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/swap.h>
   8 #include <linux/pagemap.h>
   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 #include <linux/netfs.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include "metric.h"
  21 #include <linux/ceph/osd_client.h>
  22 #include <linux/ceph/striper.h>
  23
  24 /*
  25  * Ceph address space ops.
  26  *
  27  * There are a few funny things going on here.
  28  *
  29  * The page->private field is used to reference a struct
  30  * ceph_snap_context for _every_ dirty page.  This indicates which
  31  * snapshot the page was logically dirtied in, and thus which snap
  32  * context needs to be associated with the osd write during writeback.
  33  *
  34  * Similarly, struct ceph_inode_info maintains a set of counters to
  35  * count dirty pages on the inode.  In the absence of snapshots,
  36  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  37  *
  38  * When a snapshot is taken (that is, when the client receives
  39  * notification that a snapshot was taken), each inode with caps and
  40  * with dirty pages (dirty pages implies there is a cap) gets a new
  41  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  42  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  43  * moved to capsnap->dirty. (Unless a sync write is currently in
  44  * progress.  In that case, the capsnap is said to be "pending", new
  45  * writes cannot start, and the capsnap isn't "finalized" until the
  46  * write completes (or fails) and a final size/mtime for the inode for
  47  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  48  *
  49  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  50  * we look for the first capsnap in i_cap_snaps and write out pages in
  51  * that snap context _only_.  Then we move on to the next capsnap,
  52  * eventually reaching the "live" or "head" context (i.e., pages that
  53  * are not yet snapped) and are writing the most recently dirtied
  54  * pages.
  55  *
  56  * Invalidate and so forth must take care to ensure the dirty page
  57  * accounting is preserved.
  58  */
  59
  60 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  61 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  62         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  63          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  64
  65 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  66                                         struct folio *folio, void **_fsdata);
  67
  68 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  69 {
  70         if (PagePrivate(page))
  71                 return (void *)page->private;
  72         return NULL;
  73 }
  74
  75 /*
  76  * Dirty a page.  Optimistically adjust accounting, on the assumption
  77  * that we won't race with invalidate.  If we do, readjust.
  78  */
  79 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
  80 {
  81         struct inode *inode;
  82         struct ceph_inode_info *ci;
  83         struct ceph_snap_context *snapc;
  84
  85         if (folio_test_dirty(folio)) {
  86                 dout("%p dirty_folio %p idx %lu -- already dirty\n",
  87                      mapping->host, folio, folio->index);
  88                 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
  89                 return false;
  90         }
  91
  92         inode = mapping->host;
  93         ci = ceph_inode(inode);
  94
  95         /* dirty the head */
  96         spin_lock(&ci->i_ceph_lock);
  97         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  98         if (__ceph_have_pending_cap_snap(ci)) {
  99                 struct ceph_cap_snap *capsnap =
 100                                 list_last_entry(&ci->i_cap_snaps,
 101                                                 struct ceph_cap_snap,
 102                                                 ci_item);
 103                 snapc = ceph_get_snap_context(capsnap->context);
 104                 capsnap->dirty_pages++;
 105         } else {
 106                 BUG_ON(!ci->i_head_snapc);
 107                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 108                 ++ci->i_wrbuffer_ref_head;
 109         }
 110         if (ci->i_wrbuffer_ref == 0)
 111                 ihold(inode);
 112         ++ci->i_wrbuffer_ref;
 113         dout("%p dirty_folio %p idx %lu head %d/%d -> %d/%d "
 114              "snapc %p seq %lld (%d snaps)\n",
 115              mapping->host, folio, folio->index,
 116              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 117              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 118              snapc, snapc->seq, snapc->num_snaps);
 119         spin_unlock(&ci->i_ceph_lock);
 120
 121         /*
 122          * Reference snap context in folio->private.  Also set
 123          * PagePrivate so that we get invalidate_folio callback.
 124          */
 125         VM_BUG_ON_FOLIO(folio_test_private(folio), folio);
 126         folio_attach_private(folio, snapc);
 127
 128         return ceph_fscache_dirty_folio(mapping, folio);
 129 }
 130
 131 /*
 132  * If we are truncating the full folio (i.e. offset == 0), adjust the
 133  * dirty folio counters appropriately.  Only called if there is private
 134  * data on the folio.
 135  */
 136 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 137                                 size_t length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc;
 142
 143         inode = folio->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != folio_size(folio)) {
 147                 dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
 148                      inode, folio->index, offset, length);
 149                 return;
 150         }
 151
 152         WARN_ON(!folio_test_locked(folio));
 153         if (folio_test_private(folio)) {
 154                 dout("%p invalidate_folio idx %lu full dirty page\n",
 155                      inode, folio->index);
 156
 157                 snapc = folio_detach_private(folio);
 158                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159                 ceph_put_snap_context(snapc);
 160         }
 161
 162         folio_wait_fscache(folio);
 163 }
 164
 165 static int ceph_releasepage(struct page *page, gfp_t gfp)
 166 {
 167         struct inode *inode = page->mapping->host;
 168
 169         dout("%llx:%llx releasepage %p idx %lu (%sdirty)\n",
 170              ceph_vinop(inode), page,
 171              page->index, PageDirty(page) ? "" : "not ");
 172
 173         if (PagePrivate(page))
 174                 return 0;
 175
 176         if (PageFsCache(page)) {
 177                 if (current_is_kswapd() || !(gfp & __GFP_FS))
 178                         return 0;
 179                 wait_on_page_fscache(page);
 180         }
 181         ceph_fscache_note_page_release(inode);
 182         return 1;
 183 }
 184
 185 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 186 {
 187         struct inode *inode = rreq->inode;
 188         struct ceph_inode_info *ci = ceph_inode(inode);
 189         struct ceph_file_layout *lo = &ci->i_layout;
 190         u32 blockoff;
 191         u64 blockno;
 192
 193         /* Expand the start downward */
 194         blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 195         rreq->start = blockno * lo->stripe_unit;
 196         rreq->len += blockoff;
 197
 198         /* Now, round up the length to the next block */
 199         rreq->len = roundup(rreq->len, lo->stripe_unit);
 200 }
 201
 202 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
 203 {
 204         struct inode *inode = subreq->rreq->inode;
 205         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 206         struct ceph_inode_info *ci = ceph_inode(inode);
 207         u64 objno, objoff;
 208         u32 xlen;
 209
 210         /* Truncate the extent at the end of the current block */
 211         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 212                                       &objno, &objoff, &xlen);
 213         subreq->len = min(xlen, fsc->mount_options->rsize);
 214         return true;
 215 }
 216
 217 static void finish_netfs_read(struct ceph_osd_request *req)
 218 {
 219         struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 220         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 221         struct netfs_io_subrequest *subreq = req->r_priv;
 222         int num_pages;
 223         int err = req->r_result;
 224
 225         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 226                                  req->r_end_latency, osd_data->length, err);
 227
 228         dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 229              subreq->len, i_size_read(req->r_inode));
 230
 231         /* no object means success but no data */
 232         if (err == -ENOENT)
 233                 err = 0;
 234         else if (err == -EBLOCKLISTED)
 235                 fsc->blocklisted = true;
 236
 237         if (err >= 0 && err < subreq->len)
 238                 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 239
 240         netfs_subreq_terminated(subreq, err, true);
 241
 242         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 243         ceph_put_page_vector(osd_data->pages, num_pages, false);
 244         iput(req->r_inode);
 245 }
 246
 247 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 248 {
 249         struct netfs_io_request *rreq = subreq->rreq;
 250         struct inode *inode = rreq->inode;
 251         struct ceph_mds_reply_info_parsed *rinfo;
 252         struct ceph_mds_reply_info_in *iinfo;
 253         struct ceph_mds_request *req;
 254         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 255         struct ceph_inode_info *ci = ceph_inode(inode);
 256         struct iov_iter iter;
 257         ssize_t err = 0;
 258         size_t len;
 259
 260         __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 261         __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 262
 263         if (subreq->start >= inode->i_size)
 264                 goto out;
 265
 266         /* We need to fetch the inline data. */
 267         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, USE_ANY_MDS);
 268         if (IS_ERR(req)) {
 269                 err = PTR_ERR(req);
 270                 goto out;
