12 #include <linux/signal.h>
  13 #include <linux/iversion.h>
  14
  15 #include "super.h"
  16 #include "mds_client.h"
  17 #include "cache.h"
  18 #include <linux/ceph/osd_client.h>
  19 #include <linux/ceph/striper.h>
  20
  21 /*
  22  * Ceph address space ops.
  23  *
  24  * There are a few funny things going on here.
  25  *
  26  * The page->private field is used to reference a struct
  27  * ceph_snap_context for _every_ dirty page.  This indicates which
  28  * snapshot the page was logically dirtied in, and thus which snap
  29  * context needs to be associated with the osd write during writeback.
  30  *
  31  * Similarly, struct ceph_inode_info maintains a set of counters to
  32  * count dirty pages on the inode.  In the absence of snapshots,
  33  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  34  *
  35  * When a snapshot is taken (that is, when the client receives
  36  * notification that a snapshot was taken), each inode with caps and
  37  * with dirty pages (dirty pages implies there is a cap) gets a new
  38  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  39  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  40  * moved to capsnap->dirty. (Unless a sync write is currently in
  41  * progress.  In that case, the capsnap is said to be "pending", new
  42  * writes cannot start, and the capsnap isn't "finalized" until the
  43  * write completes (or fails) and a final size/mtime for the inode for
  44  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  45  *
  46  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  47  * we look for the first capsnap in i_cap_snaps and write out pages in
  48  * that snap context _only_.  Then we move on to the next capsnap,
  49  * eventually reaching the "live" or "head" context (i.e., pages that
  50  * are not yet snapped) and are writing the most recently dirtied
  51  * pages.
  52  *
  53  * Invalidate and so forth must take care to ensure the dirty page
  54  * accounting is preserved.
  55  */
  56
  57 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  58 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  59         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  60          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  61
  62 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  63 {
  64         if (PagePrivate(page))
  65                 return (void *)page->private;
  66         return NULL;
  67 }
  68
  69 /*
  70  * Dirty a page.  Optimistically adjust accounting, on the assumption
  71  * that we won't race with invalidate.  If we do, readjust.
  72  */
  73 static int ceph_set_page_dirty(struct page *page)
  74 {
  75         struct address_space *mapping = page->mapping;
  76         struct inode *inode;
  77         struct ceph_inode_info *ci;
  78         struct ceph_snap_context *snapc;
  79         int ret;
  80
  81         if (unlikely(!mapping))
  82                 return !TestSetPageDirty(page);
  83
  84         if (PageDirty(page)) {
  85                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  86                      mapping->host, page, page->index);
  87                 BUG_ON(!PagePrivate(page));
  88                 return 0;
  89         }
  90
  91         inode = mapping->host;
  92         ci = ceph_inode(inode);
  93
  94         /* dirty the head */
  95         spin_lock(&ci->i_ceph_lock);
  96         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  97         if (__ceph_have_pending_cap_snap(ci)) {
  98                 struct ceph_cap_snap *capsnap =
  99                                 list_last_entry(&ci->i_cap_snaps,
 100                                                 struct ceph_cap_snap,
 101                                                 ci_item);
 102                 snapc = ceph_get_snap_context(capsnap->context);
 103                 capsnap->dirty_pages++;
 104         } else {
 105                 BUG_ON(!ci->i_head_snapc);
 106                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 107                 ++ci->i_wrbuffer_ref_head;
 108         }
 109         if (ci->i_wrbuffer_ref == 0)
 110                 ihold(inode);
 111         ++ci->i_wrbuffer_ref;
 112         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 113              "snapc %p seq %lld (%d snaps)\n",
 114              mapping->host, page, page->index,
 115              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 116              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 117              snapc, snapc->seq, snapc->num_snaps);
 118         spin_unlock(&ci->i_ceph_lock);
 119
 120         /*
 121          * Reference snap context in page->private.  Also set
 122          * PagePrivate so that we get invalidatepage callback.
 123          */
 124         BUG_ON(PagePrivate(page));
 125         page->private = (unsigned long)snapc;
 126         SetPagePrivate(page);
 127
 128         ret = __set_page_dirty_nobuffers(page);
 129         WARN_ON(!PageLocked(page));
 130         WARN_ON(!page->mapping);
 131
 132         return ret;
 133 }
 134
 135 /*
 136  * If we are truncating the full page (i.e. offset == 0), adjust the
 137  * dirty page counters appropriately.  Only called if there is private
 138  * data on the page.
 139  */
 140 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 141                                 unsigned int length)
 142 {
 143         struct inode *inode;
 144         struct ceph_inode_info *ci;
 145         struct ceph_snap_context *snapc = page_snap_context(page);
 146
 147         inode = page->mapping->host;
 148         ci = ceph_inode(inode);
 149
 150         if (offset != 0 || length != PAGE_SIZE) {
 151                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 152                      inode, page, page->index, offset, length);
 153                 return;
 154         }
 155
 156         ceph_invalidate_fscache_page(inode, page);
 157
 158         WARN_ON(!PageLocked(page));
 159         if (!PagePrivate(page))
 160                 return;
 161
 162         dout("%p invalidatepage %p idx %lu full dirty page\n",
 163              inode, page, page->index);
 164
 165         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 166         ceph_put_snap_context(snapc);
 167         page->private = 0;
 168         ClearPagePrivate(page);
 169 }
 170
 171 static int ceph_releasepage(struct page *page, gfp_t g)
 172 {
 173         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 174              page, page->index, PageDirty(page) ? "" : "not ");
 175
 176         /* Can we release the page from the cache? */
 177         if (!ceph_release_fscache_page(page, g))
 178                 return 0;
 179
 180         return !PagePrivate(page);
 181 }
 182
 183 /*
 184  * Read some contiguous pages.  If we cross a stripe boundary, shorten
 185  * *plen.  Return number of bytes read, or error.
 186  */
 187 static int ceph_sync_readpages(struct ceph_fs_client *fsc,
 188                                struct ceph_vino vino,
 189                                struct ceph_file_layout *layout,
 190                                u64 off, u64 *plen,
 191                                u32 truncate_seq, u64 truncate_size,
 192                                struct page **pages, int num_pages,
 193                                int page_align)
 194 {
 195         struct ceph_osd_client *osdc = &fsc->client->osdc;
 196         struct ceph_osd_request *req;
 197         int rc = 0;
 198
 199         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
 200              vino.snap, off, *plen);
 201         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
 202                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 203                                     NULL, truncate_seq, truncate_size,
 204                                     false);
 205         if (IS_ERR(req))
 206                 return PTR_ERR(req);
 207
 208         /* it may be a short read due to an object boundary */
 209         osd_req_op_extent_osd_data_pages(req, 0,
 210                                 pages, *plen, page_align, false, false);
 211
 212         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
 213              off, *plen, *plen, page_align);
 214
 215         rc = ceph_osdc_start_request(osdc, req, false);
 216         if (!rc)
 217                 rc = ceph_osdc_wait_request(osdc, req);
 218
 219         ceph_osdc_put_request(req);
 220         dout("readpages result %d\n", rc);
 221         return rc;
 222 }
 223
 224 /*
 225  * read a single page, without unlocking it.
 226  */
 227 static int ceph_do_readpage(struct file *filp, struct page *page)
 228 {
 229         struct inode *inode = file_inode(filp);
 230         struct ceph_inode_info *ci = ceph_inode(inode);
 231         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 232         int err = 0;
 233         u64 off = page_offset(page);
 234         u64 len = PAGE_SIZE;
 235
 236         if (off >= i_size_read(inode)) {
 237                 zero_user_segment(page, 0, PAGE_SIZE);
 238                 SetPageUptodate(page);
 239                 return 0;
 240         }
 241
 242         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 243                 /*
 244                  * Uptodate inline data should have been added
 245                  * into page cache while getting Fcr caps.
 246                  */
 247                 if (off == 0)
 248                         return -EINVAL;
 249                 zero_user_segment(page, 0, PAGE_SIZE);
 250                 SetPageUptodate(page);
 251                 return 0;
 252         }
 253
 254         err = ceph_readpage_from_fscache(inode, page);
 255         if (err == 0)
 256                 return -EINPROGRESS;
 257
 258         dout("readpage inode %p file %p page %p index %lu\n",
 259              inode, filp, page, page->index);
 260         err = ceph_sync_readpages(fsc, ceph_vino(inode),
 261                                   &ci->i_layout, off, &len,
 262                                   ci->i_truncate_seq, ci->i_truncate_size,
 263                                   &page, 1, 0);
 264         if (err == -ENOENT)
 265                 err = 0;
 266         if (err < 0) {
 267                 SetPageError(page);
 268                 ceph_fscache_readpage_cancel(inode, page);
 269                 if (err == -EBLACKLISTED)
 270                         fsc->blacklisted = true;
 271                 goto out;
 272         }
 273         if (err < PAGE_SIZE)
 274                 /* zero fill remainder of page */
 275                 zero_user_segment(page, err, PAGE_SIZE);
 276         else
 277                 flush_dcache_page(page);
 278
 279         SetPageUptodate(page);
 280         ceph_readpage_to_fscache(inode, page);
 281
 282 out:
 283         return err < 0 ? err : 0;
 284 }
 285
 286 static int ceph_readpage(struct file *filp, struct page *page)
 287 {
 288         int r = ceph_do_readpage(filp, page);
 289         if (r != -EINPROGRESS)
 290                 unlock_page(page);
 291         else
 292                 r = 0;
 293         return r;
 294 }
 295
 296 /*
 297  * Finish an async read(ahead) op.
