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 **foliop, 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_WARN_ON_FOLIO(folio->private, 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 bool ceph_release_folio(struct folio *folio, gfp_t gfp)
 166 {
 167         struct inode *inode = folio->mapping->host;
 168
 169         dout("%llx:%llx release_folio idx %lu (%sdirty)\n",
 170              ceph_vinop(inode),
 171              folio->index, folio_test_dirty(folio) ? "" : "not ");
 172
 173         if (folio_test_private(folio))
 174                 return false;
 175
 176         if (folio_test_fscache(folio)) {
 177                 if (current_is_kswapd() || !(gfp & __GFP_FS))
 178                         return false;
 179                 folio_wait_fscache(folio);
 180         }
 181         ceph_fscache_note_page_release(inode);
 182         return true;
 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         unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
 191         loff_t end = rreq->start + rreq->len, new_end;
 192         struct ceph_netfs_request_data *priv = rreq->netfs_priv;
 193         unsigned long max_len;
 194         u32 blockoff;
 195
 196         if (priv) {
 197                 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
 198                 if (priv->file_ra_disabled)
 199                         max_pages = 0;
 200                 else
 201                         max_pages = priv->file_ra_pages;
 202
 203         }
 204
 205         /* Readahead is disabled */
 206         if (!max_pages)
 207                 return;
 208
 209         max_len = max_pages << PAGE_SHIFT;
 210
 211         /*
 212          * Try to expand the length forward by rounding up it to the next
 213          * block, but do not exceed the file size, unless the original
 214          * request already exceeds it.
 215          */
 216         new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
 217         if (new_end > end && new_end <= rreq->start + max_len)
 218                 rreq->len = new_end - rreq->start;
 219
 220         /* Try to expand the start downward */
 221         div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 222         if (rreq->len + blockoff <= max_len) {
 223                 rreq->start -= blockoff;
 224                 rreq->len += blockoff;
 225         }
 226 }
 227
 228 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
 229 {
 230         struct inode *inode = subreq->rreq->inode;
 231         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 232         struct ceph_inode_info *ci = ceph_inode(inode);
 233         u64 objno, objoff;
 234         u32 xlen;
 235
 236         /* Truncate the extent at the end of the current block */
 237         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 238                                       &objno, &objoff, &xlen);
 239         subreq->len = min(xlen, fsc->mount_options->rsize);
 240         return true;
 241 }
 242
 243 static void finish_netfs_read(struct ceph_osd_request *req)
 244 {
 245         struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 246         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 247         struct netfs_io_subrequest *subreq = req->r_priv;
 248         int num_pages;
 249         int err = req->r_result;
 250
 251         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 252                                  req->r_end_latency, osd_data->length, err);
 253
 254         dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 255              subreq->len, i_size_read(req->r_inode));
 256
 257         /* no object means success but no data */
 258         if (err == -ENOENT)
 259                 err = 0;
 260         else if (err == -EBLOCKLISTED)
 261                 fsc->blocklisted = true;
 262
 263         if (err >= 0 && err < subreq->len)
 264                 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 265
 266         netfs_subreq_terminated(subreq, err, false);
 267
 268         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 269         ceph_put_page_vector(osd_data->pages, num_pages, false);
 270         iput(req->r_inode);
 271 }
 272
 273 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 274 {
 275         struct netfs_io_request *rreq = subreq->rreq;
 276         struct inode *inode = rreq->inode;
 277         struct ceph_mds_reply_info_parsed *rinfo;
 278         struct ceph_mds_reply_info_in *iinfo;
 279         struct ceph_mds_request *req;
 280         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 281         struct ceph_inode_info *ci = ceph_inode(inode);
 282         struct iov_iter iter;
 283         ssize_t err = 0;
 284         size_t len;
 285         int mode;
 286
 287         __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 288         __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 289
 290         if (subreq->start >= inode->i_size)
 291                 goto out;
 292
 293         /* We need to fetch the inline data. */
 294         mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
 295         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
 296         if (IS_ERR(req)) {
 297                 err = PTR_ERR(req);
 298                 goto out;
 299         }
 300         req->r_ino1 = ci->i_vino;
 301         req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 302         req->r_num_caps = 2;
 303
 304         err = ceph_mdsc_do_request(mdsc, NULL, req);
 305         if (err < 0)
 306                 goto out;
 307
 308         rinfo = &req->r_reply_info;
 309         iinfo = &rinfo->targeti;
 310         if (iinfo->inline_version == CEPH_INLINE_NONE) {
 311                 /* The data got uninlined */
 312                 ceph_mdsc_put_request(req);
 313                 return false;
 314         }
 315
 316         len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
 317         iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
 318         err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
 319         if (err == 0)
 320                 err = -EFAULT;
 321
 322         ceph_mdsc_put_request(req);
 323 out:
 324         netfs_subreq_terminated(subreq, err, false);
 325         return true;
 326 }
 327
 328 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 329 {
 330         struct netfs_io_request *rreq = subreq->rreq;
 331         struct inode *inode = rreq->inode;
 332         struct ceph_inode_info *ci = ceph_inode(inode);
 333         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 334         struct ceph_osd_request *req = NULL;
 335         struct ceph_vino vino = ceph_vino(inode);
 336         struct iov_iter iter;
 337         struct page **pages;
 338         size_t page_off;
 339         int err = 0;
 340         u64 len = subreq->len;
 341
 342         if (ceph_inode_is_shutdown(inode)) {
 343                 err = -EIO;
 344                 goto out;
 345         }
 346
 347         if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
 348                 return;
 349
 350         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 351                         0, 1, CEPH_OSD_OP_READ,
 352                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 353                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 354         if (IS_ERR(req)) {
 355                 err = PTR_ERR(req);
 356                 req = NULL;
 357                 goto out;
 358         }
 359
 360         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 361         iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
 362         err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
 363         if (err < 0) {
 364                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 365                 goto out;
 366         }
 367
 368         /* should always give us a page-aligned read */
 369         WARN_ON_ONCE(page_off);
 370         len = err;
 371         err = 0;
 372
 373         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 374         req->r_callback = finish_netfs_read;
 375         req->r_priv = subreq;
 376         req->r_inode = inode;
 377         ihold(inode);
 378
 379         ceph_osdc_start_request(req->r_osdc, req);
 380 out:
 381         ceph_osdc_put_request(req);
 382         if (err)
 383                 netfs_subreq_terminated(subreq, err, false);
 384         dout("%s: result %d\n", __func__, err);
 385 }
 386
 387 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 388 {
 389         struct inode *inode = rreq->inode;
 390         int got = 0, want = CEPH_CAP_FILE_CACHE;
 391         struct ceph_netfs_request_data *priv;
 392         int ret = 0;
 393
 394         if (rreq->origin != NETFS_READAHEAD)
 395                 return 0;
 396
 397         priv = kzalloc(sizeof(*priv), GFP_NOFS);
 398         if (!priv)
 399                 return -ENOMEM;
 400
 401         if (file) {
 402                 struct ceph_rw_context *rw_ctx;
 403                 struct ceph_file_info *fi = file->private_data;
 404
 405                 priv->file_ra_pages = file->f_ra.ra_pages;
 406                 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
 407
 408                 rw_ctx = ceph_find_rw_context(fi);
 409                 if (rw_ctx) {
 410                         rreq->netfs_priv = priv;
 411                         return 0;
 412                 }
 413         }
 414
 415         /*
 416          * readahead callers do not necessarily hold Fcb caps
 417          * (e.g. fadvise, madvise).
