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,
 866                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 867         else
 868                 kfree(osd_data->pages);
 869         ceph_osdc_put_request(req);
 870 }
 871
 872 /*
 873  * initiate async writeback
 874  */
 875 static int ceph_writepages_start(struct address_space *mapping,
 876                                  struct writeback_control *wbc)
 877 {
 878         struct inode *inode = mapping->host;
 879         struct ceph_inode_info *ci = ceph_inode(inode);
 880         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 881         struct ceph_vino vino = ceph_vino(inode);
 882         pgoff_t index, start_index, end = -1;
 883         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 884         struct pagevec pvec;
 885         int rc = 0;
 886         unsigned int wsize = i_blocksize(inode);
 887         struct ceph_osd_request *req = NULL;
 888         struct ceph_writeback_ctl ceph_wbc;
 889         bool should_loop, range_whole = false;
 890         bool done = false;
 891
 892         dout("writepages_start %p (mode=%s)\n", inode,
 893              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 894              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 895
 896         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 897                 if (ci->i_wrbuffer_ref > 0) {
 898                         pr_warn_ratelimited(
 899                                 "writepage_start %p %lld forced umount\n",
 900                                 inode, ceph_ino(inode));
 901                 }
 902                 mapping_set_error(mapping, -EIO);
 903                 return -EIO; /* we're in a forced umount, don't write! */
 904         }
 905         if (fsc->mount_options->wsize < wsize)
 906                 wsize = fsc->mount_options->wsize;
 907
 908         pagevec_init(&pvec);
 909
 910         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 911         index = start_index;
 912
 913 retry:
 914         /* find oldest snap context with dirty data */
 915         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 916         if (!snapc) {
 917                 /* hmm, why does writepages get called when there
 918                    is no dirty data? */
 919                 dout(" no snap context with dirty data?\n");
 920                 goto out;
 921         }
 922         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 923              snapc, snapc->seq, snapc->num_snaps);
 924
 925         should_loop = false;
 926         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 927                 /* where to start/end? */
 928                 if (wbc->range_cyclic) {
 929                         index = start_index;
 930                         end = -1;
 931                         if (index > 0)
 932                                 should_loop = true;
 933                         dout(" cyclic, start at %lu\n", index);
 934                 } else {
 935                         index = wbc->range_start >> PAGE_SHIFT;
 936                         end = wbc->range_end >> PAGE_SHIFT;
 937                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 938                                 range_whole = true;
 939                         dout(" not cyclic, %lu to %lu\n", index, end);
 940                 }
 941         } else if (!ceph_wbc.head_snapc) {
 942                 /* Do not respect wbc->range_{start,end}. Dirty pages
 943                  * in that range can be associated with newer snapc.
 944                  * They are not writeable until we write all dirty pages
 945                  * associated with 'snapc' get written */
 946                 if (index > 0)
 947                         should_loop = true;
 948                 dout(" non-head snapc, range whole\n");
 949         }
 950
 951         ceph_put_snap_context(last_snapc);
 952         last_snapc = snapc;
 953
 954         while (!done && index <= end) {
 955                 int num_ops = 0, op_idx;
 956                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 957                 struct page **pages = NULL, **data_pages;
 958                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 959                 struct page *page;
 960                 pgoff_t strip_unit_end = 0;
 961                 u64 offset = 0, len = 0;
 962
 963                 max_pages = wsize >> PAGE_SHIFT;
 964
 965 get_more_pages:
 966                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 967                                                 end, PAGECACHE_TAG_DIRTY,
 968                                                 max_pages - locked_pages);
 969                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 970                 if (!pvec_pages && !locked_pages)
 971                         break;
 972                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 973                         page = pvec.pages[i];
 974                         dout("? %p idx %lu\n", page, page->index);
 975                         if (locked_pages == 0)
 976                                 lock_page(page);  /* first page */
 977                         else if (!trylock_page(page))
 978                                 break;
 979
 980                         /* only dirty pages, or our accounting breaks */
 981                         if (unlikely(!PageDirty(page)) ||
 982                             unlikely(page->mapping != mapping)) {
 983                                 dout("!dirty or !mapping %p\n", page);
 984                                 unlock_page(page);
 985                                 continue;
 986                         }
 987                         /* only if matching snap context */
 988                         pgsnapc = page_snap_context(page);
 989                         if (pgsnapc != snapc) {
 990                                 dout("page snapc %p %lld != oldest %p %lld\n",
 991                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 992                                 if (!should_loop &&
 993                                     !ceph_wbc.head_snapc &&
 994                                     wbc->sync_mode != WB_SYNC_NONE)
 995                                         should_loop = true;
 996                                 unlock_page(page);
 997                                 continue;
 998                         }
 999                         if (page_offset(page) >= ceph_wbc.i_size) {
1000                                 dout("%p page eof %llu\n",
1001                                      page, ceph_wbc.i_size);
1002                                 if ((ceph_wbc.size_stable ||
1003                                     page_offset(page) >= i_size_read(inode)) &&
1004                                     clear_page_dirty_for_io(page))
1005                                         mapping->a_ops->invalidatepage(page,
1006                                                                 0, PAGE_SIZE);
1007                                 unlock_page(page);
1008                                 continue;
1009                         }
1010                         if (strip_unit_end && (page->index > strip_unit_end)) {
1011                                 dout("end of strip unit %p\n", page);
