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