fs/ceph/addr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/backing-dev.h>
   5 #include <linux/fs.h>
   6 #include <linux/mm.h>
   7 #include <linux/swap.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/slab.h>
  10 #include <linux/pagevec.h>
  11 #include <linux/task_io_accounting_ops.h>
  12 #include <linux/signal.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15 #include <linux/netfs.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include "metric.h"
  21 #include <linux/ceph/osd_client.h>
  22 #include <linux/ceph/striper.h>
  23
  24 /*
  25  * Ceph address space ops.
  26  *
  27  * There are a few funny things going on here.
  28  *
  29  * The page->private field is used to reference a struct
  30  * ceph_snap_context for _every_ dirty page.  This indicates which
  31  * snapshot the page was logically dirtied in, and thus which snap
  32  * context needs to be associated with the osd write during writeback.
  33  *
  34  * Similarly, struct ceph_inode_info maintains a set of counters to
  35  * count dirty pages on the inode.  In the absence of snapshots,
  36  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  37  *
  38  * When a snapshot is taken (that is, when the client receives
  39  * notification that a snapshot was taken), each inode with caps and
  40  * with dirty pages (dirty pages implies there is a cap) gets a new
  41  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  42  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  43  * moved to capsnap->dirty. (Unless a sync write is currently in
  44  * progress.  In that case, the capsnap is said to be "pending", new
  45  * writes cannot start, and the capsnap isn't "finalized" until the
  46  * write completes (or fails) and a final size/mtime for the inode for
  47  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  48  *
  49  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  50  * we look for the first capsnap in i_cap_snaps and write out pages in
  51  * that snap context _only_.  Then we move on to the next capsnap,
  52  * eventually reaching the "live" or "head" context (i.e., pages that
  53  * are not yet snapped) and are writing the most recently dirtied
  54  * pages.
  55  *
  56  * Invalidate and so forth must take care to ensure the dirty page
  57  * accounting is preserved.
  58  */
  59
  60 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  61 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  62         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  63          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  64
  65 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  66                                         struct folio *folio, void **_fsdata);
  67
  68 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  69 {
  70         if (PagePrivate(page))
  71                 return (void *)page->private;
  72         return NULL;
  73 }
  74
  75 /*
  76  * Dirty a page.  Optimistically adjust accounting, on the assumption
  77  * that we won't race with invalidate.  If we do, readjust.
  78  */
  79 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
  80 {
  81         struct inode *inode;
  82         struct ceph_inode_info *ci;
  83         struct ceph_snap_context *snapc;
  84
  85         if (folio_test_dirty(folio)) {
  86                 dout("%p dirty_folio %p idx %lu -- already dirty\n",
  87                      mapping->host, folio, folio->index);
  88                 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
  89                 return false;
  90         }
  91
  92         inode = mapping->host;
  93         ci = ceph_inode(inode);
  94
  95         /* dirty the head */
  96         spin_lock(&ci->i_ceph_lock);
  97         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  98         if (__ceph_have_pending_cap_snap(ci)) {
  99                 struct ceph_cap_snap *capsnap =
 100                                 list_last_entry(&ci->i_cap_snaps,
 101                                                 struct ceph_cap_snap,
 102                                                 ci_item);
 103                 snapc = ceph_get_snap_context(capsnap->context);
 104                 capsnap->dirty_pages++;
 105         } else {
 106                 BUG_ON(!ci->i_head_snapc);
 107                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 108                 ++ci->i_wrbuffer_ref_head;
 109         }
 110         if (ci->i_wrbuffer_ref == 0)
 111                 ihold(inode);
 112         ++ci->i_wrbuffer_ref;
 113         dout("%p dirty_folio %p idx %lu head %d/%d -> %d/%d "
 114              "snapc %p seq %lld (%d snaps)\n",
 115              mapping->host, folio, folio->index,
 116              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 117              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 118              snapc, snapc->seq, snapc->num_snaps);
 119         spin_unlock(&ci->i_ceph_lock);
 120
 121         /*
 122          * Reference snap context in folio->private.  Also set
 123          * PagePrivate so that we get invalidate_folio callback.
 124          */
 125         VM_BUG_ON_FOLIO(folio_test_private(folio), folio);
 126         folio_attach_private(folio, snapc);
 127
 128         return ceph_fscache_dirty_folio(mapping, folio);
 129 }
 130
 131 /*
 132  * If we are truncating the full folio (i.e. offset == 0), adjust the
 133  * dirty folio counters appropriately.  Only called if there is private
 134  * data on the folio.
 135  */
 136 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 137                                 size_t length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc;
 142
 143         inode = folio->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != folio_size(folio)) {
 147                 dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
 148                      inode, folio->index, offset, length);
 149                 return;
 150         }
 151
 152         WARN_ON(!folio_test_locked(folio));
 153         if (folio_test_private(folio)) {
 154                 dout("%p invalidate_folio idx %lu full dirty page\n",
 155                      inode, folio->index);
 156
 157                 snapc = folio_detach_private(folio);
 158                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159                 ceph_put_snap_context(snapc);
 160         }
 161
 162         folio_wait_fscache(folio);
 163 }
 164
 165 static int ceph_releasepage(struct page *page, gfp_t gfp)
 166 {
 167         struct inode *inode = page->mapping->host;
 168
 169         dout("%llx:%llx releasepage %p idx %lu (%sdirty)\n",
 170              ceph_vinop(inode), page,
 171              page->index, PageDirty(page) ? "" : "not ");
 172
 173         if (PagePrivate(page))
 174                 return 0;
 175
 176         if (PageFsCache(page)) {
 177                 if (current_is_kswapd() || !(gfp & __GFP_FS))
 178                         return 0;
 179                 wait_on_page_fscache(page);
 180         }
 181         ceph_fscache_note_page_release(inode);
 182         return 1;
 183 }
 184
 185 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 186 {
 187         struct inode *inode = rreq->inode;
 188         struct ceph_inode_info *ci = ceph_inode(inode);
 189         struct ceph_file_layout *lo = &ci->i_layout;
 190         u32 blockoff;
 191         u64 blockno;
 192
 193         /* Expand the start downward */
 194         blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 195         rreq->start = blockno * lo->stripe_unit;
 196         rreq->len += blockoff;
 197
 198         /* Now, round up the length to the next block */
 199         rreq->len = roundup(rreq->len, lo->stripe_unit);
 200 }
 201
 202 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
 203 {
 204         struct inode *inode = subreq->rreq->inode;
 205         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 206         struct ceph_inode_info *ci = ceph_inode(inode);
 207         u64 objno, objoff;
 208         u32 xlen;
 209
 210         /* Truncate the extent at the end of the current block */
 211         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 212                                       &objno, &objoff, &xlen);
 213         subreq->len = min(xlen, fsc->mount_options->rsize);
 214         return true;
 215 }
 216
 217 static void finish_netfs_read(struct ceph_osd_request *req)
 218 {
 219         struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 220         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 221         struct netfs_io_subrequest *subreq = req->r_priv;
 222         int num_pages;
 223         int err = req->r_result;
 224
 225         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 226                                  req->r_end_latency, osd_data->length, err);
 227
 228         dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 229              subreq->len, i_size_read(req->r_inode));
 230
 231         /* no object means success but no data */
 232         if (err == -ENOENT)
 233                 err = 0;
 234         else if (err == -EBLOCKLISTED)
 235                 fsc->blocklisted = true;
 236
 237         if (err >= 0 && err < subreq->len)
 238                 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 239
 240         netfs_subreq_terminated(subreq, err, true);
 241
 242         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 243         ceph_put_page_vector(osd_data->pages, num_pages, false);
 244         iput(req->r_inode);
 245 }
 246
 247 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 248 {
 249         struct netfs_io_request *rreq = subreq->rreq;
 250         struct inode *inode = rreq->inode;
 251         struct ceph_mds_reply_info_parsed *rinfo;
 252         struct ceph_mds_reply_info_in *iinfo;
 253         struct ceph_mds_request *req;
 254         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 255         struct ceph_inode_info *ci = ceph_inode(inode);
 256         struct iov_iter iter;
 257         ssize_t err = 0;
 258         size_t len;
 259         int mode;
 260
 261         __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 262         __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 263
 264         if (subreq->start >= inode->i_size)
 265                 goto out;
 266
 267         /* We need to fetch the inline data. */
 268         mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
