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