 418          */
 419         ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 420         if (ret < 0) {
 421                 dout("start_read %p, error getting cap\n", inode);
 422                 goto out;
 423         }
 424
 425         if (!(got & want)) {
 426                 dout("start_read %p, no cache cap\n", inode);
 427                 ret = -EACCES;
 428                 goto out;
 429         }
 430         if (ret == 0) {
 431                 ret = -EACCES;
 432                 goto out;
 433         }
 434
 435         priv->caps = got;
 436         rreq->netfs_priv = priv;
 437
 438 out:
 439         if (ret < 0)
 440                 kfree(priv);
 441
 442         return ret;
 443 }
 444
 445 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 446 {
 447         struct ceph_netfs_request_data *priv = rreq->netfs_priv;
 448
 449         if (!priv)
 450                 return;
 451
 452         if (priv->caps)
 453                 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
 454         kfree(priv);
 455         rreq->netfs_priv = NULL;
 456 }
 457
 458 const struct netfs_request_ops ceph_netfs_ops = {
 459         .init_request           = ceph_init_request,
 460         .free_request           = ceph_netfs_free_request,
 461         .begin_cache_operation  = ceph_begin_cache_operation,
 462         .issue_read             = ceph_netfs_issue_read,
 463         .expand_readahead       = ceph_netfs_expand_readahead,
 464         .clamp_length           = ceph_netfs_clamp_length,
 465         .check_write_begin      = ceph_netfs_check_write_begin,
 466 };
 467
 468 #ifdef CONFIG_CEPH_FSCACHE
 469 static void ceph_set_page_fscache(struct page *page)
 470 {
 471         set_page_fscache(page);
 472 }
 473
 474 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 475 {
 476         struct inode *inode = priv;
 477
 478         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 479                 ceph_fscache_invalidate(inode, false);
 480 }
 481
 482 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 483 {
 484         struct ceph_inode_info *ci = ceph_inode(inode);
 485         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 486
 487         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 488                                ceph_fscache_write_terminated, inode, caching);
 489 }
 490 #else
 491 static inline void ceph_set_page_fscache(struct page *page)
 492 {
 493 }
 494
 495 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 496 {
 497 }
 498 #endif /* CONFIG_CEPH_FSCACHE */
 499
 500 struct ceph_writeback_ctl
 501 {
 502         loff_t i_size;
 503         u64 truncate_size;
 504         u32 truncate_seq;
 505         bool size_stable;
 506         bool head_snapc;
 507 };
 508
 509 /*
 510  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 511  * only snap context we are allowed to write back.
 512  */
 513 static struct ceph_snap_context *
 514 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 515                    struct ceph_snap_context *page_snapc)
 516 {
 517         struct ceph_inode_info *ci = ceph_inode(inode);
 518         struct ceph_snap_context *snapc = NULL;
 519         struct ceph_cap_snap *capsnap = NULL;
 520
 521         spin_lock(&ci->i_ceph_lock);
 522         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 523                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 524                      capsnap->context, capsnap->dirty_pages);
 525                 if (!capsnap->dirty_pages)
 526                         continue;
 527
 528                 /* get i_size, truncate_{seq,size} for page_snapc? */
 529                 if (snapc && capsnap->context != page_snapc)
 530                         continue;
 531
 532                 if (ctl) {
 533                         if (capsnap->writing) {
 534                                 ctl->i_size = i_size_read(inode);
 535                                 ctl->size_stable = false;
 536                         } else {
 537                                 ctl->i_size = capsnap->size;
 538                                 ctl->size_stable = true;
 539                         }
 540                         ctl->truncate_size = capsnap->truncate_size;
 541                         ctl->truncate_seq = capsnap->truncate_seq;
 542                         ctl->head_snapc = false;
 543                 }
 544
 545                 if (snapc)
 546                         break;
 547
 548                 snapc = ceph_get_snap_context(capsnap->context);
 549                 if (!page_snapc ||
 550                     page_snapc == snapc ||
 551                     page_snapc->seq > snapc->seq)
 552                         break;
 553         }
 554         if (!snapc && ci->i_wrbuffer_ref_head) {
 555                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 556                 dout(" head snapc %p has %d dirty pages\n",
 557                      snapc, ci->i_wrbuffer_ref_head);
 558                 if (ctl) {
 559                         ctl->i_size = i_size_read(inode);
 560                         ctl->truncate_size = ci->i_truncate_size;
 561                         ctl->truncate_seq = ci->i_truncate_seq;
 562                         ctl->size_stable = false;
 563                         ctl->head_snapc = true;
 564                 }
 565         }
 566         spin_unlock(&ci->i_ceph_lock);
 567         return snapc;
 568 }
 569
 570 static u64 get_writepages_data_length(struct inode *inode,
 571                                       struct page *page, u64 start)
 572 {
 573         struct ceph_inode_info *ci = ceph_inode(inode);
 574         struct ceph_snap_context *snapc = page_snap_context(page);
 575         struct ceph_cap_snap *capsnap = NULL;
 576         u64 end = i_size_read(inode);
 577
 578         if (snapc != ci->i_head_snapc) {
 579                 bool found = false;
 580                 spin_lock(&ci->i_ceph_lock);
 581                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 582                         if (capsnap->context == snapc) {
 583                                 if (!capsnap->writing)
 584                                         end = capsnap->size;
 585                                 found = true;
 586                                 break;
 587                         }
 588                 }
 589                 spin_unlock(&ci->i_ceph_lock);
 590                 WARN_ON(!found);
 591         }
 592         if (end > page_offset(page) + thp_size(page))
 593                 end = page_offset(page) + thp_size(page);
 594         return end > start ? end - start : 0;
 595 }
 596
 597 /*
 598  * Write a single page, but leave the page locked.
 599  *
 600  * If we get a write error, mark the mapping for error, but still adjust the
 601  * dirty page accounting (i.e., page is no longer dirty).
 602  */
 603 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 604 {
 605         struct folio *folio = page_folio(page);
 606         struct inode *inode = page->mapping->host;
 607         struct ceph_inode_info *ci = ceph_inode(inode);
 608         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 609         struct ceph_snap_context *snapc, *oldest;
 610         loff_t page_off = page_offset(page);
 611         int err;
 612         loff_t len = thp_size(page);
 613         struct ceph_writeback_ctl ceph_wbc;
 614         struct ceph_osd_client *osdc = &fsc->client->osdc;
 615         struct ceph_osd_request *req;
 616         bool caching = ceph_is_cache_enabled(inode);
 617
 618         dout("writepage %p idx %lu\n", page, page->index);
 619
 620         if (ceph_inode_is_shutdown(inode))
 621                 return -EIO;
 622
 623         /* verify this is a writeable snap context */
 624         snapc = page_snap_context(page);
 625         if (!snapc) {
 626                 dout("writepage %p page %p not dirty?\n", inode, page);
 627                 return 0;
 628         }
 629         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 630         if (snapc->seq > oldest->seq) {
 631                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 632                      inode, page, snapc);
 633                 /* we should only noop if called by kswapd */
 634                 WARN_ON(!(current->flags & PF_MEMALLOC));
 635                 ceph_put_snap_context(oldest);
 636                 redirty_page_for_writepage(wbc, page);
 637                 return 0;
 638         }
 639         ceph_put_snap_context(oldest);
 640
 641         /* is this a partial page at end of file? */
 642         if (page_off >= ceph_wbc.i_size) {
 643                 dout("folio at %lu beyond eof %llu\n", folio->index,
 644                                 ceph_wbc.i_size);
 645                 folio_invalidate(folio, 0, folio_size(folio));
 646                 return 0;
 647         }
 648
 649         if (ceph_wbc.i_size < page_off + len)
 650                 len = ceph_wbc.i_size - page_off;
 651
 652         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 653              inode, page, page->index, page_off, len, snapc, snapc->seq);
 654
 655         if (atomic_long_inc_return(&fsc->writeback_count) >
 656             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 657                 fsc->write_congested = true;
 658
 659         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 660                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 661                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 662                                     true);
 663         if (IS_ERR(req)) {
 664                 redirty_page_for_writepage(wbc, page);
 665                 return PTR_ERR(req);
 666         }
 667
 668         set_page_writeback(page);
 669         if (caching)
 670                 ceph_set_page_fscache(page);
 671         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 672
 673         /* it may be a short write due to an object boundary */
 674         WARN_ON_ONCE(len > thp_size(page));
 675         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 676         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 677
 678         req->r_mtime = inode->i_mtime;
 679         ceph_osdc_start_request(osdc, req);
 680         err = ceph_osdc_wait_request(osdc, req);
 681
 682         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 683                                   req->r_end_latency, len, err);
 684
 685         ceph_osdc_put_request(req);
 686         if (err == 0)
 687                 err = len;
 688
 689         if (err < 0) {
 690                 struct writeback_control tmp_wbc;
 691                 if (!wbc)
 692                         wbc = &tmp_wbc;
 693                 if (err == -ERESTARTSYS) {
 694                         /* killed by SIGKILL */
 695                         dout("writepage interrupted page %p\n", page);
 696                         redirty_page_for_writepage(wbc, page);
 697                         end_page_writeback(page);
 698                         return err;
 699                 }
 700                 if (err == -EBLOCKLISTED)
 701                         fsc->blocklisted = true;
 702                 dout("writepage setting page/mapping error %d %p\n",
 703                      err, page);
 704                 mapping_set_error(&inode->i_data, err);
 705                 wbc->pages_skipped++;
 706         } else {
 707                 dout("writepage cleaned page %p\n", page);
 708                 err = 0;  /* vfs expects us to return 0 */
 709         }
 710         oldest = detach_page_private(page);
 711         WARN_ON_ONCE(oldest != snapc);
 712         end_page_writeback(page);
 713         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 714         ceph_put_snap_context(snapc);  /* page's reference */
 715
 716         if (atomic_long_dec_return(&fsc->writeback_count) <
 717             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 718                 fsc->write_congested = false;
 719
 720         return err;
 721 }
 722
 723 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 724 {
 725         int err;
 726         struct inode *inode = page->mapping->host;
 727         BUG_ON(!inode);
 728         ihold(inode);
 729
 730         if (wbc->sync_mode == WB_SYNC_NONE &&
 731             ceph_inode_to_client(inode)->write_congested)
 732                 return AOP_WRITEPAGE_ACTIVATE;
 733
 734         wait_on_page_fscache(page);
 735
 736         err = writepage_nounlock(page, wbc);
 737         if (err == -ERESTARTSYS) {
 738                 /* direct memory reclaimer was killed by SIGKILL. return 0
 739                  * to prevent caller from setting mapping/page error */
 740                 err = 0;
 741         }
 742         unlock_page(page);
 743         iput(inode);
 744         return err;
 745 }
 746
 747 /*
 748  * async writeback completion handler.