 269         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
 270         if (IS_ERR(req)) {
 271                 err = PTR_ERR(req);
 272                 goto out;
 273         }
 274         req->r_ino1 = ci->i_vino;
 275         req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 276         req->r_num_caps = 2;
 277
 278         err = ceph_mdsc_do_request(mdsc, NULL, req);
 279         if (err < 0)
 280                 goto out;
 281
 282         rinfo = &req->r_reply_info;
 283         iinfo = &rinfo->targeti;
 284         if (iinfo->inline_version == CEPH_INLINE_NONE) {
 285                 /* The data got uninlined */
 286                 ceph_mdsc_put_request(req);
 287                 return false;
 288         }
 289
 290         len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
 291         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 292         err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
 293         if (err == 0)
 294                 err = -EFAULT;
 295
 296         ceph_mdsc_put_request(req);
 297 out:
 298         netfs_subreq_terminated(subreq, err, false);
 299         return true;
 300 }
 301
 302 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 303 {
 304         struct netfs_io_request *rreq = subreq->rreq;
 305         struct inode *inode = rreq->inode;
 306         struct ceph_inode_info *ci = ceph_inode(inode);
 307         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 308         struct ceph_osd_request *req;
 309         struct ceph_vino vino = ceph_vino(inode);
 310         struct iov_iter iter;
 311         struct page **pages;
 312         size_t page_off;
 313         int err = 0;
 314         u64 len = subreq->len;
 315
 316         if (ci->i_inline_version != CEPH_INLINE_NONE &&
 317             ceph_netfs_issue_op_inline(subreq))
 318                 return;
 319
 320         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 321                         0, 1, CEPH_OSD_OP_READ,
 322                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 323                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 324         if (IS_ERR(req)) {
 325                 err = PTR_ERR(req);
 326                 req = NULL;
 327                 goto out;
 328         }
 329
 330         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 331         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 332         err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 333         if (err < 0) {
 334                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 335                 goto out;
 336         }
 337
 338         /* should always give us a page-aligned read */
 339         WARN_ON_ONCE(page_off);
 340         len = err;
 341
 342         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 343         req->r_callback = finish_netfs_read;
 344         req->r_priv = subreq;
 345         req->r_inode = inode;
 346         ihold(inode);
 347
 348         err = ceph_osdc_start_request(req->r_osdc, req, false);
 349         if (err)
 350                 iput(inode);
 351 out:
 352         ceph_osdc_put_request(req);
 353         if (err)
 354                 netfs_subreq_terminated(subreq, err, false);
 355         dout("%s: result %d\n", __func__, err);
 356 }
 357
 358 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 359 {
 360         struct inode *inode = rreq->inode;
 361         int got = 0, want = CEPH_CAP_FILE_CACHE;
 362         int ret = 0;
 363
 364         if (rreq->origin != NETFS_READAHEAD)
 365                 return 0;
 366
 367         if (file) {
 368                 struct ceph_rw_context *rw_ctx;
 369                 struct ceph_file_info *fi = file->private_data;
 370
 371                 rw_ctx = ceph_find_rw_context(fi);
 372                 if (rw_ctx)
 373                         return 0;
 374         }
 375
 376         /*
 377          * readahead callers do not necessarily hold Fcb caps
 378          * (e.g. fadvise, madvise).
 379          */
 380         ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 381         if (ret < 0) {
 382                 dout("start_read %p, error getting cap\n", inode);
 383                 return ret;
 384         }
 385
 386         if (!(got & want)) {
 387                 dout("start_read %p, no cache cap\n", inode);
 388                 return -EACCES;
 389         }
 390         if (ret == 0)
 391                 return -EACCES;
 392
 393         rreq->netfs_priv = (void *)(uintptr_t)got;
 394         return 0;
 395 }
 396
 397 static void ceph_readahead_cleanup(struct address_space *mapping, void *priv)
 398 {
 399         struct inode *inode = mapping->host;
 400         struct ceph_inode_info *ci = ceph_inode(inode);
 401         int got = (uintptr_t)priv;
 402
 403         if (got)
 404                 ceph_put_cap_refs(ci, got);
 405 }
 406
 407 const struct netfs_request_ops ceph_netfs_ops = {
 408         .init_request           = ceph_init_request,
 409         .begin_cache_operation  = ceph_begin_cache_operation,
 410         .issue_read             = ceph_netfs_issue_read,
 411         .expand_readahead       = ceph_netfs_expand_readahead,
 412         .clamp_length           = ceph_netfs_clamp_length,
 413         .check_write_begin      = ceph_netfs_check_write_begin,
 414         .cleanup                = ceph_readahead_cleanup,
 415 };
 416
 417 #ifdef CONFIG_CEPH_FSCACHE
 418 static void ceph_set_page_fscache(struct page *page)
 419 {
 420         set_page_fscache(page);
 421 }
 422
 423 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 424 {
 425         struct inode *inode = priv;
 426
 427         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 428                 ceph_fscache_invalidate(inode, false);
 429 }
 430
 431 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 432 {
 433         struct ceph_inode_info *ci = ceph_inode(inode);
 434         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 435
 436         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 437                                ceph_fscache_write_terminated, inode, caching);
 438 }
 439 #else
 440 static inline void ceph_set_page_fscache(struct page *page)
 441 {
 442 }
 443
 444 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 445 {
 446 }
 447 #endif /* CONFIG_CEPH_FSCACHE */
 448
 449 struct ceph_writeback_ctl
 450 {
 451         loff_t i_size;
 452         u64 truncate_size;
 453         u32 truncate_seq;
 454         bool size_stable;
 455         bool head_snapc;
 456 };
 457
 458 /*
 459  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 460  * only snap context we are allowed to write back.
 461  */
 462 static struct ceph_snap_context *
 463 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 464                    struct ceph_snap_context *page_snapc)
 465 {
 466         struct ceph_inode_info *ci = ceph_inode(inode);
 467         struct ceph_snap_context *snapc = NULL;
 468         struct ceph_cap_snap *capsnap = NULL;
 469
 470         spin_lock(&ci->i_ceph_lock);
 471         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 472                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 473                      capsnap->context, capsnap->dirty_pages);
 474                 if (!capsnap->dirty_pages)
 475                         continue;
 476
 477                 /* get i_size, truncate_{seq,size} for page_snapc? */
 478                 if (snapc && capsnap->context != page_snapc)
 479                         continue;
 480
 481                 if (ctl) {
 482                         if (capsnap->writing) {
 483                                 ctl->i_size = i_size_read(inode);
 484                                 ctl->size_stable = false;
 485                         } else {
 486                                 ctl->i_size = capsnap->size;
 487                                 ctl->size_stable = true;
 488                         }
 489                         ctl->truncate_size = capsnap->truncate_size;
 490                         ctl->truncate_seq = capsnap->truncate_seq;
 491                         ctl->head_snapc = false;
 492                 }
 493
 494                 if (snapc)
 495                         break;
 496
 497                 snapc = ceph_get_snap_context(capsnap->context);
 498                 if (!page_snapc ||
 499                     page_snapc == snapc ||
 500                     page_snapc->seq > snapc->seq)
 501                         break;
 502         }
 503         if (!snapc && ci->i_wrbuffer_ref_head) {
 504                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 505                 dout(" head snapc %p has %d dirty pages\n",
 506                      snapc, ci->i_wrbuffer_ref_head);
 507                 if (ctl) {
 508                         ctl->i_size = i_size_read(inode);
 509                         ctl->truncate_size = ci->i_truncate_size;
 510                         ctl->truncate_seq = ci->i_truncate_seq;
 511                         ctl->size_stable = false;
 512                         ctl->head_snapc = true;
 513                 }
 514         }
 515         spin_unlock(&ci->i_ceph_lock);
 516         return snapc;
 517 }
 518
 519 static u64 get_writepages_data_length(struct inode *inode,
 520                                       struct page *page, u64 start)
 521 {
 522         struct ceph_inode_info *ci = ceph_inode(inode);
 523         struct ceph_snap_context *snapc = page_snap_context(page);
 524         struct ceph_cap_snap *capsnap = NULL;
 525         u64 end = i_size_read(inode);
 526
 527         if (snapc != ci->i_head_snapc) {
 528                 bool found = false;
 529                 spin_lock(&ci->i_ceph_lock);
 530                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 531                         if (capsnap->context == snapc) {
 532                                 if (!capsnap->writing)
 533                                         end = capsnap->size;
 534                                 found = true;
 535                                 break;
 536                         }
 537                 }
 538                 spin_unlock(&ci->i_ceph_lock);
 539                 WARN_ON(!found);
 540         }
 541         if (end > page_offset(page) + thp_size(page))
 542                 end = page_offset(page) + thp_size(page);
 543         return end > start ? end - start : 0;
 544 }
 545
 546 /*
 547  * Write a single page, but leave the page locked.
