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 #include <linux/ktime.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "metric.h"
  20 #include <linux/ceph/osd_client.h>
  21 #include <linux/ceph/striper.h>
  22
  23 /*
  24  * Ceph address space ops.
  25  *
  26  * There are a few funny things going on here.
  27  *
  28  * The page->private field is used to reference a struct
  29  * ceph_snap_context for _every_ dirty page.  This indicates which
  30  * snapshot the page was logically dirtied in, and thus which snap
  31  * context needs to be associated with the osd write during writeback.
  32  *
  33  * Similarly, struct ceph_inode_info maintains a set of counters to
  34  * count dirty pages on the inode.  In the absence of snapshots,
  35  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  36  *
  37  * When a snapshot is taken (that is, when the client receives
  38  * notification that a snapshot was taken), each inode with caps and
  39  * with dirty pages (dirty pages implies there is a cap) gets a new
  40  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  41  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  42  * moved to capsnap->dirty. (Unless a sync write is currently in
  43  * progress.  In that case, the capsnap is said to be "pending", new
  44  * writes cannot start, and the capsnap isn't "finalized" until the
  45  * write completes (or fails) and a final size/mtime for the inode for
  46  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  47  *
  48  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  49  * we look for the first capsnap in i_cap_snaps and write out pages in
  50  * that snap context _only_.  Then we move on to the next capsnap,
  51  * eventually reaching the "live" or "head" context (i.e., pages that
  52  * are not yet snapped) and are writing the most recently dirtied
  53  * pages.
  54  *
  55  * Invalidate and so forth must take care to ensure the dirty page
  56  * accounting is preserved.
  57  */
  58
  59 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  60 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  61         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  62          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  63
  64 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  65 {
  66         if (PagePrivate(page))
  67                 return (void *)page->private;
  68         return NULL;
  69 }
  70
  71 /*
  72  * Dirty a page.  Optimistically adjust accounting, on the assumption
  73  * that we won't race with invalidate.  If we do, readjust.
  74  */
  75 static int ceph_set_page_dirty(struct page *page)
  76 {
  77         struct address_space *mapping = page->mapping;
  78         struct inode *inode;
  79         struct ceph_inode_info *ci;
  80         struct ceph_snap_context *snapc;
  81         int ret;
  82
  83         if (unlikely(!mapping))
  84                 return !TestSetPageDirty(page);
  85
  86         if (PageDirty(page)) {
  87                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  88                      mapping->host, page, page->index);
  89                 BUG_ON(!PagePrivate(page));
  90                 return 0;
  91         }
  92
  93         inode = mapping->host;
  94         ci = ceph_inode(inode);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  99         if (__ceph_have_pending_cap_snap(ci)) {
 100                 struct ceph_cap_snap *capsnap =
 101                                 list_last_entry(&ci->i_cap_snaps,
 102                                                 struct ceph_cap_snap,
 103                                                 ci_item);
 104                 snapc = ceph_get_snap_context(capsnap->context);
 105                 capsnap->dirty_pages++;
 106         } else {
 107                 BUG_ON(!ci->i_head_snapc);
 108                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 109                 ++ci->i_wrbuffer_ref_head;
 110         }
 111         if (ci->i_wrbuffer_ref == 0)
 112                 ihold(inode);
 113         ++ci->i_wrbuffer_ref;
 114         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 115              "snapc %p seq %lld (%d snaps)\n",
 116              mapping->host, page, page->index,
 117              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 118              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 119              snapc, snapc->seq, snapc->num_snaps);
 120         spin_unlock(&ci->i_ceph_lock);
 121
 122         /*
 123          * Reference snap context in page->private.  Also set
 124          * PagePrivate so that we get invalidatepage callback.
 125          */
 126         BUG_ON(PagePrivate(page));
 127         page->private = (unsigned long)snapc;
 128         SetPagePrivate(page);
 129
 130         ret = __set_page_dirty_nobuffers(page);
 131         WARN_ON(!PageLocked(page));
 132         WARN_ON(!page->mapping);
 133
 134         return ret;
 135 }
 136
 137 /*
 138  * If we are truncating the full page (i.e. offset == 0), adjust the
 139  * dirty page counters appropriately.  Only called if there is private
 140  * data on the page.
 141  */
 142 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 143                                 unsigned int length)
 144 {
 145         struct inode *inode;
 146         struct ceph_inode_info *ci;
 147         struct ceph_snap_context *snapc = page_snap_context(page);
 148
 149         inode = page->mapping->host;
 150         ci = ceph_inode(inode);
 151
 152         if (offset != 0 || length != PAGE_SIZE) {
 153                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 154                      inode, page, page->index, offset, length);
 155                 return;
 156         }
 157
 158         ceph_invalidate_fscache_page(inode, page);
 159
 160         WARN_ON(!PageLocked(page));
 161         if (!PagePrivate(page))
 162                 return;
 163
 164         dout("%p invalidatepage %p idx %lu full dirty page\n",
 165              inode, page, page->index);
 166
 167         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 168         ceph_put_snap_context(snapc);
 169         page->private = 0;
 170         ClearPagePrivate(page);
 171 }
 172
 173 static int ceph_releasepage(struct page *page, gfp_t g)
 174 {
 175         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 176              page, page->index, PageDirty(page) ? "" : "not ");
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * Read some contiguous pages.  If we cross a stripe boundary, shorten
 187  * *plen.  Return number of bytes read, or error.
 188  */
 189 static int ceph_sync_readpages(struct ceph_fs_client *fsc,
 190                                struct ceph_vino vino,
 191                                struct ceph_file_layout *layout,
 192                                u64 off, u64 *plen,
 193                                u32 truncate_seq, u64 truncate_size,
 194                                struct page **pages, int num_pages,
 195                                int page_align)
 196 {
 197         struct ceph_osd_client *osdc = &fsc->client->osdc;
 198         struct ceph_osd_request *req;
 199         int rc = 0;
 200
 201         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
 202              vino.snap, off, *plen);
 203         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
 204                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 205                                     NULL, truncate_seq, truncate_size,
 206                                     false);
 207         if (IS_ERR(req))
 208                 return PTR_ERR(req);
 209
 210         /* it may be a short read due to an object boundary */
 211         osd_req_op_extent_osd_data_pages(req, 0,
 212                                 pages, *plen, page_align, false, false);
 213
 214         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
 215              off, *plen, *plen, page_align);
 216
 217         rc = ceph_osdc_start_request(osdc, req, false);
 218         if (!rc)
 219                 rc = ceph_osdc_wait_request(osdc, req);
 220
 221         ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency,
 222                                  req->r_end_latency, rc);
 223
 224         ceph_osdc_put_request(req);
 225         dout("readpages result %d\n", rc);
 226         return rc;
 227 }
 228
 229 /*
 230  * read a single page, without unlocking it.
 231  */
 232 static int ceph_do_readpage(struct file *filp, struct page *page)
 233 {
 234         struct inode *inode = file_inode(filp);
 235         struct ceph_inode_info *ci = ceph_inode(inode);
 236         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 237         int err = 0;
 238         u64 off = page_offset(page);
 239         u64 len = PAGE_SIZE;
 240
 241         if (off >= i_size_read(inode)) {
 242                 zero_user_segment(page, 0, PAGE_SIZE);
 243                 SetPageUptodate(page);
 244                 return 0;
 245         }
 246
 247         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 248                 /*
 249                  * Uptodate inline data should have been added
 250                  * into page cache while getting Fcr caps.
 251                  */
 252                 if (off == 0)
 253                         return -EINVAL;
 254                 zero_user_segment(page, 0, PAGE_SIZE);
 255                 SetPageUptodate(page);
 256                 return 0;
 257         }
 258
 259         err = ceph_readpage_from_fscache(inode, page);
 260         if (err == 0)
 261                 return -EINPROGRESS;
 262
 263         dout("readpage inode %p file %p page %p index %lu\n",
 264              inode, filp, page, page->index);
 265         err = ceph_sync_readpages(fsc, ceph_vino(inode),
 266                                   &ci->i_layout, off, &len,
 267                                   ci->i_truncate_seq, ci->i_truncate_size,
 268                                   &page, 1, 0);
 269         if (err == -ENOENT)
 270                 err = 0;
 271         if (err < 0) {
 272                 SetPageError(page);
 273                 ceph_fscache_readpage_cancel(inode, page);
 274                 if (err == -EBLACKLISTED)
 275                         fsc->blacklisted = true;
 276                 goto out;
 277         }
 278         if (err < PAGE_SIZE)
 279                 /* zero fill remainder of page */
 280                 zero_user_segment(page, err, PAGE_SIZE);
 281         else
 282                 flush_dcache_page(page);
 283
 284         SetPageUptodate(page);
 285         ceph_readpage_to_fscache(inode, page);
 286
 287 out:
 288         return err < 0 ? err : 0;
 289 }
 290
 291 static int ceph_readpage(struct file *filp, struct page *page)
 292 {
 293         int r = ceph_do_readpage(filp, page);
 294         if (r != -EINPROGRESS)
 295                 unlock_page(page);
 296         else
 297                 r = 0;
 298         return r;
 299 }
 300
 301 /*
 302  * Finish an async read(ahead) op.
