fs/ceph/addr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/backing-dev.h>
   5 #include <linux/fs.h>
   6 #include <linux/mm.h>
   7 #include <linux/swap.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/slab.h>
  10 #include <linux/pagevec.h>
  11 #include <linux/task_io_accounting_ops.h>
  12 #include <linux/signal.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15 #include <linux/netfs.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include "metric.h"
  21 #include <linux/ceph/osd_client.h>
  22 #include <linux/ceph/striper.h>
  23
  24 /*
  25  * Ceph address space ops.
  26  *
  27  * There are a few funny things going on here.
  28  *
  29  * The page->private field is used to reference a struct
  30  * ceph_snap_context for _every_ dirty page.  This indicates which
  31  * snapshot the page was logically dirtied in, and thus which snap
  32  * context needs to be associated with the osd write during writeback.
  33  *
  34  * Similarly, struct ceph_inode_info maintains a set of counters to
  35  * count dirty pages on the inode.  In the absence of snapshots,
  36  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  37  *
  38  * When a snapshot is taken (that is, when the client receives
  39  * notification that a snapshot was taken), each inode with caps and
  40  * with dirty pages (dirty pages implies there is a cap) gets a new
  41  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  42  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  43  * moved to capsnap->dirty. (Unless a sync write is currently in
  44  * progress.  In that case, the capsnap is said to be "pending", new
  45  * writes cannot start, and the capsnap isn't "finalized" until the
  46  * write completes (or fails) and a final size/mtime for the inode for
  47  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  48  *
  49  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  50  * we look for the first capsnap in i_cap_snaps and write out pages in
  51  * that snap context _only_.  Then we move on to the next capsnap,
  52  * eventually reaching the "live" or "head" context (i.e., pages that
  53  * are not yet snapped) and are writing the most recently dirtied
  54  * pages.
  55  *
  56  * Invalidate and so forth must take care to ensure the dirty page
  57  * accounting is preserved.
  58  */
  59
  60 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  61 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  62         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  63          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  64
  65 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  66                                         struct folio *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 invalidate_folio callback.
 125          */
 126         BUG_ON(PagePrivate(page));
 127         attach_page_private(page, snapc);
 128
 129         return ceph_fscache_set_page_dirty(page);
 130 }
 131
 132 /*
 133  * If we are truncating the full folio (i.e. offset == 0), adjust the
 134  * dirty folio counters appropriately.  Only called if there is private
 135  * data on the folio.
 136  */
 137 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 138                                 size_t length)
 139 {
 140         struct inode *inode;
 141         struct ceph_inode_info *ci;
 142         struct ceph_snap_context *snapc;
 143
 144         inode = folio->mapping->host;
 145         ci = ceph_inode(inode);
 146
 147         if (offset != 0 || length != folio_size(folio)) {
 148                 dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
 149                      inode, folio->index, offset, length);
 150                 return;
 151         }
 152
 153         WARN_ON(!folio_test_locked(folio));
 154         if (folio_get_private(folio)) {
 155                 dout("%p invalidate_folio idx %lu full dirty page\n",
 156                      inode, folio->index);
 157
 158                 snapc = folio_detach_private(folio);
 159                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 160                 ceph_put_snap_context(snapc);
 161         }
 162
 163         folio_wait_fscache(folio);
 164 }
 165
 166 static int ceph_releasepage(struct page *page, gfp_t gfp)
 167 {
 168         struct inode *inode = page->mapping->host;
 169
 170         dout("%llx:%llx releasepage %p idx %lu (%sdirty)\n",
 171              ceph_vinop(inode), page,
 172              page->index, PageDirty(page) ? "" : "not ");
 173
 174         if (PagePrivate(page))
 175                 return 0;
 176
 177         if (PageFsCache(page)) {
 178                 if (current_is_kswapd() || !(gfp & __GFP_FS))
 179                         return 0;
 180                 wait_on_page_fscache(page);
 181         }
 182         ceph_fscache_note_page_release(inode);
 183         return 1;
 184 }
 185
 186 static void ceph_netfs_expand_readahead(struct netfs_read_request *rreq)
 187 {
 188         struct inode *inode = rreq->mapping->host;
 189         struct ceph_inode_info *ci = ceph_inode(inode);
 190         struct ceph_file_layout *lo = &ci->i_layout;
 191         u32 blockoff;
 192         u64 blockno;
 193
 194         /* Expand the start downward */
 195         blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 196         rreq->start = blockno * lo->stripe_unit;
 197         rreq->len += blockoff;
 198
 199         /* Now, round up the length to the next block */
 200         rreq->len = roundup(rreq->len, lo->stripe_unit);
 201 }
 202
 203 static bool ceph_netfs_clamp_length(struct netfs_read_subrequest *subreq)
 204 {