 298  */
 299 static void finish_read(struct ceph_osd_request *req)
 300 {
 301         struct inode *inode = req->r_inode;
 302         struct ceph_osd_data *osd_data;
 303         int rc = req->r_result <= 0 ? req->r_result : 0;
 304         int bytes = req->r_result >= 0 ? req->r_result : 0;
 305         int num_pages;
 306         int i;
 307
 308         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 309         if (rc == -EBLACKLISTED)
 310                 ceph_inode_to_client(inode)->blacklisted = true;
 311
 312         /* unlock all pages, zeroing any data we didn't read */
 313         osd_data = osd_req_op_extent_osd_data(req, 0);
 314         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 315         num_pages = calc_pages_for((u64)osd_data->alignment,
 316                                         (u64)osd_data->length);
 317         for (i = 0; i < num_pages; i++) {
 318                 struct page *page = osd_data->pages[i];
 319
 320                 if (rc < 0 && rc != -ENOENT) {
 321                         ceph_fscache_readpage_cancel(inode, page);
 322                         goto unlock;
 323                 }
 324                 if (bytes < (int)PAGE_SIZE) {
 325                         /* zero (remainder of) page */
 326                         int s = bytes < 0 ? 0 : bytes;
 327                         zero_user_segment(page, s, PAGE_SIZE);
 328                 }
 329                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 330                      page->index);
 331                 flush_dcache_page(page);
 332                 SetPageUptodate(page);
 333                 ceph_readpage_to_fscache(inode, page);
 334 unlock:
 335                 unlock_page(page);
 336                 put_page(page);
 337                 bytes -= PAGE_SIZE;
 338         }
 339         kfree(osd_data->pages);
 340 }
 341
 342 /*
 343  * start an async read(ahead) operation.  return nr_pages we submitted
 344  * a read for on success, or negative error code.
 345  */
 346 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
 347                       struct list_head *page_list, int max)
 348 {
 349         struct ceph_osd_client *osdc =
 350                 &ceph_inode_to_client(inode)->client->osdc;
 351         struct ceph_inode_info *ci = ceph_inode(inode);
 352         struct page *page = lru_to_page(page_list);
 353         struct ceph_vino vino;
 354         struct ceph_osd_request *req;
 355         u64 off;
 356         u64 len;
 357         int i;
 358         struct page **pages;
 359         pgoff_t next_index;
 360         int nr_pages = 0;
 361         int got = 0;
 362         int ret = 0;
 363
 364         if (!rw_ctx) {
 365                 /* caller of readpages does not hold buffer and read caps
 366                  * (fadvise, madvise and readahead cases) */
 367                 int want = CEPH_CAP_FILE_CACHE;
 368                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want,
 369                                         true, &got);
 370                 if (ret < 0) {
 371                         dout("start_read %p, error getting cap\n", inode);
 372                 } else if (!(got & want)) {
 373                         dout("start_read %p, no cache cap\n", inode);
 374                         ret = 0;
 375                 }
 376                 if (ret <= 0) {
 377                         if (got)
 378                                 ceph_put_cap_refs(ci, got);
 379                         while (!list_empty(page_list)) {
 380                                 page = lru_to_page(page_list);
 381                                 list_del(&page->lru);
 382                                 put_page(page);
 383                         }
 384                         return ret;
 385                 }
 386         }
 387
 388         off = (u64) page_offset(page);
 389
 390         /* count pages */
 391         next_index = page->index;
 392         list_for_each_entry_reverse(page, page_list, lru) {
 393                 if (page->index != next_index)
 394                         break;
 395                 nr_pages++;
 396                 next_index++;
 397                 if (max && nr_pages == max)
 398                         break;
 399         }
 400         len = nr_pages << PAGE_SHIFT;
 401         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 402              off, len);
 403         vino = ceph_vino(inode);
 404         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 405                                     0, 1, CEPH_OSD_OP_READ,
 406                                     CEPH_OSD_FLAG_READ, NULL,
 407                                     ci->i_truncate_seq, ci->i_truncate_size,
 408                                     false);
 409         if (IS_ERR(req)) {
 410                 ret = PTR_ERR(req);
 411                 goto out;
 412         }
 413
 414         /* build page vector */
 415         nr_pages = calc_pages_for(0, len);
 416         pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
 417         if (!pages) {
 418                 ret = -ENOMEM;
 419                 goto out_put;
 420         }
 421         for (i = 0; i < nr_pages; ++i) {
 422                 page = list_entry(page_list->prev, struct page, lru);
 423                 BUG_ON(PageLocked(page));
 424                 list_del(&page->lru);
 425
 426                 dout("start_read %p adding %p idx %lu\n", inode, page,
 427                      page->index);
 428                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 429                                           GFP_KERNEL)) {
 430                         ceph_fscache_uncache_page(inode, page);
 431                         put_page(page);
 432                         dout("start_read %p add_to_page_cache failed %p\n",
 433                              inode, page);
 434                         nr_pages = i;
 435                         if (nr_pages > 0) {
 436                                 len = nr_pages << PAGE_SHIFT;
 437                                 osd_req_op_extent_update(req, 0, len);
 438                                 break;
 439                         }
 440                         goto out_pages;
 441                 }
 442                 pages[i] = page;
 443         }
 444         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 445         req->r_callback = finish_read;
 446         req->r_inode = inode;
 447
 448         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 449         ret = ceph_osdc_start_request(osdc, req, false);
 450         if (ret < 0)
 451                 goto out_pages;
 452         ceph_osdc_put_request(req);
 453
 454         /* After adding locked pages to page cache, the inode holds cache cap.
 455          * So we can drop our cap refs. */
 456         if (got)
 457                 ceph_put_cap_refs(ci, got);
 458
 459         return nr_pages;
 460
 461 out_pages:
 462         for (i = 0; i < nr_pages; ++i) {
 463                 ceph_fscache_readpage_cancel(inode, pages[i]);
 464                 unlock_page(pages[i]);
 465         }
 466         ceph_put_page_vector(pages, nr_pages, false);
 467 out_put:
 468         ceph_osdc_put_request(req);
 469 out:
 470         if (got)
 471                 ceph_put_cap_refs(ci, got);
 472         return ret;
 473 }
 474
 475
 476 /*
 477  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 478  * the caller (VM) cleans them up.
 479  */
 480 static int ceph_readpages(struct file *file, struct address_space *mapping,
 481                           struct list_head *page_list, unsigned nr_pages)
 482 {
 483         struct inode *inode = file_inode(file);
 484         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 485         struct ceph_file_info *fi = file->private_data;
 486         struct ceph_rw_context *rw_ctx;
 487         int rc = 0;
 488         int max = 0;
 489
 490         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 491                 return -EINVAL;
 492
 493         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 494                                          &nr_pages);
 495
 496         if (rc == 0)
 497                 goto out;
 498
 499         rw_ctx = ceph_find_rw_context(fi);
 500         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 501         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 502              inode, file, rw_ctx, nr_pages, max);
 503         while (!list_empty(page_list)) {
 504                 rc = start_read(inode, rw_ctx, page_list, max);
 505                 if (rc < 0)
 506                         goto out;
 507         }
 508 out:
 509         ceph_fscache_readpages_cancel(inode, page_list);
 510
 511         dout("readpages %p file %p ret %d\n", inode, file, rc);
 512         return rc;
 513 }
 514
 515 struct ceph_writeback_ctl
 516 {
 517         loff_t i_size;
 518         u64 truncate_size;
 519         u32 truncate_seq;
 520         bool size_stable;
 521         bool head_snapc;
 522 };
 523
 524 /*
 525  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 526  * only snap context we are allowed to write back.