 749  *
 750  * If we get an error, set the mapping error bit, but not the individual
 751  * page error bits.
 752  */
 753 static void writepages_finish(struct ceph_osd_request *req)
 754 {
 755         struct inode *inode = req->r_inode;
 756         struct ceph_inode_info *ci = ceph_inode(inode);
 757         struct ceph_osd_data *osd_data;
 758         struct page *page;
 759         int num_pages, total_pages = 0;
 760         int i, j;
 761         int rc = req->r_result;
 762         struct ceph_snap_context *snapc = req->r_snapc;
 763         struct address_space *mapping = inode->i_mapping;
 764         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 765         unsigned int len = 0;
 766         bool remove_page;
 767
 768         dout("writepages_finish %p rc %d\n", inode, rc);
 769         if (rc < 0) {
 770                 mapping_set_error(mapping, rc);
 771                 ceph_set_error_write(ci);
 772                 if (rc == -EBLOCKLISTED)
 773                         fsc->blocklisted = true;
 774         } else {
 775                 ceph_clear_error_write(ci);
 776         }
 777
 778         /*
 779          * We lost the cache cap, need to truncate the page before
 780          * it is unlocked, otherwise we'd truncate it later in the
 781          * page truncation thread, possibly losing some data that
 782          * raced its way in
 783          */
 784         remove_page = !(ceph_caps_issued(ci) &
 785                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 786
 787         /* clean all pages */
 788         for (i = 0; i < req->r_num_ops; i++) {
 789                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 790                         pr_warn("%s incorrect op %d req %p index %d tid %llu\n",
 791                                 __func__, req->r_ops[i].op, req, i, req->r_tid);
 792                         break;
 793                 }
 794
 795                 osd_data = osd_req_op_extent_osd_data(req, i);
 796                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 797                 len += osd_data->length;
 798                 num_pages = calc_pages_for((u64)osd_data->alignment,
 799                                            (u64)osd_data->length);
 800                 total_pages += num_pages;
 801                 for (j = 0; j < num_pages; j++) {
 802                         page = osd_data->pages[j];
 803                         BUG_ON(!page);
 804                         WARN_ON(!PageUptodate(page));
 805
 806                         if (atomic_long_dec_return(&fsc->writeback_count) <
 807                              CONGESTION_OFF_THRESH(
 808                                         fsc->mount_options->congestion_kb))
 809                                 fsc->write_congested = false;
 810
 811                         ceph_put_snap_context(detach_page_private(page));
 812                         end_page_writeback(page);
 813                         dout("unlocking %p\n", page);
 814
 815                         if (remove_page)
 816                                 generic_error_remove_page(inode->i_mapping,
 817                                                           page);
 818
 819                         unlock_page(page);
 820                 }
 821                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 822                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 823
 824                 release_pages(osd_data->pages, num_pages);
 825         }
 826
 827         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 828                                   req->r_end_latency, len, rc);
 829
 830         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 831
 832         osd_data = osd_req_op_extent_osd_data(req, 0);
 833         if (osd_data->pages_from_pool)
 834                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 835         else
 836                 kfree(osd_data->pages);
 837         ceph_osdc_put_request(req);
 838 }
 839
 840 /*
 841  * initiate async writeback
 842  */
 843 static int ceph_writepages_start(struct address_space *mapping,
 844                                  struct writeback_control *wbc)
 845 {
 846         struct inode *inode = mapping->host;
 847         struct ceph_inode_info *ci = ceph_inode(inode);
 848         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 849         struct ceph_vino vino = ceph_vino(inode);
 850         pgoff_t index, start_index, end = -1;
 851         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 852         struct folio_batch fbatch;
 853         int rc = 0;
 854         unsigned int wsize = i_blocksize(inode);
 855         struct ceph_osd_request *req = NULL;
 856         struct ceph_writeback_ctl ceph_wbc;
 857         bool should_loop, range_whole = false;
 858         bool done = false;
 859         bool caching = ceph_is_cache_enabled(inode);
 860         xa_mark_t tag;
 861
 862         if (wbc->sync_mode == WB_SYNC_NONE &&
 863             fsc->write_congested)
 864                 return 0;
 865
 866         dout("writepages_start %p (mode=%s)\n", inode,
 867              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 868              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 869
 870         if (ceph_inode_is_shutdown(inode)) {
 871                 if (ci->i_wrbuffer_ref > 0) {
 872                         pr_warn_ratelimited(
 873                                 "writepage_start %p %lld forced umount\n",
 874                                 inode, ceph_ino(inode));
 875                 }
 876                 mapping_set_error(mapping, -EIO);
 877                 return -EIO; /* we're in a forced umount, don't write! */
 878         }
 879         if (fsc->mount_options->wsize < wsize)
 880                 wsize = fsc->mount_options->wsize;
 881
 882         folio_batch_init(&fbatch);
 883
 884         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 885         index = start_index;
 886
 887         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
 888                 tag = PAGECACHE_TAG_TOWRITE;
 889         } else {
 890                 tag = PAGECACHE_TAG_DIRTY;
 891         }
 892 retry:
 893         /* find oldest snap context with dirty data */
 894         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 895         if (!snapc) {
 896                 /* hmm, why does writepages get called when there
 897                    is no dirty data? */
 898                 dout(" no snap context with dirty data?\n");
 899                 goto out;
 900         }
 901         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 902              snapc, snapc->seq, snapc->num_snaps);
 903
 904         should_loop = false;
 905         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 906                 /* where to start/end? */
 907                 if (wbc->range_cyclic) {
 908                         index = start_index;
 909                         end = -1;
 910                         if (index > 0)
 911                                 should_loop = true;
 912                         dout(" cyclic, start at %lu\n", index);
 913                 } else {
 914                         index = wbc->range_start >> PAGE_SHIFT;
 915                         end = wbc->range_end >> PAGE_SHIFT;
 916                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 917                                 range_whole = true;
 918                         dout(" not cyclic, %lu to %lu\n", index, end);
 919                 }
 920         } else if (!ceph_wbc.head_snapc) {
 921                 /* Do not respect wbc->range_{start,end}. Dirty pages
 922                  * in that range can be associated with newer snapc.
