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_BUG_ON_FOLIO(folio_test_private(folio), folio);
 126         folio_attach_private(folio, snapc);
 127
 128         return ceph_fscache_dirty_folio(mapping, folio);
 129 }
 130
 131 /*
 132  * If we are truncating the full folio (i.e. offset == 0), adjust the
 133  * dirty folio counters appropriately.  Only called if there is private
 134  * data on the folio.
 135  */
 136 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 137                                 size_t length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc;
 142
 143         inode = folio->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != folio_size(folio)) {
 147                 dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
 148                      inode, folio->index, offset, length);
 149                 return;
 150         }
 151
 152         WARN_ON(!folio_test_locked(folio));
 153         if (folio_test_private(folio)) {
 154                 dout("%p invalidate_folio idx %lu full dirty page\n",
 155                      inode, folio->index);
 156
 157                 snapc = folio_detach_private(folio);
 158                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159                 ceph_put_snap_context(snapc);
 160         }
 161
 162         folio_wait_fscache(folio);
 163 }
 164
 165 static 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, true);
 241
 242         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 243         ceph_put_page_vector(osd_data->pages, num_pages, false);
 244         iput(req->r_inode);
 245 }
 246
 247 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 248 {
 249         struct netfs_io_request *rreq = subreq->rreq;
 250         struct inode *inode = rreq->inode;
 251         struct ceph_mds_reply_info_parsed *rinfo;
 252         struct ceph_mds_reply_info_in *iinfo;
 253         struct ceph_mds_request *req;
 254         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 255         struct ceph_inode_info *ci = ceph_inode(inode);
 256         struct iov_iter iter;
 257         ssize_t err = 0;
 258         size_t len;
 259         int mode;
 260
 261         __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 262         __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 263
 264         if (subreq->start >= inode->i_size)
 265                 goto out;
 266
 267         /* We need to fetch the inline data. */
 268         mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
 269         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
 270         if (IS_ERR(req)) {
 271                 err = PTR_ERR(req);
 272                 goto out;
 273         }
 274         req->r_ino1 = ci->i_vino;
 275         req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 276         req->r_num_caps = 2;
 277
 278         err = ceph_mdsc_do_request(mdsc, NULL, req);
 279         if (err < 0)
 280                 goto out;
 281
 282         rinfo = &req->r_reply_info;
 283         iinfo = &rinfo->targeti;
 284         if (iinfo->inline_version == CEPH_INLINE_NONE) {
 285                 /* The data got uninlined */
 286                 ceph_mdsc_put_request(req);
 287                 return false;
 288         }
 289
 290         len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
 291         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 292         err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
 293         if (err == 0)
 294                 err = -EFAULT;
 295
 296         ceph_mdsc_put_request(req);
 297 out:
 298         netfs_subreq_terminated(subreq, err, false);
 299         return true;
 300 }
 301
 302 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 303 {
 304         struct netfs_io_request *rreq = subreq->rreq;
 305         struct inode *inode = rreq->inode;
 306         struct ceph_inode_info *ci = ceph_inode(inode);
 307         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 308         struct ceph_osd_request *req;
 309         struct ceph_vino vino = ceph_vino(inode);
 310         struct iov_iter iter;
 311         struct page **pages;
 312         size_t page_off;
 313         int err = 0;
 314         u64 len = subreq->len;
 315
 316         if (ci->i_inline_version != CEPH_INLINE_NONE &&
 317             ceph_netfs_issue_op_inline(subreq))
 318                 return;
 319
 320         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 321                         0, 1, CEPH_OSD_OP_READ,
 322                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 323                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 324         if (IS_ERR(req)) {
 325                 err = PTR_ERR(req);
 326                 req = NULL;
 327                 goto out;
 328         }
 329
 330         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 331         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 332         err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 333         if (err < 0) {
 334                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 335                 goto out;
 336         }
 337
 338         /* should always give us a page-aligned read */
 339         WARN_ON_ONCE(page_off);
 340         len = err;
 341
 342         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 343         req->r_callback = finish_netfs_read;
 344         req->r_priv = subreq;
 345         req->r_inode = inode;
 346         ihold(inode);
 347
 348         err = ceph_osdc_start_request(req->r_osdc, req, false);
 349         if (err)
 350                 iput(inode);
 351 out:
 352         ceph_osdc_put_request(req);
 353         if (err)
 354                 netfs_subreq_terminated(subreq, err, false);
 355         dout("%s: result %d\n", __func__, err);
 356 }
 357
 358 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 359 {
 360         struct inode *inode = rreq->inode;
 361         int got = 0, want = CEPH_CAP_FILE_CACHE;
 362         int ret = 0;
 363
 364         if (rreq->origin != NETFS_READAHEAD)
 365                 return 0;
 366
 367         if (file) {
 368                 struct ceph_rw_context *rw_ctx;
 369                 struct ceph_file_info *fi = file->private_data;
 370
 371                 rw_ctx = ceph_find_rw_context(fi);
 372                 if (rw_ctx)
 373                         return 0;
 374         }
 375
 376         /*
 377          * readahead callers do not necessarily hold Fcb caps
 378          * (e.g. fadvise, madvise).
 379          */
 380         ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 381         if (ret < 0) {
 382                 dout("start_read %p, error getting cap\n", inode);
 383                 return ret;
 384         }
 385
 386         if (!(got & want)) {
 387                 dout("start_read %p, no cache cap\n", inode);
 388                 return -EACCES;
 389         }
 390         if (ret == 0)
 391                 return -EACCES;
 392
 393         rreq->netfs_priv = (void *)(uintptr_t)got;
 394         return 0;
 395 }
 396
 397 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 398 {
 399         struct ceph_inode_info *ci = ceph_inode(rreq->inode);
 400         int got = (uintptr_t)rreq->netfs_priv;
 401
 402         if (got)
 403                 ceph_put_cap_refs(ci, got);
 404 }
 405
 406 const struct netfs_request_ops ceph_netfs_ops = {
 407         .init_request           = ceph_init_request,
 408         .free_request           = ceph_netfs_free_request,
 409         .begin_cache_operation  = ceph_begin_cache_operation,
 410         .issue_read             = ceph_netfs_issue_read,
 411         .expand_readahead       = ceph_netfs_expand_readahead,
 412         .clamp_length           = ceph_netfs_clamp_length,
 413         .check_write_begin      = ceph_netfs_check_write_begin,
 414 };
 415
 416 #ifdef CONFIG_CEPH_FSCACHE
 417 static void ceph_set_page_fscache(struct page *page)
 418 {
 419         set_page_fscache(page);
 420 }
 421
 422 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 423 {
 424         struct inode *inode = priv;
 425
 426         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 427                 ceph_fscache_invalidate(inode, false);
 428 }
 429
 430 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 431 {
 432         struct ceph_inode_info *ci = ceph_inode(inode);
 433         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 434
 435         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 436                                ceph_fscache_write_terminated, inode, caching);
 437 }
 438 #else
 439 static inline void ceph_set_page_fscache(struct page *page)
 440 {
 441 }
 442
 443 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 444 {
 445 }
 446 #endif /* CONFIG_CEPH_FSCACHE */
 447
 448 struct ceph_writeback_ctl
 449 {
 450         loff_t i_size;
 451         u64 truncate_size;
 452         u32 truncate_seq;
 453         bool size_stable;
 454         bool head_snapc;
 455 };
 456
 457 /*
 458  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 459  * only snap context we are allowed to write back.