 548  *
 549  * If we get a write error, mark the mapping for error, but still adjust the
 550  * dirty page accounting (i.e., page is no longer dirty).
 551  */
 552 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 553 {
 554         struct folio *folio = page_folio(page);
 555         struct inode *inode = page->mapping->host;
 556         struct ceph_inode_info *ci = ceph_inode(inode);
 557         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 558         struct ceph_snap_context *snapc, *oldest;
 559         loff_t page_off = page_offset(page);
 560         int err;
 561         loff_t len = thp_size(page);
 562         struct ceph_writeback_ctl ceph_wbc;
 563         struct ceph_osd_client *osdc = &fsc->client->osdc;
 564         struct ceph_osd_request *req;
 565         bool caching = ceph_is_cache_enabled(inode);
 566
 567         dout("writepage %p idx %lu\n", page, page->index);
 568
 569         /* verify this is a writeable snap context */
 570         snapc = page_snap_context(page);
 571         if (!snapc) {
 572                 dout("writepage %p page %p not dirty?\n", inode, page);
 573                 return 0;
 574         }
 575         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 576         if (snapc->seq > oldest->seq) {
 577                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 578                      inode, page, snapc);
 579                 /* we should only noop if called by kswapd */
 580                 WARN_ON(!(current->flags & PF_MEMALLOC));
 581                 ceph_put_snap_context(oldest);
 582                 redirty_page_for_writepage(wbc, page);
 583                 return 0;
 584         }
 585         ceph_put_snap_context(oldest);
 586
 587         /* is this a partial page at end of file? */
 588         if (page_off >= ceph_wbc.i_size) {
 589                 dout("folio at %lu beyond eof %llu\n", folio->index,
 590                                 ceph_wbc.i_size);
 591                 folio_invalidate(folio, 0, folio_size(folio));
 592                 return 0;
 593         }
 594
 595         if (ceph_wbc.i_size < page_off + len)
 596                 len = ceph_wbc.i_size - page_off;
 597
 598         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 599              inode, page, page->index, page_off, len, snapc, snapc->seq);
 600
 601         if (atomic_long_inc_return(&fsc->writeback_count) >
 602             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 603                 fsc->write_congested = true;
 604
 605         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 606                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 607                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 608                                     true);
 609         if (IS_ERR(req)) {
 610                 redirty_page_for_writepage(wbc, page);
 611                 return PTR_ERR(req);
 612         }
 613
 614         set_page_writeback(page);
 615         if (caching)
 616                 ceph_set_page_fscache(page);
 617         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 618
 619         /* it may be a short write due to an object boundary */
 620         WARN_ON_ONCE(len > thp_size(page));
 621         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 622         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 623
 624         req->r_mtime = inode->i_mtime;
 625         err = ceph_osdc_start_request(osdc, req, true);
 626         if (!err)
 627                 err = ceph_osdc_wait_request(osdc, req);
 628
 629         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 630                                   req->r_end_latency, len, err);
 631
 632         ceph_osdc_put_request(req);
 633         if (err == 0)
 634                 err = len;
 635
 636         if (err < 0) {
 637                 struct writeback_control tmp_wbc;
 638                 if (!wbc)
 639                         wbc = &tmp_wbc;
 640                 if (err == -ERESTARTSYS) {
 641                         /* killed by SIGKILL */
 642                         dout("writepage interrupted page %p\n", page);
 643                         redirty_page_for_writepage(wbc, page);
 644                         end_page_writeback(page);
 645                         return err;
 646                 }
 647                 if (err == -EBLOCKLISTED)
 648                         fsc->blocklisted = true;
 649                 dout("writepage setting page/mapping error %d %p\n",
 650                      err, page);
 651                 mapping_set_error(&inode->i_data, err);
 652                 wbc->pages_skipped++;
 653         } else {
 654                 dout("writepage cleaned page %p\n", page);
 655                 err = 0;  /* vfs expects us to return 0 */
 656         }
 657         oldest = detach_page_private(page);
 658         WARN_ON_ONCE(oldest != snapc);
 659         end_page_writeback(page);
 660         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 661         ceph_put_snap_context(snapc);  /* page's reference */
 662
 663         if (atomic_long_dec_return(&fsc->writeback_count) <
 664             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 665                 fsc->write_congested = false;
 666
 667         return err;
 668 }
 669
 670 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 671 {
 672         int err;
 673         struct inode *inode = page->mapping->host;
 674         BUG_ON(!inode);
 675         ihold(inode);
 676
 677         if (wbc->sync_mode == WB_SYNC_NONE &&
 678             ceph_inode_to_client(inode)->write_congested)
 679                 return AOP_WRITEPAGE_ACTIVATE;
 680
 681         wait_on_page_fscache(page);
 682
 683         err = writepage_nounlock(page, wbc);
 684         if (err == -ERESTARTSYS) {
 685                 /* direct memory reclaimer was killed by SIGKILL. return 0
 686                  * to prevent caller from setting mapping/page error */
 687                 err = 0;
 688         }
 689         unlock_page(page);
 690         iput(inode);
 691         return err;
 692 }
 693
 694 /*
 695  * async writeback completion handler.
 696  *
 697  * If we get an error, set the mapping error bit, but not the individual
 698  * page error bits.
 699  */
 700 static void writepages_finish(struct ceph_osd_request *req)
 701 {
 702         struct inode *inode = req->r_inode;
 703         struct ceph_inode_info *ci = ceph_inode(inode);
 704         struct ceph_osd_data *osd_data;
 705         struct page *page;
 706         int num_pages, total_pages = 0;
 707         int i, j;
 708         int rc = req->r_result;
 709         struct ceph_snap_context *snapc = req->r_snapc;
 710         struct address_space *mapping = inode->i_mapping;
 711         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 712         unsigned int len = 0;
 713         bool remove_page;
 714
 715         dout("writepages_finish %p rc %d\n", inode, rc);
 716         if (rc < 0) {
 717                 mapping_set_error(mapping, rc);
 718                 ceph_set_error_write(ci);
 719                 if (rc == -EBLOCKLISTED)
 720                         fsc->blocklisted = true;
 721         } else {
 722                 ceph_clear_error_write(ci);
 723         }
 724
 725         /*
 726          * We lost the cache cap, need to truncate the page before
 727          * it is unlocked, otherwise we'd truncate it later in the
 728          * page truncation thread, possibly losing some data that
 729          * raced its way in
 730          */
 731         remove_page = !(ceph_caps_issued(ci) &
 732                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 733
 734         /* clean all pages */
 735         for (i = 0; i < req->r_num_ops; i++) {
 736                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 737                         pr_warn("%s incorrect op %d req %p index %d tid %llu\n",
 738                                 __func__, req->r_ops[i].op, req, i, req->r_tid);
 739                         break;
 740                 }
 741
 742                 osd_data = osd_req_op_extent_osd_data(req, i);
 743                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 744                 len += osd_data->length;
 745                 num_pages = calc_pages_for((u64)osd_data->alignment,
 746                                            (u64)osd_data->length);
 747                 total_pages += num_pages;
 748                 for (j = 0; j < num_pages; j++) {
 749                         page = osd_data->pages[j];
 750                         BUG_ON(!page);
 751                         WARN_ON(!PageUptodate(page));
 752
 753                         if (atomic_long_dec_return(&fsc->writeback_count) <
 754                              CONGESTION_OFF_THRESH(
 755                                         fsc->mount_options->congestion_kb))
 756                                 fsc->write_congested = false;
 757
 758                         ceph_put_snap_context(detach_page_private(page));
 759                         end_page_writeback(page);
 760                         dout("unlocking %p\n", page);
 761
 762                         if (remove_page)
 763                                 generic_error_remove_page(inode->i_mapping,
 764                                                           page);
 765
 766                         unlock_page(page);
 767                 }
 768                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 769                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 770
 771                 release_pages(osd_data->pages, num_pages);
 772         }
 773
 774         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 775                                   req->r_end_latency, len, rc);
 776
 777         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 778
 779         osd_data = osd_req_op_extent_osd_data(req, 0);
 780         if (osd_data->pages_from_pool)
 781                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 782         else
 783                 kfree(osd_data->pages);
 784         ceph_osdc_put_request(req);
 785 }
 786
 787 /*
 788  * initiate async writeback
 789  */
 790 static int ceph_writepages_start(struct address_space *mapping,
 791                                  struct writeback_control *wbc)
 792 {
 793         struct inode *inode = mapping->host;
 794         struct ceph_inode_info *ci = ceph_inode(inode);
 795         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 796         struct ceph_vino vino = ceph_vino(inode);
 797         pgoff_t index, start_index, end = -1;
 798         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 799         struct pagevec pvec;
 800         int rc = 0;
 801         unsigned int wsize = i_blocksize(inode);
 802         struct ceph_osd_request *req = NULL;
 803         struct ceph_writeback_ctl ceph_wbc;
 804         bool should_loop, range_whole = false;