1012                                 unlock_page(page);
1013                                 break;
1014                         }
1015                         if (PageWriteback(page)) {
1016                                 if (wbc->sync_mode == WB_SYNC_NONE) {
1017                                         dout("%p under writeback\n", page);
1018                                         unlock_page(page);
1019                                         continue;
1020                                 }
1021                                 dout("waiting on writeback %p\n", page);
1022                                 wait_on_page_writeback(page);
1023                         }
1024
1025                         if (!clear_page_dirty_for_io(page)) {
1026                                 dout("%p !clear_page_dirty_for_io\n", page);
1027                                 unlock_page(page);
1028                                 continue;
1029                         }
1030
1031                         /*
1032                          * We have something to write.  If this is
1033                          * the first locked page this time through,
1034                          * calculate max possinle write size and
1035                          * allocate a page array
1036                          */
1037                         if (locked_pages == 0) {
1038                                 u64 objnum;
1039                                 u64 objoff;
1040                                 u32 xlen;
1041
1042                                 /* prepare async write request */
1043                                 offset = (u64)page_offset(page);
1044                                 ceph_calc_file_object_mapping(&ci->i_layout,
1045                                                               offset, wsize,
1046                                                               &objnum, &objoff,
1047                                                               &xlen);
1048                                 len = xlen;
1049
1050                                 num_ops = 1;
1051                                 strip_unit_end = page->index +
1052                                         ((len - 1) >> PAGE_SHIFT);
1053
1054                                 BUG_ON(pages);
1055                                 max_pages = calc_pages_for(0, (u64)len);
1056                                 pages = kmalloc_array(max_pages,
1057                                                       sizeof(*pages),
1058                                                       GFP_NOFS);
1059                                 if (!pages) {
1060                                         pool = fsc->wb_pagevec_pool;
1061                                         pages = mempool_alloc(pool, GFP_NOFS);
1062                                         BUG_ON(!pages);
1063                                 }
1064
1065                                 len = 0;
1066                         } else if (page->index !=
1067                                    (offset + len) >> PAGE_SHIFT) {
1068                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
1069                                                         CEPH_OSD_MAX_OPS)) {
1070                                         redirty_page_for_writepage(wbc, page);
1071                                         unlock_page(page);
1072                                         break;
1073                                 }
1074
1075                                 num_ops++;
1076                                 offset = (u64)page_offset(page);
1077                                 len = 0;
1078                         }
1079
1080                         /* note position of first page in pvec */
1081                         dout("%p will write page %p idx %lu\n",
1082                              inode, page, page->index);
1083
1084                         if (atomic_long_inc_return(&fsc->writeback_count) >
1085                             CONGESTION_ON_THRESH(
1086                                     fsc->mount_options->congestion_kb)) {
1087                                 set_bdi_congested(inode_to_bdi(inode),
1088                                                   BLK_RW_ASYNC);
1089                         }
1090
1091
1092                         pages[locked_pages++] = page;
1093                         pvec.pages[i] = NULL;
1094
1095                         len += PAGE_SIZE;
1096                 }
1097
1098                 /* did we get anything? */
1099                 if (!locked_pages)
1100                         goto release_pvec_pages;
1101                 if (i) {
1102                         unsigned j, n = 0;
1103                         /* shift unused page to beginning of pvec */
1104                         for (j = 0; j < pvec_pages; j++) {
1105                                 if (!pvec.pages[j])
1106                                         continue;
1107                                 if (n < j)
1108                                         pvec.pages[n] = pvec.pages[j];
1109                                 n++;
1110                         }
1111                         pvec.nr = n;
1112
1113                         if (pvec_pages && i == pvec_pages &&
1114                             locked_pages < max_pages) {
1115                                 dout("reached end pvec, trying for more\n");
1116                                 pagevec_release(&pvec);
1117                                 goto get_more_pages;
1118                         }
1119                 }
1120
1121 new_request:
1122                 offset = page_offset(pages[0]);
1123                 len = wsize;
1124
1125                 req = ceph_osdc_new_request(&fsc->client->osdc,
1126                                         &ci->i_layout, vino,
1127                                         offset, &len, 0, num_ops,
1128                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1129                                         snapc, ceph_wbc.truncate_seq,
1130                                         ceph_wbc.truncate_size, false);
1131                 if (IS_ERR(req)) {
1132                         req = ceph_osdc_new_request(&fsc->client->osdc,
1133                                                 &ci->i_layout, vino,
1134                                                 offset, &len, 0,
1135                                                 min(num_ops,
1136                                                     CEPH_OSD_SLAB_OPS),
1137                                                 CEPH_OSD_OP_WRITE,
1138                                                 CEPH_OSD_FLAG_WRITE,
1139                                                 snapc, ceph_wbc.truncate_seq,
1140                                                 ceph_wbc.truncate_size, true);
1141                         BUG_ON(IS_ERR(req));
1142                 }
1143                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1144                              PAGE_SIZE - offset);
1145
1146                 req->r_callback = writepages_finish;
1147                 req->r_inode = inode;
1148
1149                 /* Format the osd request message and submit the write */
1150                 len = 0;
1151                 data_pages = pages;
1152                 op_idx = 0;
1153                 for (i = 0; i < locked_pages; i++) {