 460  */
 461 static struct ceph_snap_context *
 462 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 463                    struct ceph_snap_context *page_snapc)
 464 {
 465         struct ceph_inode_info *ci = ceph_inode(inode);
 466         struct ceph_snap_context *snapc = NULL;
 467         struct ceph_cap_snap *capsnap = NULL;
 468
 469         spin_lock(&ci->i_ceph_lock);
 470         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 471                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 472                      capsnap->context, capsnap->dirty_pages);
 473                 if (!capsnap->dirty_pages)
 474                         continue;
 475
 476                 /* get i_size, truncate_{seq,size} for page_snapc? */
 477                 if (snapc && capsnap->context != page_snapc)
 478                         continue;
 479
 480                 if (ctl) {
 481                         if (capsnap->writing) {
 482                                 ctl->i_size = i_size_read(inode);
 483                                 ctl->size_stable = false;
 484                         } else {
 485                                 ctl->i_size = capsnap->size;
 486                                 ctl->size_stable = true;
 487                         }
 488                         ctl->truncate_size = capsnap->truncate_size;
 489                         ctl->truncate_seq = capsnap->truncate_seq;
 490                         ctl->head_snapc = false;
 491                 }
 492
 493                 if (snapc)
 494                         break;
 495
 496                 snapc = ceph_get_snap_context(capsnap->context);
 497                 if (!page_snapc ||
 498                     page_snapc == snapc ||
 499                     page_snapc->seq > snapc->seq)
 500                         break;
 501         }
 502         if (!snapc && ci->i_wrbuffer_ref_head) {
 503                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 504                 dout(" head snapc %p has %d dirty pages\n",
 505                      snapc, ci->i_wrbuffer_ref_head);
 506                 if (ctl) {
 507                         ctl->i_size = i_size_read(inode);
 508                         ctl->truncate_size = ci->i_truncate_size;
 509                         ctl->truncate_seq = ci->i_truncate_seq;
 510                         ctl->size_stable = false;
 511                         ctl->head_snapc = true;
 512                 }
 513         }
 514         spin_unlock(&ci->i_ceph_lock);
 515         return snapc;
 516 }
 517
 518 static u64 get_writepages_data_length(struct inode *inode,
 519                                       struct page *page, u64 start)
 520 {
 521         struct ceph_inode_info *ci = ceph_inode(inode);
 522         struct ceph_snap_context *snapc = page_snap_context(page);
 523         struct ceph_cap_snap *capsnap = NULL;
 524         u64 end = i_size_read(inode);
 525
 526         if (snapc != ci->i_head_snapc) {
 527                 bool found = false;
 528                 spin_lock(&ci->i_ceph_lock);
 529                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 530                         if (capsnap->context == snapc) {
 531                                 if (!capsnap->writing)
 532                                         end = capsnap->size;
 533                                 found = true;
 534                                 break;
 535                         }
 536                 }
 537                 spin_unlock(&ci->i_ceph_lock);
 538                 WARN_ON(!found);
 539         }
 540         if (end > page_offset(page) + thp_size(page))
 541                 end = page_offset(page) + thp_size(page);
 542         return end > start ? end - start : 0;
 543 }
 544
 545 /*
 546  * Write a single page, but leave the page locked.
 547  *
 548  * If we get a write error, mark the mapping for error, but still adjust the
 549  * dirty page accounting (i.e., page is no longer dirty).
 550  */
 551 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 552 {
 553         struct folio *folio = page_folio(page);
 554         struct inode *inode = page->mapping->host;
 555         struct ceph_inode_info *ci = ceph_inode(inode);
 556         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 557         struct ceph_snap_context *snapc, *oldest;
 558         loff_t page_off = page_offset(page);
 559         int err;
 560         loff_t len = thp_size(page);
 561         struct ceph_writeback_ctl ceph_wbc;
 562         struct ceph_osd_client *osdc = &fsc->client->osdc;
 563         struct ceph_osd_request *req;
 564         bool caching = ceph_is_cache_enabled(inode);
 565
 566         dout("writepage %p idx %lu\n", page, page->index);
 567
 568         /* verify this is a writeable snap context */
 569         snapc = page_snap_context(page);
 570         if (!snapc) {
 571                 dout("writepage %p page %p not dirty?\n", inode, page);
 572                 return 0;
 573         }
 574         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 575         if (snapc->seq > oldest->seq) {
 576                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 577                      inode, page, snapc);
 578                 /* we should only noop if called by kswapd */
 579                 WARN_ON(!(current->flags & PF_MEMALLOC));
 580                 ceph_put_snap_context(oldest);
 581                 redirty_page_for_writepage(wbc, page);
 582                 return 0;
 583         }
 584         ceph_put_snap_context(oldest);
 585
 586         /* is this a partial page at end of file? */
 587         if (page_off >= ceph_wbc.i_size) {
 588                 dout("folio at %lu beyond eof %llu\n", folio->index,
 589                                 ceph_wbc.i_size);
 590                 folio_invalidate(folio, 0, folio_size(folio));
 591                 return 0;
 592         }
 593
 594         if (ceph_wbc.i_size < page_off + len)
 595                 len = ceph_wbc.i_size - page_off;
 596
 597         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 598              inode, page, page->index, page_off, len, snapc, snapc->seq);
 599
 600         if (atomic_long_inc_return(&fsc->writeback_count) >
 601             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 602                 fsc->write_congested = true;
 603
 604         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 605                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 606                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 607                                     true);
 608         if (IS_ERR(req)) {
 609                 redirty_page_for_writepage(wbc, page);
 610                 return PTR_ERR(req);
 611         }
 612
 613         set_page_writeback(page);
 614         if (caching)
 615                 ceph_set_page_fscache(page);
 616         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 617
 618         /* it may be a short write due to an object boundary */
 619         WARN_ON_ONCE(len > thp_size(page));
 620         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 621         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 622
 623         req->r_mtime = inode->i_mtime;
 624         err = ceph_osdc_start_request(osdc, req, true);
 625         if (!err)
 626                 err = ceph_osdc_wait_request(osdc, req);
 627
 628         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 629                                   req->r_end_latency, len, err);
 630
 631         ceph_osdc_put_request(req);
 632         if (err == 0)
 633                 err = len;
 634
 635         if (err < 0) {
 636                 struct writeback_control tmp_wbc;
 637                 if (!wbc)
 638                         wbc = &tmp_wbc;
 639                 if (err == -ERESTARTSYS) {
 640                         /* killed by SIGKILL */
 641                         dout("writepage interrupted page %p\n", page);
 642                         redirty_page_for_writepage(wbc, page);
 643                         end_page_writeback(page);
 644                         return err;
 645                 }
 646                 if (err == -EBLOCKLISTED)
 647                         fsc->blocklisted = true;
 648                 dout("writepage setting page/mapping error %d %p\n",
 649                      err, page);
 650                 mapping_set_error(&inode->i_data, err);
 651                 wbc->pages_skipped++;
 652         } else {
 653                 dout("writepage cleaned page %p\n", page);
 654                 err = 0;  /* vfs expects us to return 0 */
 655         }
 656         oldest = detach_page_private(page);
 657         WARN_ON_ONCE(oldest != snapc);
 658         end_page_writeback(page);
 659         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 660         ceph_put_snap_context(snapc);  /* page's reference */
 661
 662         if (atomic_long_dec_return(&fsc->writeback_count) <
 663             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 664                 fsc->write_congested = false;
 665
 666         return err;
 667 }
 668
 669 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 670 {
 671         int err;
 672         struct inode *inode = page->mapping->host;
 673         BUG_ON(!inode);
 674         ihold(inode);
 675
 676         if (wbc->sync_mode == WB_SYNC_NONE &&
 677             ceph_inode_to_client(inode)->write_congested)
 678                 return AOP_WRITEPAGE_ACTIVATE;
 679
 680         wait_on_page_fscache(page);
 681
 682         err = writepage_nounlock(page, wbc);
 683         if (err == -ERESTARTSYS) {
 684                 /* direct memory reclaimer was killed by SIGKILL. return 0
 685                  * to prevent caller from setting mapping/page error */
 686                 err = 0;
 687         }
 688         unlock_page(page);
 689         iput(inode);
 690         return err;
 691 }
 692
 693 /*
 694  * async writeback completion handler.