 923                  * They are not writeable until we write all dirty pages
 924                  * associated with 'snapc' get written */
 925                 if (index > 0)
 926                         should_loop = true;
 927                 dout(" non-head snapc, range whole\n");
 928         }
 929
 930         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 931                 tag_pages_for_writeback(mapping, index, end);
 932
 933         ceph_put_snap_context(last_snapc);
 934         last_snapc = snapc;
 935
 936         while (!done && index <= end) {
 937                 int num_ops = 0, op_idx;
 938                 unsigned i, nr_folios, max_pages, locked_pages = 0;
 939                 struct page **pages = NULL, **data_pages;
 940                 struct page *page;
 941                 pgoff_t strip_unit_end = 0;
 942                 u64 offset = 0, len = 0;
 943                 bool from_pool = false;
 944
 945                 max_pages = wsize >> PAGE_SHIFT;
 946
 947 get_more_pages:
 948                 nr_folios = filemap_get_folios_tag(mapping, &index,
 949                                                    end, tag, &fbatch);
 950                 dout("pagevec_lookup_range_tag got %d\n", nr_folios);
 951                 if (!nr_folios && !locked_pages)
 952                         break;
 953                 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
 954                         page = &fbatch.folios[i]->page;
 955                         dout("? %p idx %lu\n", page, page->index);
 956                         if (locked_pages == 0)
 957                                 lock_page(page);  /* first page */
 958                         else if (!trylock_page(page))
 959                                 break;
 960
 961                         /* only dirty pages, or our accounting breaks */
 962                         if (unlikely(!PageDirty(page)) ||
 963                             unlikely(page->mapping != mapping)) {
 964                                 dout("!dirty or !mapping %p\n", page);
 965                                 unlock_page(page);
 966                                 continue;
 967                         }
 968                         /* only if matching snap context */
 969                         pgsnapc = page_snap_context(page);
 970                         if (pgsnapc != snapc) {
 971                                 dout("page snapc %p %lld != oldest %p %lld\n",
 972                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 973                                 if (!should_loop &&
 974                                     !ceph_wbc.head_snapc &&
 975                                     wbc->sync_mode != WB_SYNC_NONE)
 976                                         should_loop = true;
 977                                 unlock_page(page);
 978                                 continue;
 979                         }
 980                         if (page_offset(page) >= ceph_wbc.i_size) {
 981                                 struct folio *folio = page_folio(page);
 982
 983                                 dout("folio at %lu beyond eof %llu\n",
 984                                      folio->index, ceph_wbc.i_size);
 985                                 if ((ceph_wbc.size_stable ||
 986                                     folio_pos(folio) >= i_size_read(inode)) &&
 987                                     folio_clear_dirty_for_io(folio))
 988                                         folio_invalidate(folio, 0,
 989                                                         folio_size(folio));
 990                                 folio_unlock(folio);
 991                                 continue;
 992                         }
 993                         if (strip_unit_end && (page->index > strip_unit_end)) {
 994                                 dout("end of strip unit %p\n", page);
 995                                 unlock_page(page);
 996                                 break;
 997                         }
 998                         if (PageWriteback(page) || PageFsCache(page)) {
 999                                 if (wbc->sync_mode == WB_SYNC_NONE) {
1000                                         dout("%p under writeback\n", page);
1001                                         unlock_page(page);
1002                                         continue;
1003                                 }
1004                                 dout("waiting on writeback %p\n", page);
1005                                 wait_on_page_writeback(page);
1006                                 wait_on_page_fscache(page);
1007                         }
1008
1009                         if (!clear_page_dirty_for_io(page)) {
1010                                 dout("%p !clear_page_dirty_for_io\n", page);
1011                                 unlock_page(page);
1012                                 continue;
1013                         }
1014
1015                         /*
1016                          * We have something to write.  If this is
1017                          * the first locked page this time through,
1018                          * calculate max possinle write size and
1019                          * allocate a page array
1020                          */
1021                         if (locked_pages == 0) {
1022                                 u64 objnum;
1023                                 u64 objoff;
1024                                 u32 xlen;
1025
1026                                 /* prepare async write request */
1027                                 offset = (u64)page_offset(page);
1028                                 ceph_calc_file_object_mapping(&ci->i_layout,
1029                                                               offset, wsize,
1030                                                               &objnum, &objoff,
1031                                                               &xlen);
1032                                 len = xlen;
1033
1034                                 num_ops = 1;
1035                                 strip_unit_end = page->index +
1036                                         ((len - 1) >> PAGE_SHIFT);
1037
1038                                 BUG_ON(pages);
1039                                 max_pages = calc_pages_for(0, (u64)len);
1040                                 pages = kmalloc_array(max_pages,
1041                                                       sizeof(*pages),
1042                                                       GFP_NOFS);
1043                                 if (!pages) {
1044                                         from_pool = true;
1045                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1046                                         BUG_ON(!pages);
1047                                 }
1048
1049                                 len = 0;
1050                         } else if (page->index !=
1051                                    (offset + len) >> PAGE_SHIFT) {
1052                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1053                                                              CEPH_OSD_MAX_OPS)) {
1054                                         redirty_page_for_writepage(wbc, page);
1055                                         unlock_page(page);
1056                                         break;
1057                                 }
1058
1059                                 num_ops++;
1060                                 offset = (u64)page_offset(page);
1061                                 len = 0;
1062                         }
1063
1064                         /* note position of first page in fbatch */
1065                         dout("%p will write page %p idx %lu\n",
1066                              inode, page, page->index);
1067
1068                         if (atomic_long_inc_return(&fsc->writeback_count) >
1069                             CONGESTION_ON_THRESH(
1070                                     fsc->mount_options->congestion_kb))
1071                                 fsc->write_congested = true;
1072
1073                         pages[locked_pages++] = page;
1074                         fbatch.folios[i] = NULL;
1075
1076                         len += thp_size(page);
1077                 }
1078
1079                 /* did we get anything? */
1080                 if (!locked_pages)
1081                         goto release_folios;
1082                 if (i) {
1083                         unsigned j, n = 0;
1084                         /* shift unused page to beginning of fbatch */
1085                         for (j = 0; j < nr_folios; j++) {
1086                                 if (!fbatch.folios[j])
1087                                         continue;
1088                                 if (n < j)
1089                                         fbatch.folios[n] = fbatch.folios[j];
1090                                 n++;
1091                         }
1092                         fbatch.nr = n;
1093
1094                         if (nr_folios && i == nr_folios &&
1095                             locked_pages < max_pages) {
1096                                 dout("reached end fbatch, trying for more\n");
1097                                 folio_batch_release(&fbatch);
1098                                 goto get_more_pages;
1099                         }
1100                 }
1101
1102 new_request:
1103                 offset = page_offset(pages[0]);
1104                 len = wsize;
1105
1106                 req = ceph_osdc_new_request(&fsc->client->osdc,
1107                                         &ci->i_layout, vino,
1108                                         offset, &len, 0, num_ops,
1109                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1110                                         snapc, ceph_wbc.truncate_seq,
1111                                         ceph_wbc.truncate_size, false);
1112                 if (IS_ERR(req)) {
1113                         req = ceph_osdc_new_request(&fsc->client->osdc,
1114                                                 &ci->i_layout, vino,
1115                                                 offset, &len, 0,
1116                                                 min(num_ops,
1117                                                     CEPH_OSD_SLAB_OPS),
1118                                                 CEPH_OSD_OP_WRITE,
1119                                                 CEPH_OSD_FLAG_WRITE,
1120                                                 snapc, ceph_wbc.truncate_seq,
1121                                                 ceph_wbc.truncate_size, true);
1122                         BUG_ON(IS_ERR(req));
1123                 }
1124                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1125                              thp_size(page) - offset);
1126
1127                 req->r_callback = writepages_finish;
1128                 req->r_inode = inode;
1129
1130                 /* Format the osd request message and submit the write */
1131                 len = 0;
1132                 data_pages = pages;
1133                 op_idx = 0;
1134                 for (i = 0; i < locked_pages; i++) {
1135                         u64 cur_offset = page_offset(pages[i]);
1136                         /*
1137                          * Discontinuity in page range? Ceph can handle that by just passing
1138                          * multiple extents in the write op.