 805         bool done = false;
 806         bool caching = ceph_is_cache_enabled(inode);
 807
 808         if (wbc->sync_mode == WB_SYNC_NONE &&
 809             fsc->write_congested)
 810                 return 0;
 811
 812         dout("writepages_start %p (mode=%s)\n", inode,
 813              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 814              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 815
 816         if (ceph_inode_is_shutdown(inode)) {
 817                 if (ci->i_wrbuffer_ref > 0) {
 818                         pr_warn_ratelimited(
 819                                 "writepage_start %p %lld forced umount\n",
 820                                 inode, ceph_ino(inode));
 821                 }
 822                 mapping_set_error(mapping, -EIO);
 823                 return -EIO; /* we're in a forced umount, don't write! */
 824         }
 825         if (fsc->mount_options->wsize < wsize)
 826                 wsize = fsc->mount_options->wsize;
 827
 828         pagevec_init(&pvec);
 829
 830         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 831         index = start_index;
 832
 833 retry:
 834         /* find oldest snap context with dirty data */
 835         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 836         if (!snapc) {
 837                 /* hmm, why does writepages get called when there
 838                    is no dirty data? */
 839                 dout(" no snap context with dirty data?\n");
 840                 goto out;
 841         }
 842         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 843              snapc, snapc->seq, snapc->num_snaps);
 844
 845         should_loop = false;
 846         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 847                 /* where to start/end? */
 848                 if (wbc->range_cyclic) {
 849                         index = start_index;
 850                         end = -1;
 851                         if (index > 0)
 852                                 should_loop = true;
 853                         dout(" cyclic, start at %lu\n", index);
 854                 } else {
 855                         index = wbc->range_start >> PAGE_SHIFT;
 856                         end = wbc->range_end >> PAGE_SHIFT;
 857                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 858                                 range_whole = true;
 859                         dout(" not cyclic, %lu to %lu\n", index, end);
 860                 }
 861         } else if (!ceph_wbc.head_snapc) {
 862                 /* Do not respect wbc->range_{start,end}. Dirty pages
 863                  * in that range can be associated with newer snapc.
 864                  * They are not writeable until we write all dirty pages
 865                  * associated with 'snapc' get written */
 866                 if (index > 0)
 867                         should_loop = true;
 868                 dout(" non-head snapc, range whole\n");
 869         }
 870
 871         ceph_put_snap_context(last_snapc);
 872         last_snapc = snapc;
 873
 874         while (!done && index <= end) {
 875                 int num_ops = 0, op_idx;
 876                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 877                 struct page **pages = NULL, **data_pages;
 878                 struct page *page;
 879                 pgoff_t strip_unit_end = 0;
 880                 u64 offset = 0, len = 0;
 881                 bool from_pool = false;
 882
 883                 max_pages = wsize >> PAGE_SHIFT;
 884
 885 get_more_pages:
 886                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 887                                                 end, PAGECACHE_TAG_DIRTY);
 888                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 889                 if (!pvec_pages && !locked_pages)
 890                         break;
 891                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 892                         page = pvec.pages[i];
 893                         dout("? %p idx %lu\n", page, page->index);
 894                         if (locked_pages == 0)
 895                                 lock_page(page);  /* first page */
 896                         else if (!trylock_page(page))
 897                                 break;
 898
 899                         /* only dirty pages, or our accounting breaks */
 900                         if (unlikely(!PageDirty(page)) ||
 901                             unlikely(page->mapping != mapping)) {
 902                                 dout("!dirty or !mapping %p\n", page);
 903                                 unlock_page(page);
 904                                 continue;
 905                         }
 906                         /* only if matching snap context */
 907                         pgsnapc = page_snap_context(page);
 908                         if (pgsnapc != snapc) {
 909                                 dout("page snapc %p %lld != oldest %p %lld\n",
 910                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 911                                 if (!should_loop &&
 912                                     !ceph_wbc.head_snapc &&
 913                                     wbc->sync_mode != WB_SYNC_NONE)
 914                                         should_loop = true;
 915                                 unlock_page(page);
 916                                 continue;
 917                         }
 918                         if (page_offset(page) >= ceph_wbc.i_size) {
 919                                 struct folio *folio = page_folio(page);
 920
 921                                 dout("folio at %lu beyond eof %llu\n",
 922                                      folio->index, ceph_wbc.i_size);
 923                                 if ((ceph_wbc.size_stable ||
 924                                     folio_pos(folio) >= i_size_read(inode)) &&
 925                                     folio_clear_dirty_for_io(folio))
 926                                         folio_invalidate(folio, 0,
 927                                                         folio_size(folio));
 928                                 folio_unlock(folio);
 929                                 continue;
 930                         }
 931                         if (strip_unit_end && (page->index > strip_unit_end)) {
 932                                 dout("end of strip unit %p\n", page);
 933                                 unlock_page(page);
 934                                 break;
 935                         }
 936                         if (PageWriteback(page) || PageFsCache(page)) {
 937                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 938                                         dout("%p under writeback\n", page);
 939                                         unlock_page(page);
 940                                         continue;
 941                                 }
 942                                 dout("waiting on writeback %p\n", page);
 943                                 wait_on_page_writeback(page);
 944                                 wait_on_page_fscache(page);
 945                         }
 946
 947                         if (!clear_page_dirty_for_io(page)) {
 948                                 dout("%p !clear_page_dirty_for_io\n", page);
 949                                 unlock_page(page);
 950                                 continue;
 951                         }
 952
 953                         /*
 954                          * We have something to write.  If this is
 955                          * the first locked page this time through,
 956                          * calculate max possinle write size and
 957                          * allocate a page array
 958                          */
 959                         if (locked_pages == 0) {
 960                                 u64 objnum;
 961                                 u64 objoff;
 962                                 u32 xlen;
 963
 964                                 /* prepare async write request */
 965                                 offset = (u64)page_offset(page);
 966                                 ceph_calc_file_object_mapping(&ci->i_layout,
 967                                                               offset, wsize,
 968                                                               &objnum, &objoff,
 969                                                               &xlen);
 970                                 len = xlen;
 971
 972                                 num_ops = 1;
 973                                 strip_unit_end = page->index +
 974                                         ((len - 1) >> PAGE_SHIFT);
 975
 976                                 BUG_ON(pages);
 977                                 max_pages = calc_pages_for(0, (u64)len);
 978                                 pages = kmalloc_array(max_pages,
 979                                                       sizeof(*pages),
 980                                                       GFP_NOFS);
 981                                 if (!pages) {
 982                                         from_pool = true;
 983                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 984                                         BUG_ON(!pages);
 985                                 }
 986
 987                                 len = 0;
 988                         } else if (page->index !=
 989                                    (offset + len) >> PAGE_SHIFT) {
 990                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 991                                                              CEPH_OSD_MAX_OPS)) {
 992                                         redirty_page_for_writepage(wbc, page);
 993                                         unlock_page(page);
 994                                         break;
 995                                 }
 996
 997                                 num_ops++;
 998                                 offset = (u64)page_offset(page);
 999                                 len = 0;
1000                         }
1001
1002                         /* note position of first page in pvec */
1003                         dout("%p will write page %p idx %lu\n",
1004                              inode, page, page->index);
1005
1006                         if (atomic_long_inc_return(&fsc->writeback_count) >
1007                             CONGESTION_ON_THRESH(
1008                                     fsc->mount_options->congestion_kb))
1009                                 fsc->write_congested = true;
1010
1011                         pages[locked_pages++] = page;
1012                         pvec.pages[i] = NULL;
1013
1014                         len += thp_size(page);
1015                 }
1016
1017                 /* did we get anything? */
1018                 if (!locked_pages)
1019                         goto release_pvec_pages;