 695  *
 696  * If we get an error, set the mapping error bit, but not the individual
 697  * page error bits.
 698  */
 699 static void writepages_finish(struct ceph_osd_request *req)
 700 {
 701         struct inode *inode = req->r_inode;
 702         struct ceph_inode_info *ci = ceph_inode(inode);
 703         struct ceph_osd_data *osd_data;
 704         struct page *page;
 705         int num_pages, total_pages = 0;
 706         int i, j;
 707         int rc = req->r_result;
 708         struct ceph_snap_context *snapc = req->r_snapc;
 709         struct address_space *mapping = inode->i_mapping;
 710         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 711         unsigned int len = 0;
 712         bool remove_page;
 713
 714         dout("writepages_finish %p rc %d\n", inode, rc);
 715         if (rc < 0) {
 716                 mapping_set_error(mapping, rc);
 717                 ceph_set_error_write(ci);
 718                 if (rc == -EBLOCKLISTED)
 719                         fsc->blocklisted = true;
 720         } else {
 721                 ceph_clear_error_write(ci);
 722         }
 723
 724         /*
 725          * We lost the cache cap, need to truncate the page before
 726          * it is unlocked, otherwise we'd truncate it later in the
 727          * page truncation thread, possibly losing some data that
 728          * raced its way in
 729          */
 730         remove_page = !(ceph_caps_issued(ci) &
 731                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 732
 733         /* clean all pages */
 734         for (i = 0; i < req->r_num_ops; i++) {
 735                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 736                         pr_warn("%s incorrect op %d req %p index %d tid %llu\n",
 737                                 __func__, req->r_ops[i].op, req, i, req->r_tid);
 738                         break;
 739                 }
 740
 741                 osd_data = osd_req_op_extent_osd_data(req, i);
 742                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 743                 len += osd_data->length;
 744                 num_pages = calc_pages_for((u64)osd_data->alignment,
 745                                            (u64)osd_data->length);
 746                 total_pages += num_pages;
 747                 for (j = 0; j < num_pages; j++) {
 748                         page = osd_data->pages[j];
 749                         BUG_ON(!page);
 750                         WARN_ON(!PageUptodate(page));
 751
 752                         if (atomic_long_dec_return(&fsc->writeback_count) <
 753                              CONGESTION_OFF_THRESH(
 754                                         fsc->mount_options->congestion_kb))
 755                                 fsc->write_congested = false;
 756
 757                         ceph_put_snap_context(detach_page_private(page));
 758                         end_page_writeback(page);
 759                         dout("unlocking %p\n", page);
 760
 761                         if (remove_page)
 762                                 generic_error_remove_page(inode->i_mapping,
 763                                                           page);
 764
 765                         unlock_page(page);
 766                 }
 767                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 768                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 769
 770                 release_pages(osd_data->pages, num_pages);
 771         }
 772
 773         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 774                                   req->r_end_latency, len, rc);
 775
 776         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 777
 778         osd_data = osd_req_op_extent_osd_data(req, 0);
 779         if (osd_data->pages_from_pool)
 780                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 781         else
 782                 kfree(osd_data->pages);
 783         ceph_osdc_put_request(req);
 784 }
 785
 786 /*
 787  * initiate async writeback
 788  */
 789 static int ceph_writepages_start(struct address_space *mapping,
 790                                  struct writeback_control *wbc)
 791 {
 792         struct inode *inode = mapping->host;
 793         struct ceph_inode_info *ci = ceph_inode(inode);
 794         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 795         struct ceph_vino vino = ceph_vino(inode);
 796         pgoff_t index, start_index, end = -1;
 797         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 798         struct pagevec pvec;
 799         int rc = 0;
 800         unsigned int wsize = i_blocksize(inode);
 801         struct ceph_osd_request *req = NULL;
 802         struct ceph_writeback_ctl ceph_wbc;
 803         bool should_loop, range_whole = false;
 804         bool done = false;
 805         bool caching = ceph_is_cache_enabled(inode);
 806
 807         if (wbc->sync_mode == WB_SYNC_NONE &&
 808             fsc->write_congested)
 809                 return 0;
 810
 811         dout("writepages_start %p (mode=%s)\n", inode,
 812              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 813              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 814
 815         if (ceph_inode_is_shutdown(inode)) {
 816                 if (ci->i_wrbuffer_ref > 0) {
 817                         pr_warn_ratelimited(
 818                                 "writepage_start %p %lld forced umount\n",
 819                                 inode, ceph_ino(inode));
 820                 }
 821                 mapping_set_error(mapping, -EIO);
 822                 return -EIO; /* we're in a forced umount, don't write! */
 823         }
 824         if (fsc->mount_options->wsize < wsize)
 825                 wsize = fsc->mount_options->wsize;
 826
 827         pagevec_init(&pvec);
 828
 829         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 830         index = start_index;
 831
 832 retry:
 833         /* find oldest snap context with dirty data */
 834         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 835         if (!snapc) {
 836                 /* hmm, why does writepages get called when there
 837                    is no dirty data? */
 838                 dout(" no snap context with dirty data?\n");
 839                 goto out;
 840         }
 841         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 842              snapc, snapc->seq, snapc->num_snaps);
 843
 844         should_loop = false;
 845         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 846                 /* where to start/end? */
 847                 if (wbc->range_cyclic) {
 848                         index = start_index;
 849                         end = -1;
 850                         if (index > 0)
 851                                 should_loop = true;
 852                         dout(" cyclic, start at %lu\n", index);
 853                 } else {
 854                         index = wbc->range_start >> PAGE_SHIFT;
 855                         end = wbc->range_end >> PAGE_SHIFT;
 856                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 857                                 range_whole = true;
 858                         dout(" not cyclic, %lu to %lu\n", index, end);
 859                 }
 860         } else if (!ceph_wbc.head_snapc) {
 861                 /* Do not respect wbc->range_{start,end}. Dirty pages
 862                  * in that range can be associated with newer snapc.