 271         }
 272         req->r_ino1 = ci->i_vino;
 273         req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 274         req->r_num_caps = 2;
 275
 276         err = ceph_mdsc_do_request(mdsc, NULL, req);
 277         if (err < 0)
 278                 goto out;
 279
 280         rinfo = &req->r_reply_info;
 281         iinfo = &rinfo->targeti;
 282         if (iinfo->inline_version == CEPH_INLINE_NONE) {
 283                 /* The data got uninlined */
 284                 ceph_mdsc_put_request(req);
 285                 return false;
 286         }
 287
 288         len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
 289         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 290         err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
 291         if (err == 0)
 292                 err = -EFAULT;
 293
 294         ceph_mdsc_put_request(req);
 295 out:
 296         netfs_subreq_terminated(subreq, err, false);
 297         return true;
 298 }
 299
 300 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 301 {
 302         struct netfs_io_request *rreq = subreq->rreq;
 303         struct inode *inode = rreq->inode;
 304         struct ceph_inode_info *ci = ceph_inode(inode);
 305         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 306         struct ceph_osd_request *req;
 307         struct ceph_vino vino = ceph_vino(inode);
 308         struct iov_iter iter;
 309         struct page **pages;
 310         size_t page_off;
 311         int err = 0;
 312         u64 len = subreq->len;
 313
 314         if (ci->i_inline_version != CEPH_INLINE_NONE &&
 315             ceph_netfs_issue_op_inline(subreq))
 316                 return;
 317
 318         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 319                         0, 1, CEPH_OSD_OP_READ,
 320                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 321                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 322         if (IS_ERR(req)) {
 323                 err = PTR_ERR(req);
 324                 req = NULL;
 325                 goto out;
 326         }
 327
 328         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 329         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 330         err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 331         if (err < 0) {
 332                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 333                 goto out;
 334         }
 335
 336         /* should always give us a page-aligned read */
 337         WARN_ON_ONCE(page_off);
 338         len = err;
 339
 340         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 341         req->r_callback = finish_netfs_read;
 342         req->r_priv = subreq;
 343         req->r_inode = inode;
 344         ihold(inode);
 345
 346         err = ceph_osdc_start_request(req->r_osdc, req, false);
 347         if (err)
 348                 iput(inode);
 349 out:
 350         ceph_osdc_put_request(req);
 351         if (err)
 352                 netfs_subreq_terminated(subreq, err, false);
 353         dout("%s: result %d\n", __func__, err);
 354 }
 355
 356 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 357 {
 358         struct inode *inode = rreq->inode;
 359         int got = 0, want = CEPH_CAP_FILE_CACHE;
 360         int ret = 0;
 361
 362         if (rreq->origin != NETFS_READAHEAD)
 363                 return 0;
 364
 365         if (file) {
 366                 struct ceph_rw_context *rw_ctx;
 367                 struct ceph_file_info *fi = file->private_data;
 368
 369                 rw_ctx = ceph_find_rw_context(fi);
 370                 if (rw_ctx)
 371                         return 0;
 372         }
 373
 374         /*
 375          * readahead callers do not necessarily hold Fcb caps
 376          * (e.g. fadvise, madvise).
 377          */
 378         ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 379         if (ret < 0) {
 380                 dout("start_read %p, error getting cap\n", inode);
 381                 return ret;
 382         }
 383
 384         if (!(got & want)) {
 385                 dout("start_read %p, no cache cap\n", inode);
 386                 return -EACCES;
 387         }
 388         if (ret == 0)
 389                 return -EACCES;
 390
 391         rreq->netfs_priv = (void *)(uintptr_t)got;
 392         return 0;
 393 }
 394
 395 static void ceph_readahead_cleanup(struct address_space *mapping, void *priv)
 396 {
 397         struct inode *inode = mapping->host;
 398         struct ceph_inode_info *ci = ceph_inode(inode);
 399         int got = (uintptr_t)priv;
 400
 401         if (got)
 402                 ceph_put_cap_refs(ci, got);
 403 }
 404
 405 const struct netfs_request_ops ceph_netfs_ops = {
 406         .init_request           = ceph_init_request,
 407         .begin_cache_operation  = ceph_begin_cache_operation,
 408         .issue_read             = ceph_netfs_issue_read,
 409         .expand_readahead       = ceph_netfs_expand_readahead,
 410         .clamp_length           = ceph_netfs_clamp_length,
 411         .check_write_begin      = ceph_netfs_check_write_begin,
 412         .cleanup                = ceph_readahead_cleanup,
 413 };
 414
 415 #ifdef CONFIG_CEPH_FSCACHE
 416 static void ceph_set_page_fscache(struct page *page)
 417 {
 418         set_page_fscache(page);
 419 }
 420
 421 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 422 {
 423         struct inode *inode = priv;
 424
 425         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 426                 ceph_fscache_invalidate(inode, false);
 427 }
 428
 429 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 430 {
 431         struct ceph_inode_info *ci = ceph_inode(inode);
 432         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 433
 434         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 435                                ceph_fscache_write_terminated, inode, caching);
 436 }
 437 #else
 438 static inline void ceph_set_page_fscache(struct page *page)
 439 {
 440 }
 441
 442 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 443 {
 444 }
 445 #endif /* CONFIG_CEPH_FSCACHE */
 446
 447 struct ceph_writeback_ctl
 448 {
 449         loff_t i_size;
 450         u64 truncate_size;
 451         u32 truncate_seq;
 452         bool size_stable;
 453         bool head_snapc;
 454 };
 455
 456 /*
 457  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 458  * only snap context we are allowed to write back.
 459  */
 460 static struct ceph_snap_context *
 461 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 462                    struct ceph_snap_context *page_snapc)
 463 {
 464         struct ceph_inode_info *ci = ceph_inode(inode);
 465         struct ceph_snap_context *snapc = NULL;
 466         struct ceph_cap_snap *capsnap = NULL;
 467
 468         spin_lock(&ci->i_ceph_lock);
 469         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 470                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 471                      capsnap->context, capsnap->dirty_pages);
 472                 if (!capsnap->dirty_pages)
 473                         continue;
 474
 475                 /* get i_size, truncate_{seq,size} for page_snapc? */
 476                 if (snapc && capsnap->context != page_snapc)
 477                         continue;
 478
 479                 if (ctl) {
 480                         if (capsnap->writing) {
 481                                 ctl->i_size = i_size_read(inode);
 482                                 ctl->size_stable = false;
 483                         } else {
 484                                 ctl->i_size = capsnap->size;
 485                                 ctl->size_stable = true;
 486                         }
 487                         ctl->truncate_size = capsnap->truncate_size;
 488                         ctl->truncate_seq = capsnap->truncate_seq;
 489                         ctl->head_snapc = false;
 490                 }
 491
 492                 if (snapc)
 493                         break;
 494
 495                 snapc = ceph_get_snap_context(capsnap->context);
 496                 if (!page_snapc ||
 497                     page_snapc == snapc ||
 498                     page_snapc->seq > snapc->seq)
 499                         break;
 500         }
 501         if (!snapc && ci->i_wrbuffer_ref_head) {
 502                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 503                 dout(" head snapc %p has %d dirty pages\n",
 504                      snapc, ci->i_wrbuffer_ref_head);
 505                 if (ctl) {
 506                         ctl->i_size = i_size_read(inode);
 507                         ctl->truncate_size = ci->i_truncate_size;
 508                         ctl->truncate_seq = ci->i_truncate_seq;
 509                         ctl->size_stable = false;
 510                         ctl->head_snapc = true;
 511                 }
 512         }
 513         spin_unlock(&ci->i_ceph_lock);
 514         return snapc;
 515 }
 516
 517 static u64 get_writepages_data_length(struct inode *inode,
 518                                       struct page *page, u64 start)
 519 {
 520         struct ceph_inode_info *ci = ceph_inode(inode);
 521         struct ceph_snap_context *snapc = page_snap_context(page);
 522         struct ceph_cap_snap *capsnap = NULL;
 523         u64 end = i_size_read(inode);
 524
 525         if (snapc != ci->i_head_snapc) {
 526                 bool found = false;
 527                 spin_lock(&ci->i_ceph_lock);
 528                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 529                         if (capsnap->context == snapc) {
 530                                 if (!capsnap->writing)
 531                                         end = capsnap->size;
 532                                 found = true;
 533                                 break;
 534                         }
 535                 }
 536                 spin_unlock(&ci->i_ceph_lock);
 537                 WARN_ON(!found);
 538         }
 539         if (end > page_offset(page) + thp_size(page))
 540                 end = page_offset(page) + thp_size(page);
 541         return end > start ? end - start : 0;
 542 }
 543
 544 /*
 545  * Write a single page, but leave the page locked.