1139                          */
1140                         if (offset + len != cur_offset) {
1141                                 /* If it's full, stop here */
1142                                 if (op_idx + 1 == req->r_num_ops)
1143                                         break;
1144
1145                                 /* Kick off an fscache write with what we have so far. */
1146                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1147
1148                                 /* Start a new extent */
1149                                 osd_req_op_extent_dup_last(req, op_idx,
1150                                                            cur_offset - offset);
1151                                 dout("writepages got pages at %llu~%llu\n",
1152                                      offset, len);
1153                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1154                                                         data_pages, len, 0,
1155                                                         from_pool, false);
1156                                 osd_req_op_extent_update(req, op_idx, len);
1157
1158                                 len = 0;
1159                                 offset = cur_offset;
1160                                 data_pages = pages + i;
1161                                 op_idx++;
1162                         }
1163
1164                         set_page_writeback(pages[i]);
1165                         if (caching)
1166                                 ceph_set_page_fscache(pages[i]);
1167                         len += thp_size(page);
1168                 }
1169                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1170
1171                 if (ceph_wbc.size_stable) {
1172                         len = min(len, ceph_wbc.i_size - offset);
1173                 } else if (i == locked_pages) {
1174                         /* writepages_finish() clears writeback pages
1175                          * according to the data length, so make sure
1176                          * data length covers all locked pages */
1177                         u64 min_len = len + 1 - thp_size(page);
1178                         len = get_writepages_data_length(inode, pages[i - 1],
1179                                                          offset);
1180                         len = max(len, min_len);
1181                 }
1182                 dout("writepages got pages at %llu~%llu\n", offset, len);
1183
1184                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1185                                                  0, from_pool, false);
1186                 osd_req_op_extent_update(req, op_idx, len);
1187
1188                 BUG_ON(op_idx + 1 != req->r_num_ops);
1189
1190                 from_pool = false;
1191                 if (i < locked_pages) {
1192                         BUG_ON(num_ops <= req->r_num_ops);
1193                         num_ops -= req->r_num_ops;
1194                         locked_pages -= i;
1195
1196                         /* allocate new pages array for next request */
1197                         data_pages = pages;
1198                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1199                                               GFP_NOFS);
1200                         if (!pages) {
1201                                 from_pool = true;
1202                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1203                                 BUG_ON(!pages);
1204                         }
1205                         memcpy(pages, data_pages + i,
1206                                locked_pages * sizeof(*pages));
1207                         memset(data_pages + i, 0,
1208                                locked_pages * sizeof(*pages));
1209                 } else {
1210                         BUG_ON(num_ops != req->r_num_ops);
1211                         index = pages[i - 1]->index + 1;
1212                         /* request message now owns the pages array */
1213                         pages = NULL;
1214                 }
1215
1216                 req->r_mtime = inode->i_mtime;
1217                 ceph_osdc_start_request(&fsc->client->osdc, req);
1218                 req = NULL;
1219
1220                 wbc->nr_to_write -= i;
1221                 if (pages)
1222                         goto new_request;
1223
1224                 /*
1225                  * We stop writing back only if we are not doing
1226                  * integrity sync. In case of integrity sync we have to
1227                  * keep going until we have written all the pages
1228                  * we tagged for writeback prior to entering this loop.
1229                  */
1230                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1231                         done = true;
1232
1233 release_folios:
1234                 dout("folio_batch release on %d folios (%p)\n", (int)fbatch.nr,
1235                      fbatch.nr ? fbatch.folios[0] : NULL);
1236                 folio_batch_release(&fbatch);
1237         }
1238
1239         if (should_loop && !done) {
1240                 /* more to do; loop back to beginning of file */
1241                 dout("writepages looping back to beginning of file\n");
1242                 end = start_index - 1; /* OK even when start_index == 0 */
1243
1244                 /* to write dirty pages associated with next snapc,
1245                  * we need to wait until current writes complete */
1246                 if (wbc->sync_mode != WB_SYNC_NONE &&
1247                     start_index == 0 && /* all dirty pages were checked */
1248                     !ceph_wbc.head_snapc) {
1249                         struct page *page;
1250                         unsigned i, nr;
1251                         index = 0;
1252                         while ((index <= end) &&
1253                                (nr = filemap_get_folios_tag(mapping, &index,
1254                                                 (pgoff_t)-1,
1255                                                 PAGECACHE_TAG_WRITEBACK,
1256                                                 &fbatch))) {
1257                                 for (i = 0; i < nr; i++) {
1258                                         page = &fbatch.folios[i]->page;
1259                                         if (page_snap_context(page) != snapc)
1260                                                 continue;
1261                                         wait_on_page_writeback(page);
1262                                 }
1263                                 folio_batch_release(&fbatch);
1264                                 cond_resched();
1265                         }
1266                 }
1267
1268                 start_index = 0;
1269                 index = 0;
1270                 goto retry;
1271         }
1272
1273         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1274                 mapping->writeback_index = index;
1275
1276 out:
1277         ceph_osdc_put_request(req);
1278         ceph_put_snap_context(last_snapc);
1279         dout("writepages dend - startone, rc = %d\n", rc);
1280         return rc;
1281 }
1282
1283
1284
1285 /*
1286  * See if a given @snapc is either writeable, or already written.
1287  */
1288 static int context_is_writeable_or_written(struct inode *inode,
1289                                            struct ceph_snap_context *snapc)
1290 {
1291         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1292         int ret = !oldest || snapc->seq <= oldest->seq;
1293
1294         ceph_put_snap_context(oldest);
1295         return ret;
1296 }
1297
1298 /**
1299  * ceph_find_incompatible - find an incompatible context and return it
1300  * @page: page being dirtied
1301  *
1302  * We are only allowed to write into/dirty a page if the page is
1303  * clean, or already dirty within the same snap context. Returns a
1304  * conflicting context if there is one, NULL if there isn't, or a
1305  * negative error code on other errors.
1306  *
1307  * Must be called with page lock held.
1308  */
1309 static struct ceph_snap_context *
1310 ceph_find_incompatible(struct page *page)
1311 {
1312         struct inode *inode = page->mapping->host;
1313         struct ceph_inode_info *ci = ceph_inode(inode);
1314
1315         if (ceph_inode_is_shutdown(inode)) {
1316                 dout(" page %p %llx:%llx is shutdown\n", page,
1317                      ceph_vinop(inode));
1318                 return ERR_PTR(-ESTALE);
1319         }
1320
1321         for (;;) {
1322                 struct ceph_snap_context *snapc, *oldest;
1323
1324                 wait_on_page_writeback(page);
1325
1326                 snapc = page_snap_context(page);
1327                 if (!snapc || snapc == ci->i_head_snapc)
1328                         break;
1329
1330                 /*
1331                  * this page is already dirty in another (older) snap
1332                  * context!  is it writeable now?