 303  */
 304 static void finish_read(struct ceph_osd_request *req)
 305 {
 306         struct inode *inode = req->r_inode;
 307         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 308         struct ceph_osd_data *osd_data;
 309         int rc = req->r_result <= 0 ? req->r_result : 0;
 310         int bytes = req->r_result >= 0 ? req->r_result : 0;
 311         int num_pages;
 312         int i;
 313
 314         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 315         if (rc == -EBLACKLISTED)
 316                 ceph_inode_to_client(inode)->blacklisted = true;
 317
 318         /* unlock all pages, zeroing any data we didn't read */
 319         osd_data = osd_req_op_extent_osd_data(req, 0);
 320         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 321         num_pages = calc_pages_for((u64)osd_data->alignment,
 322                                         (u64)osd_data->length);
 323         for (i = 0; i < num_pages; i++) {
 324                 struct page *page = osd_data->pages[i];
 325
 326                 if (rc < 0 && rc != -ENOENT) {
 327                         ceph_fscache_readpage_cancel(inode, page);
 328                         goto unlock;
 329                 }
 330                 if (bytes < (int)PAGE_SIZE) {
 331                         /* zero (remainder of) page */
 332                         int s = bytes < 0 ? 0 : bytes;
 333                         zero_user_segment(page, s, PAGE_SIZE);
 334                 }
 335                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 336                      page->index);
 337                 flush_dcache_page(page);
 338                 SetPageUptodate(page);
 339                 ceph_readpage_to_fscache(inode, page);
 340 unlock:
 341                 unlock_page(page);
 342                 put_page(page);
 343                 bytes -= PAGE_SIZE;
 344         }
 345
 346         ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency,
 347                                  req->r_end_latency, rc);
 348
 349         kfree(osd_data->pages);
 350 }
 351
 352 /*
 353  * start an async read(ahead) operation.  return nr_pages we submitted
 354  * a read for on success, or negative error code.
 355  */
 356 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
 357                       struct list_head *page_list, int max)
 358 {
 359         struct ceph_osd_client *osdc =
 360                 &ceph_inode_to_client(inode)->client->osdc;
 361         struct ceph_inode_info *ci = ceph_inode(inode);
 362         struct page *page = lru_to_page(page_list);
 363         struct ceph_vino vino;
 364         struct ceph_osd_request *req;
 365         u64 off;
 366         u64 len;
 367         int i;
 368         struct page **pages;
 369         pgoff_t next_index;
 370         int nr_pages = 0;
 371         int got = 0;
 372         int ret = 0;
 373
 374         if (!rw_ctx) {
 375                 /* caller of readpages does not hold buffer and read caps
 376                  * (fadvise, madvise and readahead cases) */
 377                 int want = CEPH_CAP_FILE_CACHE;
 378                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want,
 379                                         true, &got);
 380                 if (ret < 0) {
 381                         dout("start_read %p, error getting cap\n", inode);
 382                 } else if (!(got & want)) {
 383                         dout("start_read %p, no cache cap\n", inode);
 384                         ret = 0;
 385                 }
 386                 if (ret <= 0) {
 387                         if (got)
 388                                 ceph_put_cap_refs(ci, got);
 389                         while (!list_empty(page_list)) {
 390                                 page = lru_to_page(page_list);
 391                                 list_del(&page->lru);
 392                                 put_page(page);
 393                         }
 394                         return ret;
 395                 }
 396         }
 397
 398         off = (u64) page_offset(page);
 399
 400         /* count pages */
 401         next_index = page->index;
 402         list_for_each_entry_reverse(page, page_list, lru) {
 403                 if (page->index != next_index)
 404                         break;
 405                 nr_pages++;
 406                 next_index++;
 407                 if (max && nr_pages == max)
 408                         break;
 409         }
 410         len = nr_pages << PAGE_SHIFT;
 411         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 412              off, len);
 413         vino = ceph_vino(inode);
 414         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 415                                     0, 1, CEPH_OSD_OP_READ,
 416                                     CEPH_OSD_FLAG_READ, NULL,
 417                                     ci->i_truncate_seq, ci->i_truncate_size,
 418                                     false);
 419         if (IS_ERR(req)) {
 420                 ret = PTR_ERR(req);
 421                 goto out;
 422         }
 423
 424         /* build page vector */
 425         nr_pages = calc_pages_for(0, len);
 426         pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
 427         if (!pages) {
 428                 ret = -ENOMEM;
 429                 goto out_put;
 430         }
 431         for (i = 0; i < nr_pages; ++i) {
 432                 page = list_entry(page_list->prev, struct page, lru);
 433                 BUG_ON(PageLocked(page));
 434                 list_del(&page->lru);
 435
 436                 dout("start_read %p adding %p idx %lu\n", inode, page,
 437                      page->index);
 438                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 439                                           GFP_KERNEL)) {
 440                         ceph_fscache_uncache_page(inode, page);
 441                         put_page(page);
 442                         dout("start_read %p add_to_page_cache failed %p\n",
 443                              inode, page);
 444                         nr_pages = i;
 445                         if (nr_pages > 0) {
 446                                 len = nr_pages << PAGE_SHIFT;
 447                                 osd_req_op_extent_update(req, 0, len);
 448                                 break;
 449                         }
 450                         goto out_pages;
 451                 }
 452                 pages[i] = page;
 453         }
 454         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 455         req->r_callback = finish_read;
 456         req->r_inode = inode;
 457
 458         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 459         ret = ceph_osdc_start_request(osdc, req, false);
 460         if (ret < 0)
 461                 goto out_pages;
 462         ceph_osdc_put_request(req);
 463
 464         /* After adding locked pages to page cache, the inode holds cache cap.
 465          * So we can drop our cap refs. */
 466         if (got)
 467                 ceph_put_cap_refs(ci, got);
 468
 469         return nr_pages;
 470
 471 out_pages:
 472         for (i = 0; i < nr_pages; ++i) {
 473                 ceph_fscache_readpage_cancel(inode, pages[i]);
 474                 unlock_page(pages[i]);
 475         }
 476         ceph_put_page_vector(pages, nr_pages, false);
 477 out_put:
 478         ceph_osdc_put_request(req);
 479 out:
 480         if (got)
 481                 ceph_put_cap_refs(ci, got);
 482         return ret;
 483 }
 484
 485
 486 /*
 487  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 488  * the caller (VM) cleans them up.
 489  */
 490 static int ceph_readpages(struct file *file, struct address_space *mapping,
 491                           struct list_head *page_list, unsigned nr_pages)
 492 {
 493         struct inode *inode = file_inode(file);
 494         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 495         struct ceph_file_info *fi = file->private_data;
 496         struct ceph_rw_context *rw_ctx;
 497         int rc = 0;
 498         int max = 0;
 499
 500         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 501                 return -EINVAL;
 502
 503         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 504                                          &nr_pages);
 505
 506         if (rc == 0)
 507                 goto out;
 508
 509         rw_ctx = ceph_find_rw_context(fi);
 510         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 511         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 512              inode, file, rw_ctx, nr_pages, max);
 513         while (!list_empty(page_list)) {
 514                 rc = start_read(inode, rw_ctx, page_list, max);
 515                 if (rc < 0)
 516                         goto out;
 517         }
 518 out:
 519         ceph_fscache_readpages_cancel(inode, page_list);
 520
 521         dout("readpages %p file %p ret %d\n", inode, file, rc);
 522         return rc;
 523 }
 524
 525 struct ceph_writeback_ctl
 526 {
 527         loff_t i_size;
 528         u64 truncate_size;
 529         u32 truncate_seq;
 530         bool size_stable;
 531         bool head_snapc;
 532 };
 533
 534 /*
 535  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 536  * only snap context we are allowed to write back.