 546  *
 547  * If we get a write error, mark the mapping for error, but still adjust the
 548  * dirty page accounting (i.e., page is no longer dirty).
 549  */
 550 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 551 {
 552         struct folio *folio = page_folio(page);
 553         struct inode *inode = page->mapping->host;
 554         struct ceph_inode_info *ci = ceph_inode(inode);
 555         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 556         struct ceph_snap_context *snapc, *oldest;
 557         loff_t page_off = page_offset(page);
 558         int err;
 559         loff_t len = thp_size(page);
 560         struct ceph_writeback_ctl ceph_wbc;
 561         struct ceph_osd_client *osdc = &fsc->client->osdc;
 562         struct ceph_osd_request *req;
 563         bool caching = ceph_is_cache_enabled(inode);
 564
 565         dout("writepage %p idx %lu\n", page, page->index);
 566
 567         /* verify this is a writeable snap context */
 568         snapc = page_snap_context(page);
 569         if (!snapc) {
 570                 dout("writepage %p page %p not dirty?\n", inode, page);
 571                 return 0;
 572         }
 573         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 574         if (snapc->seq > oldest->seq) {
 575                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 576                      inode, page, snapc);
 577                 /* we should only noop if called by kswapd */
 578                 WARN_ON(!(current->flags & PF_MEMALLOC));
 579                 ceph_put_snap_context(oldest);
 580                 redirty_page_for_writepage(wbc, page);
 581                 return 0;
 582         }
 583         ceph_put_snap_context(oldest);
 584
 585         /* is this a partial page at end of file? */
 586         if (page_off >= ceph_wbc.i_size) {
 587                 dout("folio at %lu beyond eof %llu\n", folio->index,
 588                                 ceph_wbc.i_size);
 589                 folio_invalidate(folio, 0, folio_size(folio));
 590                 return 0;
 591         }
 592
 593         if (ceph_wbc.i_size < page_off + len)
 594                 len = ceph_wbc.i_size - page_off;
 595
 596         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 597              inode, page, page->index, page_off, len, snapc, snapc->seq);
 598
 599         if (atomic_long_inc_return(&fsc->writeback_count) >
 600             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 601                 fsc->write_congested = true;
 602
 603         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 604                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 605                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 606                                     true);
 607         if (IS_ERR(req))
 608                 return PTR_ERR(req);
 609
 610         set_page_writeback(page);
 611         if (caching)
 612                 ceph_set_page_fscache(page);
 613         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 614
 615         /* it may be a short write due to an object boundary */
 616         WARN_ON_ONCE(len > thp_size(page));
 617         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 618         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 619
 620         req->r_mtime = inode->i_mtime;
 621         err = ceph_osdc_start_request(osdc, req, true);
 622         if (!err)
 623                 err = ceph_osdc_wait_request(osdc, req);
 624
 625         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 626                                   req->r_end_latency, len, err);
 627
 628         ceph_osdc_put_request(req);
 629         if (err == 0)
 630                 err = len;
 631
 632         if (err < 0) {
 633                 struct writeback_control tmp_wbc;
 634                 if (!wbc)
 635                         wbc = &tmp_wbc;
 636                 if (err == -ERESTARTSYS) {
 637                         /* killed by SIGKILL */
 638                         dout("writepage interrupted page %p\n", page);
 639                         redirty_page_for_writepage(wbc, page);
 640                         end_page_writeback(page);
 641                         return err;
 642                 }
 643                 if (err == -EBLOCKLISTED)
 644                         fsc->blocklisted = true;
 645                 dout("writepage setting page/mapping error %d %p\n",
 646                      err, page);
 647                 mapping_set_error(&inode->i_data, err);
 648                 wbc->pages_skipped++;
 649         } else {
 650                 dout("writepage cleaned page %p\n", page);
 651                 err = 0;  /* vfs expects us to return 0 */
 652         }
 653         oldest = detach_page_private(page);
 654         WARN_ON_ONCE(oldest != snapc);
 655         end_page_writeback(page);
 656         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 657         ceph_put_snap_context(snapc);  /* page's reference */
 658
 659         if (atomic_long_dec_return(&fsc->writeback_count) <
 660             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 661                 fsc->write_congested = false;
 662
 663         return err;
 664 }
 665
 666 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 667 {
 668         int err;
 669         struct inode *inode = page->mapping->host;
 670         BUG_ON(!inode);
 671         ihold(inode);
 672
 673         if (wbc->sync_mode == WB_SYNC_NONE &&
 674             ceph_inode_to_client(inode)->write_congested)
 675                 return AOP_WRITEPAGE_ACTIVATE;
 676
 677         wait_on_page_fscache(page);
 678
 679         err = writepage_nounlock(page, wbc);
 680         if (err == -ERESTARTSYS) {
 681                 /* direct memory reclaimer was killed by SIGKILL. return 0
 682                  * to prevent caller from setting mapping/page error */
 683                 err = 0;
 684         }
 685         unlock_page(page);
 686         iput(inode);
 687         return err;
 688 }
 689
 690 /*
 691  * async writeback completion handler.
 692  *
 693  * If we get an error, set the mapping error bit, but not the individual
 694  * page error bits.
 695  */
 696 static void writepages_finish(struct ceph_osd_request *req)
 697 {
 698         struct inode *inode = req->r_inode;
 699         struct ceph_inode_info *ci = ceph_inode(inode);
 700         struct ceph_osd_data *osd_data;
 701         struct page *page;
 702         int num_pages, total_pages = 0;
 703         int i, j;
 704         int rc = req->r_result;
 705         struct ceph_snap_context *snapc = req->r_snapc;
 706         struct address_space *mapping = inode->i_mapping;
 707         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 708         unsigned int len = 0;
 709         bool remove_page;
 710
 711         dout("writepages_finish %p rc %d\n", inode, rc);
 712         if (rc < 0) {
 713                 mapping_set_error(mapping, rc);
 714                 ceph_set_error_write(ci);
 715                 if (rc == -EBLOCKLISTED)
 716                         fsc->blocklisted = true;
 717         } else {
 718                 ceph_clear_error_write(ci);
 719         }
 720
 721         /*
 722          * We lost the cache cap, need to truncate the page before
 723          * it is unlocked, otherwise we'd truncate it later in the
 724          * page truncation thread, possibly losing some data that
 725          * raced its way in
 726          */
 727         remove_page = !(ceph_caps_issued(ci) &
 728                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 729
 730         /* clean all pages */
 731         for (i = 0; i < req->r_num_ops; i++) {
 732                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 733                         pr_warn("%s incorrect op %d req %p index %d tid %llu\n",
 734                                 __func__, req->r_ops[i].op, req, i, req->r_tid);
 735                         break;
 736                 }
 737
 738                 osd_data = osd_req_op_extent_osd_data(req, i);
 739                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 740                 len += osd_data->length;
 741                 num_pages = calc_pages_for((u64)osd_data->alignment,
 742                                            (u64)osd_data->length);
 743                 total_pages += num_pages;
 744                 for (j = 0; j < num_pages; j++) {
 745                         page = osd_data->pages[j];
 746                         BUG_ON(!page);
 747                         WARN_ON(!PageUptodate(page));
 748
 749                         if (atomic_long_dec_return(&fsc->writeback_count) <
 750                              CONGESTION_OFF_THRESH(
 751                                         fsc->mount_options->congestion_kb))
 752                                 fsc->write_congested = false;
 753
 754                         ceph_put_snap_context(detach_page_private(page));
 755                         end_page_writeback(page);
 756                         dout("unlocking %p\n", page);
 757
 758                         if (remove_page)
 759                                 generic_error_remove_page(inode->i_mapping,
 760                                                           page);
 761
 762                         unlock_page(page);
 763                 }
 764                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 765                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 766
 767                 release_pages(osd_data->pages, num_pages);
 768         }
 769
 770         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 771                                   req->r_end_latency, len, rc);
 772
 773         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 774
 775         osd_data = osd_req_op_extent_osd_data(req, 0);
 776         if (osd_data->pages_from_pool)
 777                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 778         else
 779                 kfree(osd_data->pages);
 780         ceph_osdc_put_request(req);
 781 }
 782
 783 /*
 784  * initiate async writeback
 785  */
 786 static int ceph_writepages_start(struct address_space *mapping,
 787                                  struct writeback_control *wbc)
 788 {
 789         struct inode *inode = mapping->host;
 790         struct ceph_inode_info *ci = ceph_inode(inode);
 791         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 792         struct ceph_vino vino = ceph_vino(inode);
 793         pgoff_t index, start_index, end = -1;
 794         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 795         struct pagevec pvec;
 796         int rc = 0;
 797         unsigned int wsize = i_blocksize(inode);
 798         struct ceph_osd_request *req = NULL;
 799         struct ceph_writeback_ctl ceph_wbc;
 800         bool should_loop, range_whole = false;
 801         bool done = false;
 802         bool caching = ceph_is_cache_enabled(inode);
 803
 804         if (wbc->sync_mode == WB_SYNC_NONE &&
 805             fsc->write_congested)
 806                 return 0;
 807
 808         dout("writepages_start %p (mode=%s)\n", inode,
 809              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 810              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 811
 812         if (ceph_inode_is_shutdown(inode)) {
 813                 if (ci->i_wrbuffer_ref > 0) {
 814                         pr_warn_ratelimited(
 815                                 "writepage_start %p %lld forced umount\n",
 816                                 inode, ceph_ino(inode));
 817                 }
 818                 mapping_set_error(mapping, -EIO);
 819                 return -EIO; /* we're in a forced umount, don't write! */
 820         }
 821         if (fsc->mount_options->wsize < wsize)
 822                 wsize = fsc->mount_options->wsize;
 823
 824         pagevec_init(&pvec);
 825
 826         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 827         index = start_index;
 828
 829 retry:
 830         /* find oldest snap context with dirty data */
 831         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 832         if (!snapc) {
 833                 /* hmm, why does writepages get called when there
 834                    is no dirty data? */
 835                 dout(" no snap context with dirty data?\n");
 836                 goto out;
 837         }