1154                         u64 cur_offset = page_offset(pages[i]);
1155                         if (offset + len != cur_offset) {
1156                                 if (op_idx + 1 == req->r_num_ops)
1157                                         break;
1158                                 osd_req_op_extent_dup_last(req, op_idx,
1159                                                            cur_offset - offset);
1160                                 dout("writepages got pages at %llu~%llu\n",
1161                                      offset, len);
1162                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1163                                                         data_pages, len, 0,
1164                                                         !!pool, false);
1165                                 osd_req_op_extent_update(req, op_idx, len);
1166
1167                                 len = 0;
1168                                 offset = cur_offset;
1169                                 data_pages = pages + i;
1170                                 op_idx++;
1171                         }
1172
1173                         set_page_writeback(pages[i]);
1174                         len += PAGE_SIZE;
1175                 }
1176
1177                 if (ceph_wbc.size_stable) {
1178                         len = min(len, ceph_wbc.i_size - offset);
1179                 } else if (i == locked_pages) {
1180                         /* writepages_finish() clears writeback pages
1181                          * according to the data length, so make sure
1182                          * data length covers all locked pages */
1183                         u64 min_len = len + 1 - PAGE_SIZE;
1184                         len = get_writepages_data_length(inode, pages[i - 1],
1185                                                          offset);
1186                         len = max(len, min_len);
1187                 }
1188                 dout("writepages got pages at %llu~%llu\n", offset, len);
1189
1190                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1191                                                  0, !!pool, false);
1192                 osd_req_op_extent_update(req, op_idx, len);
1193
1194                 BUG_ON(op_idx + 1 != req->r_num_ops);
1195
1196                 pool = NULL;
1197                 if (i < locked_pages) {
1198                         BUG_ON(num_ops <= req->r_num_ops);
1199                         num_ops -= req->r_num_ops;
1200                         locked_pages -= i;
1201
1202                         /* allocate new pages array for next request */
1203                         data_pages = pages;
1204                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1205                                               GFP_NOFS);
1206                         if (!pages) {
1207                                 pool = fsc->wb_pagevec_pool;
1208                                 pages = mempool_alloc(pool, GFP_NOFS);
1209                                 BUG_ON(!pages);
1210                         }
1211                         memcpy(pages, data_pages + i,
1212                                locked_pages * sizeof(*pages));
1213                         memset(data_pages + i, 0,
1214                                locked_pages * sizeof(*pages));
1215                 } else {
1216                         BUG_ON(num_ops != req->r_num_ops);
1217                         index = pages[i - 1]->index + 1;
1218                         /* request message now owns the pages array */
1219                         pages = NULL;
1220                 }
1221
1222                 req->r_mtime = inode->i_mtime;
1223                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1224                 BUG_ON(rc);
1225                 req = NULL;
1226
1227                 wbc->nr_to_write -= i;
1228                 if (pages)
1229                         goto new_request;
1230
1231                 /*
1232                  * We stop writing back only if we are not doing
1233                  * integrity sync. In case of integrity sync we have to
1234                  * keep going until we have written all the pages
1235                  * we tagged for writeback prior to entering this loop.
1236                  */
1237                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1238                         done = true;
1239
1240 release_pvec_pages:
1241                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1242                      pvec.nr ? pvec.pages[0] : NULL);
1243                 pagevec_release(&pvec);
1244         }
1245
1246         if (should_loop && !done) {
1247                 /* more to do; loop back to beginning of file */
1248                 dout("writepages looping back to beginning of file\n");
1249                 end = start_index - 1; /* OK even when start_index == 0 */
1250
1251                 /* to write dirty pages associated with next snapc,
1252                  * we need to wait until current writes complete */
1253                 if (wbc->sync_mode != WB_SYNC_NONE &&
1254                     start_index == 0 && /* all dirty pages were checked */
1255                     !ceph_wbc.head_snapc) {
1256                         struct page *page;
1257                         unsigned i, nr;
1258                         index = 0;
1259                         while ((index <= end) &&
1260                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1261                                                 PAGECACHE_TAG_WRITEBACK))) {
1262                                 for (i = 0; i < nr; i++) {
1263                                         page = pvec.pages[i];
1264                                         if (page_snap_context(page) != snapc)
1265                                                 continue;
1266                                         wait_on_page_writeback(page);
1267                                 }
1268                                 pagevec_release(&pvec);
1269                                 cond_resched();
1270                         }
1271                 }
1272
1273                 start_index = 0;
1274                 index = 0;
1275                 goto retry;
1276         }
1277
1278         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1279                 mapping->writeback_index = index;
1280
1281 out:
1282         ceph_osdc_put_request(req);
1283         ceph_put_snap_context(last_snapc);
1284         dout("writepages dend - startone, rc = %d\n", rc);
1285         return rc;
1286 }
1287
1288
1289
1290 /*
1291  * See if a given @snapc is either writeable, or already written.
1292  */
1293 static int context_is_writeable_or_written(struct inode *inode,
1294                                            struct ceph_snap_context *snapc)
1295 {
1296         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1297         int ret = !oldest || snapc->seq <= oldest->seq;
1298
1299         ceph_put_snap_context(oldest);
1300         return ret;
1301 }
1302
1303 /*
1304  * We are only allowed to write into/dirty the page if the page is
1305  * clean, or already dirty within the same snap context.
1306  *
1307  * called with page locked.
1308  * return success with page locked,
1309  * or any failure (incl -EAGAIN) with page unlocked.