 527  */
 528 static struct ceph_snap_context *
 529 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 530                    struct ceph_snap_context *page_snapc)
 531 {
 532         struct ceph_inode_info *ci = ceph_inode(inode);
 533         struct ceph_snap_context *snapc = NULL;
 534         struct ceph_cap_snap *capsnap = NULL;
 535
 536         spin_lock(&ci->i_ceph_lock);
 537         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 538                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 539                      capsnap->context, capsnap->dirty_pages);
 540                 if (!capsnap->dirty_pages)
 541                         continue;
 542
 543                 /* get i_size, truncate_{seq,size} for page_snapc? */
 544                 if (snapc && capsnap->context != page_snapc)
 545                         continue;
 546
 547                 if (ctl) {
 548                         if (capsnap->writing) {
 549                                 ctl->i_size = i_size_read(inode);
 550                                 ctl->size_stable = false;
 551                         } else {
 552                                 ctl->i_size = capsnap->size;
 553                                 ctl->size_stable = true;
 554                         }
 555                         ctl->truncate_size = capsnap->truncate_size;
 556                         ctl->truncate_seq = capsnap->truncate_seq;
 557                         ctl->head_snapc = false;
 558                 }
 559
 560                 if (snapc)
 561                         break;
 562
 563                 snapc = ceph_get_snap_context(capsnap->context);
 564                 if (!page_snapc ||
 565                     page_snapc == snapc ||
 566                     page_snapc->seq > snapc->seq)
 567                         break;
 568         }
 569         if (!snapc && ci->i_wrbuffer_ref_head) {
 570                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 571                 dout(" head snapc %p has %d dirty pages\n",
 572                      snapc, ci->i_wrbuffer_ref_head);
 573                 if (ctl) {
 574                         ctl->i_size = i_size_read(inode);
 575                         ctl->truncate_size = ci->i_truncate_size;
 576                         ctl->truncate_seq = ci->i_truncate_seq;
 577                         ctl->size_stable = false;
 578                         ctl->head_snapc = true;
 579                 }
 580         }
 581         spin_unlock(&ci->i_ceph_lock);
 582         return snapc;
 583 }
 584
 585 static u64 get_writepages_data_length(struct inode *inode,
 586                                       struct page *page, u64 start)
 587 {
 588         struct ceph_inode_info *ci = ceph_inode(inode);
 589         struct ceph_snap_context *snapc = page_snap_context(page);
 590         struct ceph_cap_snap *capsnap = NULL;
 591         u64 end = i_size_read(inode);
 592
 593         if (snapc != ci->i_head_snapc) {
 594                 bool found = false;
 595                 spin_lock(&ci->i_ceph_lock);
 596                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 597                         if (capsnap->context == snapc) {
 598                                 if (!capsnap->writing)
 599                                         end = capsnap->size;
 600                                 found = true;
 601                                 break;
 602                         }
 603                 }
 604                 spin_unlock(&ci->i_ceph_lock);
 605                 WARN_ON(!found);
 606         }
 607         if (end > page_offset(page) + PAGE_SIZE)
 608                 end = page_offset(page) + PAGE_SIZE;
 609         return end > start ? end - start : 0;
 610 }
 611
 612 /*
 613  * do a synchronous write on N pages
 614  */
 615 static int ceph_sync_writepages(struct ceph_fs_client *fsc,
 616                                 struct ceph_vino vino,
 617                                 struct ceph_file_layout *layout,
 618                                 struct ceph_snap_context *snapc,
 619                                 u64 off, u64 len,
 620                                 u32 truncate_seq, u64 truncate_size,
 621                                 struct timespec64 *mtime,
 622                                 struct page **pages, int num_pages)
 623 {
 624         struct ceph_osd_client *osdc = &fsc->client->osdc;
 625         struct ceph_osd_request *req;
 626         int rc = 0;
 627         int page_align = off & ~PAGE_MASK;
 628
 629         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
 630                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
 631                                     snapc, truncate_seq, truncate_size,
 632                                     true);
 633         if (IS_ERR(req))
 634                 return PTR_ERR(req);
 635
 636         /* it may be a short write due to an object boundary */
 637         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
 638                                 false, false);
 639         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
 640
 641         req->r_mtime = *mtime;
 642         rc = ceph_osdc_start_request(osdc, req, true);
 643         if (!rc)
 644                 rc = ceph_osdc_wait_request(osdc, req);
 645
 646         ceph_osdc_put_request(req);
 647         if (rc == 0)
 648                 rc = len;
 649         dout("writepages result %d\n", rc);
 650         return rc;
 651 }
 652
 653 /*
 654  * Write a single page, but leave the page locked.
 655  *
 656  * If we get a write error, mark the mapping for error, but still adjust the
 657  * dirty page accounting (i.e., page is no longer dirty).
 658  */
 659 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 660 {
 661         struct inode *inode;
 662         struct ceph_inode_info *ci;
 663         struct ceph_fs_client *fsc;
 664         struct ceph_snap_context *snapc, *oldest;
 665         loff_t page_off = page_offset(page);
 666         int err, len = PAGE_SIZE;
 667         struct ceph_writeback_ctl ceph_wbc;
 668
 669         dout("writepage %p idx %lu\n", page, page->index);
 670
 671         inode = page->mapping->host;
 672         ci = ceph_inode(inode);
 673         fsc = ceph_inode_to_client(inode);
 674
 675         /* verify this is a writeable snap context */
 676         snapc = page_snap_context(page);
 677         if (!snapc) {
 678                 dout("writepage %p page %p not dirty?\n", inode, page);
 679                 return 0;
 680         }
 681         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 682         if (snapc->seq > oldest->seq) {
 683                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 684                      inode, page, snapc);
 685                 /* we should only noop if called by kswapd */
 686                 WARN_ON(!(current->flags & PF_MEMALLOC));
 687                 ceph_put_snap_context(oldest);
 688                 redirty_page_for_writepage(wbc, page);
 689                 return 0;
 690         }
 691         ceph_put_snap_context(oldest);
 692
 693         /* is this a partial page at end of file? */
 694         if (page_off >= ceph_wbc.i_size) {
 695                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 696                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 697                 return 0;
 698         }
 699
 700         if (ceph_wbc.i_size < page_off + len)
 701                 len = ceph_wbc.i_size - page_off;
 702
 703         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 704              inode, page, page->index, page_off, len, snapc, snapc->seq);
 705
 706         if (atomic_long_inc_return(&fsc->writeback_count) >
 707             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 708                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 709
 710         set_page_writeback(page);
 711         err = ceph_sync_writepages(fsc, ceph_vino(inode),
 712                                    &ci->i_layout, snapc, page_off, len,
 713                                    ceph_wbc.truncate_seq,
 714                                    ceph_wbc.truncate_size,
 715                                    &inode->i_mtime, &page, 1);
 716         if (err < 0) {
 717                 struct writeback_control tmp_wbc;
 718                 if (!wbc)
 719                         wbc = &tmp_wbc;
 720                 if (err == -ERESTARTSYS) {
 721                         /* killed by SIGKILL */
 722                         dout("writepage interrupted page %p\n", page);
 723                         redirty_page_for_writepage(wbc, page);
 724                         end_page_writeback(page);
 725                         return err;
 726                 }
 727                 if (err == -EBLACKLISTED)
 728                         fsc->blacklisted = true;
 729                 dout("writepage setting page/mapping error %d %p\n",
 730                      err, page);
 731                 mapping_set_error(&inode->i_data, err);
 732                 wbc->pages_skipped++;
 733         } else {
 734                 dout("writepage cleaned page %p\n", page);
 735                 err = 0;  /* vfs expects us to return 0 */
 736         }
 737         page->private = 0;
 738         ClearPagePrivate(page);
 739         end_page_writeback(page);
 740         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 741         ceph_put_snap_context(snapc);  /* page's reference */
 742
 743         if (atomic_long_dec_return(&fsc->writeback_count) <
 744             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 745                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 746
 747         return err;
 748 }
 749
 750 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 751 {
 752         int err;
 753         struct inode *inode = page->mapping->host;
 754         BUG_ON(!inode);
 755         ihold(inode);
 756         err = writepage_nounlock(page, wbc);
 757         if (err == -ERESTARTSYS) {
 758                 /* direct memory reclaimer was killed by SIGKILL. return 0
 759                  * to prevent caller from setting mapping/page error */
 760                 err = 0;
 761         }
 762         unlock_page(page);
 763         iput(inode);
 764         return err;
 765 }
 766
 767 /*
 768  * async writeback completion handler.
 769  *
 770  * If we get an error, set the mapping error bit, but not the individual
 771  * page error bits.