1020                 if (i) {
1021                         unsigned j, n = 0;
1022                         /* shift unused page to beginning of pvec */
1023                         for (j = 0; j < pvec_pages; j++) {
1024                                 if (!pvec.pages[j])
1025                                         continue;
1026                                 if (n < j)
1027                                         pvec.pages[n] = pvec.pages[j];
1028                                 n++;
1029                         }
1030                         pvec.nr = n;
1031
1032                         if (pvec_pages && i == pvec_pages &&
1033                             locked_pages < max_pages) {
1034                                 dout("reached end pvec, trying for more\n");
1035                                 pagevec_release(&pvec);
1036                                 goto get_more_pages;
1037                         }
1038                 }
1039
1040 new_request:
1041                 offset = page_offset(pages[0]);
1042                 len = wsize;
1043
1044                 req = ceph_osdc_new_request(&fsc->client->osdc,
1045                                         &ci->i_layout, vino,
1046                                         offset, &len, 0, num_ops,
1047                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1048                                         snapc, ceph_wbc.truncate_seq,
1049                                         ceph_wbc.truncate_size, false);
1050                 if (IS_ERR(req)) {
1051                         req = ceph_osdc_new_request(&fsc->client->osdc,
1052                                                 &ci->i_layout, vino,
1053                                                 offset, &len, 0,
1054                                                 min(num_ops,
1055                                                     CEPH_OSD_SLAB_OPS),
1056                                                 CEPH_OSD_OP_WRITE,
1057                                                 CEPH_OSD_FLAG_WRITE,
1058                                                 snapc, ceph_wbc.truncate_seq,
1059                                                 ceph_wbc.truncate_size, true);
1060                         BUG_ON(IS_ERR(req));
1061                 }
1062                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1063                              thp_size(page) - offset);
1064
1065                 req->r_callback = writepages_finish;
1066                 req->r_inode = inode;
1067
1068                 /* Format the osd request message and submit the write */
1069                 len = 0;
1070                 data_pages = pages;
1071                 op_idx = 0;
1072                 for (i = 0; i < locked_pages; i++) {
1073                         u64 cur_offset = page_offset(pages[i]);
1074                         /*
1075                          * Discontinuity in page range? Ceph can handle that by just passing
1076                          * multiple extents in the write op.
1077                          */
1078                         if (offset + len != cur_offset) {
1079                                 /* If it's full, stop here */
1080                                 if (op_idx + 1 == req->r_num_ops)
1081                                         break;
1082
1083                                 /* Kick off an fscache write with what we have so far. */
1084                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1085
1086                                 /* Start a new extent */
1087                                 osd_req_op_extent_dup_last(req, op_idx,
1088                                                            cur_offset - offset);
1089                                 dout("writepages got pages at %llu~%llu\n",
1090                                      offset, len);
1091                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1092                                                         data_pages, len, 0,
1093                                                         from_pool, false);
1094                                 osd_req_op_extent_update(req, op_idx, len);
1095
1096                                 len = 0;
1097                                 offset = cur_offset;
1098                                 data_pages = pages + i;
1099                                 op_idx++;
1100                         }
1101
1102                         set_page_writeback(pages[i]);
1103                         if (caching)
1104                                 ceph_set_page_fscache(pages[i]);
1105                         len += thp_size(page);
1106                 }
1107                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1108
1109                 if (ceph_wbc.size_stable) {
1110                         len = min(len, ceph_wbc.i_size - offset);
1111                 } else if (i == locked_pages) {
1112                         /* writepages_finish() clears writeback pages
1113                          * according to the data length, so make sure
1114                          * data length covers all locked pages */
1115                         u64 min_len = len + 1 - thp_size(page);
1116                         len = get_writepages_data_length(inode, pages[i - 1],
1117                                                          offset);
1118                         len = max(len, min_len);
1119                 }
1120                 dout("writepages got pages at %llu~%llu\n", offset, len);
1121
1122                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1123                                                  0, from_pool, false);
1124                 osd_req_op_extent_update(req, op_idx, len);
1125
1126                 BUG_ON(op_idx + 1 != req->r_num_ops);
1127
1128                 from_pool = false;
1129                 if (i < locked_pages) {
1130                         BUG_ON(num_ops <= req->r_num_ops);
1131                         num_ops -= req->r_num_ops;
1132                         locked_pages -= i;
1133
1134                         /* allocate new pages array for next request */
1135                         data_pages = pages;
1136                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1137                                               GFP_NOFS);
1138                         if (!pages) {
1139                                 from_pool = true;
1140                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1141                                 BUG_ON(!pages);
1142                         }
1143                         memcpy(pages, data_pages + i,
1144                                locked_pages * sizeof(*pages));
1145                         memset(data_pages + i, 0,
1146                                locked_pages * sizeof(*pages));
1147                 } else {
1148                         BUG_ON(num_ops != req->r_num_ops);
1149                         index = pages[i - 1]->index + 1;
1150                         /* request message now owns the pages array */
1151                         pages = NULL;
1152                 }
1153
1154                 req->r_mtime = inode->i_mtime;
1155                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1156                 BUG_ON(rc);
1157                 req = NULL;
1158
1159                 wbc->nr_to_write -= i;
1160                 if (pages)
1161                         goto new_request;
1162
1163                 /*
1164                  * We stop writing back only if we are not doing
1165                  * integrity sync. In case of integrity sync we have to
1166                  * keep going until we have written all the pages
1167                  * we tagged for writeback prior to entering this loop.
1168                  */
1169                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1170                         done = true;
1171
1172 release_pvec_pages:
1173                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1174                      pvec.nr ? pvec.pages[0] : NULL);
1175                 pagevec_release(&pvec);
1176         }
1177
1178         if (should_loop && !done) {
1179                 /* more to do; loop back to beginning of file */
1180                 dout("writepages looping back to beginning of file\n");
1181                 end = start_index - 1; /* OK even when start_index == 0 */
1182
1183                 /* to write dirty pages associated with next snapc,
1184                  * we need to wait until current writes complete */
1185                 if (wbc->sync_mode != WB_SYNC_NONE &&
1186                     start_index == 0 && /* all dirty pages were checked */
1187                     !ceph_wbc.head_snapc) {
1188                         struct page *page;
1189                         unsigned i, nr;
1190                         index = 0;
1191                         while ((index <= end) &&
1192                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1193                                                 PAGECACHE_TAG_WRITEBACK))) {
1194                                 for (i = 0; i < nr; i++) {
1195                                         page = pvec.pages[i];
1196                                         if (page_snap_context(page) != snapc)
1197                                                 continue;
1198                                         wait_on_page_writeback(page);
1199                                 }
1200                                 pagevec_release(&pvec);
1201                                 cond_resched();
1202                         }
1203                 }
1204
1205                 start_index = 0;
1206                 index = 0;
1207                 goto retry;
1208         }
1209
1210         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1211                 mapping->writeback_index = index;
1212
1213 out:
1214         ceph_osdc_put_request(req);
1215         ceph_put_snap_context(last_snapc);
1216         dout("writepages dend - startone, rc = %d\n", rc);
1217         return rc;
1218 }
1219
1220
1221
1222 /*
1223  * See if a given @snapc is either writeable, or already written.
1224  */
1225 static int context_is_writeable_or_written(struct inode *inode,
1226                                            struct ceph_snap_context *snapc)
1227 {
1228         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1229         int ret = !oldest || snapc->seq <= oldest->seq;
1230
1231         ceph_put_snap_context(oldest);
1232         return ret;
1233 }
1234
1235 /**
1236  * ceph_find_incompatible - find an incompatible context and return it
1237  * @page: page being dirtied
1238  *
1239  * We are only allowed to write into/dirty a page if the page is
1240  * clean, or already dirty within the same snap context. Returns a
1241  * conflicting context if there is one, NULL if there isn't, or a
1242  * negative error code on other errors.
1243  *
1244  * Must be called with page lock held.