1310  */
1311 static int ceph_update_writeable_page(struct file *file,
1312                             loff_t pos, unsigned len,
1313                             struct page *page)
1314 {
1315         struct inode *inode = file_inode(file);
1316         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1317         struct ceph_inode_info *ci = ceph_inode(inode);
1318         loff_t page_off = pos & PAGE_MASK;
1319         int pos_in_page = pos & ~PAGE_MASK;
1320         int end_in_page = pos_in_page + len;
1321         loff_t i_size;
1322         int r;
1323         struct ceph_snap_context *snapc, *oldest;
1324
1325         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1326                 dout(" page %p forced umount\n", page);
1327                 unlock_page(page);
1328                 return -EIO;
1329         }
1330
1331 retry_locked:
1332         /* writepages currently holds page lock, but if we change that later, */
1333         wait_on_page_writeback(page);
1334
1335         snapc = page_snap_context(page);
1336         if (snapc && snapc != ci->i_head_snapc) {
1337                 /*
1338                  * this page is already dirty in another (older) snap
1339                  * context!  is it writeable now?
1340                  */
1341                 oldest = get_oldest_context(inode, NULL, NULL);
1342                 if (snapc->seq > oldest->seq) {
1343                         ceph_put_snap_context(oldest);
1344                         dout(" page %p snapc %p not current or oldest\n",
1345                              page, snapc);
1346                         /*
1347                          * queue for writeback, and wait for snapc to
1348                          * be writeable or written
1349                          */
1350                         snapc = ceph_get_snap_context(snapc);
1351                         unlock_page(page);
1352                         ceph_queue_writeback(inode);
1353                         r = wait_event_killable(ci->i_cap_wq,
1354                                context_is_writeable_or_written(inode, snapc));
1355                         ceph_put_snap_context(snapc);
1356                         if (r == -ERESTARTSYS)
1357                                 return r;
1358                         return -EAGAIN;
1359                 }
1360                 ceph_put_snap_context(oldest);
1361
1362                 /* yay, writeable, do it now (without dropping page lock) */
1363                 dout(" page %p snapc %p not current, but oldest\n",
1364                      page, snapc);
1365                 if (!clear_page_dirty_for_io(page))
1366                         goto retry_locked;
1367                 r = writepage_nounlock(page, NULL);
1368                 if (r < 0)
1369                         goto fail_unlock;
1370                 goto retry_locked;
1371         }
1372
1373         if (PageUptodate(page)) {
1374                 dout(" page %p already uptodate\n", page);
1375                 return 0;
1376         }
1377
1378         /* full page? */
1379         if (pos_in_page == 0 && len == PAGE_SIZE)
1380                 return 0;
1381
1382         /* past end of file? */
1383         i_size = i_size_read(inode);
1384
1385         if (page_off >= i_size ||
1386             (pos_in_page == 0 && (pos+len) >= i_size &&
1387              end_in_page - pos_in_page != PAGE_SIZE)) {
1388                 dout(" zeroing %p 0 - %d and %d - %d\n",
1389                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1390                 zero_user_segments(page,
1391                                    0, pos_in_page,
1392                                    end_in_page, PAGE_SIZE);
1393                 return 0;
1394         }
1395
1396         /* we need to read it. */
1397         r = ceph_do_readpage(file, page);
1398         if (r < 0) {
1399                 if (r == -EINPROGRESS)
1400                         return -EAGAIN;
1401                 goto fail_unlock;
1402         }
1403         goto retry_locked;
1404 fail_unlock:
1405         unlock_page(page);
1406         return r;
1407 }
1408
1409 /*
1410  * We are only allowed to write into/dirty the page if the page is
1411  * clean, or already dirty within the same snap context.
1412  */
1413 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1414                             loff_t pos, unsigned len, unsigned flags,
1415                             struct page **pagep, void **fsdata)
1416 {
1417         struct inode *inode = file_inode(file);
1418         struct page *page;
1419         pgoff_t index = pos >> PAGE_SHIFT;
1420         int r;
1421
1422         do {
1423                 /* get a page */
1424                 page = grab_cache_page_write_begin(mapping, index, 0);
1425                 if (!page)
1426                         return -ENOMEM;
1427
1428                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1429                      inode, page, (int)pos, (int)len);
1430
1431                 r = ceph_update_writeable_page(file, pos, len, page);
1432                 if (r < 0)
1433                         put_page(page);
1434                 else
1435                         *pagep = page;
1436         } while (r == -EAGAIN);
1437
1438         return r;
1439 }
1440
1441 /*
1442  * we don't do anything in here that simple_write_end doesn't do
1443  * except adjust dirty page accounting
1444  */
1445 static int ceph_write_end(struct file *file, struct address_space *mapping,
1446                           loff_t pos, unsigned len, unsigned copied,
1447                           struct page *page, void *fsdata)
1448 {
1449         struct inode *inode = file_inode(file);
1450         bool check_cap = false;
1451
1452         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1453              inode, page, (int)pos, (int)copied, (int)len);
1454
1455         /* zero the stale part of the page if we did a short copy */
1456         if (!PageUptodate(page)) {
1457                 if (copied < len) {
1458                         copied = 0;
1459                         goto out;
1460                 }
1461                 SetPageUptodate(page);
1462         }
1463
1464         /* did file size increase? */
1465         if (pos+copied > i_size_read(inode))
1466                 check_cap = ceph_inode_set_size(inode, pos+copied);
1467
1468         set_page_dirty(page);
1469
1470 out:
1471         unlock_page(page);
1472         put_page(page);
1473
1474         if (check_cap)
1475                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1476
1477         return copied;
1478 }
1479
1480 /*
1481  * we set .direct_IO to indicate direct io is supported, but since we
1482  * intercept O_DIRECT reads and writes early, this function should
1483  * never get called.