1333                  */
1334                 oldest = get_oldest_context(inode, NULL, NULL);
1335                 if (snapc->seq > oldest->seq) {
1336                         /* not writeable -- return it for the caller to deal with */
1337                         ceph_put_snap_context(oldest);
1338                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1339                         return ceph_get_snap_context(snapc);
1340                 }
1341                 ceph_put_snap_context(oldest);
1342
1343                 /* yay, writeable, do it now (without dropping page lock) */
1344                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1345                 if (clear_page_dirty_for_io(page)) {
1346                         int r = writepage_nounlock(page, NULL);
1347                         if (r < 0)
1348                                 return ERR_PTR(r);
1349                 }
1350         }
1351         return NULL;
1352 }
1353
1354 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1355                                         struct folio **foliop, void **_fsdata)
1356 {
1357         struct inode *inode = file_inode(file);
1358         struct ceph_inode_info *ci = ceph_inode(inode);
1359         struct ceph_snap_context *snapc;
1360
1361         snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1362         if (snapc) {
1363                 int r;
1364
1365                 folio_unlock(*foliop);
1366                 folio_put(*foliop);
1367                 *foliop = NULL;
1368                 if (IS_ERR(snapc))
1369                         return PTR_ERR(snapc);
1370
1371                 ceph_queue_writeback(inode);
1372                 r = wait_event_killable(ci->i_cap_wq,
1373                                         context_is_writeable_or_written(inode, snapc));
1374                 ceph_put_snap_context(snapc);
1375                 return r == 0 ? -EAGAIN : r;
1376         }
1377         return 0;
1378 }
1379
1380 /*
1381  * We are only allowed to write into/dirty the page if the page is
1382  * clean, or already dirty within the same snap context.
1383  */
1384 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1385                             loff_t pos, unsigned len,
1386                             struct page **pagep, void **fsdata)
1387 {
1388         struct inode *inode = file_inode(file);
1389         struct ceph_inode_info *ci = ceph_inode(inode);
1390         struct folio *folio = NULL;
1391         int r;
1392
1393         r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1394         if (r < 0)
1395                 return r;
1396
1397         folio_wait_fscache(folio);
1398         WARN_ON_ONCE(!folio_test_locked(folio));
1399         *pagep = &folio->page;
1400         return 0;
1401 }
1402
1403 /*
1404  * we don't do anything in here that simple_write_end doesn't do
1405  * except adjust dirty page accounting
1406  */
1407 static int ceph_write_end(struct file *file, struct address_space *mapping,
1408                           loff_t pos, unsigned len, unsigned copied,
1409                           struct page *subpage, void *fsdata)
1410 {
1411         struct folio *folio = page_folio(subpage);
1412         struct inode *inode = file_inode(file);
1413         bool check_cap = false;
1414
1415         dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1416              inode, folio, (int)pos, (int)copied, (int)len);
1417
1418         if (!folio_test_uptodate(folio)) {
1419                 /* just return that nothing was copied on a short copy */
1420                 if (copied < len) {
1421                         copied = 0;
1422                         goto out;
1423                 }
1424                 folio_mark_uptodate(folio);
1425         }
1426
1427         /* did file size increase? */
1428         if (pos+copied > i_size_read(inode))
1429                 check_cap = ceph_inode_set_size(inode, pos+copied);
1430
1431         folio_mark_dirty(folio);
1432
1433 out:
1434         folio_unlock(folio);
1435         folio_put(folio);
1436
1437         if (check_cap)
1438                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1439
1440         return copied;
1441 }
1442
1443 const struct address_space_operations ceph_aops = {
1444         .read_folio = netfs_read_folio,
1445         .readahead = netfs_readahead,
1446         .writepage = ceph_writepage,
1447         .writepages = ceph_writepages_start,
1448         .write_begin = ceph_write_begin,
1449         .write_end = ceph_write_end,
1450         .dirty_folio = ceph_dirty_folio,
1451         .invalidate_folio = ceph_invalidate_folio,
1452         .release_folio = ceph_release_folio,
1453         .direct_IO = noop_direct_IO,
1454 };
1455
1456 static void ceph_block_sigs(sigset_t *oldset)
1457 {
1458         sigset_t mask;
1459         siginitsetinv(&mask, sigmask(SIGKILL));
1460         sigprocmask(SIG_BLOCK, &mask, oldset);
1461 }
1462
1463 static void ceph_restore_sigs(sigset_t *oldset)
1464 {
1465         sigprocmask(SIG_SETMASK, oldset, NULL);
1466 }
1467
1468 /*
1469  * vm ops
1470  */
1471 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1472 {
1473         struct vm_area_struct *vma = vmf->vma;
1474         struct inode *inode = file_inode(vma->vm_file);
1475         struct ceph_inode_info *ci = ceph_inode(inode);
1476         struct ceph_file_info *fi = vma->vm_file->private_data;
1477         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1478         int want, got, err;
1479         sigset_t oldset;
1480         vm_fault_t ret = VM_FAULT_SIGBUS;
1481
1482         if (ceph_inode_is_shutdown(inode))
1483                 return ret;
1484
1485         ceph_block_sigs(&oldset);
1486
1487         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1488              inode, ceph_vinop(inode), off);
1489         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1490                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1491         else
1492                 want = CEPH_CAP_FILE_CACHE;
1493
1494         got = 0;
1495         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1496         if (err < 0)
1497                 goto out_restore;
1498
1499         dout("filemap_fault %p %llu got cap refs on %s\n",
1500              inode, off, ceph_cap_string(got));
1501
1502         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1503             !ceph_has_inline_data(ci)) {
1504                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1505                 ceph_add_rw_context(fi, &rw_ctx);
1506                 ret = filemap_fault(vmf);
1507                 ceph_del_rw_context(fi, &rw_ctx);
1508                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1509                      inode, off, ceph_cap_string(got), ret);
1510         } else
1511                 err = -EAGAIN;
1512
1513         ceph_put_cap_refs(ci, got);
1514
1515         if (err != -EAGAIN)
1516                 goto out_restore;
1517
1518         /* read inline data */
1519         if (off >= PAGE_SIZE) {
1520                 /* does not support inline data > PAGE_SIZE */
1521                 ret = VM_FAULT_SIGBUS;
1522         } else {
1523                 struct address_space *mapping = inode->i_mapping;
1524                 struct page *page;
1525
1526                 filemap_invalidate_lock_shared(mapping);
1527                 page = find_or_create_page(mapping, 0,
1528                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1529                 if (!page) {
1530                         ret = VM_FAULT_OOM;
1531                         goto out_inline;
1532                 }
1533                 err = __ceph_do_getattr(inode, page,
1534                                          CEPH_STAT_CAP_INLINE_DATA, true);
1535                 if (err < 0 || off >= i_size_read(inode)) {
1536                         unlock_page(page);
1537                         put_page(page);
1538                         ret = vmf_error(err);
1539                         goto out_inline;
1540                 }
1541                 if (err < PAGE_SIZE)
1542                         zero_user_segment(page, err, PAGE_SIZE);
1543                 else
1544                         flush_dcache_page(page);
1545                 SetPageUptodate(page);
1546                 vmf->page = page;
1547                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1548 out_inline:
1549                 filemap_invalidate_unlock_shared(mapping);
1550                 dout("filemap_fault %p %llu read inline data ret %x\n",
1551                      inode, off, ret);
1552         }
1553 out_restore:
1554         ceph_restore_sigs(&oldset);
1555         if (err < 0)
1556                 ret = vmf_error(err);
1557
1558         return ret;
1559 }
1560
1561 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1562 {
1563         struct vm_area_struct *vma = vmf->vma;
1564         struct inode *inode = file_inode(vma->vm_file);
1565         struct ceph_inode_info *ci = ceph_inode(inode);
1566         struct ceph_file_info *fi = vma->vm_file->private_data;
1567         struct ceph_cap_flush *prealloc_cf;
1568         struct page *page = vmf->page;
1569         loff_t off = page_offset(page);
1570         loff_t size = i_size_read(inode);
1571         size_t len;
1572         int want, got, err;
1573         sigset_t oldset;
1574         vm_fault_t ret = VM_FAULT_SIGBUS;
1575
1576         if (ceph_inode_is_shutdown(inode))
1577                 return ret;
1578
1579         prealloc_cf = ceph_alloc_cap_flush();
1580         if (!prealloc_cf)
1581                 return VM_FAULT_OOM;
1582
1583         sb_start_pagefault(inode->i_sb);
1584         ceph_block_sigs(&oldset);
1585
1586         if (off + thp_size(page) <= size)
1587                 len = thp_size(page);
1588         else
1589                 len = offset_in_thp(page, size);
1590
1591         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1592              inode, ceph_vinop(inode), off, len, size);