 838         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 839              snapc, snapc->seq, snapc->num_snaps);
 840
 841         should_loop = false;
 842         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 843                 /* where to start/end? */
 844                 if (wbc->range_cyclic) {
 845                         index = start_index;
 846                         end = -1;
 847                         if (index > 0)
 848                                 should_loop = true;
 849                         dout(" cyclic, start at %lu\n", index);
 850                 } else {
 851                         index = wbc->range_start >> PAGE_SHIFT;
 852                         end = wbc->range_end >> PAGE_SHIFT;
 853                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 854                                 range_whole = true;
 855                         dout(" not cyclic, %lu to %lu\n", index, end);
 856                 }
 857         } else if (!ceph_wbc.head_snapc) {
 858                 /* Do not respect wbc->range_{start,end}. Dirty pages
 859                  * in that range can be associated with newer snapc.
 860                  * They are not writeable until we write all dirty pages
 861                  * associated with 'snapc' get written */
 862                 if (index > 0)
 863                         should_loop = true;
 864                 dout(" non-head snapc, range whole\n");
 865         }
 866
 867         ceph_put_snap_context(last_snapc);
 868         last_snapc = snapc;
 869
 870         while (!done && index <= end) {
 871                 int num_ops = 0, op_idx;
 872                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 873                 struct page **pages = NULL, **data_pages;
 874                 struct page *page;
 875                 pgoff_t strip_unit_end = 0;
 876                 u64 offset = 0, len = 0;
 877                 bool from_pool = false;
 878
 879                 max_pages = wsize >> PAGE_SHIFT;
 880
 881 get_more_pages:
 882                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 883                                                 end, PAGECACHE_TAG_DIRTY);
 884                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 885                 if (!pvec_pages && !locked_pages)
 886                         break;
 887                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 888                         page = pvec.pages[i];
 889                         dout("? %p idx %lu\n", page, page->index);
 890                         if (locked_pages == 0)
 891                                 lock_page(page);  /* first page */
 892                         else if (!trylock_page(page))
 893                                 break;
 894
 895                         /* only dirty pages, or our accounting breaks */
 896                         if (unlikely(!PageDirty(page)) ||
 897                             unlikely(page->mapping != mapping)) {
 898                                 dout("!dirty or !mapping %p\n", page);
 899                                 unlock_page(page);
 900                                 continue;
 901                         }
 902                         /* only if matching snap context */
 903                         pgsnapc = page_snap_context(page);
 904                         if (pgsnapc != snapc) {
 905                                 dout("page snapc %p %lld != oldest %p %lld\n",
 906                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 907                                 if (!should_loop &&
 908                                     !ceph_wbc.head_snapc &&
 909                                     wbc->sync_mode != WB_SYNC_NONE)
 910                                         should_loop = true;
 911                                 unlock_page(page);
 912                                 continue;
 913                         }
 914                         if (page_offset(page) >= ceph_wbc.i_size) {
 915                                 struct folio *folio = page_folio(page);
 916
 917                                 dout("folio at %lu beyond eof %llu\n",
 918                                      folio->index, ceph_wbc.i_size);
 919                                 if ((ceph_wbc.size_stable ||
 920                                     folio_pos(folio) >= i_size_read(inode)) &&
 921                                     folio_clear_dirty_for_io(folio))
 922                                         folio_invalidate(folio, 0,
 923                                                         folio_size(folio));
 924                                 folio_unlock(folio);
 925                                 continue;
 926                         }
 927                         if (strip_unit_end && (page->index > strip_unit_end)) {
 928                                 dout("end of strip unit %p\n", page);
 929                                 unlock_page(page);
 930                                 break;
 931                         }
 932                         if (PageWriteback(page) || PageFsCache(page)) {
 933                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 934                                         dout("%p under writeback\n", page);
 935                                         unlock_page(page);
 936                                         continue;
 937                                 }
 938                                 dout("waiting on writeback %p\n", page);
 939                                 wait_on_page_writeback(page);
 940                                 wait_on_page_fscache(page);
 941                         }
 942
 943                         if (!clear_page_dirty_for_io(page)) {
 944                                 dout("%p !clear_page_dirty_for_io\n", page);
 945                                 unlock_page(page);
 946                                 continue;
 947                         }
 948
 949                         /*
 950                          * We have something to write.  If this is
 951                          * the first locked page this time through,
 952                          * calculate max possinle write size and
 953                          * allocate a page array
 954                          */
 955                         if (locked_pages == 0) {
 956                                 u64 objnum;
 957                                 u64 objoff;
 958                                 u32 xlen;
 959
 960                                 /* prepare async write request */
 961                                 offset = (u64)page_offset(page);
 962                                 ceph_calc_file_object_mapping(&ci->i_layout,
 963                                                               offset, wsize,
 964                                                               &objnum, &objoff,
 965                                                               &xlen);
 966                                 len = xlen;
 967
 968                                 num_ops = 1;
 969                                 strip_unit_end = page->index +
 970                                         ((len - 1) >> PAGE_SHIFT);
 971
 972                                 BUG_ON(pages);
 973                                 max_pages = calc_pages_for(0, (u64)len);
 974                                 pages = kmalloc_array(max_pages,
 975                                                       sizeof(*pages),
 976                                                       GFP_NOFS);
 977                                 if (!pages) {
 978                                         from_pool = true;
 979                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 980                                         BUG_ON(!pages);
 981                                 }
 982
 983                                 len = 0;
 984                         } else if (page->index !=
 985                                    (offset + len) >> PAGE_SHIFT) {
 986                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 987                                                              CEPH_OSD_MAX_OPS)) {
 988                                         redirty_page_for_writepage(wbc, page);
 989                                         unlock_page(page);
 990                                         break;
 991                                 }
 992
 993                                 num_ops++;
 994                                 offset = (u64)page_offset(page);
 995                                 len = 0;
 996                         }
 997
 998                         /* note position of first page in pvec */
 999                         dout("%p will write page %p idx %lu\n",
1000                              inode, page, page->index);
1001
1002                         if (atomic_long_inc_return(&fsc->writeback_count) >
1003                             CONGESTION_ON_THRESH(
1004                                     fsc->mount_options->congestion_kb))
1005                                 fsc->write_congested = true;
1006
1007                         pages[locked_pages++] = page;
1008                         pvec.pages[i] = NULL;
1009
1010                         len += thp_size(page);
1011                 }
1012
1013                 /* did we get anything? */
1014                 if (!locked_pages)
1015                         goto release_pvec_pages;
1016                 if (i) {
1017                         unsigned j, n = 0;
1018                         /* shift unused page to beginning of pvec */
1019                         for (j = 0; j < pvec_pages; j++) {
1020                                 if (!pvec.pages[j])
1021                                         continue;
1022                                 if (n < j)
1023                                         pvec.pages[n] = pvec.pages[j];
1024                                 n++;
1025                         }
1026                         pvec.nr = n;
1027
1028                         if (pvec_pages && i == pvec_pages &&
1029                             locked_pages < max_pages) {
1030                                 dout("reached end pvec, trying for more\n");
1031                                 pagevec_release(&pvec);
1032                                 goto get_more_pages;
1033                         }
1034                 }
1035
1036 new_request:
1037                 offset = page_offset(pages[0]);
1038                 len = wsize;
1039
1040                 req = ceph_osdc_new_request(&fsc->client->osdc,
1041                                         &ci->i_layout, vino,
1042                                         offset, &len, 0, num_ops,
1043                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1044                                         snapc, ceph_wbc.truncate_seq,
1045                                         ceph_wbc.truncate_size, false);
1046                 if (IS_ERR(req)) {
1047                         req = ceph_osdc_new_request(&fsc->client->osdc,
1048                                                 &ci->i_layout, vino,
1049                                                 offset, &len, 0,
1050                                                 min(num_ops,
1051                                                     CEPH_OSD_SLAB_OPS),
1052                                                 CEPH_OSD_OP_WRITE,
1053                                                 CEPH_OSD_FLAG_WRITE,
1054                                                 snapc, ceph_wbc.truncate_seq,
1055                                                 ceph_wbc.truncate_size, true);
1056                         BUG_ON(IS_ERR(req));
1057                 }
1058                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1059                              thp_size(page) - offset);
1060
1061                 req->r_callback = writepages_finish;
1062                 req->r_inode = inode;
1063
1064                 /* Format the osd request message and submit the write */
1065                 len = 0;
1066                 data_pages = pages;
1067                 op_idx = 0;
1068                 for (i = 0; i < locked_pages; i++) {
1069                         u64 cur_offset = page_offset(pages[i]);
1070                         /*
1071                          * Discontinuity in page range? Ceph can handle that by just passing
1072                          * multiple extents in the write op.