 537  */
 538 static struct ceph_snap_context *
 539 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 540                    struct ceph_snap_context *page_snapc)
 541 {
 542         struct ceph_inode_info *ci = ceph_inode(inode);
 543         struct ceph_snap_context *snapc = NULL;
 544         struct ceph_cap_snap *capsnap = NULL;
 545
 546         spin_lock(&ci->i_ceph_lock);
 547         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 548                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 549                      capsnap->context, capsnap->dirty_pages);
 550                 if (!capsnap->dirty_pages)
 551                         continue;
 552
 553                 /* get i_size, truncate_{seq,size} for page_snapc? */
 554                 if (snapc && capsnap->context != page_snapc)
 555                         continue;
 556
 557                 if (ctl) {
 558                         if (capsnap->writing) {
 559                                 ctl->i_size = i_size_read(inode);
 560                                 ctl->size_stable = false;
 561                         } else {
 562                                 ctl->i_size = capsnap->size;
 563                                 ctl->size_stable = true;
 564                         }
 565                         ctl->truncate_size = capsnap->truncate_size;
 566                         ctl->truncate_seq = capsnap->truncate_seq;
 567                         ctl->head_snapc = false;
 568                 }
 569
 570                 if (snapc)
 571                         break;
 572
 573                 snapc = ceph_get_snap_context(capsnap->context);
 574                 if (!page_snapc ||
 575                     page_snapc == snapc ||
 576                     page_snapc->seq > snapc->seq)
 577                         break;
 578         }
 579         if (!snapc && ci->i_wrbuffer_ref_head) {
 580                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 581                 dout(" head snapc %p has %d dirty pages\n",
 582                      snapc, ci->i_wrbuffer_ref_head);
 583                 if (ctl) {
 584                         ctl->i_size = i_size_read(inode);
 585                         ctl->truncate_size = ci->i_truncate_size;
 586                         ctl->truncate_seq = ci->i_truncate_seq;
 587                         ctl->size_stable = false;
 588                         ctl->head_snapc = true;
 589                 }
 590         }
 591         spin_unlock(&ci->i_ceph_lock);
 592         return snapc;
 593 }
 594
 595 static u64 get_writepages_data_length(struct inode *inode,
 596                                       struct page *page, u64 start)
 597 {
 598         struct ceph_inode_info *ci = ceph_inode(inode);
 599         struct ceph_snap_context *snapc = page_snap_context(page);
 600         struct ceph_cap_snap *capsnap = NULL;
 601         u64 end = i_size_read(inode);
 602
 603         if (snapc != ci->i_head_snapc) {
 604                 bool found = false;
 605                 spin_lock(&ci->i_ceph_lock);
 606                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 607                         if (capsnap->context == snapc) {
 608                                 if (!capsnap->writing)
 609                                         end = capsnap->size;
 610                                 found = true;
 611                                 break;
 612                         }
 613                 }
 614                 spin_unlock(&ci->i_ceph_lock);
 615                 WARN_ON(!found);
 616         }
 617         if (end > page_offset(page) + PAGE_SIZE)
 618                 end = page_offset(page) + PAGE_SIZE;
 619         return end > start ? end - start : 0;
 620 }
 621
 622 /*
 623  * do a synchronous write on N pages
 624  */
 625 static int ceph_sync_writepages(struct ceph_fs_client *fsc,
 626                                 struct ceph_vino vino,
 627                                 struct ceph_file_layout *layout,
 628                                 struct ceph_snap_context *snapc,
 629                                 u64 off, u64 len,
 630                                 u32 truncate_seq, u64 truncate_size,
 631                                 struct timespec64 *mtime,
 632                                 struct page **pages, int num_pages)
 633 {
 634         struct ceph_osd_client *osdc = &fsc->client->osdc;
 635         struct ceph_osd_request *req;
 636         int rc = 0;
 637         int page_align = off & ~PAGE_MASK;
 638
 639         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
 640                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
 641                                     snapc, truncate_seq, truncate_size,
 642                                     true);
 643         if (IS_ERR(req))
 644                 return PTR_ERR(req);
 645
 646         /* it may be a short write due to an object boundary */
 647         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
 648                                 false, false);
 649         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
 650
 651         req->r_mtime = *mtime;
 652         rc = ceph_osdc_start_request(osdc, req, true);
 653         if (!rc)
 654                 rc = ceph_osdc_wait_request(osdc, req);
 655
 656         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
 657                                   req->r_end_latency, rc);
 658
 659         ceph_osdc_put_request(req);
 660         if (rc == 0)
 661                 rc = len;
 662         dout("writepages result %d\n", rc);
 663         return rc;
 664 }
 665
 666 /*
 667  * Write a single page, but leave the page locked.
 668  *
 669  * If we get a write error, mark the mapping for error, but still adjust the
 670  * dirty page accounting (i.e., page is no longer dirty).
 671  */
 672 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 673 {
 674         struct inode *inode;
 675         struct ceph_inode_info *ci;
 676         struct ceph_fs_client *fsc;
 677         struct ceph_snap_context *snapc, *oldest;
 678         loff_t page_off = page_offset(page);
 679         int err, len = PAGE_SIZE;
 680         struct ceph_writeback_ctl ceph_wbc;
 681
 682         dout("writepage %p idx %lu\n", page, page->index);
 683
 684         inode = page->mapping->host;
 685         ci = ceph_inode(inode);
 686         fsc = ceph_inode_to_client(inode);
 687
 688         /* verify this is a writeable snap context */
 689         snapc = page_snap_context(page);
 690         if (!snapc) {
 691                 dout("writepage %p page %p not dirty?\n", inode, page);
 692                 return 0;
 693         }
 694         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 695         if (snapc->seq > oldest->seq) {
 696                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 697                      inode, page, snapc);
 698                 /* we should only noop if called by kswapd */
 699                 WARN_ON(!(current->flags & PF_MEMALLOC));
 700                 ceph_put_snap_context(oldest);
 701                 redirty_page_for_writepage(wbc, page);
 702                 return 0;
 703         }
 704         ceph_put_snap_context(oldest);
 705
 706         /* is this a partial page at end of file? */
 707         if (page_off >= ceph_wbc.i_size) {
 708                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 709                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 710                 return 0;
 711         }
 712
 713         if (ceph_wbc.i_size < page_off + len)
 714                 len = ceph_wbc.i_size - page_off;
 715
 716         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 717              inode, page, page->index, page_off, len, snapc, snapc->seq);
 718
 719         if (atomic_long_inc_return(&fsc->writeback_count) >
 720             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 721                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 722
 723         set_page_writeback(page);
 724         err = ceph_sync_writepages(fsc, ceph_vino(inode),
 725                                    &ci->i_layout, snapc, page_off, len,
 726                                    ceph_wbc.truncate_seq,
 727                                    ceph_wbc.truncate_size,
 728                                    &inode->i_mtime, &page, 1);
 729         if (err < 0) {
 730                 struct writeback_control tmp_wbc;
 731                 if (!wbc)
 732                         wbc = &tmp_wbc;
 733                 if (err == -ERESTARTSYS) {
 734                         /* killed by SIGKILL */
 735                         dout("writepage interrupted page %p\n", page);
 736                         redirty_page_for_writepage(wbc, page);
 737                         end_page_writeback(page);
 738                         return err;
 739                 }
 740                 if (err == -EBLACKLISTED)
 741                         fsc->blacklisted = true;
 742                 dout("writepage setting page/mapping error %d %p\n",
 743                      err, page);
 744                 mapping_set_error(&inode->i_data, err);
 745                 wbc->pages_skipped++;
 746         } else {
 747                 dout("writepage cleaned page %p\n", page);
 748                 err = 0;  /* vfs expects us to return 0 */
 749         }
 750         page->private = 0;
 751         ClearPagePrivate(page);
 752         end_page_writeback(page);
 753         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 754         ceph_put_snap_context(snapc);  /* page's reference */
 755
 756         if (atomic_long_dec_return(&fsc->writeback_count) <
 757             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 758                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 759
 760         return err;
 761 }
 762
 763 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 764 {
 765         int err;
 766         struct inode *inode = page->mapping->host;
 767         BUG_ON(!inode);
 768         ihold(inode);
 769         err = writepage_nounlock(page, wbc);
 770         if (err == -ERESTARTSYS) {
 771                 /* direct memory reclaimer was killed by SIGKILL. return 0
 772                  * to prevent caller from setting mapping/page error */
 773                 err = 0;
 774         }
 775         unlock_page(page);
 776         iput(inode);
 777         return err;
 778 }
 779
 780 /*
 781  * async writeback completion handler.
 782  *
 783  * If we get an error, set the mapping error bit, but not the individual
 784  * page error bits.