1593         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1594                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1595         else
1596                 want = CEPH_CAP_FILE_BUFFER;
1597
1598         got = 0;
1599         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1600         if (err < 0)
1601                 goto out_free;
1602
1603         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1604              inode, off, len, ceph_cap_string(got));
1605
1606         /* Update time before taking page lock */
1607         file_update_time(vma->vm_file);
1608         inode_inc_iversion_raw(inode);
1609
1610         do {
1611                 struct ceph_snap_context *snapc;
1612
1613                 lock_page(page);
1614
1615                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1616                         unlock_page(page);
1617                         ret = VM_FAULT_NOPAGE;
1618                         break;
1619                 }
1620
1621                 snapc = ceph_find_incompatible(page);
1622                 if (!snapc) {
1623                         /* success.  we'll keep the page locked. */
1624                         set_page_dirty(page);
1625                         ret = VM_FAULT_LOCKED;
1626                         break;
1627                 }
1628
1629                 unlock_page(page);
1630
1631                 if (IS_ERR(snapc)) {
1632                         ret = VM_FAULT_SIGBUS;
1633                         break;
1634                 }
1635
1636                 ceph_queue_writeback(inode);
1637                 err = wait_event_killable(ci->i_cap_wq,
1638                                 context_is_writeable_or_written(inode, snapc));
1639                 ceph_put_snap_context(snapc);
1640         } while (err == 0);
1641
1642         if (ret == VM_FAULT_LOCKED) {
1643                 int dirty;
1644                 spin_lock(&ci->i_ceph_lock);
1645                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1646                                                &prealloc_cf);
1647                 spin_unlock(&ci->i_ceph_lock);
1648                 if (dirty)
1649                         __mark_inode_dirty(inode, dirty);
1650         }
1651
1652         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1653              inode, off, len, ceph_cap_string(got), ret);
1654         ceph_put_cap_refs_async(ci, got);
1655 out_free:
1656         ceph_restore_sigs(&oldset);
1657         sb_end_pagefault(inode->i_sb);
1658         ceph_free_cap_flush(prealloc_cf);
1659         if (err < 0)
1660                 ret = vmf_error(err);
1661         return ret;
1662 }
1663
1664 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1665                            char *data, size_t len)
1666 {
1667         struct address_space *mapping = inode->i_mapping;
1668         struct page *page;
1669
1670         if (locked_page) {
1671                 page = locked_page;
1672         } else {
1673                 if (i_size_read(inode) == 0)
1674                         return;
1675                 page = find_or_create_page(mapping, 0,
1676                                            mapping_gfp_constraint(mapping,
1677                                            ~__GFP_FS));
1678                 if (!page)
1679                         return;
1680                 if (PageUptodate(page)) {
1681                         unlock_page(page);
1682                         put_page(page);
1683                         return;
1684                 }
1685         }
1686
1687         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1688              inode, ceph_vinop(inode), len, locked_page);
1689
1690         if (len > 0) {
1691                 void *kaddr = kmap_atomic(page);
1692                 memcpy(kaddr, data, len);
1693                 kunmap_atomic(kaddr);
1694         }
1695
1696         if (page != locked_page) {
1697                 if (len < PAGE_SIZE)
1698                         zero_user_segment(page, len, PAGE_SIZE);
1699                 else
1700                         flush_dcache_page(page);
1701
1702                 SetPageUptodate(page);
1703                 unlock_page(page);
1704                 put_page(page);
1705         }
1706 }
1707
1708 int ceph_uninline_data(struct file *file)
1709 {
1710         struct inode *inode = file_inode(file);
1711         struct ceph_inode_info *ci = ceph_inode(inode);
1712         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1713         struct ceph_osd_request *req = NULL;
1714         struct ceph_cap_flush *prealloc_cf = NULL;
1715         struct folio *folio = NULL;
1716         u64 inline_version = CEPH_INLINE_NONE;
1717         struct page *pages[1];
1718         int err = 0;
1719         u64 len;
1720
1721         spin_lock(&ci->i_ceph_lock);
1722         inline_version = ci->i_inline_version;
1723         spin_unlock(&ci->i_ceph_lock);
1724
1725         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1726              inode, ceph_vinop(inode), inline_version);
1727
1728         if (ceph_inode_is_shutdown(inode)) {
1729                 err = -EIO;
1730                 goto out;
1731         }
1732
1733         if (inline_version == CEPH_INLINE_NONE)
1734                 return 0;
1735
1736         prealloc_cf = ceph_alloc_cap_flush();
1737         if (!prealloc_cf)
1738                 return -ENOMEM;
1739
1740         if (inline_version == 1) /* initial version, no data */
1741                 goto out_uninline;
1742
1743         folio = read_mapping_folio(inode->i_mapping, 0, file);
1744         if (IS_ERR(folio)) {
1745                 err = PTR_ERR(folio);
1746                 goto out;
1747         }
1748
1749         folio_lock(folio);
1750
1751         len = i_size_read(inode);
1752         if (len > folio_size(folio))
1753                 len = folio_size(folio);
1754
1755         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1756                                     ceph_vino(inode), 0, &len, 0, 1,
1757                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1758                                     NULL, 0, 0, false);
1759         if (IS_ERR(req)) {
1760                 err = PTR_ERR(req);
1761                 goto out_unlock;
1762         }
1763
1764         req->r_mtime = inode->i_mtime;
1765         ceph_osdc_start_request(&fsc->client->osdc, req);
1766         err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1767         ceph_osdc_put_request(req);
1768         if (err < 0)
1769                 goto out_unlock;
1770
1771         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1772                                     ceph_vino(inode), 0, &len, 1, 3,
1773                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1774                                     NULL, ci->i_truncate_seq,
1775                                     ci->i_truncate_size, false);
1776         if (IS_ERR(req)) {
1777                 err = PTR_ERR(req);
1778                 goto out_unlock;
1779         }
1780
1781         pages[0] = folio_page(folio, 0);
1782         osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1783
1784         {
1785                 __le64 xattr_buf = cpu_to_le64(inline_version);
1786                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1787                                             "inline_version", &xattr_buf,
1788                                             sizeof(xattr_buf),
1789                                             CEPH_OSD_CMPXATTR_OP_GT,
1790                                             CEPH_OSD_CMPXATTR_MODE_U64);
1791                 if (err)
1792                         goto out_put_req;
1793         }
1794
1795         {
1796                 char xattr_buf[32];
1797                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1798                                          "%llu", inline_version);
1799                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1800                                             "inline_version",
1801                                             xattr_buf, xattr_len, 0, 0);
1802                 if (err)
1803                         goto out_put_req;
1804         }
1805
1806         req->r_mtime = inode->i_mtime;
1807         ceph_osdc_start_request(&fsc->client->osdc, req);
1808         err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1809
1810         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1811                                   req->r_end_latency, len, err);
1812
1813 out_uninline:
1814         if (!err) {
1815                 int dirty;
1816
1817                 /* Set to CAP_INLINE_NONE and dirty the caps */
1818                 down_read(&fsc->mdsc->snap_rwsem);
1819                 spin_lock(&ci->i_ceph_lock);
1820                 ci->i_inline_version = CEPH_INLINE_NONE;
1821                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1822                 spin_unlock(&ci->i_ceph_lock);
1823                 up_read(&fsc->mdsc->snap_rwsem);
1824                 if (dirty)
1825                         __mark_inode_dirty(inode, dirty);
1826         }
1827 out_put_req:
1828         ceph_osdc_put_request(req);
1829         if (err == -ECANCELED)
1830                 err = 0;
1831 out_unlock:
1832         if (folio) {
1833                 folio_unlock(folio);
1834                 folio_put(folio);
1835         }
1836 out:
1837         ceph_free_cap_flush(prealloc_cf);
1838         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1839              inode, ceph_vinop(inode), inline_version, err);
1840         return err;
1841 }
1842
1843 static const struct vm_operations_struct ceph_vmops = {
1844         .fault          = ceph_filemap_fault,
1845         .page_mkwrite   = ceph_page_mkwrite,
1846 };
1847
1848 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1849 {
1850         struct address_space *mapping = file->f_mapping;
1851
1852         if (!mapping->a_ops->read_folio)
1853                 return -ENOEXEC;
1854         vma->vm_ops = &ceph_vmops;
1855         return 0;
1856 }
1857
1858 enum {
1859         POOL_READ       = 1,