 863                  * They are not writeable until we write all dirty pages
 864                  * associated with 'snapc' get written */
 865                 if (index > 0)
 866                         should_loop = true;
 867                 dout(" non-head snapc, range whole\n");
 868         }
 869
 870         ceph_put_snap_context(last_snapc);
 871         last_snapc = snapc;
 872
 873         while (!done && index <= end) {
 874                 int num_ops = 0, op_idx;
 875                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 876                 struct page **pages = NULL, **data_pages;
 877                 struct page *page;
 878                 pgoff_t strip_unit_end = 0;
 879                 u64 offset = 0, len = 0;
 880                 bool from_pool = false;
 881
 882                 max_pages = wsize >> PAGE_SHIFT;
 883
 884 get_more_pages:
 885                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 886                                                 end, PAGECACHE_TAG_DIRTY);
 887                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 888                 if (!pvec_pages && !locked_pages)
 889                         break;
 890                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 891                         page = pvec.pages[i];
 892                         dout("? %p idx %lu\n", page, page->index);
 893                         if (locked_pages == 0)
 894                                 lock_page(page);  /* first page */
 895                         else if (!trylock_page(page))
 896                                 break;
 897
 898                         /* only dirty pages, or our accounting breaks */
 899                         if (unlikely(!PageDirty(page)) ||
 900                             unlikely(page->mapping != mapping)) {
 901                                 dout("!dirty or !mapping %p\n", page);
 902                                 unlock_page(page);
 903                                 continue;
 904                         }
 905                         /* only if matching snap context */
 906                         pgsnapc = page_snap_context(page);
 907                         if (pgsnapc != snapc) {
 908                                 dout("page snapc %p %lld != oldest %p %lld\n",
 909                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 910                                 if (!should_loop &&
 911                                     !ceph_wbc.head_snapc &&
 912                                     wbc->sync_mode != WB_SYNC_NONE)
 913                                         should_loop = true;
 914                                 unlock_page(page);
 915                                 continue;
 916                         }
 917                         if (page_offset(page) >= ceph_wbc.i_size) {
 918                                 struct folio *folio = page_folio(page);
 919
 920                                 dout("folio at %lu beyond eof %llu\n",
 921                                      folio->index, ceph_wbc.i_size);
 922                                 if ((ceph_wbc.size_stable ||
 923                                     folio_pos(folio) >= i_size_read(inode)) &&
 924                                     folio_clear_dirty_for_io(folio))
 925                                         folio_invalidate(folio, 0,
 926                                                         folio_size(folio));
 927                                 folio_unlock(folio);
 928                                 continue;
 929                         }
 930                         if (strip_unit_end && (page->index > strip_unit_end)) {
 931                                 dout("end of strip unit %p\n", page);
 932                                 unlock_page(page);
 933                                 break;
 934                         }
 935                         if (PageWriteback(page) || PageFsCache(page)) {
 936                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 937                                         dout("%p under writeback\n", page);
 938                                         unlock_page(page);
 939                                         continue;
 940                                 }
 941                                 dout("waiting on writeback %p\n", page);
 942                                 wait_on_page_writeback(page);
 943                                 wait_on_page_fscache(page);
 944                         }
 945
 946                         if (!clear_page_dirty_for_io(page)) {
 947                                 dout("%p !clear_page_dirty_for_io\n", page);
 948                                 unlock_page(page);
 949                                 continue;
 950                         }
 951
 952                         /*
 953                          * We have something to write.  If this is
 954                          * the first locked page this time through,
 955                          * calculate max possinle write size and
 956                          * allocate a page array
 957                          */
 958                         if (locked_pages == 0) {
 959                                 u64 objnum;
 960                                 u64 objoff;
 961                                 u32 xlen;
 962
 963                                 /* prepare async write request */
 964                                 offset = (u64)page_offset(page);
 965                                 ceph_calc_file_object_mapping(&ci->i_layout,
 966                                                               offset, wsize,
 967                                                               &objnum, &objoff,
 968                                                               &xlen);
 969                                 len = xlen;
 970
 971                                 num_ops = 1;
 972                                 strip_unit_end = page->index +
 973                                         ((len - 1) >> PAGE_SHIFT);
 974
 975                                 BUG_ON(pages);
 976                                 max_pages = calc_pages_for(0, (u64)len);
 977                                 pages = kmalloc_array(max_pages,
 978                                                       sizeof(*pages),
 979                                                       GFP_NOFS);
 980                                 if (!pages) {
 981                                         from_pool = true;
 982                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 983                                         BUG_ON(!pages);
 984                                 }
 985
 986                                 len = 0;
 987                         } else if (page->index !=
 988                                    (offset + len) >> PAGE_SHIFT) {
 989                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 990                                                              CEPH_OSD_MAX_OPS)) {
 991                                         redirty_page_for_writepage(wbc, page);
 992                                         unlock_page(page);
 993                                         break;
 994                                 }
 995
 996                                 num_ops++;
 997                                 offset = (u64)page_offset(page);
 998                                 len = 0;
 999                         }
1000
1001                         /* note position of first page in pvec */
1002                         dout("%p will write page %p idx %lu\n",
1003                              inode, page, page->index);
1004
1005                         if (atomic_long_inc_return(&fsc->writeback_count) >
1006                             CONGESTION_ON_THRESH(
1007                                     fsc->mount_options->congestion_kb))
1008                                 fsc->write_congested = true;
1009
1010                         pages[locked_pages++] = page;
1011                         pvec.pages[i] = NULL;
1012
1013                         len += thp_size(page);
1014                 }
1015
1016                 /* did we get anything? */
1017                 if (!locked_pages)
1018                         goto release_pvec_pages;
1019                 if (i) {
1020                         unsigned j, n = 0;
1021                         /* shift unused page to beginning of pvec */
1022                         for (j = 0; j < pvec_pages; j++) {
1023                                 if (!pvec.pages[j])
1024                                         continue;
1025                                 if (n < j)
1026                                         pvec.pages[n] = pvec.pages[j];
1027                                 n++;
1028                         }
1029                         pvec.nr = n;
1030
1031                         if (pvec_pages && i == pvec_pages &&
1032                             locked_pages < max_pages) {
1033                                 dout("reached end pvec, trying for more\n");
1034                                 pagevec_release(&pvec);
1035                                 goto get_more_pages;
1036                         }
1037                 }
1038
1039 new_request:
1040                 offset = page_offset(pages[0]);
1041                 len = wsize;
1042
1043                 req = ceph_osdc_new_request(&fsc->client->osdc,
1044                                         &ci->i_layout, vino,
1045                                         offset, &len, 0, num_ops,
1046                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1047                                         snapc, ceph_wbc.truncate_seq,
1048                                         ceph_wbc.truncate_size, false);
1049                 if (IS_ERR(req)) {
1050                         req = ceph_osdc_new_request(&fsc->client->osdc,
1051                                                 &ci->i_layout, vino,
1052                                                 offset, &len, 0,
1053                                                 min(num_ops,
1054                                                     CEPH_OSD_SLAB_OPS),
1055                                                 CEPH_OSD_OP_WRITE,
1056                                                 CEPH_OSD_FLAG_WRITE,
1057                                                 snapc, ceph_wbc.truncate_seq,
1058                                                 ceph_wbc.truncate_size, true);
1059                         BUG_ON(IS_ERR(req));
1060                 }
1061                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1062                              thp_size(page) - offset);
1063
1064                 req->r_callback = writepages_finish;
1065                 req->r_inode = inode;
1066
1067                 /* Format the osd request message and submit the write */
1068                 len = 0;
1069                 data_pages = pages;
1070                 op_idx = 0;
1071                 for (i = 0; i < locked_pages; i++) {
1072                         u64 cur_offset = page_offset(pages[i]);
1073                         /*
1074                          * Discontinuity in page range? Ceph can handle that by just passing
1075                          * multiple extents in the write op.