 785  */
 786 static void writepages_finish(struct ceph_osd_request *req)
 787 {
 788         struct inode *inode = req->r_inode;
 789         struct ceph_inode_info *ci = ceph_inode(inode);
 790         struct ceph_osd_data *osd_data;
 791         struct page *page;
 792         int num_pages, total_pages = 0;
 793         int i, j;
 794         int rc = req->r_result;
 795         struct ceph_snap_context *snapc = req->r_snapc;
 796         struct address_space *mapping = inode->i_mapping;
 797         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 798         bool remove_page;
 799
 800         dout("writepages_finish %p rc %d\n", inode, rc);
 801         if (rc < 0) {
 802                 mapping_set_error(mapping, rc);
 803                 ceph_set_error_write(ci);
 804                 if (rc == -EBLACKLISTED)
 805                         fsc->blacklisted = true;
 806         } else {
 807                 ceph_clear_error_write(ci);
 808         }
 809
 810         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
 811                                   req->r_end_latency, rc);
 812
 813         /*
 814          * We lost the cache cap, need to truncate the page before
 815          * it is unlocked, otherwise we'd truncate it later in the
 816          * page truncation thread, possibly losing some data that
 817          * raced its way in
 818          */
 819         remove_page = !(ceph_caps_issued(ci) &
 820                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 821
 822         /* clean all pages */
 823         for (i = 0; i < req->r_num_ops; i++) {
 824                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 825                         break;
 826
 827                 osd_data = osd_req_op_extent_osd_data(req, i);
 828                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 829                 num_pages = calc_pages_for((u64)osd_data->alignment,
 830                                            (u64)osd_data->length);
 831                 total_pages += num_pages;
 832                 for (j = 0; j < num_pages; j++) {
 833                         page = osd_data->pages[j];
 834                         BUG_ON(!page);
 835                         WARN_ON(!PageUptodate(page));
 836
 837                         if (atomic_long_dec_return(&fsc->writeback_count) <
 838                              CONGESTION_OFF_THRESH(
 839                                         fsc->mount_options->congestion_kb))
 840                                 clear_bdi_congested(inode_to_bdi(inode),
 841                                                     BLK_RW_ASYNC);
 842
 843                         ceph_put_snap_context(page_snap_context(page));
 844                         page->private = 0;
 845                         ClearPagePrivate(page);
 846                         dout("unlocking %p\n", page);
 847                         end_page_writeback(page);
 848
 849                         if (remove_page)
 850                                 generic_error_remove_page(inode->i_mapping,
 851                                                           page);
 852
 853                         unlock_page(page);
 854                 }
 855                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 856                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 857
 858                 release_pages(osd_data->pages, num_pages);
 859         }
 860
 861         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 862
 863         osd_data = osd_req_op_extent_osd_data(req, 0);
 864         if (osd_data->pages_from_pool)
 865                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 866         else
 867                 kfree(osd_data->pages);
 868         ceph_osdc_put_request(req);
 869 }
 870
 871 /*
 872  * initiate async writeback
 873  */
 874 static int ceph_writepages_start(struct address_space *mapping,
 875                                  struct writeback_control *wbc)
 876 {
 877         struct inode *inode = mapping->host;
 878         struct ceph_inode_info *ci = ceph_inode(inode);
 879         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 880         struct ceph_vino vino = ceph_vino(inode);
 881         pgoff_t index, start_index, end = -1;
 882         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 883         struct pagevec pvec;
 884         int rc = 0;
 885         unsigned int wsize = i_blocksize(inode);
 886         struct ceph_osd_request *req = NULL;
 887         struct ceph_writeback_ctl ceph_wbc;
 888         bool should_loop, range_whole = false;
 889         bool done = false;
 890
 891         dout("writepages_start %p (mode=%s)\n", inode,
 892              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 893              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 894
 895         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 896                 if (ci->i_wrbuffer_ref > 0) {
 897                         pr_warn_ratelimited(
 898                                 "writepage_start %p %lld forced umount\n",
 899                                 inode, ceph_ino(inode));
 900                 }
 901                 mapping_set_error(mapping, -EIO);
 902                 return -EIO; /* we're in a forced umount, don't write! */
 903         }
 904         if (fsc->mount_options->wsize < wsize)
 905                 wsize = fsc->mount_options->wsize;
 906
 907         pagevec_init(&pvec);
 908
 909         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 910         index = start_index;
 911
 912 retry:
 913         /* find oldest snap context with dirty data */
 914         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 915         if (!snapc) {
 916                 /* hmm, why does writepages get called when there
 917                    is no dirty data? */
 918                 dout(" no snap context with dirty data?\n");
 919                 goto out;
 920         }
 921         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 922              snapc, snapc->seq, snapc->num_snaps);
 923
 924         should_loop = false;
 925         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 926                 /* where to start/end? */
 927                 if (wbc->range_cyclic) {
 928                         index = start_index;
 929                         end = -1;
 930                         if (index > 0)
 931                                 should_loop = true;
 932                         dout(" cyclic, start at %lu\n", index);
 933                 } else {
 934                         index = wbc->range_start >> PAGE_SHIFT;
 935                         end = wbc->range_end >> PAGE_SHIFT;
 936                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 937                                 range_whole = true;
 938                         dout(" not cyclic, %lu to %lu\n", index, end);
 939                 }
 940         } else if (!ceph_wbc.head_snapc) {
 941                 /* Do not respect wbc->range_{start,end}. Dirty pages
 942                  * in that range can be associated with newer snapc.
 943                  * They are not writeable until we write all dirty pages
 944                  * associated with 'snapc' get written */
 945                 if (index > 0)
 946                         should_loop = true;
 947                 dout(" non-head snapc, range whole\n");
 948         }
 949
 950         ceph_put_snap_context(last_snapc);
 951         last_snapc = snapc;
 952
 953         while (!done && index <= end) {
 954                 int num_ops = 0, op_idx;
 955                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 956                 struct page **pages = NULL, **data_pages;
 957                 struct page *page;
 958                 pgoff_t strip_unit_end = 0;
 959                 u64 offset = 0, len = 0;
 960                 bool from_pool = false;
 961
 962                 max_pages = wsize >> PAGE_SHIFT;
 963
 964 get_more_pages:
 965                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 966                                                 end, PAGECACHE_TAG_DIRTY,
 967                                                 max_pages - locked_pages);
 968                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 969                 if (!pvec_pages && !locked_pages)
 970                         break;
 971                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 972                         page = pvec.pages[i];
 973                         dout("? %p idx %lu\n", page, page->index);
 974                         if (locked_pages == 0)
 975                                 lock_page(page);  /* first page */
 976                         else if (!trylock_page(page))
 977                                 break;
 978
 979                         /* only dirty pages, or our accounting breaks */
 980                         if (unlikely(!PageDirty(page)) ||
 981                             unlikely(page->mapping != mapping)) {
 982                                 dout("!dirty or !mapping %p\n", page);
 983                                 unlock_page(page);
 984                                 continue;
 985                         }
 986                         /* only if matching snap context */
 987                         pgsnapc = page_snap_context(page);
 988                         if (pgsnapc != snapc) {
 989                                 dout("page snapc %p %lld != oldest %p %lld\n",
 990                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 991                                 if (!should_loop &&
 992                                     !ceph_wbc.head_snapc &&
 993                                     wbc->sync_mode != WB_SYNC_NONE)
 994                                         should_loop = true;
 995                                 unlock_page(page);
 996                                 continue;
 997                         }
 998                         if (page_offset(page) >= ceph_wbc.i_size) {
 999                                 dout("%p page eof %llu\n",
1000                                      page, ceph_wbc.i_size);
1001                                 if ((ceph_wbc.size_stable ||
1002                                     page_offset(page) >= i_size_read(inode)) &&
1003                                     clear_page_dirty_for_io(page))
1004                                         mapping->a_ops->invalidatepage(page,
1005                                                                 0, PAGE_SIZE);
1006                                 unlock_page(page);
1007                                 continue;
1008                         }
1009                         if (strip_unit_end && (page->index > strip_unit_end)) {
1010                                 dout("end of strip unit %p\n", page);
1011                                 unlock_page(page);
1012                                 break;
1013                         }
1014                         if (PageWriteback(page)) {
1015                                 if (wbc->sync_mode == WB_SYNC_NONE) {
1016                                         dout("%p under writeback\n", page);
1017                                         unlock_page(page);
1018                                         continue;
1019                                 }
1020                                 dout("waiting on writeback %p\n", page);
1021                                 wait_on_page_writeback(page);
1022                         }
1023
1024                         if (!clear_page_dirty_for_io(page)) {
1025                                 dout("%p !clear_page_dirty_for_io\n", page);
1026                                 unlock_page(page);
1027                                 continue;
1028                         }
1029
1030                         /*
1031                          * We have something to write.  If this is
1032                          * the first locked page this time through,
1033                          * calculate max possinle write size and
1034                          * allocate a page array
1035                          */
1036                         if (locked_pages == 0) {
1037                                 u64 objnum;
1038                                 u64 objoff;
1039                                 u32 xlen;
1040
1041                                 /* prepare async write request */
1042                                 offset = (u64)page_offset(page);
1043                                 ceph_calc_file_object_mapping(&ci->i_layout,
1044                                                               offset, wsize,
1045                                                               &objnum, &objoff,