1484  */
1485 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1486 {
1487         WARN_ON(1);
1488         return -EINVAL;
1489 }
1490
1491 const struct address_space_operations ceph_aops = {
1492         .readpage = ceph_readpage,
1493         .readpages = ceph_readpages,
1494         .writepage = ceph_writepage,
1495         .writepages = ceph_writepages_start,
1496         .write_begin = ceph_write_begin,
1497         .write_end = ceph_write_end,
1498         .set_page_dirty = ceph_set_page_dirty,
1499         .invalidatepage = ceph_invalidatepage,
1500         .releasepage = ceph_releasepage,
1501         .direct_IO = ceph_direct_io,
1502 };
1503
1504 static void ceph_block_sigs(sigset_t *oldset)
1505 {
1506         sigset_t mask;
1507         siginitsetinv(&mask, sigmask(SIGKILL));
1508         sigprocmask(SIG_BLOCK, &mask, oldset);
1509 }
1510
1511 static void ceph_restore_sigs(sigset_t *oldset)
1512 {
1513         sigprocmask(SIG_SETMASK, oldset, NULL);
1514 }
1515
1516 /*
1517  * vm ops
1518  */
1519 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1520 {
1521         struct vm_area_struct *vma = vmf->vma;
1522         struct inode *inode = file_inode(vma->vm_file);
1523         struct ceph_inode_info *ci = ceph_inode(inode);
1524         struct ceph_file_info *fi = vma->vm_file->private_data;
1525         struct page *pinned_page = NULL;
1526         loff_t off = vmf->pgoff << PAGE_SHIFT;
1527         int want, got, err;
1528         sigset_t oldset;
1529         vm_fault_t ret = VM_FAULT_SIGBUS;
1530
1531         ceph_block_sigs(&oldset);
1532
1533         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1534              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1535         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1536                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1537         else
1538                 want = CEPH_CAP_FILE_CACHE;
1539
1540         got = 0;
1541         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1,
1542                             &got, &pinned_page);
1543         if (err < 0)
1544                 goto out_restore;
1545
1546         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1547              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1548
1549         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1550             ci->i_inline_version == CEPH_INLINE_NONE) {
1551                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1552                 ceph_add_rw_context(fi, &rw_ctx);
1553                 ret = filemap_fault(vmf);
1554                 ceph_del_rw_context(fi, &rw_ctx);
1555                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1556                         inode, off, (size_t)PAGE_SIZE,
1557                                 ceph_cap_string(got), ret);
1558         } else
1559                 err = -EAGAIN;
1560
1561         if (pinned_page)
1562                 put_page(pinned_page);
1563         ceph_put_cap_refs(ci, got);
1564
1565         if (err != -EAGAIN)
1566                 goto out_restore;
1567
1568         /* read inline data */
1569         if (off >= PAGE_SIZE) {
1570                 /* does not support inline data > PAGE_SIZE */
1571                 ret = VM_FAULT_SIGBUS;
1572         } else {
1573                 struct address_space *mapping = inode->i_mapping;
1574                 struct page *page = find_or_create_page(mapping, 0,
1575                                                 mapping_gfp_constraint(mapping,
1576                                                 ~__GFP_FS));
1577                 if (!page) {
1578                         ret = VM_FAULT_OOM;
1579                         goto out_inline;
1580                 }
1581                 err = __ceph_do_getattr(inode, page,
1582                                          CEPH_STAT_CAP_INLINE_DATA, true);
1583                 if (err < 0 || off >= i_size_read(inode)) {
1584                         unlock_page(page);
1585                         put_page(page);
1586                         ret = vmf_error(err);
1587                         goto out_inline;
1588                 }
1589                 if (err < PAGE_SIZE)
1590                         zero_user_segment(page, err, PAGE_SIZE);
1591                 else
1592                         flush_dcache_page(page);
1593                 SetPageUptodate(page);
1594                 vmf->page = page;
1595                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1596 out_inline:
1597                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1598                      inode, off, (size_t)PAGE_SIZE, ret);
1599         }
1600 out_restore:
1601         ceph_restore_sigs(&oldset);
1602         if (err < 0)
1603                 ret = vmf_error(err);
1604
1605         return ret;
1606 }
1607
1608 /*
1609  * Reuse write_begin here for simplicity.
