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