1046                                                               &xlen);
1047                                 len = xlen;
1048
1049                                 num_ops = 1;
1050                                 strip_unit_end = page->index +
1051                                         ((len - 1) >> PAGE_SHIFT);
1052
1053                                 BUG_ON(pages);
1054                                 max_pages = calc_pages_for(0, (u64)len);
1055                                 pages = kmalloc_array(max_pages,
1056                                                       sizeof(*pages),
1057                                                       GFP_NOFS);
1058                                 if (!pages) {
1059                                         from_pool = true;
1060                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1061                                         BUG_ON(!pages);
1062                                 }
1063
1064                                 len = 0;
1065                         } else if (page->index !=
1066                                    (offset + len) >> PAGE_SHIFT) {
1067                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1068                                                              CEPH_OSD_MAX_OPS)) {
1069                                         redirty_page_for_writepage(wbc, page);
1070                                         unlock_page(page);
1071                                         break;
1072                                 }
1073
1074                                 num_ops++;
1075                                 offset = (u64)page_offset(page);
1076                                 len = 0;
1077                         }
1078
1079                         /* note position of first page in pvec */
1080                         dout("%p will write page %p idx %lu\n",
1081                              inode, page, page->index);
1082
1083                         if (atomic_long_inc_return(&fsc->writeback_count) >
1084                             CONGESTION_ON_THRESH(
1085                                     fsc->mount_options->congestion_kb)) {
1086                                 set_bdi_congested(inode_to_bdi(inode),
1087                                                   BLK_RW_ASYNC);
1088                         }
1089
1090
1091                         pages[locked_pages++] = page;
1092                         pvec.pages[i] = NULL;
1093
1094                         len += PAGE_SIZE;
1095                 }
1096
1097                 /* did we get anything? */
1098                 if (!locked_pages)
1099                         goto release_pvec_pages;
1100                 if (i) {
1101                         unsigned j, n = 0;
1102                         /* shift unused page to beginning of pvec */
1103                         for (j = 0; j < pvec_pages; j++) {
1104                                 if (!pvec.pages[j])
1105                                         continue;
1106                                 if (n < j)
1107                                         pvec.pages[n] = pvec.pages[j];
1108                                 n++;
1109                         }
1110                         pvec.nr = n;
1111
1112                         if (pvec_pages && i == pvec_pages &&
1113                             locked_pages < max_pages) {
1114                                 dout("reached end pvec, trying for more\n");
1115                                 pagevec_release(&pvec);
1116                                 goto get_more_pages;
1117                         }
1118                 }
1119
1120 new_request:
1121                 offset = page_offset(pages[0]);
1122                 len = wsize;
1123
1124                 req = ceph_osdc_new_request(&fsc->client->osdc,
1125                                         &ci->i_layout, vino,
1126                                         offset, &len, 0, num_ops,
1127                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1128                                         snapc, ceph_wbc.truncate_seq,
1129                                         ceph_wbc.truncate_size, false);
1130                 if (IS_ERR(req)) {
1131                         req = ceph_osdc_new_request(&fsc->client->osdc,
1132                                                 &ci->i_layout, vino,
1133                                                 offset, &len, 0,
1134                                                 min(num_ops,
1135                                                     CEPH_OSD_SLAB_OPS),
1136                                                 CEPH_OSD_OP_WRITE,
1137                                                 CEPH_OSD_FLAG_WRITE,
1138                                                 snapc, ceph_wbc.truncate_seq,
1139                                                 ceph_wbc.truncate_size, true);
1140                         BUG_ON(IS_ERR(req));
1141                 }
1142                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1143                              PAGE_SIZE - offset);
1144
1145                 req->r_callback = writepages_finish;
1146                 req->r_inode = inode;
1147
1148                 /* Format the osd request message and submit the write */
1149                 len = 0;
1150                 data_pages = pages;
1151                 op_idx = 0;
1152                 for (i = 0; i < locked_pages; i++) {
1153                         u64 cur_offset = page_offset(pages[i]);
1154                         if (offset + len != cur_offset) {
1155                                 if (op_idx + 1 == req->r_num_ops)
1156                                         break;
1157                                 osd_req_op_extent_dup_last(req, op_idx,
1158                                                            cur_offset - offset);
1159                                 dout("writepages got pages at %llu~%llu\n",
1160                                      offset, len);
1161                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1162                                                         data_pages, len, 0,
1163                                                         from_pool, false);
1164                                 osd_req_op_extent_update(req, op_idx, len);
1165
1166                                 len = 0;
1167                                 offset = cur_offset;
1168                                 data_pages = pages + i;
1169                                 op_idx++;
1170                         }
1171
1172                         set_page_writeback(pages[i]);
1173                         len += PAGE_SIZE;
1174                 }
1175
1176                 if (ceph_wbc.size_stable) {
1177                         len = min(len, ceph_wbc.i_size - offset);
1178                 } else if (i == locked_pages) {
1179                         /* writepages_finish() clears writeback pages
1180                          * according to the data length, so make sure
1181                          * data length covers all locked pages */
1182                         u64 min_len = len + 1 - PAGE_SIZE;
1183                         len = get_writepages_data_length(inode, pages[i - 1],
1184                                                          offset);
1185                         len = max(len, min_len);
1186                 }
1187                 dout("writepages got pages at %llu~%llu\n", offset, len);
1188
1189                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1190                                                  0, from_pool, false);
1191                 osd_req_op_extent_update(req, op_idx, len);
1192
1193                 BUG_ON(op_idx + 1 != req->r_num_ops);
1194
1195                 from_pool = false;
1196                 if (i < locked_pages) {
1197                         BUG_ON(num_ops <= req->r_num_ops);
1198                         num_ops -= req->r_num_ops;
1199                         locked_pages -= i;
1200
1201                         /* allocate new pages array for next request */
1202                         data_pages = pages;
1203                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1204                                               GFP_NOFS);
1205                         if (!pages) {
1206                                 from_pool = true;
1207                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1208                                 BUG_ON(!pages);
1209                         }
1210                         memcpy(pages, data_pages + i,
1211                                locked_pages * sizeof(*pages));
1212                         memset(data_pages + i, 0,
1213                                locked_pages * sizeof(*pages));
1214                 } else {
1215                         BUG_ON(num_ops != req->r_num_ops);
1216                         index = pages[i - 1]->index + 1;
1217                         /* request message now owns the pages array */
1218                         pages = NULL;
1219                 }
1220
1221                 req->r_mtime = inode->i_mtime;
1222                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1223                 BUG_ON(rc);
1224                 req = NULL;
1225
1226                 wbc->nr_to_write -= i;
1227                 if (pages)
1228                         goto new_request;
1229
1230                 /*
1231                  * We stop writing back only if we are not doing
1232                  * integrity sync. In case of integrity sync we have to
1233                  * keep going until we have written all the pages
1234                  * we tagged for writeback prior to entering this loop.
1235                  */
1236                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1237                         done = true;
1238
1239 release_pvec_pages:
1240                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1241                      pvec.nr ? pvec.pages[0] : NULL);
1242                 pagevec_release(&pvec);
1243         }
1244
1245         if (should_loop && !done) {
1246                 /* more to do; loop back to beginning of file */
1247                 dout("writepages looping back to beginning of file\n");
1248                 end = start_index - 1; /* OK even when start_index == 0 */
1249
1250                 /* to write dirty pages associated with next snapc,
1251                  * we need to wait until current writes complete */
1252                 if (wbc->sync_mode != WB_SYNC_NONE &&
1253                     start_index == 0 && /* all dirty pages were checked */
1254                     !ceph_wbc.head_snapc) {
1255                         struct page *page;
1256                         unsigned i, nr;
1257                         index = 0;
1258                         while ((index <= end) &&
1259                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1260                                                 PAGECACHE_TAG_WRITEBACK))) {
1261                                 for (i = 0; i < nr; i++) {
1262                                         page = pvec.pages[i];
1263                                         if (page_snap_context(page) != snapc)
1264                                                 continue;
1265                                         wait_on_page_writeback(page);
1266                                 }
1267                                 pagevec_release(&pvec);
1268                                 cond_resched();
1269                         }
1270                 }
1271
1272                 start_index = 0;
1273                 index = 0;
1274                 goto retry;
1275         }
1276
1277         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1278                 mapping->writeback_index = index;
1279
1280 out:
1281         ceph_osdc_put_request(req);
1282         ceph_put_snap_context(last_snapc);
1283         dout("writepages dend - startone, rc = %d\n", rc);
1284         return rc;
1285 }
1286
1287
1288
1289 /*
1290  * See if a given @snapc is either writeable, or already written.
1291  */
1292 static int context_is_writeable_or_written(struct inode *inode,
1293                                            struct ceph_snap_context *snapc)
1294 {
1295         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1296         int ret = !oldest || snapc->seq <= oldest->seq;
1297
1298         ceph_put_snap_context(oldest);
1299         return ret;
1300 }
1301
1302 /*
1303  * We are only allowed to write into/dirty the page if the page is
1304  * clean, or already dirty within the same snap context.
1305  *
1306  * called with page locked.
1307  * return success with page locked,
1308  * or any failure (incl -EAGAIN) with page unlocked.