1076                          */
1077                         if (offset + len != cur_offset) {
1078                                 /* If it's full, stop here */
1079                                 if (op_idx + 1 == req->r_num_ops)
1080                                         break;
1081
1082                                 /* Kick off an fscache write with what we have so far. */
1083                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1084
1085                                 /* Start a new extent */
1086                                 osd_req_op_extent_dup_last(req, op_idx,
1087                                                            cur_offset - offset);
1088                                 dout("writepages got pages at %llu~%llu\n",
1089                                      offset, len);
1090                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1091                                                         data_pages, len, 0,
1092                                                         from_pool, false);
1093                                 osd_req_op_extent_update(req, op_idx, len);
1094
1095                                 len = 0;
1096                                 offset = cur_offset;
1097                                 data_pages = pages + i;
1098                                 op_idx++;
1099                         }
1100
1101                         set_page_writeback(pages[i]);
1102                         if (caching)
1103                                 ceph_set_page_fscache(pages[i]);
1104                         len += thp_size(page);
1105                 }
1106                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1107
1108                 if (ceph_wbc.size_stable) {
1109                         len = min(len, ceph_wbc.i_size - offset);
1110                 } else if (i == locked_pages) {
1111                         /* writepages_finish() clears writeback pages
1112                          * according to the data length, so make sure
1113                          * data length covers all locked pages */
1114                         u64 min_len = len + 1 - thp_size(page);
1115                         len = get_writepages_data_length(inode, pages[i - 1],
1116                                                          offset);
1117                         len = max(len, min_len);
1118                 }
1119                 dout("writepages got pages at %llu~%llu\n", offset, len);
1120
1121                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1122                                                  0, from_pool, false);
1123                 osd_req_op_extent_update(req, op_idx, len);
1124
1125                 BUG_ON(op_idx + 1 != req->r_num_ops);
1126
1127                 from_pool = false;
1128                 if (i < locked_pages) {
1129                         BUG_ON(num_ops <= req->r_num_ops);
1130                         num_ops -= req->r_num_ops;
1131                         locked_pages -= i;
1132
1133                         /* allocate new pages array for next request */
1134                         data_pages = pages;
1135                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1136                                               GFP_NOFS);
1137                         if (!pages) {
1138                                 from_pool = true;
1139                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1140                                 BUG_ON(!pages);
1141                         }
1142                         memcpy(pages, data_pages + i,
1143                                locked_pages * sizeof(*pages));
1144                         memset(data_pages + i, 0,
1145                                locked_pages * sizeof(*pages));
1146                 } else {
1147                         BUG_ON(num_ops != req->r_num_ops);
1148                         index = pages[i - 1]->index + 1;
1149                         /* request message now owns the pages array */
1150                         pages = NULL;
1151                 }
1152
1153                 req->r_mtime = inode->i_mtime;
1154                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1155                 BUG_ON(rc);
1156                 req = NULL;
1157
1158                 wbc->nr_to_write -= i;
1159                 if (pages)
1160                         goto new_request;
1161
1162                 /*
1163                  * We stop writing back only if we are not doing
1164                  * integrity sync. In case of integrity sync we have to
1165                  * keep going until we have written all the pages
1166                  * we tagged for writeback prior to entering this loop.
1167                  */
1168                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1169                         done = true;
1170
1171 release_pvec_pages:
1172                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1173                      pvec.nr ? pvec.pages[0] : NULL);
1174                 pagevec_release(&pvec);
1175         }
1176
1177         if (should_loop && !done) {
1178                 /* more to do; loop back to beginning of file */
1179                 dout("writepages looping back to beginning of file\n");
1180                 end = start_index - 1; /* OK even when start_index == 0 */
1181
1182                 /* to write dirty pages associated with next snapc,
1183                  * we need to wait until current writes complete */
1184                 if (wbc->sync_mode != WB_SYNC_NONE &&
1185                     start_index == 0 && /* all dirty pages were checked */
1186                     !ceph_wbc.head_snapc) {
1187                         struct page *page;
1188                         unsigned i, nr;
1189                         index = 0;
1190                         while ((index <= end) &&
1191                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1192                                                 PAGECACHE_TAG_WRITEBACK))) {
1193                                 for (i = 0; i < nr; i++) {
1194                                         page = pvec.pages[i];
1195                                         if (page_snap_context(page) != snapc)
1196                                                 continue;
1197                                         wait_on_page_writeback(page);
1198                                 }
1199                                 pagevec_release(&pvec);
1200                                 cond_resched();
1201                         }
1202                 }
1203
1204                 start_index = 0;
1205                 index = 0;
1206                 goto retry;
1207         }
1208
1209         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1210                 mapping->writeback_index = index;
1211
1212 out:
1213         ceph_osdc_put_request(req);
1214         ceph_put_snap_context(last_snapc);
1215         dout("writepages dend - startone, rc = %d\n", rc);
1216         return rc;
1217 }
1218
1219
1220
1221 /*
1222  * See if a given @snapc is either writeable, or already written.
1223  */
1224 static int context_is_writeable_or_written(struct inode *inode,
1225                                            struct ceph_snap_context *snapc)
1226 {
1227         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1228         int ret = !oldest || snapc->seq <= oldest->seq;
1229
1230         ceph_put_snap_context(oldest);
1231         return ret;
1232 }
1233
1234 /**
1235  * ceph_find_incompatible - find an incompatible context and return it
1236  * @page: page being dirtied
1237  *
1238  * We are only allowed to write into/dirty a page if the page is
1239  * clean, or already dirty within the same snap context. Returns a
1240  * conflicting context if there is one, NULL if there isn't, or a
1241  * negative error code on other errors.
1242  *
1243  * Must be called with page lock held.
1244  */
1245 static struct ceph_snap_context *
1246 ceph_find_incompatible(struct page *page)
1247 {
1248         struct inode *inode = page->mapping->host;
1249         struct ceph_inode_info *ci = ceph_inode(inode);
1250
1251         if (ceph_inode_is_shutdown(inode)) {
1252                 dout(" page %p %llx:%llx is shutdown\n", page,
1253                      ceph_vinop(inode));
1254                 return ERR_PTR(-ESTALE);
1255         }
1256
1257         for (;;) {
1258                 struct ceph_snap_context *snapc, *oldest;
1259
1260                 wait_on_page_writeback(page);
1261
1262                 snapc = page_snap_context(page);
1263                 if (!snapc || snapc == ci->i_head_snapc)
1264                         break;
1265
1266                 /*
1267                  * this page is already dirty in another (older) snap
1268                  * context!  is it writeable now?
1269                  */
1270                 oldest = get_oldest_context(inode, NULL, NULL);
1271                 if (snapc->seq > oldest->seq) {
1272                         /* not writeable -- return it for the caller to deal with */
1273                         ceph_put_snap_context(oldest);
1274                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1275                         return ceph_get_snap_context(snapc);
1276                 }
1277                 ceph_put_snap_context(oldest);
1278
1279                 /* yay, writeable, do it now (without dropping page lock) */
1280                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1281                 if (clear_page_dirty_for_io(page)) {
1282                         int r = writepage_nounlock(page, NULL);
1283                         if (r < 0)
1284                                 return ERR_PTR(r);
1285                 }
1286         }
1287         return NULL;
1288 }
1289
1290 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1291                                         struct folio **foliop, void **_fsdata)
1292 {
1293         struct inode *inode = file_inode(file);
1294         struct ceph_inode_info *ci = ceph_inode(inode);
1295         struct ceph_snap_context *snapc;
1296
1297         snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1298         if (snapc) {
1299                 int r;
1300
1301                 folio_unlock(*foliop);
1302                 folio_put(*foliop);
1303                 *foliop = NULL;
1304                 if (IS_ERR(snapc))
1305                         return PTR_ERR(snapc);
1306
1307                 ceph_queue_writeback(inode);
1308                 r = wait_event_killable(ci->i_cap_wq,
1309                                         context_is_writeable_or_written(inode, snapc));
1310                 ceph_put_snap_context(snapc);
1311                 return r == 0 ? -EAGAIN : r;
1312         }
1313         return 0;
1314 }
1315
1316 /*
1317  * We are only allowed to write into/dirty the page if the page is
1318  * clean, or already dirty within the same snap context.