1245  */
1246 static struct ceph_snap_context *
1247 ceph_find_incompatible(struct page *page)
1248 {
1249         struct inode *inode = page->mapping->host;
1250         struct ceph_inode_info *ci = ceph_inode(inode);
1251
1252         if (ceph_inode_is_shutdown(inode)) {
1253                 dout(" page %p %llx:%llx is shutdown\n", page,
1254                      ceph_vinop(inode));
1255                 return ERR_PTR(-ESTALE);
1256         }
1257
1258         for (;;) {
1259                 struct ceph_snap_context *snapc, *oldest;
1260
1261                 wait_on_page_writeback(page);
1262
1263                 snapc = page_snap_context(page);
1264                 if (!snapc || snapc == ci->i_head_snapc)
1265                         break;
1266
1267                 /*
1268                  * this page is already dirty in another (older) snap
1269                  * context!  is it writeable now?
1270                  */
1271                 oldest = get_oldest_context(inode, NULL, NULL);
1272                 if (snapc->seq > oldest->seq) {
1273                         /* not writeable -- return it for the caller to deal with */
1274                         ceph_put_snap_context(oldest);
1275                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1276                         return ceph_get_snap_context(snapc);
1277                 }
1278                 ceph_put_snap_context(oldest);
1279
1280                 /* yay, writeable, do it now (without dropping page lock) */
1281                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1282                 if (clear_page_dirty_for_io(page)) {
1283                         int r = writepage_nounlock(page, NULL);
1284                         if (r < 0)
1285                                 return ERR_PTR(r);
1286                 }
1287         }
1288         return NULL;
1289 }
1290
1291 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1292                                         struct folio *folio, void **_fsdata)
1293 {
1294         struct inode *inode = file_inode(file);
1295         struct ceph_inode_info *ci = ceph_inode(inode);
1296         struct ceph_snap_context *snapc;
1297
1298         snapc = ceph_find_incompatible(folio_page(folio, 0));
1299         if (snapc) {
1300                 int r;
1301
1302                 folio_unlock(folio);
1303                 folio_put(folio);
1304                 if (IS_ERR(snapc))
1305                         return PTR_ERR(snapc);
1306
1307                 ceph_queue_writeback(inode);
1308                 r = wait_event_killable(ci->i_cap_wq,
1309                                         context_is_writeable_or_written(inode, snapc));
1310                 ceph_put_snap_context(snapc);
1311                 return r == 0 ? -EAGAIN : r;
1312         }
1313         return 0;
1314 }
1315
1316 /*
1317  * We are only allowed to write into/dirty the page if the page is
1318  * clean, or already dirty within the same snap context.
1319  */
1320 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1321                             loff_t pos, unsigned len, unsigned aop_flags,
1322                             struct page **pagep, void **fsdata)
1323 {
1324         struct inode *inode = file_inode(file);
1325         struct folio *folio = NULL;
1326         int r;
1327
1328         r = netfs_write_begin(file, inode->i_mapping, pos, len, 0, &folio, NULL);
1329         if (r == 0)
1330                 folio_wait_fscache(folio);
1331         if (r < 0) {
1332                 if (folio)
1333                         folio_put(folio);
1334         } else {
1335                 WARN_ON_ONCE(!folio_test_locked(folio));
1336                 *pagep = &folio->page;
1337         }
1338         return r;
1339 }
1340
1341 /*
1342  * we don't do anything in here that simple_write_end doesn't do
1343  * except adjust dirty page accounting
1344  */
1345 static int ceph_write_end(struct file *file, struct address_space *mapping,
1346                           loff_t pos, unsigned len, unsigned copied,
1347                           struct page *subpage, void *fsdata)
1348 {
1349         struct folio *folio = page_folio(subpage);
1350         struct inode *inode = file_inode(file);
1351         bool check_cap = false;
1352
1353         dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1354              inode, folio, (int)pos, (int)copied, (int)len);
1355
1356         if (!folio_test_uptodate(folio)) {
1357                 /* just return that nothing was copied on a short copy */
1358                 if (copied < len) {
1359                         copied = 0;
1360                         goto out;
1361                 }
1362                 folio_mark_uptodate(folio);
1363         }
1364
1365         /* did file size increase? */
1366         if (pos+copied > i_size_read(inode))
1367                 check_cap = ceph_inode_set_size(inode, pos+copied);
1368
1369         folio_mark_dirty(folio);
1370
1371 out:
1372         folio_unlock(folio);
1373         folio_put(folio);
1374
1375         if (check_cap)
1376                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1377
1378         return copied;
1379 }
1380
1381 const struct address_space_operations ceph_aops = {
1382         .readpage = netfs_readpage,
1383         .readahead = netfs_readahead,
1384         .writepage = ceph_writepage,
1385         .writepages = ceph_writepages_start,
1386         .write_begin = ceph_write_begin,
1387         .write_end = ceph_write_end,
1388         .dirty_folio = ceph_dirty_folio,
1389         .invalidate_folio = ceph_invalidate_folio,
1390         .releasepage = ceph_releasepage,
1391         .direct_IO = noop_direct_IO,
1392 };
1393
1394 static void ceph_block_sigs(sigset_t *oldset)
1395 {
1396         sigset_t mask;
1397         siginitsetinv(&mask, sigmask(SIGKILL));
1398         sigprocmask(SIG_BLOCK, &mask, oldset);
1399 }
1400
1401 static void ceph_restore_sigs(sigset_t *oldset)
1402 {
1403         sigprocmask(SIG_SETMASK, oldset, NULL);
1404 }
1405
1406 /*
1407  * vm ops
1408  */
1409 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1410 {
1411         struct vm_area_struct *vma = vmf->vma;
1412         struct inode *inode = file_inode(vma->vm_file);
1413         struct ceph_inode_info *ci = ceph_inode(inode);
1414         struct ceph_file_info *fi = vma->vm_file->private_data;
1415         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1416         int want, got, err;
1417         sigset_t oldset;
1418         vm_fault_t ret = VM_FAULT_SIGBUS;
1419
1420         if (ceph_inode_is_shutdown(inode))
1421                 return ret;
1422
1423         ceph_block_sigs(&oldset);
1424
1425         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1426              inode, ceph_vinop(inode), off);
1427         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1428                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1429         else
1430                 want = CEPH_CAP_FILE_CACHE;
1431
1432         got = 0;
1433         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1434         if (err < 0)
1435                 goto out_restore;
1436
1437         dout("filemap_fault %p %llu got cap refs on %s\n",
1438              inode, off, ceph_cap_string(got));
1439
1440         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1441             ci->i_inline_version == CEPH_INLINE_NONE) {
1442                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1443                 ceph_add_rw_context(fi, &rw_ctx);
1444                 ret = filemap_fault(vmf);
1445                 ceph_del_rw_context(fi, &rw_ctx);
1446                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1447                      inode, off, ceph_cap_string(got), ret);
1448         } else
1449                 err = -EAGAIN;
1450
1451         ceph_put_cap_refs(ci, got);
1452
1453         if (err != -EAGAIN)
1454                 goto out_restore;
1455
1456         /* read inline data */
1457         if (off >= PAGE_SIZE) {
1458                 /* does not support inline data > PAGE_SIZE */
1459                 ret = VM_FAULT_SIGBUS;
1460         } else {
1461                 struct address_space *mapping = inode->i_mapping;
1462                 struct page *page;
1463
1464                 filemap_invalidate_lock_shared(mapping);
1465                 page = find_or_create_page(mapping, 0,
1466                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1467                 if (!page) {
1468                         ret = VM_FAULT_OOM;
1469                         goto out_inline;
1470                 }
1471                 err = __ceph_do_getattr(inode, page,
1472                                          CEPH_STAT_CAP_INLINE_DATA, true);
1473                 if (err < 0 || off >= i_size_read(inode)) {
1474                         unlock_page(page);
1475                         put_page(page);
1476                         ret = vmf_error(err);
1477                         goto out_inline;
1478                 }
1479                 if (err < PAGE_SIZE)
1480                         zero_user_segment(page, err, PAGE_SIZE);
1481                 else
1482                         flush_dcache_page(page);
1483                 SetPageUptodate(page);
1484                 vmf->page = page;
1485                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1486 out_inline:
1487                 filemap_invalidate_unlock_shared(mapping);
1488                 dout("filemap_fault %p %llu read inline data ret %x\n",
1489                      inode, off, ret);
1490         }
1491 out_restore:
1492         ceph_restore_sigs(&oldset);
1493         if (err < 0)
1494                 ret = vmf_error(err);
1495
1496         return ret;
1497 }
1498
1499 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1500 {
1501         struct vm_area_struct *vma = vmf->vma;
1502         struct inode *inode = file_inode(vma->vm_file);
1503         struct ceph_inode_info *ci = ceph_inode(inode);
1504         struct ceph_file_info *fi = vma->vm_file->private_data;
1505         struct ceph_cap_flush *prealloc_cf;
1506         struct page *page = vmf->page;
1507         loff_t off = page_offset(page);
1508         loff_t size = i_size_read(inode);
1509         size_t len;
1510         int want, got, err;
1511         sigset_t oldset;