1073                          */
1074                         if (offset + len != cur_offset) {
1075                                 /* If it's full, stop here */
1076                                 if (op_idx + 1 == req->r_num_ops)
1077                                         break;
1078
1079                                 /* Kick off an fscache write with what we have so far. */
1080                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1081
1082                                 /* Start a new extent */
1083                                 osd_req_op_extent_dup_last(req, op_idx,
1084                                                            cur_offset - offset);
1085                                 dout("writepages got pages at %llu~%llu\n",
1086                                      offset, len);
1087                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1088                                                         data_pages, len, 0,
1089                                                         from_pool, false);
1090                                 osd_req_op_extent_update(req, op_idx, len);
1091
1092                                 len = 0;
1093                                 offset = cur_offset;
1094                                 data_pages = pages + i;
1095                                 op_idx++;
1096                         }
1097
1098                         set_page_writeback(pages[i]);
1099                         if (caching)
1100                                 ceph_set_page_fscache(pages[i]);
1101                         len += thp_size(page);
1102                 }
1103                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1104
1105                 if (ceph_wbc.size_stable) {
1106                         len = min(len, ceph_wbc.i_size - offset);
1107                 } else if (i == locked_pages) {
1108                         /* writepages_finish() clears writeback pages
1109                          * according to the data length, so make sure
1110                          * data length covers all locked pages */
1111                         u64 min_len = len + 1 - thp_size(page);
1112                         len = get_writepages_data_length(inode, pages[i - 1],
1113                                                          offset);
1114                         len = max(len, min_len);
1115                 }
1116                 dout("writepages got pages at %llu~%llu\n", offset, len);
1117
1118                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1119                                                  0, from_pool, false);
1120                 osd_req_op_extent_update(req, op_idx, len);
1121
1122                 BUG_ON(op_idx + 1 != req->r_num_ops);
1123
1124                 from_pool = false;
1125                 if (i < locked_pages) {
1126                         BUG_ON(num_ops <= req->r_num_ops);
1127                         num_ops -= req->r_num_ops;
1128                         locked_pages -= i;
1129
1130                         /* allocate new pages array for next request */
1131                         data_pages = pages;
1132                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1133                                               GFP_NOFS);
1134                         if (!pages) {
1135                                 from_pool = true;
1136                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1137                                 BUG_ON(!pages);
1138                         }
1139                         memcpy(pages, data_pages + i,
1140                                locked_pages * sizeof(*pages));
1141                         memset(data_pages + i, 0,
1142                                locked_pages * sizeof(*pages));
1143                 } else {
1144                         BUG_ON(num_ops != req->r_num_ops);
1145                         index = pages[i - 1]->index + 1;
1146                         /* request message now owns the pages array */
1147                         pages = NULL;
1148                 }
1149
1150                 req->r_mtime = inode->i_mtime;
1151                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1152                 BUG_ON(rc);
1153                 req = NULL;
1154
1155                 wbc->nr_to_write -= i;
1156                 if (pages)
1157                         goto new_request;
1158
1159                 /*
1160                  * We stop writing back only if we are not doing
1161                  * integrity sync. In case of integrity sync we have to
1162                  * keep going until we have written all the pages
1163                  * we tagged for writeback prior to entering this loop.
1164                  */
1165                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1166                         done = true;
1167
1168 release_pvec_pages:
1169                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1170                      pvec.nr ? pvec.pages[0] : NULL);
1171                 pagevec_release(&pvec);
1172         }
1173
1174         if (should_loop && !done) {
1175                 /* more to do; loop back to beginning of file */
1176                 dout("writepages looping back to beginning of file\n");
1177                 end = start_index - 1; /* OK even when start_index == 0 */
1178
1179                 /* to write dirty pages associated with next snapc,
1180                  * we need to wait until current writes complete */
1181                 if (wbc->sync_mode != WB_SYNC_NONE &&
1182                     start_index == 0 && /* all dirty pages were checked */
1183                     !ceph_wbc.head_snapc) {
1184                         struct page *page;
1185                         unsigned i, nr;
1186                         index = 0;
1187                         while ((index <= end) &&
1188                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1189                                                 PAGECACHE_TAG_WRITEBACK))) {
1190                                 for (i = 0; i < nr; i++) {
1191                                         page = pvec.pages[i];
1192                                         if (page_snap_context(page) != snapc)
1193                                                 continue;
1194                                         wait_on_page_writeback(page);
1195                                 }
1196                                 pagevec_release(&pvec);
1197                                 cond_resched();
1198                         }
1199                 }
1200
1201                 start_index = 0;
1202                 index = 0;
1203                 goto retry;
1204         }
1205
1206         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1207                 mapping->writeback_index = index;
1208
1209 out:
1210         ceph_osdc_put_request(req);
1211         ceph_put_snap_context(last_snapc);
1212         dout("writepages dend - startone, rc = %d\n", rc);
1213         return rc;
1214 }
1215
1216
1217
1218 /*
1219  * See if a given @snapc is either writeable, or already written.
1220  */
1221 static int context_is_writeable_or_written(struct inode *inode,
1222                                            struct ceph_snap_context *snapc)
1223 {
1224         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1225         int ret = !oldest || snapc->seq <= oldest->seq;
1226
1227         ceph_put_snap_context(oldest);
1228         return ret;
1229 }
1230
1231 /**
1232  * ceph_find_incompatible - find an incompatible context and return it
1233  * @page: page being dirtied
1234  *
1235  * We are only allowed to write into/dirty a page if the page is
1236  * clean, or already dirty within the same snap context. Returns a
1237  * conflicting context if there is one, NULL if there isn't, or a
1238  * negative error code on other errors.
1239  *
1240  * Must be called with page lock held.
