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