1512         vm_fault_t ret = VM_FAULT_SIGBUS;
1513
1514         if (ceph_inode_is_shutdown(inode))
1515                 return ret;
1516
1517         prealloc_cf = ceph_alloc_cap_flush();
1518         if (!prealloc_cf)
1519                 return VM_FAULT_OOM;
1520
1521         sb_start_pagefault(inode->i_sb);
1522         ceph_block_sigs(&oldset);
1523
1524         if (off + thp_size(page) <= size)
1525                 len = thp_size(page);
1526         else
1527                 len = offset_in_thp(page, size);
1528
1529         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1530              inode, ceph_vinop(inode), off, len, size);
1531         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1532                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1533         else
1534                 want = CEPH_CAP_FILE_BUFFER;
1535
1536         got = 0;
1537         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1538         if (err < 0)
1539                 goto out_free;
1540
1541         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1542              inode, off, len, ceph_cap_string(got));
1543
1544         /* Update time before taking page lock */
1545         file_update_time(vma->vm_file);
1546         inode_inc_iversion_raw(inode);
1547
1548         do {
1549                 struct ceph_snap_context *snapc;
1550
1551                 lock_page(page);
1552
1553                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1554                         unlock_page(page);
1555                         ret = VM_FAULT_NOPAGE;
1556                         break;
1557                 }
1558
1559                 snapc = ceph_find_incompatible(page);
1560                 if (!snapc) {
1561                         /* success.  we'll keep the page locked. */
1562                         set_page_dirty(page);
1563                         ret = VM_FAULT_LOCKED;
1564                         break;
1565                 }
1566
1567                 unlock_page(page);
1568
1569                 if (IS_ERR(snapc)) {
1570                         ret = VM_FAULT_SIGBUS;
1571                         break;
1572                 }
1573
1574                 ceph_queue_writeback(inode);
1575                 err = wait_event_killable(ci->i_cap_wq,
1576                                 context_is_writeable_or_written(inode, snapc));
1577                 ceph_put_snap_context(snapc);
1578         } while (err == 0);
1579
1580         if (ret == VM_FAULT_LOCKED) {
1581                 int dirty;
1582                 spin_lock(&ci->i_ceph_lock);
1583                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1584                                                &prealloc_cf);
1585                 spin_unlock(&ci->i_ceph_lock);
1586                 if (dirty)
1587                         __mark_inode_dirty(inode, dirty);
1588         }
1589
1590         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1591              inode, off, len, ceph_cap_string(got), ret);
1592         ceph_put_cap_refs_async(ci, got);
1593 out_free:
1594         ceph_restore_sigs(&oldset);
1595         sb_end_pagefault(inode->i_sb);
1596         ceph_free_cap_flush(prealloc_cf);
1597         if (err < 0)
1598                 ret = vmf_error(err);
1599         return ret;
1600 }
1601
1602 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1603                            char *data, size_t len)
1604 {
1605         struct address_space *mapping = inode->i_mapping;
1606         struct page *page;
1607
1608         if (locked_page) {
1609                 page = locked_page;
1610         } else {
1611                 if (i_size_read(inode) == 0)
1612                         return;
1613                 page = find_or_create_page(mapping, 0,
1614                                            mapping_gfp_constraint(mapping,
1615                                            ~__GFP_FS));
1616                 if (!page)
1617                         return;
1618                 if (PageUptodate(page)) {
1619                         unlock_page(page);
1620                         put_page(page);
1621                         return;
1622                 }
1623         }
1624
1625         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1626              inode, ceph_vinop(inode), len, locked_page);
1627
1628         if (len > 0) {
1629                 void *kaddr = kmap_atomic(page);
1630                 memcpy(kaddr, data, len);
1631                 kunmap_atomic(kaddr);
1632         }
1633
1634         if (page != locked_page) {
1635                 if (len < PAGE_SIZE)
1636                         zero_user_segment(page, len, PAGE_SIZE);
1637                 else
1638                         flush_dcache_page(page);
1639
1640                 SetPageUptodate(page);
1641                 unlock_page(page);
1642                 put_page(page);
1643         }
1644 }
1645
1646 int ceph_uninline_data(struct file *file)
1647 {
1648         struct inode *inode = file_inode(file);
1649         struct ceph_inode_info *ci = ceph_inode(inode);
1650         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1651         struct ceph_osd_request *req;
1652         struct ceph_cap_flush *prealloc_cf;
1653         struct folio *folio = NULL;
1654         u64 inline_version = CEPH_INLINE_NONE;
1655         struct page *pages[1];
1656         int err = 0;
1657         u64 len;
1658
1659         prealloc_cf = ceph_alloc_cap_flush();
1660         if (!prealloc_cf)
1661                 return -ENOMEM;
1662
1663         folio = read_mapping_folio(inode->i_mapping, 0, file);
1664         if (IS_ERR(folio)) {
1665                 err = PTR_ERR(folio);
1666                 goto out;
1667         }
1668
1669         folio_lock(folio);
1670
1671         spin_lock(&ci->i_ceph_lock);
1672         inline_version = ci->i_inline_version;
1673         spin_unlock(&ci->i_ceph_lock);
1674
1675         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1676              inode, ceph_vinop(inode), inline_version);
1677
1678         if (inline_version == 1 || /* initial version, no data */
1679             inline_version == CEPH_INLINE_NONE)
1680                 goto out_unlock;
1681
1682         len = i_size_read(inode);
1683         if (len > folio_size(folio))
1684                 len = folio_size(folio);
1685
1686         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1687                                     ceph_vino(inode), 0, &len, 0, 1,
1688                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1689                                     NULL, 0, 0, false);
1690         if (IS_ERR(req)) {
1691                 err = PTR_ERR(req);
1692                 goto out_unlock;
1693         }
1694
1695         req->r_mtime = inode->i_mtime;
1696         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1697         if (!err)
1698                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1699         ceph_osdc_put_request(req);
1700         if (err < 0)
1701                 goto out_unlock;
1702
1703         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1704                                     ceph_vino(inode), 0, &len, 1, 3,
1705                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1706                                     NULL, ci->i_truncate_seq,
1707                                     ci->i_truncate_size, false);
1708         if (IS_ERR(req)) {
1709                 err = PTR_ERR(req);
1710                 goto out_unlock;
1711         }
1712
1713         pages[0] = folio_page(folio, 0);
1714         osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1715
1716         {
1717                 __le64 xattr_buf = cpu_to_le64(inline_version);
1718                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1719                                             "inline_version", &xattr_buf,
1720                                             sizeof(xattr_buf),
1721                                             CEPH_OSD_CMPXATTR_OP_GT,
1722                                             CEPH_OSD_CMPXATTR_MODE_U64);
1723                 if (err)
1724                         goto out_put_req;
1725         }
1726
1727         {
1728                 char xattr_buf[32];
1729                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1730                                          "%llu", inline_version);
1731                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1732                                             "inline_version",
1733                                             xattr_buf, xattr_len, 0, 0);
1734                 if (err)
1735                         goto out_put_req;
1736         }
1737
1738         req->r_mtime = inode->i_mtime;
1739         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1740         if (!err)
1741                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1742
1743         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1744                                   req->r_end_latency, len, err);
1745
1746         if (!err) {
1747                 int dirty;
1748
1749                 /* Set to CAP_INLINE_NONE and dirty the caps */
1750                 down_read(&fsc->mdsc->snap_rwsem);
1751                 spin_lock(&ci->i_ceph_lock);
1752                 ci->i_inline_version = CEPH_INLINE_NONE;
1753                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1754                 spin_unlock(&ci->i_ceph_lock);
1755                 up_read(&fsc->mdsc->snap_rwsem);
1756                 if (dirty)
1757                         __mark_inode_dirty(inode, dirty);
1758         }
1759 out_put_req:
1760         ceph_osdc_put_request(req);
1761         if (err == -ECANCELED)
1762                 err = 0;
1763 out_unlock:
1764         folio_unlock(folio);
1765         folio_put(folio);
1766 out:
1767         ceph_free_cap_flush(prealloc_cf);
1768         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1769              inode, ceph_vinop(inode), inline_version, err);
1770         return err;
1771 }
1772
1773 static const struct vm_operations_struct ceph_vmops = {
1774         .fault          = ceph_filemap_fault,
1775         .page_mkwrite   = ceph_page_mkwrite,
1776 };
1777
1778 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1779 {
1780         struct address_space *mapping = file->f_mapping;