1610  */
1611 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1612 {
1613         struct vm_area_struct *vma = vmf->vma;
1614         struct inode *inode = file_inode(vma->vm_file);
1615         struct ceph_inode_info *ci = ceph_inode(inode);
1616         struct ceph_file_info *fi = vma->vm_file->private_data;
1617         struct ceph_cap_flush *prealloc_cf;
1618         struct page *page = vmf->page;
1619         loff_t off = page_offset(page);
1620         loff_t size = i_size_read(inode);
1621         size_t len;
1622         int want, got, err;
1623         sigset_t oldset;
1624         vm_fault_t ret = VM_FAULT_SIGBUS;
1625
1626         prealloc_cf = ceph_alloc_cap_flush();
1627         if (!prealloc_cf)
1628                 return VM_FAULT_OOM;
1629
1630         sb_start_pagefault(inode->i_sb);
1631         ceph_block_sigs(&oldset);
1632
1633         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1634                 struct page *locked_page = NULL;
1635                 if (off == 0) {
1636                         lock_page(page);
1637                         locked_page = page;
1638                 }
1639                 err = ceph_uninline_data(vma->vm_file, locked_page);
1640                 if (locked_page)
1641                         unlock_page(locked_page);
1642                 if (err < 0)
1643                         goto out_free;
1644         }
1645
1646         if (off + PAGE_SIZE <= size)
1647                 len = PAGE_SIZE;
1648         else
1649                 len = size & ~PAGE_MASK;
1650
1651         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1652              inode, ceph_vinop(inode), off, len, size);
1653         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1654                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1655         else
1656                 want = CEPH_CAP_FILE_BUFFER;
1657
1658         got = 0;
1659         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len,
1660                             &got, NULL);
1661         if (err < 0)
1662                 goto out_free;
1663
1664         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1665              inode, off, len, ceph_cap_string(got));
1666
1667         /* Update time before taking page lock */
1668         file_update_time(vma->vm_file);
1669         inode_inc_iversion_raw(inode);
1670
1671         do {
1672                 lock_page(page);
1673
1674                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1675                         unlock_page(page);
1676                         ret = VM_FAULT_NOPAGE;
1677                         break;
1678                 }
1679
1680                 err = ceph_update_writeable_page(vma->vm_file, off, len, page);
1681                 if (err >= 0) {
1682                         /* success.  we'll keep the page locked. */
1683                         set_page_dirty(page);
1684                         ret = VM_FAULT_LOCKED;
1685                 }
1686         } while (err == -EAGAIN);
1687
1688         if (ret == VM_FAULT_LOCKED ||
1689             ci->i_inline_version != CEPH_INLINE_NONE) {
1690                 int dirty;
1691                 spin_lock(&ci->i_ceph_lock);
1692                 ci->i_inline_version = CEPH_INLINE_NONE;
1693                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1694                                                &prealloc_cf);
1695                 spin_unlock(&ci->i_ceph_lock);
1696                 if (dirty)
1697                         __mark_inode_dirty(inode, dirty);
1698         }
1699
1700         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1701              inode, off, len, ceph_cap_string(got), ret);
1702         ceph_put_cap_refs(ci, got);
1703 out_free:
1704         ceph_restore_sigs(&oldset);
1705         sb_end_pagefault(inode->i_sb);
1706         ceph_free_cap_flush(prealloc_cf);
1707         if (err < 0)
1708                 ret = vmf_error(err);
1709         return ret;
1710 }
1711
1712 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1713                            char *data, size_t len)
1714 {
1715         struct address_space *mapping = inode->i_mapping;
1716         struct page *page;
1717
1718         if (locked_page) {
1719                 page = locked_page;
1720         } else {
1721                 if (i_size_read(inode) == 0)
1722                         return;
1723                 page = find_or_create_page(mapping, 0,
1724                                            mapping_gfp_constraint(mapping,
1725                                            ~__GFP_FS));
1726                 if (!page)
1727                         return;
1728                 if (PageUptodate(page)) {
1729                         unlock_page(page);
1730                         put_page(page);
1731                         return;
1732                 }
1733         }
1734
1735         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1736              inode, ceph_vinop(inode), len, locked_page);
1737
1738         if (len > 0) {
1739                 void *kaddr = kmap_atomic(page);
1740                 memcpy(kaddr, data, len);
1741                 kunmap_atomic(kaddr);
1742         }
1743
1744         if (page != locked_page) {
1745                 if (len < PAGE_SIZE)
1746                         zero_user_segment(page, len, PAGE_SIZE);
1747                 else
1748                         flush_dcache_page(page);
1749
1750                 SetPageUptodate(page);
1751                 unlock_page(page);
1752                 put_page(page);
1753         }
1754 }
1755
1756 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1757 {
1758         struct inode *inode = file_inode(filp);
1759         struct ceph_inode_info *ci = ceph_inode(inode);
1760         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1761         struct ceph_osd_request *req;
1762         struct page *page = NULL;
1763         u64 len, inline_version;
1764         int err = 0;
1765         bool from_pagecache = false;
1766
1767         spin_lock(&ci->i_ceph_lock);
1768         inline_version = ci->i_inline_version;
1769         spin_unlock(&ci->i_ceph_lock);
1770
1771         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1772              inode, ceph_vinop(inode), inline_version);
1773
1774         if (inline_version == 1 || /* initial version, no data */
1775             inline_version == CEPH_INLINE_NONE)
1776                 goto out;
1777
1778         if (locked_page) {
1779                 page = locked_page;
1780                 WARN_ON(!PageUptodate(page));
1781         } else if (ceph_caps_issued(ci) &
1782                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1783                 page = find_get_page(inode->i_mapping, 0);
1784                 if (page) {
1785                         if (PageUptodate(page)) {
1786                                 from_pagecache = true;
1787                                 lock_page(page);
1788                         } else {
1789                                 put_page(page);
1790                                 page = NULL;
1791                         }
1792                 }
1793         }
1794
1795         if (page) {
1796                 len = i_size_read(inode);