1319  */
1320 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1321                             loff_t pos, unsigned len,
1322                             struct page **pagep, void **fsdata)
1323 {
1324         struct inode *inode = file_inode(file);
1325         struct ceph_inode_info *ci = ceph_inode(inode);
1326         struct folio *folio = NULL;
1327         int r;
1328
1329         r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1330         if (r == 0)
1331                 folio_wait_fscache(folio);
1332         if (r < 0) {
1333                 if (folio)
1334                         folio_put(folio);
1335         } else {
1336                 WARN_ON_ONCE(!folio_test_locked(folio));
1337                 *pagep = &folio->page;
1338         }
1339         return r;
1340 }
1341
1342 /*
1343  * we don't do anything in here that simple_write_end doesn't do
1344  * except adjust dirty page accounting
1345  */
1346 static int ceph_write_end(struct file *file, struct address_space *mapping,
1347                           loff_t pos, unsigned len, unsigned copied,
1348                           struct page *subpage, void *fsdata)
1349 {
1350         struct folio *folio = page_folio(subpage);
1351         struct inode *inode = file_inode(file);
1352         bool check_cap = false;
1353
1354         dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1355              inode, folio, (int)pos, (int)copied, (int)len);
1356
1357         if (!folio_test_uptodate(folio)) {
1358                 /* just return that nothing was copied on a short copy */
1359                 if (copied < len) {
1360                         copied = 0;
1361                         goto out;
1362                 }
1363                 folio_mark_uptodate(folio);
1364         }
1365
1366         /* did file size increase? */
1367         if (pos+copied > i_size_read(inode))
1368                 check_cap = ceph_inode_set_size(inode, pos+copied);
1369
1370         folio_mark_dirty(folio);
1371
1372 out:
1373         folio_unlock(folio);
1374         folio_put(folio);
1375
1376         if (check_cap)
1377                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1378
1379         return copied;
1380 }
1381
1382 const struct address_space_operations ceph_aops = {
1383         .read_folio = netfs_read_folio,
1384         .readahead = netfs_readahead,
1385         .writepage = ceph_writepage,
1386         .writepages = ceph_writepages_start,
1387         .write_begin = ceph_write_begin,
1388         .write_end = ceph_write_end,
1389         .dirty_folio = ceph_dirty_folio,
1390         .invalidate_folio = ceph_invalidate_folio,
1391         .release_folio = ceph_release_folio,
1392         .direct_IO = noop_direct_IO,
1393 };
1394
1395 static void ceph_block_sigs(sigset_t *oldset)
1396 {
1397         sigset_t mask;
1398         siginitsetinv(&mask, sigmask(SIGKILL));
1399         sigprocmask(SIG_BLOCK, &mask, oldset);
1400 }
1401
1402 static void ceph_restore_sigs(sigset_t *oldset)
1403 {
1404         sigprocmask(SIG_SETMASK, oldset, NULL);
1405 }
1406
1407 /*
1408  * vm ops
1409  */
1410 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1411 {
1412         struct vm_area_struct *vma = vmf->vma;
1413         struct inode *inode = file_inode(vma->vm_file);
1414         struct ceph_inode_info *ci = ceph_inode(inode);
1415         struct ceph_file_info *fi = vma->vm_file->private_data;
1416         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1417         int want, got, err;
1418         sigset_t oldset;
1419         vm_fault_t ret = VM_FAULT_SIGBUS;
1420
1421         if (ceph_inode_is_shutdown(inode))
1422                 return ret;
1423
1424         ceph_block_sigs(&oldset);
1425
1426         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1427              inode, ceph_vinop(inode), off);
1428         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1429                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1430         else
1431                 want = CEPH_CAP_FILE_CACHE;
1432
1433         got = 0;
1434         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1435         if (err < 0)
1436                 goto out_restore;
1437
1438         dout("filemap_fault %p %llu got cap refs on %s\n",
1439              inode, off, ceph_cap_string(got));
1440
1441         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1442             ci->i_inline_version == CEPH_INLINE_NONE) {
1443                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1444                 ceph_add_rw_context(fi, &rw_ctx);
1445                 ret = filemap_fault(vmf);
1446                 ceph_del_rw_context(fi, &rw_ctx);
1447                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1448                      inode, off, ceph_cap_string(got), ret);
1449         } else
1450                 err = -EAGAIN;
1451
1452         ceph_put_cap_refs(ci, got);
1453
1454         if (err != -EAGAIN)
1455                 goto out_restore;
1456
1457         /* read inline data */
1458         if (off >= PAGE_SIZE) {
1459                 /* does not support inline data > PAGE_SIZE */
1460                 ret = VM_FAULT_SIGBUS;
1461         } else {
1462                 struct address_space *mapping = inode->i_mapping;
1463                 struct page *page;
1464
1465                 filemap_invalidate_lock_shared(mapping);
1466                 page = find_or_create_page(mapping, 0,
1467                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1468                 if (!page) {
1469                         ret = VM_FAULT_OOM;
1470                         goto out_inline;
1471                 }
1472                 err = __ceph_do_getattr(inode, page,
1473                                          CEPH_STAT_CAP_INLINE_DATA, true);
1474                 if (err < 0 || off >= i_size_read(inode)) {
1475                         unlock_page(page);
1476                         put_page(page);
1477                         ret = vmf_error(err);
1478                         goto out_inline;
1479                 }
1480                 if (err < PAGE_SIZE)
1481                         zero_user_segment(page, err, PAGE_SIZE);
1482                 else
1483                         flush_dcache_page(page);
1484                 SetPageUptodate(page);
1485                 vmf->page = page;
1486                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1487 out_inline:
1488                 filemap_invalidate_unlock_shared(mapping);
1489                 dout("filemap_fault %p %llu read inline data ret %x\n",
1490                      inode, off, ret);
1491         }
1492 out_restore:
1493         ceph_restore_sigs(&oldset);
1494         if (err < 0)
1495                 ret = vmf_error(err);
1496
1497         return ret;
1498 }
1499
1500 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1501 {
1502         struct vm_area_struct *vma = vmf->vma;
1503         struct inode *inode = file_inode(vma->vm_file);
1504         struct ceph_inode_info *ci = ceph_inode(inode);
1505         struct ceph_file_info *fi = vma->vm_file->private_data;
1506         struct ceph_cap_flush *prealloc_cf;
1507         struct page *page = vmf->page;
1508         loff_t off = page_offset(page);
1509         loff_t size = i_size_read(inode);
1510         size_t len;
1511         int want, got, err;
1512         sigset_t oldset;
1513         vm_fault_t ret = VM_FAULT_SIGBUS;
1514
1515         if (ceph_inode_is_shutdown(inode))
1516                 return ret;
1517
1518         prealloc_cf = ceph_alloc_cap_flush();
1519         if (!prealloc_cf)
1520                 return VM_FAULT_OOM;
1521
1522         sb_start_pagefault(inode->i_sb);
1523         ceph_block_sigs(&oldset);
1524
1525         if (off + thp_size(page) <= size)
1526                 len = thp_size(page);
1527         else
1528                 len = offset_in_thp(page, size);
1529
1530         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1531              inode, ceph_vinop(inode), off, len, size);
1532         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1533                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1534         else
1535                 want = CEPH_CAP_FILE_BUFFER;
1536
1537         got = 0;
1538         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1539         if (err < 0)
1540                 goto out_free;
1541
1542         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1543              inode, off, len, ceph_cap_string(got));
1544
1545         /* Update time before taking page lock */
1546         file_update_time(vma->vm_file);
1547         inode_inc_iversion_raw(inode);
1548
1549         do {
1550                 struct ceph_snap_context *snapc;
1551
1552                 lock_page(page);
1553
1554                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1555                         unlock_page(page);
1556                         ret = VM_FAULT_NOPAGE;
1557                         break;
1558                 }
1559
1560                 snapc = ceph_find_incompatible(page);
1561                 if (!snapc) {
1562                         /* success.  we'll keep the page locked. */
1563                         set_page_dirty(page);
1564                         ret = VM_FAULT_LOCKED;
1565                         break;
1566                 }
1567
1568                 unlock_page(page);