1781
1782         if (!mapping->a_ops->readpage)
1783                 return -ENOEXEC;
1784         vma->vm_ops = &ceph_vmops;
1785         return 0;
1786 }
1787
1788 enum {
1789         POOL_READ       = 1,
1790         POOL_WRITE      = 2,
1791 };
1792
1793 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1794                                 s64 pool, struct ceph_string *pool_ns)
1795 {
1796         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1797         struct ceph_mds_client *mdsc = fsc->mdsc;
1798         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1799         struct rb_node **p, *parent;
1800         struct ceph_pool_perm *perm;
1801         struct page **pages;
1802         size_t pool_ns_len;
1803         int err = 0, err2 = 0, have = 0;
1804
1805         down_read(&mdsc->pool_perm_rwsem);
1806         p = &mdsc->pool_perm_tree.rb_node;
1807         while (*p) {
1808                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1809                 if (pool < perm->pool)
1810                         p = &(*p)->rb_left;
1811                 else if (pool > perm->pool)
1812                         p = &(*p)->rb_right;
1813                 else {
1814                         int ret = ceph_compare_string(pool_ns,
1815                                                 perm->pool_ns,
1816                                                 perm->pool_ns_len);
1817                         if (ret < 0)
1818                                 p = &(*p)->rb_left;
1819                         else if (ret > 0)
1820                                 p = &(*p)->rb_right;
1821                         else {
1822                                 have = perm->perm;
1823                                 break;
1824                         }
1825                 }
1826         }
1827         up_read(&mdsc->pool_perm_rwsem);
1828         if (*p)
1829                 goto out;
1830
1831         if (pool_ns)
1832                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1833                      pool, (int)pool_ns->len, pool_ns->str);
1834         else
1835                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1836
1837         down_write(&mdsc->pool_perm_rwsem);
1838         p = &mdsc->pool_perm_tree.rb_node;
1839         parent = NULL;
1840         while (*p) {
1841                 parent = *p;
1842                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1843                 if (pool < perm->pool)
1844                         p = &(*p)->rb_left;
1845                 else if (pool > perm->pool)
1846                         p = &(*p)->rb_right;
1847                 else {
1848                         int ret = ceph_compare_string(pool_ns,
1849                                                 perm->pool_ns,
1850                                                 perm->pool_ns_len);
1851                         if (ret < 0)
1852                                 p = &(*p)->rb_left;
1853                         else if (ret > 0)
1854                                 p = &(*p)->rb_right;
1855                         else {
1856                                 have = perm->perm;
1857                                 break;
1858                         }
1859                 }
1860         }
1861         if (*p) {
1862                 up_write(&mdsc->pool_perm_rwsem);
1863                 goto out;
1864         }
1865
1866         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1867                                          1, false, GFP_NOFS);
1868         if (!rd_req) {
1869                 err = -ENOMEM;
1870                 goto out_unlock;
1871         }
1872
1873         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1874         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1875         rd_req->r_base_oloc.pool = pool;
1876         if (pool_ns)
1877                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1878         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1879
1880         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1881         if (err)
1882                 goto out_unlock;
1883
1884         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1885                                          1, false, GFP_NOFS);
1886         if (!wr_req) {
1887                 err = -ENOMEM;
1888                 goto out_unlock;
1889         }
1890
1891         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1892         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1893         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1894         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1895
1896         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1897         if (err)
1898                 goto out_unlock;
1899
1900         /* one page should be large enough for STAT data */
1901         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1902         if (IS_ERR(pages)) {
1903                 err = PTR_ERR(pages);
1904                 goto out_unlock;
1905         }
1906
1907         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1908                                      0, false, true);
1909         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1910
1911         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1912         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1913
1914         if (!err)
1915                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1916         if (!err2)
1917                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1918
1919         if (err >= 0 || err == -ENOENT)
1920                 have |= POOL_READ;
1921         else if (err != -EPERM) {
1922                 if (err == -EBLOCKLISTED)
1923                         fsc->blocklisted = true;
1924                 goto out_unlock;
1925         }
1926
1927         if (err2 == 0 || err2 == -EEXIST)
1928                 have |= POOL_WRITE;
1929         else if (err2 != -EPERM) {
1930                 if (err2 == -EBLOCKLISTED)
1931                         fsc->blocklisted = true;
1932                 err = err2;
1933                 goto out_unlock;
1934         }
1935
1936         pool_ns_len = pool_ns ? pool_ns->len : 0;
1937         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1938         if (!perm) {
1939                 err = -ENOMEM;
1940                 goto out_unlock;
1941         }
1942
1943         perm->pool = pool;
1944         perm->perm = have;
1945         perm->pool_ns_len = pool_ns_len;
1946         if (pool_ns_len > 0)
1947                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1948         perm->pool_ns[pool_ns_len] = 0;
1949
1950         rb_link_node(&perm->node, parent, p);
1951         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1952         err = 0;
1953 out_unlock:
1954         up_write(&mdsc->pool_perm_rwsem);
1955
1956         ceph_osdc_put_request(rd_req);
1957         ceph_osdc_put_request(wr_req);
1958 out:
1959         if (!err)
1960                 err = have;
1961         if (pool_ns)
1962                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1963                      pool, (int)pool_ns->len, pool_ns->str, err);
1964         else
1965                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1966         return err;
1967 }
1968
1969 int ceph_pool_perm_check(struct inode *inode, int need)
1970 {
1971         struct ceph_inode_info *ci = ceph_inode(inode);
1972         struct ceph_string *pool_ns;
1973         s64 pool;
1974         int ret, flags;
1975
1976         /* Only need to do this for regular files */
1977         if (!S_ISREG(inode->i_mode))
1978                 return 0;
1979
1980         if (ci->i_vino.snap != CEPH_NOSNAP) {
1981                 /*
1982                  * Pool permission check needs to write to the first object.
1983                  * But for snapshot, head of the first object may have alread
1984                  * been deleted. Skip check to avoid creating orphan object.
1985                  */
1986                 return 0;
1987         }
1988
1989         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
1990                                 NOPOOLPERM))
1991                 return 0;
1992
1993         spin_lock(&ci->i_ceph_lock);
1994         flags = ci->i_ceph_flags;
1995         pool = ci->i_layout.pool_id;
1996         spin_unlock(&ci->i_ceph_lock);
1997 check:
1998         if (flags & CEPH_I_POOL_PERM) {
1999                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2000                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2001                              pool);
2002                         return -EPERM;
2003                 }
2004                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2005                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2006                              pool);
2007                         return -EPERM;
2008                 }
2009                 return 0;
2010         }
2011
2012         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2013         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2014         ceph_put_string(pool_ns);
2015         if (ret < 0)
2016                 return ret;
2017
2018         flags = CEPH_I_POOL_PERM;
2019         if (ret & POOL_READ)
2020                 flags |= CEPH_I_POOL_RD;
2021         if (ret & POOL_WRITE)
2022                 flags |= CEPH_I_POOL_WR;
2023
2024         spin_lock(&ci->i_ceph_lock);
2025         if (pool == ci->i_layout.pool_id &&
2026             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2027                 ci->i_ceph_flags |= flags;
2028         } else {
2029                 pool = ci->i_layout.pool_id;
2030                 flags = ci->i_ceph_flags;
2031         }
2032         spin_unlock(&ci->i_ceph_lock);
2033         goto check;
2034 }
2035
2036 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2037 {
2038         struct ceph_pool_perm *perm;
2039         struct rb_node *n;
2040
2041         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2042                 n = rb_first(&mdsc->pool_perm_tree);
2043                 perm = rb_entry(n, struct ceph_pool_perm, node);
2044                 rb_erase(n, &mdsc->pool_perm_tree);
2045                 kfree(perm);
2046         }
2047 }