1241  */
1242 static struct ceph_snap_context *
1243 ceph_find_incompatible(struct page *page)
1244 {
1245         struct inode *inode = page->mapping->host;
1246         struct ceph_inode_info *ci = ceph_inode(inode);
1247
1248         if (ceph_inode_is_shutdown(inode)) {
1249                 dout(" page %p %llx:%llx is shutdown\n", page,
1250                      ceph_vinop(inode));
1251                 return ERR_PTR(-ESTALE);
1252         }
1253
1254         for (;;) {
1255                 struct ceph_snap_context *snapc, *oldest;
1256
1257                 wait_on_page_writeback(page);
1258
1259                 snapc = page_snap_context(page);
1260                 if (!snapc || snapc == ci->i_head_snapc)
1261                         break;
1262
1263                 /*
1264                  * this page is already dirty in another (older) snap
1265                  * context!  is it writeable now?
1266                  */
1267                 oldest = get_oldest_context(inode, NULL, NULL);
1268                 if (snapc->seq > oldest->seq) {
1269                         /* not writeable -- return it for the caller to deal with */
1270                         ceph_put_snap_context(oldest);
1271                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1272                         return ceph_get_snap_context(snapc);
1273                 }
1274                 ceph_put_snap_context(oldest);
1275
1276                 /* yay, writeable, do it now (without dropping page lock) */
1277                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1278                 if (clear_page_dirty_for_io(page)) {
1279                         int r = writepage_nounlock(page, NULL);
1280                         if (r < 0)
1281                                 return ERR_PTR(r);
1282                 }
1283         }
1284         return NULL;
1285 }
1286
1287 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1288                                         struct folio *folio, void **_fsdata)
1289 {
1290         struct inode *inode = file_inode(file);
1291         struct ceph_inode_info *ci = ceph_inode(inode);
1292         struct ceph_snap_context *snapc;
1293
1294         snapc = ceph_find_incompatible(folio_page(folio, 0));
1295         if (snapc) {
1296                 int r;
1297
1298                 folio_unlock(folio);
1299                 folio_put(folio);
1300                 if (IS_ERR(snapc))
1301                         return PTR_ERR(snapc);
1302
1303                 ceph_queue_writeback(inode);
1304                 r = wait_event_killable(ci->i_cap_wq,
1305                                         context_is_writeable_or_written(inode, snapc));
1306                 ceph_put_snap_context(snapc);
1307                 return r == 0 ? -EAGAIN : r;
1308         }
1309         return 0;
1310 }
1311
1312 /*
1313  * We are only allowed to write into/dirty the page if the page is
1314  * clean, or already dirty within the same snap context.
1315  */
1316 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1317                             loff_t pos, unsigned len, unsigned aop_flags,
1318                             struct page **pagep, void **fsdata)
1319 {
1320         struct inode *inode = file_inode(file);
1321         struct folio *folio = NULL;
1322         int r;
1323
1324         r = netfs_write_begin(file, inode->i_mapping, pos, len, 0, &folio, NULL);
1325         if (r == 0)
1326                 folio_wait_fscache(folio);
1327         if (r < 0) {
1328                 if (folio)
1329                         folio_put(folio);
1330         } else {
1331                 WARN_ON_ONCE(!folio_test_locked(folio));
1332                 *pagep = &folio->page;
1333         }
1334         return r;
1335 }
1336
1337 /*
1338  * we don't do anything in here that simple_write_end doesn't do
1339  * except adjust dirty page accounting
1340  */
1341 static int ceph_write_end(struct file *file, struct address_space *mapping,
1342                           loff_t pos, unsigned len, unsigned copied,
1343                           struct page *subpage, void *fsdata)
1344 {
1345         struct folio *folio = page_folio(subpage);
1346         struct inode *inode = file_inode(file);
1347         bool check_cap = false;
1348
1349         dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1350              inode, folio, (int)pos, (int)copied, (int)len);
1351
1352         if (!folio_test_uptodate(folio)) {
1353                 /* just return that nothing was copied on a short copy */
1354                 if (copied < len) {
1355                         copied = 0;
1356                         goto out;
1357                 }
1358                 folio_mark_uptodate(folio);
1359         }
1360
1361         /* did file size increase? */
1362         if (pos+copied > i_size_read(inode))
1363                 check_cap = ceph_inode_set_size(inode, pos+copied);
1364
1365         folio_mark_dirty(folio);
1366
1367 out:
1368         folio_unlock(folio);
1369         folio_put(folio);
1370
1371         if (check_cap)
1372                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1373
1374         return copied;
1375 }
1376
1377 const struct address_space_operations ceph_aops = {
1378         .readpage = netfs_readpage,
1379         .readahead = netfs_readahead,
1380         .writepage = ceph_writepage,
1381         .writepages = ceph_writepages_start,
1382         .write_begin = ceph_write_begin,
1383         .write_end = ceph_write_end,
1384         .dirty_folio = ceph_dirty_folio,
1385         .invalidate_folio = ceph_invalidate_folio,
1386         .releasepage = ceph_releasepage,
1387         .direct_IO = noop_direct_IO,
1388 };
1389
1390 static void ceph_block_sigs(sigset_t *oldset)
1391 {
1392         sigset_t mask;
1393         siginitsetinv(&mask, sigmask(SIGKILL));
1394         sigprocmask(SIG_BLOCK, &mask, oldset);
1395 }
1396
1397 static void ceph_restore_sigs(sigset_t *oldset)
1398 {
1399         sigprocmask(SIG_SETMASK, oldset, NULL);
1400 }
1401
1402 /*
1403  * vm ops
1404  */
1405 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1406 {
1407         struct vm_area_struct *vma = vmf->vma;
1408         struct inode *inode = file_inode(vma->vm_file);
1409         struct ceph_inode_info *ci = ceph_inode(inode);
1410         struct ceph_file_info *fi = vma->vm_file->private_data;
1411         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1412         int want, got, err;
1413         sigset_t oldset;
1414         vm_fault_t ret = VM_FAULT_SIGBUS;
1415
1416         if (ceph_inode_is_shutdown(inode))
1417                 return ret;
1418
1419         ceph_block_sigs(&oldset);
1420
1421         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1422              inode, ceph_vinop(inode), off);
1423         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1424                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1425         else
1426                 want = CEPH_CAP_FILE_CACHE;
1427
1428         got = 0;
1429         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1430         if (err < 0)
1431                 goto out_restore;
1432
1433         dout("filemap_fault %p %llu got cap refs on %s\n",
1434              inode, off, ceph_cap_string(got));
1435
1436         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1437             ci->i_inline_version == CEPH_INLINE_NONE) {
1438                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1439                 ceph_add_rw_context(fi, &rw_ctx);
1440                 ret = filemap_fault(vmf);
1441                 ceph_del_rw_context(fi, &rw_ctx);
1442                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1443                      inode, off, ceph_cap_string(got), ret);
1444         } else
1445                 err = -EAGAIN;
1446
1447         ceph_put_cap_refs(ci, got);
1448
1449         if (err != -EAGAIN)
1450                 goto out_restore;
1451
1452         /* read inline data */
1453         if (off >= PAGE_SIZE) {
1454                 /* does not support inline data > PAGE_SIZE */
1455                 ret = VM_FAULT_SIGBUS;
1456         } else {
1457                 struct address_space *mapping = inode->i_mapping;
1458                 struct page *page;
1459
1460                 filemap_invalidate_lock_shared(mapping);
1461                 page = find_or_create_page(mapping, 0,
1462                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1463                 if (!page) {
1464                         ret = VM_FAULT_OOM;
1465                         goto out_inline;
1466                 }
1467                 err = __ceph_do_getattr(inode, page,
1468                                          CEPH_STAT_CAP_INLINE_DATA, true);
1469                 if (err < 0 || off >= i_size_read(inode)) {
1470                         unlock_page(page);
1471                         put_page(page);
1472                         ret = vmf_error(err);
1473                         goto out_inline;
1474                 }
1475                 if (err < PAGE_SIZE)
1476                         zero_user_segment(page, err, PAGE_SIZE);
1477                 else
1478                         flush_dcache_page(page);
1479                 SetPageUptodate(page);
1480                 vmf->page = page;
1481                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1482 out_inline:
1483                 filemap_invalidate_unlock_shared(mapping);
1484                 dout("filemap_fault %p %llu read inline data ret %x\n",
1485                      inode, off, ret);
1486         }
1487 out_restore:
1488         ceph_restore_sigs(&oldset);
1489         if (err < 0)
1490                 ret = vmf_error(err);
1491
1492         return ret;
1493 }
1494
1495 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1496 {
1497         struct vm_area_struct *vma = vmf->vma;
1498         struct inode *inode = file_inode(vma->vm_file);
1499         struct ceph_inode_info *ci = ceph_inode(inode);
1500         struct ceph_file_info *fi = vma->vm_file->private_data;
1501         struct ceph_cap_flush *prealloc_cf;
1502         struct page *page = vmf->page;
1503         loff_t off = page_offset(page);
1504         loff_t size = i_size_read(inode);
1505         size_t len;
1506         int want, got, err;
1507         sigset_t oldset;