1569
1570                 if (IS_ERR(snapc)) {
1571                         ret = VM_FAULT_SIGBUS;
1572                         break;
1573                 }
1574
1575                 ceph_queue_writeback(inode);
1576                 err = wait_event_killable(ci->i_cap_wq,
1577                                 context_is_writeable_or_written(inode, snapc));
1578                 ceph_put_snap_context(snapc);
1579         } while (err == 0);
1580
1581         if (ret == VM_FAULT_LOCKED) {
1582                 int dirty;
1583                 spin_lock(&ci->i_ceph_lock);
1584                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1585                                                &prealloc_cf);
1586                 spin_unlock(&ci->i_ceph_lock);
1587                 if (dirty)
1588                         __mark_inode_dirty(inode, dirty);
1589         }
1590
1591         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1592              inode, off, len, ceph_cap_string(got), ret);
1593         ceph_put_cap_refs_async(ci, got);
1594 out_free:
1595         ceph_restore_sigs(&oldset);
1596         sb_end_pagefault(inode->i_sb);
1597         ceph_free_cap_flush(prealloc_cf);
1598         if (err < 0)
1599                 ret = vmf_error(err);
1600         return ret;
1601 }
1602
1603 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1604                            char *data, size_t len)
1605 {
1606         struct address_space *mapping = inode->i_mapping;
1607         struct page *page;
1608
1609         if (locked_page) {
1610                 page = locked_page;
1611         } else {
1612                 if (i_size_read(inode) == 0)
1613                         return;
1614                 page = find_or_create_page(mapping, 0,
1615                                            mapping_gfp_constraint(mapping,
1616                                            ~__GFP_FS));
1617                 if (!page)
1618                         return;
1619                 if (PageUptodate(page)) {
1620                         unlock_page(page);
1621                         put_page(page);
1622                         return;
1623                 }
1624         }
1625
1626         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1627              inode, ceph_vinop(inode), len, locked_page);
1628
1629         if (len > 0) {
1630                 void *kaddr = kmap_atomic(page);
1631                 memcpy(kaddr, data, len);
1632                 kunmap_atomic(kaddr);
1633         }
1634
1635         if (page != locked_page) {
1636                 if (len < PAGE_SIZE)
1637                         zero_user_segment(page, len, PAGE_SIZE);
1638                 else
1639                         flush_dcache_page(page);
1640
1641                 SetPageUptodate(page);
1642                 unlock_page(page);
1643                 put_page(page);
1644         }
1645 }
1646
1647 int ceph_uninline_data(struct file *file)
1648 {
1649         struct inode *inode = file_inode(file);
1650         struct ceph_inode_info *ci = ceph_inode(inode);
1651         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1652         struct ceph_osd_request *req = NULL;
1653         struct ceph_cap_flush *prealloc_cf;
1654         struct folio *folio = NULL;
1655         u64 inline_version = CEPH_INLINE_NONE;
1656         struct page *pages[1];
1657         int err = 0;
1658         u64 len;
1659
1660         spin_lock(&ci->i_ceph_lock);
1661         inline_version = ci->i_inline_version;
1662         spin_unlock(&ci->i_ceph_lock);
1663
1664         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1665              inode, ceph_vinop(inode), inline_version);
1666
1667         if (inline_version == CEPH_INLINE_NONE)
1668                 return 0;
1669
1670         prealloc_cf = ceph_alloc_cap_flush();
1671         if (!prealloc_cf)
1672                 return -ENOMEM;
1673
1674         if (inline_version == 1) /* initial version, no data */
1675                 goto out_uninline;
1676
1677         folio = read_mapping_folio(inode->i_mapping, 0, file);
1678         if (IS_ERR(folio)) {
1679                 err = PTR_ERR(folio);
1680                 goto out;
1681         }
1682
1683         folio_lock(folio);
1684
1685         len = i_size_read(inode);
1686         if (len > folio_size(folio))
1687                 len = folio_size(folio);
1688
1689         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1690                                     ceph_vino(inode), 0, &len, 0, 1,
1691                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1692                                     NULL, 0, 0, false);
1693         if (IS_ERR(req)) {
1694                 err = PTR_ERR(req);
1695                 goto out_unlock;
1696         }
1697
1698         req->r_mtime = inode->i_mtime;
1699         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1700         if (!err)
1701                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1702         ceph_osdc_put_request(req);
1703         if (err < 0)
1704                 goto out_unlock;
1705
1706         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1707                                     ceph_vino(inode), 0, &len, 1, 3,
1708                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1709                                     NULL, ci->i_truncate_seq,
1710                                     ci->i_truncate_size, false);
1711         if (IS_ERR(req)) {
1712                 err = PTR_ERR(req);
1713                 goto out_unlock;
1714         }
1715
1716         pages[0] = folio_page(folio, 0);
1717         osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1718
1719         {
1720                 __le64 xattr_buf = cpu_to_le64(inline_version);
1721                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1722                                             "inline_version", &xattr_buf,
1723                                             sizeof(xattr_buf),
1724                                             CEPH_OSD_CMPXATTR_OP_GT,
1725                                             CEPH_OSD_CMPXATTR_MODE_U64);
1726                 if (err)
1727                         goto out_put_req;
1728         }
1729
1730         {
1731                 char xattr_buf[32];
1732                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1733                                          "%llu", inline_version);
1734                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1735                                             "inline_version",
1736                                             xattr_buf, xattr_len, 0, 0);
1737                 if (err)
1738                         goto out_put_req;
1739         }
1740
1741         req->r_mtime = inode->i_mtime;
1742         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1743         if (!err)
1744                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1745
1746         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1747                                   req->r_end_latency, len, err);
1748
1749 out_uninline:
1750         if (!err) {
1751                 int dirty;
1752
1753                 /* Set to CAP_INLINE_NONE and dirty the caps */
1754                 down_read(&fsc->mdsc->snap_rwsem);
1755                 spin_lock(&ci->i_ceph_lock);
1756                 ci->i_inline_version = CEPH_INLINE_NONE;
1757                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1758                 spin_unlock(&ci->i_ceph_lock);
1759                 up_read(&fsc->mdsc->snap_rwsem);
1760                 if (dirty)
1761                         __mark_inode_dirty(inode, dirty);
1762         }
1763 out_put_req:
1764         ceph_osdc_put_request(req);
1765         if (err == -ECANCELED)
1766                 err = 0;
1767 out_unlock:
1768         if (folio) {
1769                 folio_unlock(folio);
1770                 folio_put(folio);
1771         }
1772 out:
1773         ceph_free_cap_flush(prealloc_cf);
1774         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1775              inode, ceph_vinop(inode), inline_version, err);
1776         return err;
1777 }
1778
1779 static const struct vm_operations_struct ceph_vmops = {
1780         .fault          = ceph_filemap_fault,
1781         .page_mkwrite   = ceph_page_mkwrite,
1782 };
1783
1784 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1785 {
1786         struct address_space *mapping = file->f_mapping;
1787
1788         if (!mapping->a_ops->read_folio)
1789                 return -ENOEXEC;
1790         vma->vm_ops = &ceph_vmops;
1791         return 0;
1792 }
1793
1794 enum {
1795         POOL_READ       = 1,
1796         POOL_WRITE      = 2,
1797 };
1798
1799 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1800                                 s64 pool, struct ceph_string *pool_ns)
1801 {
1802         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->netfs.inode);
1803         struct ceph_mds_client *mdsc = fsc->mdsc;
1804         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1805         struct rb_node **p, *parent;
1806         struct ceph_pool_perm *perm;
1807         struct page **pages;
1808         size_t pool_ns_len;
1809         int err = 0, err2 = 0, have = 0;
1810
1811         down_read(&mdsc->pool_perm_rwsem);
1812         p = &mdsc->pool_perm_tree.rb_node;
1813         while (*p) {