1797                 if (len > PAGE_SIZE)
1798                         len = PAGE_SIZE;
1799         } else {
1800                 page = __page_cache_alloc(GFP_NOFS);
1801                 if (!page) {
1802                         err = -ENOMEM;
1803                         goto out;
1804                 }
1805                 err = __ceph_do_getattr(inode, page,
1806                                         CEPH_STAT_CAP_INLINE_DATA, true);
1807                 if (err < 0) {
1808                         /* no inline data */
1809                         if (err == -ENODATA)
1810                                 err = 0;
1811                         goto out;
1812                 }
1813                 len = err;
1814         }
1815
1816         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1817                                     ceph_vino(inode), 0, &len, 0, 1,
1818                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1819                                     NULL, 0, 0, false);
1820         if (IS_ERR(req)) {
1821                 err = PTR_ERR(req);
1822                 goto out;
1823         }
1824
1825         req->r_mtime = inode->i_mtime;
1826         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1827         if (!err)
1828                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1829         ceph_osdc_put_request(req);
1830         if (err < 0)
1831                 goto out;
1832
1833         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1834                                     ceph_vino(inode), 0, &len, 1, 3,
1835                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1836                                     NULL, ci->i_truncate_seq,
1837                                     ci->i_truncate_size, false);
1838         if (IS_ERR(req)) {
1839                 err = PTR_ERR(req);
1840                 goto out;
1841         }
1842
1843         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1844
1845         {
1846                 __le64 xattr_buf = cpu_to_le64(inline_version);
1847                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1848                                             "inline_version", &xattr_buf,
1849                                             sizeof(xattr_buf),
1850                                             CEPH_OSD_CMPXATTR_OP_GT,
1851                                             CEPH_OSD_CMPXATTR_MODE_U64);
1852                 if (err)
1853                         goto out_put;
1854         }
1855
1856         {
1857                 char xattr_buf[32];
1858                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1859                                          "%llu", inline_version);
1860                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1861                                             "inline_version",
1862                                             xattr_buf, xattr_len, 0, 0);
1863                 if (err)
1864                         goto out_put;
1865         }
1866
1867         req->r_mtime = inode->i_mtime;
1868         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1869         if (!err)
1870                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1871
1872         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
1873                                   req->r_end_latency, err);
1874
1875 out_put:
1876         ceph_osdc_put_request(req);
1877         if (err == -ECANCELED)
1878                 err = 0;
1879 out:
1880         if (page && page != locked_page) {
1881                 if (from_pagecache) {
1882                         unlock_page(page);
1883                         put_page(page);
1884                 } else
1885                         __free_pages(page, 0);
1886         }
1887
1888         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1889              inode, ceph_vinop(inode), inline_version, err);
1890         return err;
1891 }
1892
1893 static const struct vm_operations_struct ceph_vmops = {
1894         .fault          = ceph_filemap_fault,
1895         .page_mkwrite   = ceph_page_mkwrite,
1896 };
1897
1898 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1899 {
1900         struct address_space *mapping = file->f_mapping;
1901
1902         if (!mapping->a_ops->readpage)
1903                 return -ENOEXEC;
1904         file_accessed(file);
1905         vma->vm_ops = &ceph_vmops;
1906         return 0;
1907 }
1908
1909 enum {
1910         POOL_READ       = 1,
1911         POOL_WRITE      = 2,
1912 };
1913
1914 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1915                                 s64 pool, struct ceph_string *pool_ns)
1916 {
1917         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1918         struct ceph_mds_client *mdsc = fsc->mdsc;
1919         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1920         struct rb_node **p, *parent;
1921         struct ceph_pool_perm *perm;
1922         struct page **pages;
1923         size_t pool_ns_len;
1924         int err = 0, err2 = 0, have = 0;
1925
1926         down_read(&mdsc->pool_perm_rwsem);
1927         p = &mdsc->pool_perm_tree.rb_node;
1928         while (*p) {
1929                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1930                 if (pool < perm->pool)
1931                         p = &(*p)->rb_left;
1932                 else if (pool > perm->pool)
1933                         p = &(*p)->rb_right;
1934                 else {
1935                         int ret = ceph_compare_string(pool_ns,
1936                                                 perm->pool_ns,
1937                                                 perm->pool_ns_len);
1938                         if (ret < 0)
1939                                 p = &(*p)->rb_left;
1940                         else if (ret > 0)
1941                                 p = &(*p)->rb_right;
1942                         else {
1943                                 have = perm->perm;
1944                                 break;
1945                         }
1946                 }
1947         }
1948         up_read(&mdsc->pool_perm_rwsem);
1949         if (*p)
1950                 goto out;
1951
1952         if (pool_ns)
1953                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1954                      pool, (int)pool_ns->len, pool_ns->str);
1955         else
1956                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1957
1958         down_write(&mdsc->pool_perm_rwsem);
1959         p = &mdsc->pool_perm_tree.rb_node;
1960         parent = NULL;
1961         while (*p) {
1962                 parent = *p;
1963                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1964                 if (pool < perm->pool)
1965                         p = &(*p)->rb_left;
1966                 else if (pool > perm->pool)
1967                         p = &(*p)->rb_right;
1968                 else {
1969                         int ret = ceph_compare_string(pool_ns,
1970                                                 perm->pool_ns,
1971                                                 perm->pool_ns_len);
1972                         if (ret < 0)
1973                                 p = &(*p)->rb_left;