1508         vm_fault_t ret = VM_FAULT_SIGBUS;
1509
1510         if (ceph_inode_is_shutdown(inode))
1511                 return ret;
1512
1513         prealloc_cf = ceph_alloc_cap_flush();
1514         if (!prealloc_cf)
1515                 return VM_FAULT_OOM;
1516
1517         sb_start_pagefault(inode->i_sb);
1518         ceph_block_sigs(&oldset);
1519
1520         if (off + thp_size(page) <= size)
1521                 len = thp_size(page);
1522         else
1523                 len = offset_in_thp(page, size);
1524
1525         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1526              inode, ceph_vinop(inode), off, len, size);
1527         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1528                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1529         else
1530                 want = CEPH_CAP_FILE_BUFFER;
1531
1532         got = 0;
1533         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1534         if (err < 0)
1535                 goto out_free;
1536
1537         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1538              inode, off, len, ceph_cap_string(got));
1539
1540         /* Update time before taking page lock */
1541         file_update_time(vma->vm_file);
1542         inode_inc_iversion_raw(inode);
1543
1544         do {
1545                 struct ceph_snap_context *snapc;
1546
1547                 lock_page(page);
1548
1549                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1550                         unlock_page(page);
1551                         ret = VM_FAULT_NOPAGE;
1552                         break;
1553                 }
1554
1555                 snapc = ceph_find_incompatible(page);
1556                 if (!snapc) {
1557                         /* success.  we'll keep the page locked. */
1558                         set_page_dirty(page);
1559                         ret = VM_FAULT_LOCKED;
1560                         break;
1561                 }
1562
1563                 unlock_page(page);
1564
1565                 if (IS_ERR(snapc)) {
1566                         ret = VM_FAULT_SIGBUS;
1567                         break;
1568                 }
1569
1570                 ceph_queue_writeback(inode);
1571                 err = wait_event_killable(ci->i_cap_wq,
1572                                 context_is_writeable_or_written(inode, snapc));
1573                 ceph_put_snap_context(snapc);
1574         } while (err == 0);
1575
1576         if (ret == VM_FAULT_LOCKED) {
1577                 int dirty;
1578                 spin_lock(&ci->i_ceph_lock);
1579                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1580                                                &prealloc_cf);
1581                 spin_unlock(&ci->i_ceph_lock);
1582                 if (dirty)
1583                         __mark_inode_dirty(inode, dirty);
1584         }
1585
1586         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1587              inode, off, len, ceph_cap_string(got), ret);
1588         ceph_put_cap_refs_async(ci, got);
1589 out_free:
1590         ceph_restore_sigs(&oldset);
1591         sb_end_pagefault(inode->i_sb);
1592         ceph_free_cap_flush(prealloc_cf);
1593         if (err < 0)
1594                 ret = vmf_error(err);
1595         return ret;
1596 }
1597
1598 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1599                            char *data, size_t len)
1600 {
1601         struct address_space *mapping = inode->i_mapping;
1602         struct page *page;
1603
1604         if (locked_page) {
1605                 page = locked_page;
1606         } else {
1607                 if (i_size_read(inode) == 0)
1608                         return;
1609                 page = find_or_create_page(mapping, 0,
1610                                            mapping_gfp_constraint(mapping,
1611                                            ~__GFP_FS));
1612                 if (!page)
1613                         return;
1614                 if (PageUptodate(page)) {
1615                         unlock_page(page);
1616                         put_page(page);
1617                         return;
1618                 }
1619         }
1620
1621         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1622              inode, ceph_vinop(inode), len, locked_page);
1623
1624         if (len > 0) {
1625                 void *kaddr = kmap_atomic(page);
1626                 memcpy(kaddr, data, len);
1627                 kunmap_atomic(kaddr);
1628         }
1629
1630         if (page != locked_page) {
1631                 if (len < PAGE_SIZE)
1632                         zero_user_segment(page, len, PAGE_SIZE);
1633                 else
1634                         flush_dcache_page(page);
1635
1636                 SetPageUptodate(page);
1637                 unlock_page(page);
1638                 put_page(page);
1639         }
1640 }
1641
1642 int ceph_uninline_data(struct file *file)
1643 {
1644         struct inode *inode = file_inode(file);
1645         struct ceph_inode_info *ci = ceph_inode(inode);
1646         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1647         struct ceph_osd_request *req;
1648         struct ceph_cap_flush *prealloc_cf;
1649         struct folio *folio = NULL;
1650         u64 inline_version = CEPH_INLINE_NONE;
1651         struct page *pages[1];
1652         int err = 0;
1653         u64 len;
1654
1655         prealloc_cf = ceph_alloc_cap_flush();
1656         if (!prealloc_cf)
1657                 return -ENOMEM;
1658
1659         folio = read_mapping_folio(inode->i_mapping, 0, file);
1660         if (IS_ERR(folio)) {
1661                 err = PTR_ERR(folio);
1662                 goto out;
1663         }
1664
1665         folio_lock(folio);
1666
1667         spin_lock(&ci->i_ceph_lock);
1668         inline_version = ci->i_inline_version;
1669         spin_unlock(&ci->i_ceph_lock);
1670
1671         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1672              inode, ceph_vinop(inode), inline_version);
1673
1674         if (inline_version == 1 || /* initial version, no data */
1675             inline_version == CEPH_INLINE_NONE)
1676                 goto out_unlock;
1677
1678         len = i_size_read(inode);
1679         if (len > folio_size(folio))
1680                 len = folio_size(folio);
1681
1682         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1683                                     ceph_vino(inode), 0, &len, 0, 1,
1684                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1685                                     NULL, 0, 0, false);
1686         if (IS_ERR(req)) {
1687                 err = PTR_ERR(req);
1688                 goto out_unlock;
1689         }
1690
1691         req->r_mtime = inode->i_mtime;
1692         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1693         if (!err)
1694                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1695         ceph_osdc_put_request(req);
1696         if (err < 0)
1697                 goto out_unlock;
1698
1699         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1700                                     ceph_vino(inode), 0, &len, 1, 3,
1701                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1702                                     NULL, ci->i_truncate_seq,
1703                                     ci->i_truncate_size, false);
1704         if (IS_ERR(req)) {
1705                 err = PTR_ERR(req);
1706                 goto out_unlock;
1707         }
1708
1709         pages[0] = folio_page(folio, 0);
1710         osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1711
1712         {
1713                 __le64 xattr_buf = cpu_to_le64(inline_version);
1714                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1715                                             "inline_version", &xattr_buf,
1716                                             sizeof(xattr_buf),
1717                                             CEPH_OSD_CMPXATTR_OP_GT,
1718                                             CEPH_OSD_CMPXATTR_MODE_U64);
1719                 if (err)
1720                         goto out_put_req;
1721         }
1722
1723         {
1724                 char xattr_buf[32];
1725                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1726                                          "%llu", inline_version);
1727                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1728                                             "inline_version",
1729                                             xattr_buf, xattr_len, 0, 0);
1730                 if (err)
1731                         goto out_put_req;
1732         }
1733
1734         req->r_mtime = inode->i_mtime;
1735         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1736         if (!err)
1737                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1738
1739         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1740                                   req->r_end_latency, len, err);
1741
1742         if (!err) {
1743                 int dirty;
1744
1745                 /* Set to CAP_INLINE_NONE and dirty the caps */
1746                 down_read(&fsc->mdsc->snap_rwsem);
1747                 spin_lock(&ci->i_ceph_lock);
1748                 ci->i_inline_version = CEPH_INLINE_NONE;
1749                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1750                 spin_unlock(&ci->i_ceph_lock);
1751                 up_read(&fsc->mdsc->snap_rwsem);
1752                 if (dirty)
1753                         __mark_inode_dirty(inode, dirty);
1754         }
1755 out_put_req:
1756         ceph_osdc_put_request(req);
1757         if (err == -ECANCELED)
1758                 err = 0;
1759 out_unlock:
1760         folio_unlock(folio);
1761         folio_put(folio);
1762 out:
1763         ceph_free_cap_flush(prealloc_cf);
1764         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1765              inode, ceph_vinop(inode), inline_version, err);
1766         return err;
1767 }
1768
1769 static const struct vm_operations_struct ceph_vmops = {
1770         .fault          = ceph_filemap_fault,
1771         .page_mkwrite   = ceph_page_mkwrite,
1772 };
1773
1774 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1775 {
1776         struct address_space *mapping = file->f_mapping;