1309  */
1310 static int ceph_update_writeable_page(struct file *file,
1311                             loff_t pos, unsigned len,
1312                             struct page *page)
1313 {
1314         struct inode *inode = file_inode(file);
1315         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1316         struct ceph_inode_info *ci = ceph_inode(inode);
1317         loff_t page_off = pos & PAGE_MASK;
1318         int pos_in_page = pos & ~PAGE_MASK;
1319         int end_in_page = pos_in_page + len;
1320         loff_t i_size;
1321         int r;
1322         struct ceph_snap_context *snapc, *oldest;
1323
1324         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1325                 dout(" page %p forced umount\n", page);
1326                 unlock_page(page);
1327                 return -EIO;
1328         }
1329
1330 retry_locked:
1331         /* writepages currently holds page lock, but if we change that later, */
1332         wait_on_page_writeback(page);
1333
1334         snapc = page_snap_context(page);
1335         if (snapc && snapc != ci->i_head_snapc) {
1336                 /*
1337                  * this page is already dirty in another (older) snap
1338                  * context!  is it writeable now?
1339                  */
1340                 oldest = get_oldest_context(inode, NULL, NULL);
1341                 if (snapc->seq > oldest->seq) {
1342                         ceph_put_snap_context(oldest);
1343                         dout(" page %p snapc %p not current or oldest\n",
1344                              page, snapc);
1345                         /*
1346                          * queue for writeback, and wait for snapc to
1347                          * be writeable or written
1348                          */
1349                         snapc = ceph_get_snap_context(snapc);
1350                         unlock_page(page);
1351                         ceph_queue_writeback(inode);
1352                         r = wait_event_killable(ci->i_cap_wq,
1353                                context_is_writeable_or_written(inode, snapc));
1354                         ceph_put_snap_context(snapc);
1355                         if (r == -ERESTARTSYS)
1356                                 return r;
1357                         return -EAGAIN;
1358                 }
1359                 ceph_put_snap_context(oldest);
1360
1361                 /* yay, writeable, do it now (without dropping page lock) */
1362                 dout(" page %p snapc %p not current, but oldest\n",
1363                      page, snapc);
1364                 if (!clear_page_dirty_for_io(page))
1365                         goto retry_locked;
1366                 r = writepage_nounlock(page, NULL);
1367                 if (r < 0)
1368                         goto fail_unlock;
1369                 goto retry_locked;
1370         }
1371
1372         if (PageUptodate(page)) {
1373                 dout(" page %p already uptodate\n", page);
1374                 return 0;
1375         }
1376
1377         /* full page? */
1378         if (pos_in_page == 0 && len == PAGE_SIZE)
1379                 return 0;
1380
1381         /* past end of file? */
1382         i_size = i_size_read(inode);
1383
1384         if (page_off >= i_size ||
1385             (pos_in_page == 0 && (pos+len) >= i_size &&
1386              end_in_page - pos_in_page != PAGE_SIZE)) {
1387                 dout(" zeroing %p 0 - %d and %d - %d\n",
1388                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1389                 zero_user_segments(page,
1390                                    0, pos_in_page,
1391                                    end_in_page, PAGE_SIZE);
1392                 return 0;
1393         }
1394
1395         /* we need to read it. */
1396         r = ceph_do_readpage(file, page);
1397         if (r < 0) {
1398                 if (r == -EINPROGRESS)
1399                         return -EAGAIN;
1400                 goto fail_unlock;
1401         }
1402         goto retry_locked;
1403 fail_unlock:
1404         unlock_page(page);
1405         return r;
1406 }
1407
1408 /*
1409  * We are only allowed to write into/dirty the page if the page is
1410  * clean, or already dirty within the same snap context.
1411  */
1412 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1413                             loff_t pos, unsigned len, unsigned flags,
1414                             struct page **pagep, void **fsdata)
1415 {
1416         struct inode *inode = file_inode(file);
1417         struct page *page;
1418         pgoff_t index = pos >> PAGE_SHIFT;
1419         int r;
1420
1421         do {
1422                 /* get a page */
1423                 page = grab_cache_page_write_begin(mapping, index, 0);
1424                 if (!page)
1425                         return -ENOMEM;
1426
1427                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1428                      inode, page, (int)pos, (int)len);
1429
1430                 r = ceph_update_writeable_page(file, pos, len, page);
1431                 if (r < 0)
1432                         put_page(page);
1433                 else
1434                         *pagep = page;
1435         } while (r == -EAGAIN);
1436
1437         return r;
1438 }
1439
1440 /*
1441  * we don't do anything in here that simple_write_end doesn't do
1442  * except adjust dirty page accounting
1443  */
1444 static int ceph_write_end(struct file *file, struct address_space *mapping,
1445                           loff_t pos, unsigned len, unsigned copied,
1446                           struct page *page, void *fsdata)
1447 {
1448         struct inode *inode = file_inode(file);
1449         bool check_cap = false;
1450
1451         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1452              inode, page, (int)pos, (int)copied, (int)len);
1453
1454         /* zero the stale part of the page if we did a short copy */
1455         if (!PageUptodate(page)) {
1456                 if (copied < len) {
1457                         copied = 0;
1458                         goto out;
1459                 }
1460                 SetPageUptodate(page);
1461         }
1462
1463         /* did file size increase? */
1464         if (pos+copied > i_size_read(inode))
1465                 check_cap = ceph_inode_set_size(inode, pos+copied);
1466
1467         set_page_dirty(page);
1468
1469 out:
1470         unlock_page(page);
1471         put_page(page);
1472
1473         if (check_cap)
1474                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1475
1476         return copied;
1477 }
1478
1479 /*
1480  * we set .direct_IO to indicate direct io is supported, but since we
1481  * intercept O_DIRECT reads and writes early, this function should
1482  * never get called.
1483  */
1484 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1485 {
1486         WARN_ON(1);
1487         return -EINVAL;
1488 }
1489
1490 const struct address_space_operations ceph_aops = {
1491         .readpage = ceph_readpage,
1492         .readpages = ceph_readpages,
1493         .writepage = ceph_writepage,
1494         .writepages = ceph_writepages_start,
1495         .write_begin = ceph_write_begin,
1496         .write_end = ceph_write_end,
1497         .set_page_dirty = ceph_set_page_dirty,
1498         .invalidatepage = ceph_invalidatepage,
1499         .releasepage = ceph_releasepage,
1500         .direct_IO = ceph_direct_io,
1501 };
1502
1503 static void ceph_block_sigs(sigset_t *oldset)
1504 {
1505         sigset_t mask;
1506         siginitsetinv(&mask, sigmask(SIGKILL));
1507         sigprocmask(SIG_BLOCK, &mask, oldset);
1508 }
1509
1510 static void ceph_restore_sigs(sigset_t *oldset)
1511 {
1512         sigprocmask(SIG_SETMASK, oldset, NULL);
1513 }
1514
1515 /*
1516  * vm ops
1517  */
1518 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1519 {
1520         struct vm_area_struct *vma = vmf->vma;
1521         struct inode *inode = file_inode(vma->vm_file);
1522         struct ceph_inode_info *ci = ceph_inode(inode);
1523         struct ceph_file_info *fi = vma->vm_file->private_data;
1524         struct page *pinned_page = NULL;
1525         loff_t off = vmf->pgoff << PAGE_SHIFT;
1526         int want, got, err;
1527         sigset_t oldset;
1528         vm_fault_t ret = VM_FAULT_SIGBUS;
1529
1530         ceph_block_sigs(&oldset);
1531
1532         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1533              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1534         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1535                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1536         else
1537                 want = CEPH_CAP_FILE_CACHE;
1538
1539         got = 0;
1540         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1,
1541                             &got, &pinned_page);
1542         if (err < 0)
1543                 goto out_restore;
1544
1545         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1546              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1547
1548         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1549             ci->i_inline_version == CEPH_INLINE_NONE) {
1550                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1551                 ceph_add_rw_context(fi, &rw_ctx);
1552                 ret = filemap_fault(vmf);
1553                 ceph_del_rw_context(fi, &rw_ctx);
1554                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1555                         inode, off, (size_t)PAGE_SIZE,
1556                                 ceph_cap_string(got), ret);
1557         } else
1558                 err = -EAGAIN;
1559
1560         if (pinned_page)
1561                 put_page(pinned_page);
1562         ceph_put_cap_refs(ci, got);
1563
1564         if (err != -EAGAIN)
1565                 goto out_restore;
1566
1567         /* read inline data */
1568         if (off >= PAGE_SIZE) {
1569                 /* does not support inline data > PAGE_SIZE */
1570                 ret = VM_FAULT_SIGBUS;
1571         } else {
1572                 struct address_space *mapping = inode->i_mapping;
1573                 struct page *page = find_or_create_page(mapping, 0,
1574                                                 mapping_gfp_constraint(mapping,
1575                                                 ~__GFP_FS));
1576                 if (!page) {
1577                         ret = VM_FAULT_OOM;
1578                         goto out_inline;
1579                 }
1580                 err = __ceph_do_getattr(inode, page,
1581                                          CEPH_STAT_CAP_INLINE_DATA, true);
1582                 if (err < 0 || off >= i_size_read(inode)) {
1583                         unlock_page(page);
1584                         put_page(page);
1585                         ret = vmf_error(err);
1586                         goto out_inline;
1587                 }
1588                 if (err < PAGE_SIZE)
1589                         zero_user_segment(page, err, PAGE_SIZE);
1590                 else
1591                         flush_dcache_page(page);
1592                 SetPageUptodate(page);
1593                 vmf->page = page;
1594                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1595 out_inline:
1596                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1597                      inode, off, (size_t)PAGE_SIZE, ret);
1598         }
1599 out_restore:
1600         ceph_restore_sigs(&oldset);
1601         if (err < 0)
1602                 ret = vmf_error(err);
1603
1604         return ret;
1605 }
1606
1607 /*
1608  * Reuse write_begin here for simplicity.