1860         POOL_WRITE      = 2,
1861 };
1862
1863 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1864                                 s64 pool, struct ceph_string *pool_ns)
1865 {
1866         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->netfs.inode);
1867         struct ceph_mds_client *mdsc = fsc->mdsc;
1868         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1869         struct rb_node **p, *parent;
1870         struct ceph_pool_perm *perm;
1871         struct page **pages;
1872         size_t pool_ns_len;
1873         int err = 0, err2 = 0, have = 0;
1874
1875         down_read(&mdsc->pool_perm_rwsem);
1876         p = &mdsc->pool_perm_tree.rb_node;
1877         while (*p) {
1878                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1879                 if (pool < perm->pool)
1880                         p = &(*p)->rb_left;
1881                 else if (pool > perm->pool)
1882                         p = &(*p)->rb_right;
1883                 else {
1884                         int ret = ceph_compare_string(pool_ns,
1885                                                 perm->pool_ns,
1886                                                 perm->pool_ns_len);
1887                         if (ret < 0)
1888                                 p = &(*p)->rb_left;
1889                         else if (ret > 0)
1890                                 p = &(*p)->rb_right;
1891                         else {
1892                                 have = perm->perm;
1893                                 break;
1894                         }
1895                 }
1896         }
1897         up_read(&mdsc->pool_perm_rwsem);
1898         if (*p)
1899                 goto out;
1900
1901         if (pool_ns)
1902                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1903                      pool, (int)pool_ns->len, pool_ns->str);
1904         else
1905                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1906
1907         down_write(&mdsc->pool_perm_rwsem);
1908         p = &mdsc->pool_perm_tree.rb_node;
1909         parent = NULL;
1910         while (*p) {
1911                 parent = *p;
1912                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1913                 if (pool < perm->pool)
1914                         p = &(*p)->rb_left;
1915                 else if (pool > perm->pool)
1916                         p = &(*p)->rb_right;
1917                 else {
1918                         int ret = ceph_compare_string(pool_ns,
1919                                                 perm->pool_ns,
1920                                                 perm->pool_ns_len);
1921                         if (ret < 0)
1922                                 p = &(*p)->rb_left;
1923                         else if (ret > 0)
1924                                 p = &(*p)->rb_right;
1925                         else {
1926                                 have = perm->perm;
1927                                 break;
1928                         }
1929                 }
1930         }
1931         if (*p) {
1932                 up_write(&mdsc->pool_perm_rwsem);
1933                 goto out;
1934         }
1935
1936         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1937                                          1, false, GFP_NOFS);
1938         if (!rd_req) {
1939                 err = -ENOMEM;
1940                 goto out_unlock;
1941         }
1942
1943         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1944         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1945         rd_req->r_base_oloc.pool = pool;
1946         if (pool_ns)
1947                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1948         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1949
1950         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1951         if (err)
1952                 goto out_unlock;
1953
1954         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1955                                          1, false, GFP_NOFS);
1956         if (!wr_req) {
1957                 err = -ENOMEM;
1958                 goto out_unlock;
1959         }
1960
1961         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1962         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1963         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1964         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1965
1966         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1967         if (err)
1968                 goto out_unlock;
1969
1970         /* one page should be large enough for STAT data */
1971         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1972         if (IS_ERR(pages)) {
1973                 err = PTR_ERR(pages);
1974                 goto out_unlock;
1975         }
1976
1977         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1978                                      0, false, true);
1979         ceph_osdc_start_request(&fsc->client->osdc, rd_req);
1980
1981         wr_req->r_mtime = ci->netfs.inode.i_mtime;
1982         ceph_osdc_start_request(&fsc->client->osdc, wr_req);
1983
1984         err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1985         err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1986
1987         if (err >= 0 || err == -ENOENT)
1988                 have |= POOL_READ;
1989         else if (err != -EPERM) {
1990                 if (err == -EBLOCKLISTED)
1991                         fsc->blocklisted = true;
1992                 goto out_unlock;
1993         }
1994
1995         if (err2 == 0 || err2 == -EEXIST)
1996                 have |= POOL_WRITE;
1997         else if (err2 != -EPERM) {
1998                 if (err2 == -EBLOCKLISTED)
1999                         fsc->blocklisted = true;
2000                 err = err2;
2001                 goto out_unlock;
2002         }
2003
2004         pool_ns_len = pool_ns ? pool_ns->len : 0;
2005         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2006         if (!perm) {
2007                 err = -ENOMEM;
2008                 goto out_unlock;
2009         }
2010
2011         perm->pool = pool;
2012         perm->perm = have;
2013         perm->pool_ns_len = pool_ns_len;
2014         if (pool_ns_len > 0)
2015                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2016         perm->pool_ns[pool_ns_len] = 0;
2017
2018         rb_link_node(&perm->node, parent, p);
2019         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2020         err = 0;
2021 out_unlock:
2022         up_write(&mdsc->pool_perm_rwsem);
2023
2024         ceph_osdc_put_request(rd_req);
2025         ceph_osdc_put_request(wr_req);
2026 out:
2027         if (!err)
2028                 err = have;
2029         if (pool_ns)
2030                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2031                      pool, (int)pool_ns->len, pool_ns->str, err);
2032         else
2033                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2034         return err;
2035 }
2036
2037 int ceph_pool_perm_check(struct inode *inode, int need)
2038 {
2039         struct ceph_inode_info *ci = ceph_inode(inode);
2040         struct ceph_string *pool_ns;
2041         s64 pool;
2042         int ret, flags;
2043
2044         /* Only need to do this for regular files */
2045         if (!S_ISREG(inode->i_mode))
2046                 return 0;
2047
2048         if (ci->i_vino.snap != CEPH_NOSNAP) {
2049                 /*
2050                  * Pool permission check needs to write to the first object.
2051                  * But for snapshot, head of the first object may have alread
2052                  * been deleted. Skip check to avoid creating orphan object.
2053                  */
2054                 return 0;
2055         }
2056
2057         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2058                                 NOPOOLPERM))
2059                 return 0;
2060
2061         spin_lock(&ci->i_ceph_lock);
2062         flags = ci->i_ceph_flags;
2063         pool = ci->i_layout.pool_id;
2064         spin_unlock(&ci->i_ceph_lock);
2065 check:
2066         if (flags & CEPH_I_POOL_PERM) {
2067                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2068                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2069                              pool);
2070                         return -EPERM;
2071                 }
2072                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2073                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2074                              pool);
2075                         return -EPERM;
2076                 }
2077                 return 0;
2078         }
2079
2080         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2081         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2082         ceph_put_string(pool_ns);
2083         if (ret < 0)
2084                 return ret;
2085
2086         flags = CEPH_I_POOL_PERM;
2087         if (ret & POOL_READ)
2088                 flags |= CEPH_I_POOL_RD;
2089         if (ret & POOL_WRITE)
2090                 flags |= CEPH_I_POOL_WR;
2091
2092         spin_lock(&ci->i_ceph_lock);
2093         if (pool == ci->i_layout.pool_id &&
2094             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2095                 ci->i_ceph_flags |= flags;
2096         } else {
2097                 pool = ci->i_layout.pool_id;
2098                 flags = ci->i_ceph_flags;
2099         }
2100         spin_unlock(&ci->i_ceph_lock);
2101         goto check;
2102 }
2103
2104 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2105 {
2106         struct ceph_pool_perm *perm;
2107         struct rb_node *n;
2108
2109         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2110                 n = rb_first(&mdsc->pool_perm_tree);
2111                 perm = rb_entry(n, struct ceph_pool_perm, node);
2112                 rb_erase(n, &mdsc->pool_perm_tree);
2113                 kfree(perm);
2114         }
2115 }