 772  */
 773 static void writepages_finish(struct ceph_osd_request *req)
 774 {
 775         struct inode *inode = req->r_inode;
 776         struct ceph_inode_info *ci = ceph_inode(inode);
 777         struct ceph_osd_data *osd_data;
 778         struct page *page;
 779         int num_pages, total_pages = 0;
 780         int i, j;
 781         int rc = req->r_result;
 782         struct ceph_snap_context *snapc = req->r_snapc;
 783         struct address_space *mapping = inode->i_mapping;
 784         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 785         bool remove_page;
 786
 787         dout("writepages_finish %p rc %d\n", inode, rc);
 788         if (rc < 0) {
 789                 mapping_set_error(mapping, rc);
 790                 ceph_set_error_write(ci);
 791                 if (rc == -EBLACKLISTED)
 792                         fsc->blacklisted = true;
 793         } else {
 794                 ceph_clear_error_write(ci);
 795         }
 796
 797         /*
 798          * We lost the cache cap, need to truncate the page before
 799          * it is unlocked, otherwise we'd truncate it later in the
 800          * page truncation thread, possibly losing some data that
 801          * raced its way in
 802          */
 803         remove_page = !(ceph_caps_issued(ci) &
 804                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 805
 806         /* clean all pages */
 807         for (i = 0; i < req->r_num_ops; i++) {
 808                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 809                         break;
 810
 811                 osd_data = osd_req_op_extent_osd_data(req, i);
 812                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 813                 num_pages = calc_pages_for((u64)osd_data->alignment,
 814                                            (u64)osd_data->length);
 815                 total_pages += num_pages;
 816                 for (j = 0; j < num_pages; j++) {
 817                         page = osd_data->pages[j];
 818                         BUG_ON(!page);
 819                         WARN_ON(!PageUptodate(page));
 820
 821                         if (atomic_long_dec_return(&fsc->writeback_count) <
 822                              CONGESTION_OFF_THRESH(
 823                                         fsc->mount_options->congestion_kb))
 824                                 clear_bdi_congested(inode_to_bdi(inode),
 825                                                     BLK_RW_ASYNC);
 826
 827                         ceph_put_snap_context(page_snap_context(page));
 828                         page->private = 0;
 829                         ClearPagePrivate(page);
 830                         dout("unlocking %p\n", page);
 831                         end_page_writeback(page);
 832
 833                         if (remove_page)
 834                                 generic_error_remove_page(inode->i_mapping,
 835                                                           page);
 836
 837                         unlock_page(page);
 838                 }
 839                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 840                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 841
 842                 release_pages(osd_data->pages, num_pages);
 843         }
 844
 845         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 846
 847         osd_data = osd_req_op_extent_osd_data(req, 0);
 848         if (osd_data->pages_from_pool)
 849                 mempool_free(osd_data->pages,
 850                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 851         else
 852                 kfree(osd_data->pages);
 853         ceph_osdc_put_request(req);
 854 }
 855
 856 /*
 857  * initiate async writeback
 858  */
 859 static int ceph_writepages_start(struct address_space *mapping,
 860                                  struct writeback_control *wbc)
 861 {
 862         struct inode *inode = mapping->host;
 863         struct ceph_inode_info *ci = ceph_inode(inode);
 864         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 865         struct ceph_vino vino = ceph_vino(inode);
 866         pgoff_t index, start_index, end = -1;
 867         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 868         struct pagevec pvec;
 869         int rc = 0;
 870         unsigned int wsize = i_blocksize(inode);
 871         struct ceph_osd_request *req = NULL;
 872         struct ceph_writeback_ctl ceph_wbc;
 873         bool should_loop, range_whole = false;
 874         bool done = false;
 875
 876         dout("writepages_start %p (mode=%s)\n", inode,
 877              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 878              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 879
 880         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 881                 if (ci->i_wrbuffer_ref > 0) {
 882                         pr_warn_ratelimited(
 883                                 "writepage_start %p %lld forced umount\n",
 884                                 inode, ceph_ino(inode));
 885                 }
 886                 mapping_set_error(mapping, -EIO);
 887                 return -EIO; /* we're in a forced umount, don't write! */
 888         }
 889         if (fsc->mount_options->wsize < wsize)
 890                 wsize = fsc->mount_options->wsize;
 891
 892         pagevec_init(&pvec);
 893
 894         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 895         index = start_index;
 896
 897 retry:
 898         /* find oldest snap context with dirty data */
 899         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 900         if (!snapc) {
 901                 /* hmm, why does writepages get called when there
 902                    is no dirty data? */
 903                 dout(" no snap context with dirty data?\n");
 904                 goto out;
 905         }
 906         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 907              snapc, snapc->seq, snapc->num_snaps);
 908
 909         should_loop = false;
 910         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 911                 /* where to start/end? */
 912                 if (wbc->range_cyclic) {
 913                         index = start_index;
 914                         end = -1;
 915                         if (index > 0)
 916                                 should_loop = true;
 917                         dout(" cyclic, start at %lu\n", index);
 918                 } else {
 919                         index = wbc->range_start >> PAGE_SHIFT;
 920                         end = wbc->range_end >> PAGE_SHIFT;
 921                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 922                                 range_whole = true;
 923                         dout(" not cyclic, %lu to %lu\n", index, end);
 924                 }
 925         } else if (!ceph_wbc.head_snapc) {
 926                 /* Do not respect wbc->range_{start,end}. Dirty pages
 927                  * in that range can be associated with newer snapc.
 928                  * They are not writeable until we write all dirty pages
 929                  * associated with 'snapc' get written */
 930                 if (index > 0)
 931                         should_loop = true;
 932                 dout(" non-head snapc, range whole\n");
 933         }
 934
 935         ceph_put_snap_context(last_snapc);
 936         last_snapc = snapc;
 937
 938         while (!done && index <= end) {
 939                 int num_ops = 0, op_idx;
 940                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 941                 struct page **pages = NULL, **data_pages;
 942                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 943                 struct page *page;
 944                 pgoff_t strip_unit_end = 0;
 945                 u64 offset = 0, len = 0;
 946
 947                 max_pages = wsize >> PAGE_SHIFT;
 948
 949 get_more_pages:
 950                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 951                                                 end, PAGECACHE_TAG_DIRTY,
 952                                                 max_pages - locked_pages);
 953                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 954                 if (!pvec_pages && !locked_pages)
 955                         break;
 956                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 957                         page = pvec.pages[i];
 958                         dout("? %p idx %lu\n", page, page->index);
 959                         if (locked_pages == 0)
 960                                 lock_page(page);  /* first page */
 961                         else if (!trylock_page(page))
 962                                 break;
 963
 964                         /* only dirty pages, or our accounting breaks */
 965                         if (unlikely(!PageDirty(page)) ||
 966                             unlikely(page->mapping != mapping)) {
 967                                 dout("!dirty or !mapping %p\n", page);
 968                                 unlock_page(page);
 969                                 continue;
 970                         }
 971                         /* only if matching snap context */
 972                         pgsnapc = page_snap_context(page);
 973                         if (pgsnapc != snapc) {
 974                                 dout("page snapc %p %lld != oldest %p %lld\n",
 975                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 976                                 if (!should_loop &&
 977                                     !ceph_wbc.head_snapc &&
 978                                     wbc->sync_mode != WB_SYNC_NONE)
 979                                         should_loop = true;
 980                                 unlock_page(page);
 981                                 continue;
 982                         }
 983                         if (page_offset(page) >= ceph_wbc.i_size) {
 984                                 dout("%p page eof %llu\n",
 985                                      page, ceph_wbc.i_size);
 986                                 if ((ceph_wbc.size_stable ||
 987                                     page_offset(page) >= i_size_read(inode)) &&
 988                                     clear_page_dirty_for_io(page))
 989                                         mapping->a_ops->invalidatepage(page,
 990                                                                 0, PAGE_SIZE);
 991                                 unlock_page(page);
 992                                 continue;
 993                         }
 994                         if (strip_unit_end && (page->index > strip_unit_end)) {
 995                                 dout("end of strip unit %p\n", page);
 996                                 unlock_page(page);
 997                                 break;
 998                         }
 999                         if (PageWriteback(page)) {
1000                                 if (wbc->sync_mode == WB_SYNC_NONE) {
1001                                         dout("%p under writeback\n", page);
1002                                         unlock_page(page);
1003                                         continue;
1004                                 }
1005                                 dout("waiting on writeback %p\n", page);
1006                                 wait_on_page_writeback(page);
1007                         }
1008
1009                         if (!clear_page_dirty_for_io(page)) {
1010                                 dout("%p !clear_page_dirty_for_io\n", page);
1011                                 unlock_page(page);
1012                                 continue;
1013                         }
1014
1015                         /*
1016                          * We have something to write.  If this is
1017                          * the first locked page this time through,
1018                          * calculate max possinle write size and
1019                          * allocate a page array
1020                          */
1021                         if (locked_pages == 0) {
1022                                 u64 objnum;
1023                                 u64 objoff;
1024                                 u32 xlen;
1025
1026                                 /* prepare async write request */
1027                                 offset = (u64)page_offset(page);
1028                                 ceph_calc_file_object_mapping(&ci->i_layout,
1029                                                               offset, wsize,
1030                                                               &objnum, &objoff,