1609  */
1610 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1611 {
1612         struct vm_area_struct *vma = vmf->vma;
1613         struct inode *inode = file_inode(vma->vm_file);
1614         struct ceph_inode_info *ci = ceph_inode(inode);
1615         struct ceph_file_info *fi = vma->vm_file->private_data;
1616         struct ceph_cap_flush *prealloc_cf;
1617         struct page *page = vmf->page;
1618         loff_t off = page_offset(page);
1619         loff_t size = i_size_read(inode);
1620         size_t len;
1621         int want, got, err;
1622         sigset_t oldset;
1623         vm_fault_t ret = VM_FAULT_SIGBUS;
1624
1625         prealloc_cf = ceph_alloc_cap_flush();
1626         if (!prealloc_cf)
1627                 return VM_FAULT_OOM;
1628
1629         sb_start_pagefault(inode->i_sb);
1630         ceph_block_sigs(&oldset);
1631
1632         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1633                 struct page *locked_page = NULL;
1634                 if (off == 0) {
1635                         lock_page(page);
1636                         locked_page = page;
1637                 }
1638                 err = ceph_uninline_data(vma->vm_file, locked_page);
1639                 if (locked_page)
1640                         unlock_page(locked_page);
1641                 if (err < 0)
1642                         goto out_free;
1643         }
1644
1645         if (off + PAGE_SIZE <= size)
1646                 len = PAGE_SIZE;
1647         else
1648                 len = size & ~PAGE_MASK;
1649
1650         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1651              inode, ceph_vinop(inode), off, len, size);
1652         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1653                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1654         else
1655                 want = CEPH_CAP_FILE_BUFFER;
1656
1657         got = 0;
1658         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len,
1659                             &got, NULL);
1660         if (err < 0)
1661                 goto out_free;
1662
1663         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1664              inode, off, len, ceph_cap_string(got));
1665
1666         /* Update time before taking page lock */
1667         file_update_time(vma->vm_file);
1668         inode_inc_iversion_raw(inode);
1669
1670         do {
1671                 lock_page(page);
1672
1673                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1674                         unlock_page(page);
1675                         ret = VM_FAULT_NOPAGE;
1676                         break;
1677                 }
1678
1679                 err = ceph_update_writeable_page(vma->vm_file, off, len, page);
1680                 if (err >= 0) {
1681                         /* success.  we'll keep the page locked. */
1682                         set_page_dirty(page);
1683                         ret = VM_FAULT_LOCKED;
1684                 }
1685         } while (err == -EAGAIN);
1686
1687         if (ret == VM_FAULT_LOCKED ||
1688             ci->i_inline_version != CEPH_INLINE_NONE) {
1689                 int dirty;
1690                 spin_lock(&ci->i_ceph_lock);
1691                 ci->i_inline_version = CEPH_INLINE_NONE;
1692                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1693                                                &prealloc_cf);
1694                 spin_unlock(&ci->i_ceph_lock);
1695                 if (dirty)
1696                         __mark_inode_dirty(inode, dirty);
1697         }
1698
1699         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1700              inode, off, len, ceph_cap_string(got), ret);
1701         ceph_put_cap_refs(ci, got);
1702 out_free:
1703         ceph_restore_sigs(&oldset);
1704         sb_end_pagefault(inode->i_sb);
1705         ceph_free_cap_flush(prealloc_cf);
1706         if (err < 0)
1707                 ret = vmf_error(err);
1708         return ret;
1709 }
1710
1711 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1712                            char *data, size_t len)
1713 {
1714         struct address_space *mapping = inode->i_mapping;
1715         struct page *page;
1716
1717         if (locked_page) {
1718                 page = locked_page;
1719         } else {
1720                 if (i_size_read(inode) == 0)
1721                         return;
1722                 page = find_or_create_page(mapping, 0,
1723                                            mapping_gfp_constraint(mapping,
1724                                            ~__GFP_FS));
1725                 if (!page)
1726                         return;
1727                 if (PageUptodate(page)) {
1728                         unlock_page(page);
1729                         put_page(page);
1730                         return;
1731                 }
1732         }
1733
1734         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1735              inode, ceph_vinop(inode), len, locked_page);
1736
1737         if (len > 0) {
1738                 void *kaddr = kmap_atomic(page);
1739                 memcpy(kaddr, data, len);
1740                 kunmap_atomic(kaddr);
1741         }
1742
1743         if (page != locked_page) {
1744                 if (len < PAGE_SIZE)
1745                         zero_user_segment(page, len, PAGE_SIZE);
1746                 else
1747                         flush_dcache_page(page);
1748
1749                 SetPageUptodate(page);
1750                 unlock_page(page);
1751                 put_page(page);
1752         }
1753 }
1754
1755 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1756 {
1757         struct inode *inode = file_inode(filp);
1758         struct ceph_inode_info *ci = ceph_inode(inode);
1759         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1760         struct ceph_osd_request *req;
1761         struct page *page = NULL;
1762         u64 len, inline_version;
1763         int err = 0;
1764         bool from_pagecache = false;
1765
1766         spin_lock(&ci->i_ceph_lock);
1767         inline_version = ci->i_inline_version;
1768         spin_unlock(&ci->i_ceph_lock);
1769
1770         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1771              inode, ceph_vinop(inode), inline_version);
1772
1773         if (inline_version == 1 || /* initial version, no data */
1774             inline_version == CEPH_INLINE_NONE)
1775                 goto out;
1776
1777         if (locked_page) {
1778                 page = locked_page;
1779                 WARN_ON(!PageUptodate(page));
1780         } else if (ceph_caps_issued(ci) &
1781                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1782                 page = find_get_page(inode->i_mapping, 0);
1783                 if (page) {
1784                         if (PageUptodate(page)) {
1785                                 from_pagecache = true;
1786                                 lock_page(page);
1787                         } else {
1788                                 put_page(page);
1789                                 page = NULL;
1790                         }
1791                 }
1792         }
1793
1794         if (page) {
1795                 len = i_size_read(inode);
1796                 if (len > PAGE_SIZE)
1797                         len = PAGE_SIZE;
1798         } else {
1799                 page = __page_cache_alloc(GFP_NOFS);
1800                 if (!page) {
1801                         err = -ENOMEM;
1802                         goto out;
1803                 }
1804                 err = __ceph_do_getattr(inode, page,
1805                                         CEPH_STAT_CAP_INLINE_DATA, true);
1806                 if (err < 0) {
1807                         /* no inline data */
1808                         if (err == -ENODATA)
1809                                 err = 0;
1810                         goto out;
1811                 }
1812                 len = err;
1813         }
1814
1815         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1816                                     ceph_vino(inode), 0, &len, 0, 1,
1817                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1818                                     NULL, 0, 0, false);
1819         if (IS_ERR(req)) {
1820                 err = PTR_ERR(req);
1821                 goto out;
1822         }
1823
1824         req->r_mtime = inode->i_mtime;
1825         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1826         if (!err)
1827                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1828         ceph_osdc_put_request(req);
1829         if (err < 0)
1830                 goto out;
1831
1832         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1833                                     ceph_vino(inode), 0, &len, 1, 3,
1834                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1835                                     NULL, ci->i_truncate_seq,
1836                                     ci->i_truncate_size, false);
1837         if (IS_ERR(req)) {
1838                 err = PTR_ERR(req);
1839                 goto out;
1840         }
1841
1842         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1843
1844         {
1845                 __le64 xattr_buf = cpu_to_le64(inline_version);
1846                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1847                                             "inline_version", &xattr_buf,
1848                                             sizeof(xattr_buf),
1849                                             CEPH_OSD_CMPXATTR_OP_GT,
1850                                             CEPH_OSD_CMPXATTR_MODE_U64);
1851                 if (err)
1852                         goto out_put;
1853         }
1854
1855         {
1856                 char xattr_buf[32];
1857                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1858                                          "%llu", inline_version);
1859                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1860                                             "inline_version",
1861                                             xattr_buf, xattr_len, 0, 0);
1862                 if (err)
1863                         goto out_put;
1864         }
1865
1866         req->r_mtime = inode->i_mtime;
1867         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1868         if (!err)
1869                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1870
1871         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
1872                                   req->r_end_latency, err);
1873
1874 out_put:
1875         ceph_osdc_put_request(req);
1876         if (err == -ECANCELED)
1877                 err = 0;
1878 out:
1879         if (page && page != locked_page) {
1880                 if (from_pagecache) {