1777
1778         if (!mapping->a_ops->readpage)
1779                 return -ENOEXEC;
1780         file_accessed(file);
1781         vma->vm_ops = &ceph_vmops;
1782         return 0;
1783 }
1784
1785 enum {
1786         POOL_READ       = 1,
1787         POOL_WRITE      = 2,
1788 };
1789
1790 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1791                                 s64 pool, struct ceph_string *pool_ns)
1792 {
1793         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1794         struct ceph_mds_client *mdsc = fsc->mdsc;
1795         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1796         struct rb_node **p, *parent;
1797         struct ceph_pool_perm *perm;
1798         struct page **pages;
1799         size_t pool_ns_len;
1800         int err = 0, err2 = 0, have = 0;
1801
1802         down_read(&mdsc->pool_perm_rwsem);
1803         p = &mdsc->pool_perm_tree.rb_node;
1804         while (*p) {
1805                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1806                 if (pool < perm->pool)
1807                         p = &(*p)->rb_left;
1808                 else if (pool > perm->pool)
1809                         p = &(*p)->rb_right;
1810                 else {
1811                         int ret = ceph_compare_string(pool_ns,
1812                                                 perm->pool_ns,
1813                                                 perm->pool_ns_len);
1814                         if (ret < 0)
1815                                 p = &(*p)->rb_left;
1816                         else if (ret > 0)
1817                                 p = &(*p)->rb_right;
1818                         else {
1819                                 have = perm->perm;
1820                                 break;
1821                         }
1822                 }
1823         }
1824         up_read(&mdsc->pool_perm_rwsem);
1825         if (*p)
1826                 goto out;
1827
1828         if (pool_ns)
1829                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1830                      pool, (int)pool_ns->len, pool_ns->str);
1831         else
1832                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1833
1834         down_write(&mdsc->pool_perm_rwsem);
1835         p = &mdsc->pool_perm_tree.rb_node;
1836         parent = NULL;
1837         while (*p) {
1838                 parent = *p;
1839                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1840                 if (pool < perm->pool)
1841                         p = &(*p)->rb_left;
1842                 else if (pool > perm->pool)
1843                         p = &(*p)->rb_right;
1844                 else {
1845                         int ret = ceph_compare_string(pool_ns,
1846                                                 perm->pool_ns,
1847                                                 perm->pool_ns_len);
1848                         if (ret < 0)
1849                                 p = &(*p)->rb_left;
1850                         else if (ret > 0)
1851                                 p = &(*p)->rb_right;
1852                         else {
1853                                 have = perm->perm;
1854                                 break;
1855                         }
1856                 }
1857         }
1858         if (*p) {
1859                 up_write(&mdsc->pool_perm_rwsem);
1860                 goto out;
1861         }
1862
1863         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1864                                          1, false, GFP_NOFS);
1865         if (!rd_req) {
1866                 err = -ENOMEM;
1867                 goto out_unlock;
1868         }
1869
1870         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1871         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1872         rd_req->r_base_oloc.pool = pool;
1873         if (pool_ns)
1874                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1875         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1876
1877         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1878         if (err)
1879                 goto out_unlock;
1880
1881         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1882                                          1, false, GFP_NOFS);
1883         if (!wr_req) {
1884                 err = -ENOMEM;
1885                 goto out_unlock;
1886         }
1887
1888         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1889         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1890         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1891         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1892
1893         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1894         if (err)
1895                 goto out_unlock;
1896
1897         /* one page should be large enough for STAT data */
1898         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1899         if (IS_ERR(pages)) {
1900                 err = PTR_ERR(pages);
1901                 goto out_unlock;
1902         }
1903
1904         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1905                                      0, false, true);
1906         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1907
1908         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1909         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1910
1911         if (!err)
1912                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1913         if (!err2)
1914                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1915
1916         if (err >= 0 || err == -ENOENT)
1917                 have |= POOL_READ;
1918         else if (err != -EPERM) {
1919                 if (err == -EBLOCKLISTED)
1920                         fsc->blocklisted = true;
1921                 goto out_unlock;
1922         }
1923
1924         if (err2 == 0 || err2 == -EEXIST)
1925                 have |= POOL_WRITE;
1926         else if (err2 != -EPERM) {
1927                 if (err2 == -EBLOCKLISTED)
1928                         fsc->blocklisted = true;
1929                 err = err2;
1930                 goto out_unlock;
1931         }
1932
1933         pool_ns_len = pool_ns ? pool_ns->len : 0;
1934         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1935         if (!perm) {
1936                 err = -ENOMEM;
1937                 goto out_unlock;
1938         }
1939
1940         perm->pool = pool;
1941         perm->perm = have;
1942         perm->pool_ns_len = pool_ns_len;
1943         if (pool_ns_len > 0)
1944                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1945         perm->pool_ns[pool_ns_len] = 0;
1946
1947         rb_link_node(&perm->node, parent, p);
1948         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1949         err = 0;
1950 out_unlock:
1951         up_write(&mdsc->pool_perm_rwsem);
1952
1953         ceph_osdc_put_request(rd_req);
1954         ceph_osdc_put_request(wr_req);
1955 out:
1956         if (!err)
1957                 err = have;
1958         if (pool_ns)
1959                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1960                      pool, (int)pool_ns->len, pool_ns->str, err);
1961         else
1962                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1963         return err;
1964 }
1965
1966 int ceph_pool_perm_check(struct inode *inode, int need)
1967 {
1968         struct ceph_inode_info *ci = ceph_inode(inode);
1969         struct ceph_string *pool_ns;
1970         s64 pool;
1971         int ret, flags;
1972
1973         /* Only need to do this for regular files */
1974         if (!S_ISREG(inode->i_mode))
1975                 return 0;
1976
1977         if (ci->i_vino.snap != CEPH_NOSNAP) {
1978                 /*
1979                  * Pool permission check needs to write to the first object.
1980                  * But for snapshot, head of the first object may have alread
1981                  * been deleted. Skip check to avoid creating orphan object.
1982                  */
1983                 return 0;
1984         }
1985
1986         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
1987                                 NOPOOLPERM))
1988                 return 0;
1989
1990         spin_lock(&ci->i_ceph_lock);
1991         flags = ci->i_ceph_flags;
1992         pool = ci->i_layout.pool_id;
1993         spin_unlock(&ci->i_ceph_lock);
1994 check:
1995         if (flags & CEPH_I_POOL_PERM) {
1996                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1997                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1998                              pool);
1999                         return -EPERM;
2000                 }
2001                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2002                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2003                              pool);
2004                         return -EPERM;
2005                 }
2006                 return 0;
2007         }
2008
2009         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2010         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2011         ceph_put_string(pool_ns);
2012         if (ret < 0)
2013                 return ret;
2014
2015         flags = CEPH_I_POOL_PERM;
2016         if (ret & POOL_READ)
2017                 flags |= CEPH_I_POOL_RD;
2018         if (ret & POOL_WRITE)
2019                 flags |= CEPH_I_POOL_WR;
2020
2021         spin_lock(&ci->i_ceph_lock);
2022         if (pool == ci->i_layout.pool_id &&
2023             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2024                 ci->i_ceph_flags |= flags;
2025         } else {
2026                 pool = ci->i_layout.pool_id;
2027                 flags = ci->i_ceph_flags;
2028         }
2029         spin_unlock(&ci->i_ceph_lock);
2030         goto check;
2031 }
2032
2033 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2034 {
2035         struct ceph_pool_perm *perm;
2036         struct rb_node *n;
2037
2038         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2039                 n = rb_first(&mdsc->pool_perm_tree);
2040                 perm = rb_entry(n, struct ceph_pool_perm, node);
2041                 rb_erase(n, &mdsc->pool_perm_tree);
2042                 kfree(perm);
2043         }
2044 }