1031                                                               &xlen);
1032                                 len = xlen;
1033
1034                                 num_ops = 1;
1035                                 strip_unit_end = page->index +
1036                                         ((len - 1) >> PAGE_SHIFT);
1037
1038                                 BUG_ON(pages);
1039                                 max_pages = calc_pages_for(0, (u64)len);
1040                                 pages = kmalloc_array(max_pages,
1041                                                       sizeof(*pages),
1042                                                       GFP_NOFS);
1043                                 if (!pages) {
1044                                         pool = fsc->wb_pagevec_pool;
1045                                         pages = mempool_alloc(pool, GFP_NOFS);
1046                                         BUG_ON(!pages);
1047                                 }
1048
1049                                 len = 0;
1050                         } else if (page->index !=
1051                                    (offset + len) >> PAGE_SHIFT) {
1052                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
1053                                                         CEPH_OSD_MAX_OPS)) {
1054                                         redirty_page_for_writepage(wbc, page);
1055                                         unlock_page(page);
1056                                         break;
1057                                 }
1058
1059                                 num_ops++;
1060                                 offset = (u64)page_offset(page);
1061                                 len = 0;
1062                         }
1063
1064                         /* note position of first page in pvec */
1065                         dout("%p will write page %p idx %lu\n",
1066                              inode, page, page->index);
1067
1068                         if (atomic_long_inc_return(&fsc->writeback_count) >
1069                             CONGESTION_ON_THRESH(
1070                                     fsc->mount_options->congestion_kb)) {
1071                                 set_bdi_congested(inode_to_bdi(inode),
1072                                                   BLK_RW_ASYNC);
1073                         }
1074
1075
1076                         pages[locked_pages++] = page;
1077                         pvec.pages[i] = NULL;
1078
1079                         len += PAGE_SIZE;
1080                 }
1081
1082                 /* did we get anything? */
1083                 if (!locked_pages)
1084                         goto release_pvec_pages;
1085                 if (i) {
1086                         unsigned j, n = 0;
1087                         /* shift unused page to beginning of pvec */
1088                         for (j = 0; j < pvec_pages; j++) {
1089                                 if (!pvec.pages[j])
1090                                         continue;
1091                                 if (n < j)
1092                                         pvec.pages[n] = pvec.pages[j];
1093                                 n++;
1094                         }
1095                         pvec.nr = n;
1096
1097                         if (pvec_pages && i == pvec_pages &&
1098                             locked_pages < max_pages) {
1099                                 dout("reached end pvec, trying for more\n");
1100                                 pagevec_release(&pvec);
1101                                 goto get_more_pages;
1102                         }
1103                 }
1104
1105 new_request:
1106                 offset = page_offset(pages[0]);
1107                 len = wsize;
1108
1109                 req = ceph_osdc_new_request(&fsc->client->osdc,
1110                                         &ci->i_layout, vino,
1111                                         offset, &len, 0, num_ops,
1112                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1113                                         snapc, ceph_wbc.truncate_seq,
1114                                         ceph_wbc.truncate_size, false);
1115                 if (IS_ERR(req)) {
1116                         req = ceph_osdc_new_request(&fsc->client->osdc,
1117                                                 &ci->i_layout, vino,
1118                                                 offset, &len, 0,
1119                                                 min(num_ops,
1120                                                     CEPH_OSD_SLAB_OPS),
1121                                                 CEPH_OSD_OP_WRITE,
1122                                                 CEPH_OSD_FLAG_WRITE,
1123                                                 snapc, ceph_wbc.truncate_seq,
1124                                                 ceph_wbc.truncate_size, true);
1125                         BUG_ON(IS_ERR(req));
1126                 }
1127                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1128                              PAGE_SIZE - offset);
1129
1130                 req->r_callback = writepages_finish;
1131                 req->r_inode = inode;
1132
1133                 /* Format the osd request message and submit the write */
1134                 len = 0;
1135                 data_pages = pages;
1136                 op_idx = 0;
1137                 for (i = 0; i < locked_pages; i++) {
1138                         u64 cur_offset = page_offset(pages[i]);
1139                         if (offset + len != cur_offset) {
1140                                 if (op_idx + 1 == req->r_num_ops)
1141                                         break;
1142                                 osd_req_op_extent_dup_last(req, op_idx,
1143                                                            cur_offset - offset);
1144                                 dout("writepages got pages at %llu~%llu\n",
1145                                      offset, len);
1146                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1147                                                         data_pages, len, 0,
1148                                                         !!pool, false);
1149                                 osd_req_op_extent_update(req, op_idx, len);
1150
1151                                 len = 0;
1152                                 offset = cur_offset;
1153                                 data_pages = pages + i;
1154                                 op_idx++;
1155                         }
1156
1157                         set_page_writeback(pages[i]);
1158                         len += PAGE_SIZE;
1159                 }
1160
1161                 if (ceph_wbc.size_stable) {
1162                         len = min(len, ceph_wbc.i_size - offset);
1163                 } else if (i == locked_pages) {
1164                         /* writepages_finish() clears writeback pages
1165                          * according to the data length, so make sure
1166                          * data length covers all locked pages */
1167                         u64 min_len = len + 1 - PAGE_SIZE;
1168                         len = get_writepages_data_length(inode, pages[i - 1],
1169                                                          offset);
1170                         len = max(len, min_len);
1171                 }
1172                 dout("writepages got pages at %llu~%llu\n", offset, len);
1173
1174                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1175                                                  0, !!pool, false);
1176                 osd_req_op_extent_update(req, op_idx, len);
1177
1178                 BUG_ON(op_idx + 1 != req->r_num_ops);
1179
1180                 pool = NULL;
1181                 if (i < locked_pages) {
1182                         BUG_ON(num_ops <= req->r_num_ops);
1183                         num_ops -= req->r_num_ops;
1184                         locked_pages -= i;
1185
1186                         /* allocate new pages array for next request */
1187                         data_pages = pages;
1188                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1189                                               GFP_NOFS);
1190                         if (!pages) {
1191                                 pool = fsc->wb_pagevec_pool;
1192                                 pages = mempool_alloc(pool, GFP_NOFS);
1193                                 BUG_ON(!pages);
1194                         }
1195                         memcpy(pages, data_pages + i,
1196                                locked_pages * sizeof(*pages));
1197                         memset(data_pages + i, 0,
1198                                locked_pages * sizeof(*pages));
1199                 } else {
1200                         BUG_ON(num_ops != req->r_num_ops);
1201                         index = pages[i - 1]->index + 1;
1202                         /* request message now owns the pages array */
1203                         pages = NULL;
1204                 }
1205
1206                 req->r_mtime = inode->i_mtime;
1207                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1208                 BUG_ON(rc);
1209                 req = NULL;
1210
1211                 wbc->nr_to_write -= i;
1212                 if (pages)
1213                         goto new_request;
1214
1215                 /*
1216                  * We stop writing back only if we are not doing
1217                  * integrity sync. In case of integrity sync we have to
1218                  * keep going until we have written all the pages
1219                  * we tagged for writeback prior to entering this loop.
1220                  */
1221                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1222                         done = true;
1223
1224 release_pvec_pages:
1225                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1226                      pvec.nr ? pvec.pages[0] : NULL);
1227                 pagevec_release(&pvec);
1228         }
1229
1230         if (should_loop && !done) {
1231                 /* more to do; loop back to beginning of file */
1232                 dout("writepages looping back to beginning of file\n");
1233                 end = start_index - 1; /* OK even when start_index == 0 */
1234
1235                 /* to write dirty pages associated with next snapc,
1236                  * we need to wait until current writes complete */
1237                 if (wbc->sync_mode != WB_SYNC_NONE &&
1238                     start_index == 0 && /* all dirty pages were checked */
1239                     !ceph_wbc.head_snapc) {
1240                         struct page *page;
1241                         unsigned i, nr;
1242                         index = 0;
1243                         while ((index <= end) &&
1244                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1245                                                 PAGECACHE_TAG_WRITEBACK))) {
1246                                 for (i = 0; i < nr; i++) {
1247                                         page = pvec.pages[i];
1248                                         if (page_snap_context(page) != snapc)
1249                                                 continue;
1250                                         wait_on_page_writeback(page);
1251                                 }
1252                                 pagevec_release(&pvec);
1253                                 cond_resched();
1254                         }
1255                 }
1256
1257                 start_index = 0;
1258                 index = 0;
1259                 goto retry;
1260         }
1261
1262         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1263                 mapping->writeback_index = index;
1264
1265 out:
1266         ceph_osdc_put_request(req);
1267         ceph_put_snap_context(last_snapc);
1268         dout("writepages dend - startone, rc = %d\n", rc);
1269         return rc;
1270 }
1271
1272
1273
1274 /*
1275  * See if a given @snapc is either writeable, or already written.
1276  */
1277 static int context_is_writeable_or_written(struct inode *inode,
1278                                            struct ceph_snap_context *snapc)
1279 {
1280         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1281         int ret = !oldest || snapc->seq <= oldest->seq;
1282
1283         ceph_put_snap_context(oldest);
1284         return ret;
1285 }
1286
1287 /*
1288  * We are only allowed to write into/dirty the page if the page is
1289  * clean, or already dirty within the same snap context.
1290  *
1291  * called with page locked.