1974                         else if (ret > 0)
1975                                 p = &(*p)->rb_right;
1976                         else {
1977                                 have = perm->perm;
1978                                 break;
1979                         }
1980                 }
1981         }
1982         if (*p) {
1983                 up_write(&mdsc->pool_perm_rwsem);
1984                 goto out;
1985         }
1986
1987         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1988                                          1, false, GFP_NOFS);
1989         if (!rd_req) {
1990                 err = -ENOMEM;
1991                 goto out_unlock;
1992         }
1993
1994         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1995         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1996         rd_req->r_base_oloc.pool = pool;
1997         if (pool_ns)
1998                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1999         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2000
2001         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2002         if (err)
2003                 goto out_unlock;
2004
2005         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2006                                          1, false, GFP_NOFS);
2007         if (!wr_req) {
2008                 err = -ENOMEM;
2009                 goto out_unlock;
2010         }
2011
2012         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2013         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2014         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2015         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2016
2017         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2018         if (err)
2019                 goto out_unlock;
2020
2021         /* one page should be large enough for STAT data */
2022         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2023         if (IS_ERR(pages)) {
2024                 err = PTR_ERR(pages);
2025                 goto out_unlock;
2026         }
2027
2028         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2029                                      0, false, true);
2030         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
2031
2032         wr_req->r_mtime = ci->vfs_inode.i_mtime;
2033         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
2034
2035         if (!err)
2036                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2037         if (!err2)
2038                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2039
2040         if (err >= 0 || err == -ENOENT)
2041                 have |= POOL_READ;
2042         else if (err != -EPERM) {
2043                 if (err == -EBLACKLISTED)
2044                         fsc->blacklisted = true;
2045                 goto out_unlock;
2046         }
2047
2048         if (err2 == 0 || err2 == -EEXIST)
2049                 have |= POOL_WRITE;
2050         else if (err2 != -EPERM) {
2051                 if (err2 == -EBLACKLISTED)
2052                         fsc->blacklisted = true;
2053                 err = err2;
2054                 goto out_unlock;
2055         }
2056
2057         pool_ns_len = pool_ns ? pool_ns->len : 0;
2058         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2059         if (!perm) {
2060                 err = -ENOMEM;
2061                 goto out_unlock;
2062         }
2063
2064         perm->pool = pool;
2065         perm->perm = have;
2066         perm->pool_ns_len = pool_ns_len;
2067         if (pool_ns_len > 0)
2068                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2069         perm->pool_ns[pool_ns_len] = 0;
2070
2071         rb_link_node(&perm->node, parent, p);
2072         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2073         err = 0;
2074 out_unlock:
2075         up_write(&mdsc->pool_perm_rwsem);
2076
2077         ceph_osdc_put_request(rd_req);
2078         ceph_osdc_put_request(wr_req);
2079 out:
2080         if (!err)
2081                 err = have;
2082         if (pool_ns)
2083                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2084                      pool, (int)pool_ns->len, pool_ns->str, err);
2085         else
2086                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2087         return err;
2088 }
2089
2090 int ceph_pool_perm_check(struct inode *inode, int need)
2091 {
2092         struct ceph_inode_info *ci = ceph_inode(inode);
2093         struct ceph_string *pool_ns;
2094         s64 pool;
2095         int ret, flags;
2096
2097         if (ci->i_vino.snap != CEPH_NOSNAP) {
2098                 /*
2099                  * Pool permission check needs to write to the first object.
2100                  * But for snapshot, head of the first object may have alread
2101                  * been deleted. Skip check to avoid creating orphan object.
2102                  */
2103                 return 0;
2104         }
2105
2106         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2107                                 NOPOOLPERM))
2108                 return 0;
2109
2110         spin_lock(&ci->i_ceph_lock);
2111         flags = ci->i_ceph_flags;
2112         pool = ci->i_layout.pool_id;
2113         spin_unlock(&ci->i_ceph_lock);
2114 check:
2115         if (flags & CEPH_I_POOL_PERM) {
2116                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2117                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2118                              pool);
2119                         return -EPERM;
2120                 }
2121                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2122                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2123                              pool);
2124                         return -EPERM;
2125                 }
2126                 return 0;
2127         }
2128
2129         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2130         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2131         ceph_put_string(pool_ns);
2132         if (ret < 0)
2133                 return ret;
2134
2135         flags = CEPH_I_POOL_PERM;
2136         if (ret & POOL_READ)
2137                 flags |= CEPH_I_POOL_RD;
2138         if (ret & POOL_WRITE)
2139                 flags |= CEPH_I_POOL_WR;
2140
2141         spin_lock(&ci->i_ceph_lock);
2142         if (pool == ci->i_layout.pool_id &&
2143             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2144                 ci->i_ceph_flags |= flags;
2145         } else {
2146                 pool = ci->i_layout.pool_id;
2147                 flags = ci->i_ceph_flags;
2148         }
2149         spin_unlock(&ci->i_ceph_lock);
2150         goto check;
2151 }
2152
2153 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2154 {
2155         struct ceph_pool_perm *perm;
2156         struct rb_node *n;
2157
2158         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2159                 n = rb_first(&mdsc->pool_perm_tree);
2160                 perm = rb_entry(n, struct ceph_pool_perm, node);
2161                 rb_erase(n, &mdsc->pool_perm_tree);
2162                 kfree(perm);
2163         }
2164 }