1881                         unlock_page(page);
1882                         put_page(page);
1883                 } else
1884                         __free_pages(page, 0);
1885         }
1886
1887         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1888              inode, ceph_vinop(inode), inline_version, err);
1889         return err;
1890 }
1891
1892 static const struct vm_operations_struct ceph_vmops = {
1893         .fault          = ceph_filemap_fault,
1894         .page_mkwrite   = ceph_page_mkwrite,
1895 };
1896
1897 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1898 {
1899         struct address_space *mapping = file->f_mapping;
1900
1901         if (!mapping->a_ops->readpage)
1902                 return -ENOEXEC;
1903         file_accessed(file);
1904         vma->vm_ops = &ceph_vmops;
1905         return 0;
1906 }
1907
1908 enum {
1909         POOL_READ       = 1,
1910         POOL_WRITE      = 2,
1911 };
1912
1913 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1914                                 s64 pool, struct ceph_string *pool_ns)
1915 {
1916         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1917         struct ceph_mds_client *mdsc = fsc->mdsc;
1918         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1919         struct rb_node **p, *parent;
1920         struct ceph_pool_perm *perm;
1921         struct page **pages;
1922         size_t pool_ns_len;
1923         int err = 0, err2 = 0, have = 0;
1924
1925         down_read(&mdsc->pool_perm_rwsem);
1926         p = &mdsc->pool_perm_tree.rb_node;
1927         while (*p) {
1928                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1929                 if (pool < perm->pool)
1930                         p = &(*p)->rb_left;
1931                 else if (pool > perm->pool)
1932                         p = &(*p)->rb_right;
1933                 else {
1934                         int ret = ceph_compare_string(pool_ns,
1935                                                 perm->pool_ns,
1936                                                 perm->pool_ns_len);
1937                         if (ret < 0)
1938                                 p = &(*p)->rb_left;
1939                         else if (ret > 0)
1940                                 p = &(*p)->rb_right;
1941                         else {
1942                                 have = perm->perm;
1943                                 break;
1944                         }
1945                 }
1946         }
1947         up_read(&mdsc->pool_perm_rwsem);
1948         if (*p)
1949                 goto out;
1950
1951         if (pool_ns)
1952                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1953                      pool, (int)pool_ns->len, pool_ns->str);
1954         else
1955                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1956
1957         down_write(&mdsc->pool_perm_rwsem);
1958         p = &mdsc->pool_perm_tree.rb_node;
1959         parent = NULL;
1960         while (*p) {
1961                 parent = *p;
1962                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1963                 if (pool < perm->pool)
1964                         p = &(*p)->rb_left;
1965                 else if (pool > perm->pool)
1966                         p = &(*p)->rb_right;
1967                 else {
1968                         int ret = ceph_compare_string(pool_ns,
1969                                                 perm->pool_ns,
1970                                                 perm->pool_ns_len);
1971                         if (ret < 0)
1972                                 p = &(*p)->rb_left;
1973                         else if (ret > 0)
1974                                 p = &(*p)->rb_right;
1975                         else {
1976                                 have = perm->perm;
1977                                 break;
1978                         }
1979                 }
1980         }
1981         if (*p) {
1982                 up_write(&mdsc->pool_perm_rwsem);
1983                 goto out;
1984         }
1985
1986         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1987                                          1, false, GFP_NOFS);
1988         if (!rd_req) {
1989                 err = -ENOMEM;
1990                 goto out_unlock;
1991         }
1992
1993         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1994         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1995         rd_req->r_base_oloc.pool = pool;
1996         if (pool_ns)
1997                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1998         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1999
2000         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2001         if (err)
2002                 goto out_unlock;
2003
2004         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2005                                          1, false, GFP_NOFS);
2006         if (!wr_req) {
2007                 err = -ENOMEM;
2008                 goto out_unlock;
2009         }
2010
2011         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2012         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2013         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2014         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2015
2016         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2017         if (err)
2018                 goto out_unlock;
2019
2020         /* one page should be large enough for STAT data */
2021         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2022         if (IS_ERR(pages)) {
2023                 err = PTR_ERR(pages);
2024                 goto out_unlock;
2025         }
2026
2027         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2028                                      0, false, true);
2029         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
2030
2031         wr_req->r_mtime = ci->vfs_inode.i_mtime;
2032         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
2033
2034         if (!err)
2035                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2036         if (!err2)
2037                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2038
2039         if (err >= 0 || err == -ENOENT)
2040                 have |= POOL_READ;
2041         else if (err != -EPERM) {
2042                 if (err == -EBLACKLISTED)
2043                         fsc->blacklisted = true;
2044                 goto out_unlock;
2045         }
2046
2047         if (err2 == 0 || err2 == -EEXIST)
2048                 have |= POOL_WRITE;
2049         else if (err2 != -EPERM) {
2050                 if (err2 == -EBLACKLISTED)
2051                         fsc->blacklisted = true;
2052                 err = err2;
2053                 goto out_unlock;
2054         }
2055
2056         pool_ns_len = pool_ns ? pool_ns->len : 0;
2057         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2058         if (!perm) {
2059                 err = -ENOMEM;
2060                 goto out_unlock;
2061         }
2062
2063         perm->pool = pool;
2064         perm->perm = have;
2065         perm->pool_ns_len = pool_ns_len;
2066         if (pool_ns_len > 0)
2067                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2068         perm->pool_ns[pool_ns_len] = 0;
2069
2070         rb_link_node(&perm->node, parent, p);
2071         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2072         err = 0;
2073 out_unlock:
2074         up_write(&mdsc->pool_perm_rwsem);
2075
2076         ceph_osdc_put_request(rd_req);
2077         ceph_osdc_put_request(wr_req);
2078 out:
2079         if (!err)
2080                 err = have;
2081         if (pool_ns)
2082                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2083                      pool, (int)pool_ns->len, pool_ns->str, err);
2084         else
2085                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2086         return err;
2087 }
2088
2089 int ceph_pool_perm_check(struct inode *inode, int need)
2090 {
2091         struct ceph_inode_info *ci = ceph_inode(inode);
2092         struct ceph_string *pool_ns;
2093         s64 pool;
2094         int ret, flags;
2095
2096         if (ci->i_vino.snap != CEPH_NOSNAP) {
2097                 /*
2098                  * Pool permission check needs to write to the first object.
2099                  * But for snapshot, head of the first object may have alread
2100                  * been deleted. Skip check to avoid creating orphan object.
2101                  */
2102                 return 0;
2103         }
2104
2105         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2106                                 NOPOOLPERM))
2107                 return 0;
2108
2109         spin_lock(&ci->i_ceph_lock);
2110         flags = ci->i_ceph_flags;
2111         pool = ci->i_layout.pool_id;
2112         spin_unlock(&ci->i_ceph_lock);
2113 check:
2114         if (flags & CEPH_I_POOL_PERM) {
2115                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2116                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2117                              pool);
2118                         return -EPERM;
2119                 }
2120                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2121                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2122                              pool);
2123                         return -EPERM;
2124                 }
2125                 return 0;
2126         }
2127
2128         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2129         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2130         ceph_put_string(pool_ns);
2131         if (ret < 0)
2132                 return ret;
2133
2134         flags = CEPH_I_POOL_PERM;
2135         if (ret & POOL_READ)
2136                 flags |= CEPH_I_POOL_RD;
2137         if (ret & POOL_WRITE)
2138                 flags |= CEPH_I_POOL_WR;
2139
2140         spin_lock(&ci->i_ceph_lock);
2141         if (pool == ci->i_layout.pool_id &&
2142             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2143                 ci->i_ceph_flags |= flags;
2144         } else {
2145                 pool = ci->i_layout.pool_id;
2146                 flags = ci->i_ceph_flags;
2147         }
2148         spin_unlock(&ci->i_ceph_lock);
2149         goto check;
2150 }
2151
2152 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2153 {
2154         struct ceph_pool_perm *perm;
2155         struct rb_node *n;
2156
2157         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2158                 n = rb_first(&mdsc->pool_perm_tree);
2159                 perm = rb_entry(n, struct ceph_pool_perm, node);
2160                 rb_erase(n, &mdsc->pool_perm_tree);
2161                 kfree(perm);
2162         }
2163 }