1292  * return success with page locked,
1293  * or any failure (incl -EAGAIN) with page unlocked.
1294  */
1295 static int ceph_update_writeable_page(struct file *file,
1296                             loff_t pos, unsigned len,
1297                             struct page *page)
1298 {
1299         struct inode *inode = file_inode(file);
1300         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1301         struct ceph_inode_info *ci = ceph_inode(inode);
1302         loff_t page_off = pos & PAGE_MASK;
1303         int pos_in_page = pos & ~PAGE_MASK;
1304         int end_in_page = pos_in_page + len;
1305         loff_t i_size;
1306         int r;
1307         struct ceph_snap_context *snapc, *oldest;
1308
1309         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1310                 dout(" page %p forced umount\n", page);
1311                 unlock_page(page);
1312                 return -EIO;
1313         }
1314
1315 retry_locked:
1316         /* writepages currently holds page lock, but if we change that later, */
1317         wait_on_page_writeback(page);
1318
1319         snapc = page_snap_context(page);
1320         if (snapc && snapc != ci->i_head_snapc) {
1321                 /*
1322                  * this page is already dirty in another (older) snap
1323                  * context!  is it writeable now?
1324                  */
1325                 oldest = get_oldest_context(inode, NULL, NULL);
1326                 if (snapc->seq > oldest->seq) {
1327                         ceph_put_snap_context(oldest);
1328                         dout(" page %p snapc %p not current or oldest\n",
1329                              page, snapc);
1330                         /*
1331                          * queue for writeback, and wait for snapc to
1332                          * be writeable or written
1333                          */
1334                         snapc = ceph_get_snap_context(snapc);
1335                         unlock_page(page);
1336                         ceph_queue_writeback(inode);
1337                         r = wait_event_killable(ci->i_cap_wq,
1338                                context_is_writeable_or_written(inode, snapc));
1339                         ceph_put_snap_context(snapc);
1340                         if (r == -ERESTARTSYS)
1341                                 return r;
1342                         return -EAGAIN;
1343                 }
1344                 ceph_put_snap_context(oldest);
1345
1346                 /* yay, writeable, do it now (without dropping page lock) */
1347                 dout(" page %p snapc %p not current, but oldest\n",
1348                      page, snapc);
1349                 if (!clear_page_dirty_for_io(page))
1350                         goto retry_locked;
1351                 r = writepage_nounlock(page, NULL);
1352                 if (r < 0)
1353                         goto fail_unlock;
1354                 goto retry_locked;
1355         }
1356
1357         if (PageUptodate(page)) {
1358                 dout(" page %p already uptodate\n", page);
1359                 return 0;
1360         }
1361
1362         /* full page? */
1363         if (pos_in_page == 0 && len == PAGE_SIZE)
1364                 return 0;
1365
1366         /* past end of file? */
1367         i_size = i_size_read(inode);
1368
1369         if (page_off >= i_size ||
1370             (pos_in_page == 0 && (pos+len) >= i_size &&
1371              end_in_page - pos_in_page != PAGE_SIZE)) {
1372                 dout(" zeroing %p 0 - %d and %d - %d\n",
1373                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1374                 zero_user_segments(page,
1375                                    0, pos_in_page,
1376                                    end_in_page, PAGE_SIZE);
1377                 return 0;
1378         }
1379
1380         /* we need to read it. */
1381         r = ceph_do_readpage(file, page);
1382         if (r < 0) {
1383                 if (r == -EINPROGRESS)
1384                         return -EAGAIN;
1385                 goto fail_unlock;
1386         }
1387         goto retry_locked;
1388 fail_unlock:
1389         unlock_page(page);
1390         return r;
1391 }
1392
1393 /*
1394  * We are only allowed to write into/dirty the page if the page is
1395  * clean, or already dirty within the same snap context.
1396  */
1397 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1398                             loff_t pos, unsigned len, unsigned flags,
1399                             struct page **pagep, void **fsdata)
1400 {
1401         struct inode *inode = file_inode(file);
1402         struct page *page;
1403         pgoff_t index = pos >> PAGE_SHIFT;
1404         int r;
1405
1406         do {
1407                 /* get a page */
1408                 page = grab_cache_page_write_begin(mapping, index, 0);
1409                 if (!page)
1410                         return -ENOMEM;
1411
1412                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1413                      inode, page, (int)pos, (int)len);
1414
1415                 r = ceph_update_writeable_page(file, pos, len, page);
1416                 if (r < 0)
1417                         put_page(page);
1418                 else
1419                         *pagep = page;
1420         } while (r == -EAGAIN);
1421
1422         return r;
1423 }
1424
1425 /*
1426  * we don't do anything in here that simple_write_end doesn't do
1427  * except adjust dirty page accounting
1428  */
1429 static int ceph_write_end(struct file *file, struct address_space *mapping,
1430                           loff_t pos, unsigned len, unsigned copied,
1431                           struct page *page, void *fsdata)
1432 {
1433         struct inode *inode = file_inode(file);
1434         bool check_cap = false;
1435
1436         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1437              inode, page, (int)pos, (int)copied, (int)len);
1438
1439         /* zero the stale part of the page if we did a short copy */
1440         if (!PageUptodate(page)) {
1441                 if (copied < len) {
1442                         copied = 0;
1443                         goto out;
1444                 }
1445                 SetPageUptodate(page);
1446         }
1447
1448         /* did file size increase? */
1449         if (pos+copied > i_size_read(inode))
1450                 check_cap = ceph_inode_set_size(inode, pos+copied);
1451
1452         set_page_dirty(page);
1453
1454 out:
1455         unlock_page(page);
1456         put_page(page);
1457
1458         if (check_cap)
1459                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1460
1461         return copied;
1462 }
1463
1464 /*
1465  * we set .direct_IO to indicate direct io is supported, but since we
1466  * intercept O_DIRECT reads and writes early, this function should
1467  * never get called.
1468  */
1469 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1470 {
1471         WARN_ON(1);
1472         return -EINVAL;
1473 }
1474
1475 const struct address_space_operations ceph_aops = {
1476         .readpage = ceph_readpage,
1477         .readpages = ceph_readpages,
1478         .writepage = ceph_writepage,
1479         .writepages = ceph_writepages_start,
1480         .write_begin = ceph_write_begin,
1481         .write_end = ceph_write_end,
1482         .set_page_dirty = ceph_set_page_dirty,
1483         .invalidatepage = ceph_invalidatepage,
1484         .releasepage = ceph_releasepage,
1485         .direct_IO = ceph_direct_io,
1486 };
1487
1488 static void ceph_block_sigs(sigset_t *oldset)
1489 {
1490         sigset_t mask;
1491         siginitsetinv(&mask, sigmask(SIGKILL));
1492         sigprocmask(SIG_BLOCK, &mask, oldset);
1493 }
1494
1495 static void ceph_restore_sigs(sigset_t *oldset)
1496 {
1497         sigprocmask(SIG_SETMASK, oldset, NULL);
1498 }
1499
1500 /*
1501  * vm ops
1502  */
1503 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1504 {
1505         struct vm_area_struct *vma = vmf->vma;
1506         struct inode *inode = file_inode(vma->vm_file);
1507         struct ceph_inode_info *ci = ceph_inode(inode);
1508         struct ceph_file_info *fi = vma->vm_file->private_data;
1509         struct page *pinned_page = NULL;
1510         loff_t off = vmf->pgoff << PAGE_SHIFT;
1511         int want, got, err;
1512         sigset_t oldset;
1513         vm_fault_t ret = VM_FAULT_SIGBUS;
1514
1515         ceph_block_sigs(&oldset);
1516
1517         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1518              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1519         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1520                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1521         else
1522                 want = CEPH_CAP_FILE_CACHE;
1523
1524         got = 0;
1525         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1,
1526                             &got, &pinned_page);
1527         if (err < 0)
1528                 goto out_restore;
1529
1530         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1531              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1532
1533         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1534             ci->i_inline_version == CEPH_INLINE_NONE) {
1535                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1536                 ceph_add_rw_context(fi, &rw_ctx);
1537                 ret = filemap_fault(vmf);
1538                 ceph_del_rw_context(fi, &rw_ctx);
1539                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1540                         inode, off, (size_t)PAGE_SIZE,
1541                                 ceph_cap_string(got), ret);
1542         } else
1543                 err = -EAGAIN;
1544
1545         if (pinned_page)
1546                 put_page(pinned_page);
1547         ceph_put_cap_refs(ci, got);
1548
1549         if (err != -EAGAIN)
1550                 goto out_restore;
1551
1552         /* read inline data */
1553         if (off >= PAGE_SIZE) {
1554                 /* does not support inline data > PAGE_SIZE */
1555                 ret = VM_FAULT_SIGBUS;
1556         } else {
1557                 struct address_space *mapping = inode->i_mapping;
1558                 struct page *page = find_or_create_page(mapping, 0,
1559                                                 mapping_gfp_constraint(mapping,
1560                                                 ~__GFP_FS));
1561                 if (!page) {
1562                         ret = VM_FAULT_OOM;
1563                         goto out_inline;
1564                 }
1565                 err = __ceph_do_getattr(inode, page,
1566                                          CEPH_STAT_CAP_INLINE_DATA, true);
1567                 if (err < 0 || off >= i_size_read(inode)) {
1568                         unlock_page(page);
1569                         put_page(page);
1570                         ret = vmf_error(err);
1571                         goto out_inline;
1572                 }
1573                 if (err < PAGE_SIZE)
1574                         zero_user_segment(page, err, PAGE_SIZE);
1575                 else
1576                         flush_dcache_page(page);
1577                 SetPageUptodate(page);
1578                 vmf->page = page;
1579                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1580 out_inline:
1581                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1582                      inode, off, (size_t)PAGE_SIZE, ret);
1583         }
1584 out_restore:
1585         ceph_restore_sigs(&oldset);
1586         if (err < 0)
1587                 ret = vmf_error(err);
1588
1589         return ret;
1590 }
1591
1592 /*
1593  * Reuse write_begin here for simplicity.