1814                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1815                 if (pool < perm->pool)
1816                         p = &(*p)->rb_left;
1817                 else if (pool > perm->pool)
1818                         p = &(*p)->rb_right;
1819                 else {
1820                         int ret = ceph_compare_string(pool_ns,
1821                                                 perm->pool_ns,
1822                                                 perm->pool_ns_len);
1823                         if (ret < 0)
1824                                 p = &(*p)->rb_left;
1825                         else if (ret > 0)
1826                                 p = &(*p)->rb_right;
1827                         else {
1828                                 have = perm->perm;
1829                                 break;
1830                         }
1831                 }
1832         }
1833         up_read(&mdsc->pool_perm_rwsem);
1834         if (*p)
1835                 goto out;
1836
1837         if (pool_ns)
1838                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1839                      pool, (int)pool_ns->len, pool_ns->str);
1840         else
1841                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1842
1843         down_write(&mdsc->pool_perm_rwsem);
1844         p = &mdsc->pool_perm_tree.rb_node;
1845         parent = NULL;
1846         while (*p) {
1847                 parent = *p;
1848                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1849                 if (pool < perm->pool)
1850                         p = &(*p)->rb_left;
1851                 else if (pool > perm->pool)
1852                         p = &(*p)->rb_right;
1853                 else {
1854                         int ret = ceph_compare_string(pool_ns,
1855                                                 perm->pool_ns,
1856                                                 perm->pool_ns_len);
1857                         if (ret < 0)
1858                                 p = &(*p)->rb_left;
1859                         else if (ret > 0)
1860                                 p = &(*p)->rb_right;
1861                         else {
1862                                 have = perm->perm;
1863                                 break;
1864                         }
1865                 }
1866         }
1867         if (*p) {
1868                 up_write(&mdsc->pool_perm_rwsem);
1869                 goto out;
1870         }
1871
1872         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1873                                          1, false, GFP_NOFS);
1874         if (!rd_req) {
1875                 err = -ENOMEM;
1876                 goto out_unlock;
1877         }
1878
1879         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1880         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1881         rd_req->r_base_oloc.pool = pool;
1882         if (pool_ns)
1883                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1884         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1885
1886         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1887         if (err)
1888                 goto out_unlock;
1889
1890         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1891                                          1, false, GFP_NOFS);
1892         if (!wr_req) {
1893                 err = -ENOMEM;
1894                 goto out_unlock;
1895         }
1896
1897         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1898         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1899         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1900         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1901
1902         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1903         if (err)
1904                 goto out_unlock;
1905
1906         /* one page should be large enough for STAT data */
1907         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1908         if (IS_ERR(pages)) {
1909                 err = PTR_ERR(pages);
1910                 goto out_unlock;
1911         }
1912
1913         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1914                                      0, false, true);
1915         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1916
1917         wr_req->r_mtime = ci->netfs.inode.i_mtime;
1918         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1919
1920         if (!err)
1921                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1922         if (!err2)
1923                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1924
1925         if (err >= 0 || err == -ENOENT)
1926                 have |= POOL_READ;
1927         else if (err != -EPERM) {
1928                 if (err == -EBLOCKLISTED)
1929                         fsc->blocklisted = true;
1930                 goto out_unlock;
1931         }
1932
1933         if (err2 == 0 || err2 == -EEXIST)
1934                 have |= POOL_WRITE;
1935         else if (err2 != -EPERM) {
1936                 if (err2 == -EBLOCKLISTED)
1937                         fsc->blocklisted = true;
1938                 err = err2;
1939                 goto out_unlock;
1940         }
1941
1942         pool_ns_len = pool_ns ? pool_ns->len : 0;
1943         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1944         if (!perm) {
1945                 err = -ENOMEM;
1946                 goto out_unlock;
1947         }
1948
1949         perm->pool = pool;
1950         perm->perm = have;
1951         perm->pool_ns_len = pool_ns_len;
1952         if (pool_ns_len > 0)
1953                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1954         perm->pool_ns[pool_ns_len] = 0;
1955
1956         rb_link_node(&perm->node, parent, p);
1957         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1958         err = 0;
1959 out_unlock:
1960         up_write(&mdsc->pool_perm_rwsem);
1961
1962         ceph_osdc_put_request(rd_req);
1963         ceph_osdc_put_request(wr_req);
1964 out:
1965         if (!err)
1966                 err = have;
1967         if (pool_ns)
1968                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1969                      pool, (int)pool_ns->len, pool_ns->str, err);
1970         else
1971                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1972         return err;
1973 }
1974
1975 int ceph_pool_perm_check(struct inode *inode, int need)
1976 {
1977         struct ceph_inode_info *ci = ceph_inode(inode);
1978         struct ceph_string *pool_ns;
1979         s64 pool;
1980         int ret, flags;
1981
1982         /* Only need to do this for regular files */
1983         if (!S_ISREG(inode->i_mode))
1984                 return 0;
1985
1986         if (ci->i_vino.snap != CEPH_NOSNAP) {
1987                 /*
1988                  * Pool permission check needs to write to the first object.
1989                  * But for snapshot, head of the first object may have alread
1990                  * been deleted. Skip check to avoid creating orphan object.
1991                  */
1992                 return 0;
1993         }
1994
1995         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
1996                                 NOPOOLPERM))
1997                 return 0;
1998
1999         spin_lock(&ci->i_ceph_lock);
2000         flags = ci->i_ceph_flags;
2001         pool = ci->i_layout.pool_id;
2002         spin_unlock(&ci->i_ceph_lock);
2003 check:
2004         if (flags & CEPH_I_POOL_PERM) {
2005                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2006                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2007                              pool);
2008                         return -EPERM;
2009                 }
2010                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2011                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2012                              pool);
2013                         return -EPERM;
2014                 }
2015                 return 0;
2016         }
2017
2018         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2019         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2020         ceph_put_string(pool_ns);
2021         if (ret < 0)
2022                 return ret;
2023
2024         flags = CEPH_I_POOL_PERM;
2025         if (ret & POOL_READ)
2026                 flags |= CEPH_I_POOL_RD;
2027         if (ret & POOL_WRITE)
2028                 flags |= CEPH_I_POOL_WR;
2029
2030         spin_lock(&ci->i_ceph_lock);
2031         if (pool == ci->i_layout.pool_id &&
2032             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2033                 ci->i_ceph_flags |= flags;
2034         } else {
2035                 pool = ci->i_layout.pool_id;
2036                 flags = ci->i_ceph_flags;
2037         }
2038         spin_unlock(&ci->i_ceph_lock);
2039         goto check;
2040 }
2041
2042 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2043 {
2044         struct ceph_pool_perm *perm;
2045         struct rb_node *n;
2046
2047         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2048                 n = rb_first(&mdsc->pool_perm_tree);
2049                 perm = rb_entry(n, struct ceph_pool_perm, node);
2050                 rb_erase(n, &mdsc->pool_perm_tree);
2051                 kfree(perm);
2052         }
2053 }