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