1594  */
1595 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1596 {
1597         struct vm_area_struct *vma = vmf->vma;
1598         struct inode *inode = file_inode(vma->vm_file);
1599         struct ceph_inode_info *ci = ceph_inode(inode);
1600         struct ceph_file_info *fi = vma->vm_file->private_data;
1601         struct ceph_cap_flush *prealloc_cf;
1602         struct page *page = vmf->page;
1603         loff_t off = page_offset(page);
1604         loff_t size = i_size_read(inode);
1605         size_t len;
1606         int want, got, err;
1607         sigset_t oldset;
1608         vm_fault_t ret = VM_FAULT_SIGBUS;
1609
1610         prealloc_cf = ceph_alloc_cap_flush();
1611         if (!prealloc_cf)
1612                 return VM_FAULT_OOM;
1613
1614         sb_start_pagefault(inode->i_sb);
1615         ceph_block_sigs(&oldset);
1616
1617         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1618                 struct page *locked_page = NULL;
1619                 if (off == 0) {
1620                         lock_page(page);
1621                         locked_page = page;
1622                 }
1623                 err = ceph_uninline_data(vma->vm_file, locked_page);
1624                 if (locked_page)
1625                         unlock_page(locked_page);
1626                 if (err < 0)
1627                         goto out_free;
1628         }
1629
1630         if (off + PAGE_SIZE <= size)
1631                 len = PAGE_SIZE;
1632         else
1633                 len = size & ~PAGE_MASK;
1634
1635         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1636              inode, ceph_vinop(inode), off, len, size);
1637         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1638                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1639         else
1640                 want = CEPH_CAP_FILE_BUFFER;
1641
1642         got = 0;
1643         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len,
1644                             &got, NULL);
1645         if (err < 0)
1646                 goto out_free;
1647
1648         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1649              inode, off, len, ceph_cap_string(got));
1650
1651         /* Update time before taking page lock */
1652         file_update_time(vma->vm_file);
1653         inode_inc_iversion_raw(inode);
1654
1655         do {
1656                 lock_page(page);
1657
1658                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1659                         unlock_page(page);
1660                         ret = VM_FAULT_NOPAGE;
1661                         break;
1662                 }
1663
1664                 err = ceph_update_writeable_page(vma->vm_file, off, len, page);
1665                 if (err >= 0) {
1666                         /* success.  we'll keep the page locked. */
1667                         set_page_dirty(page);
1668                         ret = VM_FAULT_LOCKED;
1669                 }
1670         } while (err == -EAGAIN);
1671
1672         if (ret == VM_FAULT_LOCKED ||
1673             ci->i_inline_version != CEPH_INLINE_NONE) {
1674                 int dirty;
1675                 spin_lock(&ci->i_ceph_lock);
1676                 ci->i_inline_version = CEPH_INLINE_NONE;
1677                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1678                                                &prealloc_cf);
1679                 spin_unlock(&ci->i_ceph_lock);
1680                 if (dirty)
1681                         __mark_inode_dirty(inode, dirty);
1682         }
1683
1684         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1685              inode, off, len, ceph_cap_string(got), ret);
1686         ceph_put_cap_refs(ci, got);
1687 out_free:
1688         ceph_restore_sigs(&oldset);
1689         sb_end_pagefault(inode->i_sb);
1690         ceph_free_cap_flush(prealloc_cf);
1691         if (err < 0)
1692                 ret = vmf_error(err);
1693         return ret;
1694 }
1695
1696 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1697                            char *data, size_t len)
1698 {
1699         struct address_space *mapping = inode->i_mapping;
1700         struct page *page;
1701
1702         if (locked_page) {
1703                 page = locked_page;
1704         } else {
1705                 if (i_size_read(inode) == 0)
1706                         return;
1707                 page = find_or_create_page(mapping, 0,
1708                                            mapping_gfp_constraint(mapping,
1709                                            ~__GFP_FS));
1710                 if (!page)
1711                         return;
1712                 if (PageUptodate(page)) {
1713                         unlock_page(page);
1714                         put_page(page);
1715                         return;
1716                 }
1717         }
1718
1719         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1720              inode, ceph_vinop(inode), len, locked_page);
1721
1722         if (len > 0) {
1723                 void *kaddr = kmap_atomic(page);
1724                 memcpy(kaddr, data, len);
1725                 kunmap_atomic(kaddr);
1726         }
1727
1728         if (page != locked_page) {
1729                 if (len < PAGE_SIZE)
1730                         zero_user_segment(page, len, PAGE_SIZE);
1731                 else
1732                         flush_dcache_page(page);
1733
1734                 SetPageUptodate(page);
1735                 unlock_page(page);
1736                 put_page(page);
1737         }
1738 }
1739
1740 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1741 {
1742         struct inode *inode = file_inode(filp);
1743         struct ceph_inode_info *ci = ceph_inode(inode);
1744         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1745         struct ceph_osd_request *req;
1746         struct page *page = NULL;
1747         u64 len, inline_version;
1748         int err = 0;
1749         bool from_pagecache = false;
1750
1751         spin_lock(&ci->i_ceph_lock);
1752         inline_version = ci->i_inline_version;
1753         spin_unlock(&ci->i_ceph_lock);
1754
1755         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1756              inode, ceph_vinop(inode), inline_version);
1757
1758         if (inline_version == 1 || /* initial version, no data */
1759             inline_version == CEPH_INLINE_NONE)
1760                 goto out;
1761
1762         if (locked_page) {
1763                 page = locked_page;
1764                 WARN_ON(!PageUptodate(page));
1765         } else if (ceph_caps_issued(ci) &
1766                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1767                 page = find_get_page(inode->i_mapping, 0);
1768                 if (page) {
1769                         if (PageUptodate(page)) {
1770                                 from_pagecache = true;
1771                                 lock_page(page);
1772                         } else {
1773                                 put_page(page);
1774                                 page = NULL;
1775                         }
1776                 }
1777         }
1778
1779         if (page) {
1780                 len = i_size_read(inode);
1781                 if (len > PAGE_SIZE)
1782                         len = PAGE_SIZE;
1783         } else {
1784                 page = __page_cache_alloc(GFP_NOFS);
1785                 if (!page) {
1786                         err = -ENOMEM;
1787                         goto out;
1788                 }
1789                 err = __ceph_do_getattr(inode, page,
1790                                         CEPH_STAT_CAP_INLINE_DATA, true);
1791                 if (err < 0) {
1792                         /* no inline data */
1793                         if (err == -ENODATA)
1794                                 err = 0;
1795                         goto out;
1796                 }
1797                 len = err;
1798         }
1799
1800         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1801                                     ceph_vino(inode), 0, &len, 0, 1,
1802                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1803                                     NULL, 0, 0, false);
1804         if (IS_ERR(req)) {
1805                 err = PTR_ERR(req);
1806                 goto out;
1807         }
1808
1809         req->r_mtime = inode->i_mtime;
1810         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1811         if (!err)
1812                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1813         ceph_osdc_put_request(req);
1814         if (err < 0)
1815                 goto out;
1816
1817         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1818                                     ceph_vino(inode), 0, &len, 1, 3,
1819                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1820                                     NULL, ci->i_truncate_seq,
1821                                     ci->i_truncate_size, false);
1822         if (IS_ERR(req)) {
1823                 err = PTR_ERR(req);
1824                 goto out;
1825         }
1826
1827         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1828
1829         {
1830                 __le64 xattr_buf = cpu_to_le64(inline_version);
1831                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1832                                             "inline_version", &xattr_buf,
1833                                             sizeof(xattr_buf),
1834                                             CEPH_OSD_CMPXATTR_OP_GT,
1835                                             CEPH_OSD_CMPXATTR_MODE_U64);
1836                 if (err)
1837                         goto out_put;
1838         }
1839
1840         {
1841                 char xattr_buf[32];
1842                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1843                                          "%llu", inline_version);
1844                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1845                                             "inline_version",
1846                                             xattr_buf, xattr_len, 0, 0);
1847                 if (err)
1848                         goto out_put;
1849         }
1850
1851         req->r_mtime = inode->i_mtime;
1852         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1853         if (!err)
1854                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1855 out_put:
1856         ceph_osdc_put_request(req);
1857         if (err == -ECANCELED)
1858                 err = 0;
1859 out:
1860         if (page && page != locked_page) {
1861                 if (from_pagecache) {
1862                         unlock_page(page);
1863                         put_page(page);
1864                 } else
1865                         __free_pages(page, 0);
1866         }
1867
1868         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1869              inode, ceph_vinop(inode), inline_version, err);
1870         return err;
1871 }
1872
1873 static const struct vm_operations_struct ceph_vmops = {
1874         .fault          = ceph_filemap_fault,
1875         .page_mkwrite   = ceph_page_mkwrite,
1876 };
1877
1878 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1879 {
1880         struct address_space *mapping = file->f_mapping;
1881
1882         if (!mapping->a_ops->readpage)
1883                 return -ENOEXEC;
1884         file_accessed(file);
1885         vma->vm_ops = &ceph_vmops;
1886         return 0;
1887 }
1888
1889 enum {
1890         POOL_READ       = 1,
1891         POOL_WRITE      = 2,
1892 };
1893
1894 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1895                                 s64 pool, struct ceph_string *pool_ns)
1896 {
1897         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1898         struct ceph_mds_client *mdsc = fsc->mdsc;
1899         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1900         struct rb_node **p, *parent;
1901         struct ceph_pool_perm *perm;
1902         struct page **pages;
1903         size_t pool_ns_len;
1904         int err = 0, err2 = 0, have = 0;
1905
1906         down_read(&mdsc->pool_perm_rwsem);
1907         p = &mdsc->pool_perm_tree.rb_node;
1908         while (*p) {
1909                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1910                 if (pool < perm->pool)
1911                         p = &(*p)->rb_left;
1912                 else if (pool > perm->pool)
1913                         p = &(*p)->rb_right;
1914                 else {
1915                         int ret = ceph_compare_string(pool_ns,
1916                                                 perm->pool_ns,
1917                                                 perm->pool_ns_len);
1918                         if (ret < 0)
1919                                 p = &(*p)->rb_left;
1920                         else if (ret > 0)
1921                                 p = &(*p)->rb_right;
1922                         else {
1923                                 have = perm->perm;
1924                                 break;
1925                         }
1926                 }
1927         }
1928         up_read(&mdsc->pool_perm_rwsem);
1929         if (*p)
1930                 goto out;
1931
1932         if (pool_ns)
1933                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1934                      pool, (int)pool_ns->len, pool_ns->str);
1935         else
1936                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1937
1938         down_write(&mdsc->pool_perm_rwsem);
1939         p = &mdsc->pool_perm_tree.rb_node;
1940         parent = NULL;
1941         while (*p) {
1942                 parent = *p;
1943                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1944                 if (pool < perm->pool)
1945                         p = &(*p)->rb_left;
1946                 else if (pool > perm->pool)
1947                         p = &(*p)->rb_right;
1948                 else {
1949                         int ret = ceph_compare_string(pool_ns,
1950                                                 perm->pool_ns,
1951                                                 perm->pool_ns_len);
1952                         if (ret < 0)
1953                                 p = &(*p)->rb_left;
1954                         else if (ret > 0)
1955                                 p = &(*p)->rb_right;
1956                         else {
1957                                 have = perm->perm;
1958                                 break;
1959                         }
1960                 }
1961         }
1962         if (*p) {
1963                 up_write(&mdsc->pool_perm_rwsem);
1964                 goto out;
1965         }
1966
1967         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1968                                          1, false, GFP_NOFS);
1969         if (!rd_req) {
1970                 err = -ENOMEM;
1971                 goto out_unlock;
1972         }
1973
1974         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1975         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1976         rd_req->r_base_oloc.pool = pool;
1977         if (pool_ns)
1978                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1979         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1980
1981         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1982         if (err)
1983                 goto out_unlock;
1984
1985         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1986                                          1, false, GFP_NOFS);
1987         if (!wr_req) {
1988                 err = -ENOMEM;
1989                 goto out_unlock;
1990         }
1991
1992         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1993         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1994         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1995         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1996
1997         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1998         if (err)
1999                 goto out_unlock;
2000
2001         /* one page should be large enough for STAT data */
2002         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2003         if (IS_ERR(pages)) {
2004                 err = PTR_ERR(pages);
2005                 goto out_unlock;
2006         }
2007
2008         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2009                                      0, false, true);
2010         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
2011
2012         wr_req->r_mtime = ci->vfs_inode.i_mtime;
2013         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
2014
2015         if (!err)
2016                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2017         if (!err2)
2018                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2019
2020         if (err >= 0 || err == -ENOENT)
2021                 have |= POOL_READ;
2022         else if (err != -EPERM) {
2023                 if (err == -EBLACKLISTED)
2024                         fsc->blacklisted = true;
2025                 goto out_unlock;
2026         }
2027
2028         if (err2 == 0 || err2 == -EEXIST)
2029                 have |= POOL_WRITE;
2030         else if (err2 != -EPERM) {
2031                 if (err2 == -EBLACKLISTED)
2032                         fsc->blacklisted = true;
2033                 err = err2;
2034                 goto out_unlock;
2035         }
2036
2037         pool_ns_len = pool_ns ? pool_ns->len : 0;
2038         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2039         if (!perm) {
2040                 err = -ENOMEM;
2041                 goto out_unlock;
2042         }
2043
2044         perm->pool = pool;
2045         perm->perm = have;
2046         perm->pool_ns_len = pool_ns_len;
2047         if (pool_ns_len > 0)
2048                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2049         perm->pool_ns[pool_ns_len] = 0;
2050
2051         rb_link_node(&perm->node, parent, p);
2052         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2053         err = 0;
2054 out_unlock:
2055         up_write(&mdsc->pool_perm_rwsem);
2056
2057         ceph_osdc_put_request(rd_req);
2058         ceph_osdc_put_request(wr_req);
2059 out:
2060         if (!err)
2061                 err = have;
2062         if (pool_ns)
2063                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2064                      pool, (int)pool_ns->len, pool_ns->str, err);
2065         else
2066                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2067         return err;
2068 }
2069
2070 int ceph_pool_perm_check(struct inode *inode, int need)
2071 {
2072         struct ceph_inode_info *ci = ceph_inode(inode);
2073         struct ceph_string *pool_ns;
2074         s64 pool;
2075         int ret, flags;
2076
2077         if (ci->i_vino.snap != CEPH_NOSNAP) {
2078                 /*
2079                  * Pool permission check needs to write to the first object.
2080                  * But for snapshot, head of the first object may have alread
2081                  * been deleted. Skip check to avoid creating orphan object.
2082                  */
2083                 return 0;
2084         }
2085
2086         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2087                                 NOPOOLPERM))
2088                 return 0;
2089
2090         spin_lock(&ci->i_ceph_lock);
2091         flags = ci->i_ceph_flags;
2092         pool = ci->i_layout.pool_id;
2093         spin_unlock(&ci->i_ceph_lock);
2094 check:
2095         if (flags & CEPH_I_POOL_PERM) {
2096                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2097                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2098                              pool);
2099                         return -EPERM;
2100                 }
2101                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2102                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2103                              pool);
2104                         return -EPERM;
2105                 }
2106                 return 0;
2107         }
2108
2109         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2110         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2111         ceph_put_string(pool_ns);
2112         if (ret < 0)
2113                 return ret;
2114
2115         flags = CEPH_I_POOL_PERM;
2116         if (ret & POOL_READ)
2117                 flags |= CEPH_I_POOL_RD;
2118         if (ret & POOL_WRITE)
2119                 flags |= CEPH_I_POOL_WR;
2120
2121         spin_lock(&ci->i_ceph_lock);
2122         if (pool == ci->i_layout.pool_id &&
2123             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2124                 ci->i_ceph_flags |= flags;
2125         } else {
2126                 pool = ci->i_layout.pool_id;
2127                 flags = ci->i_ceph_flags;
2128         }
2129         spin_unlock(&ci->i_ceph_lock);
2130         goto check;
2131 }
2132
2133 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2134 {
2135         struct ceph_pool_perm *perm;
2136         struct rb_node *n;
2137
2138         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2139                 n = rb_first(&mdsc->pool_perm_tree);
2140                 perm = rb_entry(n, struct ceph_pool_perm, node);
2141                 rb_erase(n, &mdsc->pool_perm_tree);
2142                 kfree(perm);
2143         }
2144 }