 205         struct inode *inode = subreq->rreq->mapping->host;
 206         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 207         struct ceph_inode_info *ci = ceph_inode(inode);
 208         u64 objno, objoff;
 209         u32 xlen;
 210
 211         /* Truncate the extent at the end of the current block */
 212         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 213                                       &objno, &objoff, &xlen);
 214         subreq->len = min(xlen, fsc->mount_options->rsize);
 215         return true;
 216 }
 217
 218 static void finish_netfs_read(struct ceph_osd_request *req)
 219 {
 220         struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 221         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 222         struct netfs_read_subrequest *subreq = req->r_priv;
 223         int num_pages;
 224         int err = req->r_result;
 225
 226         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 227                                  req->r_end_latency, osd_data->length, err);
 228
 229         dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 230              subreq->len, i_size_read(req->r_inode));
 231
 232         /* no object means success but no data */
 233         if (err == -ENOENT)
 234                 err = 0;
 235         else if (err == -EBLOCKLISTED)
 236                 fsc->blocklisted = true;
 237
 238         if (err >= 0 && err < subreq->len)
 239                 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 240
 241         netfs_subreq_terminated(subreq, err, true);
 242
 243         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 244         ceph_put_page_vector(osd_data->pages, num_pages, false);
 245         iput(req->r_inode);
 246 }
 247
 248 static void ceph_netfs_issue_op(struct netfs_read_subrequest *subreq)
 249 {
 250         struct netfs_read_request *rreq = subreq->rreq;
 251         struct inode *inode = rreq->mapping->host;
 252         struct ceph_inode_info *ci = ceph_inode(inode);
 253         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 254         struct ceph_osd_request *req;
 255         struct ceph_vino vino = ceph_vino(inode);
 256         struct iov_iter iter;
 257         struct page **pages;
 258         size_t page_off;
 259         int err = 0;
 260         u64 len = subreq->len;
 261
 262         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 263                         0, 1, CEPH_OSD_OP_READ,
 264                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 265                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 266         if (IS_ERR(req)) {
 267                 err = PTR_ERR(req);
 268                 req = NULL;
 269                 goto out;
 270         }
 271
 272         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 273         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 274         err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 275         if (err < 0) {
 276                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 277                 goto out;
 278         }
 279
 280         /* should always give us a page-aligned read */
 281         WARN_ON_ONCE(page_off);
 282         len = err;
 283
 284         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 285         req->r_callback = finish_netfs_read;
 286         req->r_priv = subreq;
 287         req->r_inode = inode;
 288         ihold(inode);
 289
 290         err = ceph_osdc_start_request(req->r_osdc, req, false);
 291         if (err)
 292                 iput(inode);
 293 out:
 294         ceph_osdc_put_request(req);
 295         if (err)
 296                 netfs_subreq_terminated(subreq, err, false);
 297         dout("%s: result %d\n", __func__, err);
 298 }
 299
 300 static void ceph_readahead_cleanup(struct address_space *mapping, void *priv)
 301 {
 302         struct inode *inode = mapping->host;
 303         struct ceph_inode_info *ci = ceph_inode(inode);
 304         int got = (uintptr_t)priv;
 305
 306         if (got)
 307                 ceph_put_cap_refs(ci, got);
 308 }
 309
 310 static const struct netfs_read_request_ops ceph_netfs_read_ops = {
 311         .is_cache_enabled       = ceph_is_cache_enabled,
 312         .begin_cache_operation  = ceph_begin_cache_operation,
 313         .issue_op               = ceph_netfs_issue_op,
 314         .expand_readahead       = ceph_netfs_expand_readahead,
 315         .clamp_length           = ceph_netfs_clamp_length,
 316         .check_write_begin      = ceph_netfs_check_write_begin,
 317         .cleanup                = ceph_readahead_cleanup,
 318 };
 319
 320 /* read a single page, without unlocking it. */
 321 static int ceph_readpage(struct file *file, struct page *subpage)
 322 {
 323         struct folio *folio = page_folio(subpage);
 324         struct inode *inode = file_inode(file);
 325         struct ceph_inode_info *ci = ceph_inode(inode);
 326         struct ceph_vino vino = ceph_vino(inode);
 327         size_t len = folio_size(folio);
 328         u64 off = folio_file_pos(folio);
 329
 330         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 331                 /*
 332                  * Uptodate inline data should have been added
 333                  * into page cache while getting Fcr caps.
 334                  */
 335                 if (off == 0) {
 336                         folio_unlock(folio);
 337                         return -EINVAL;
 338                 }
 339                 zero_user_segment(&folio->page, 0, folio_size(folio));
 340                 folio_mark_uptodate(folio);
 341                 folio_unlock(folio);
 342                 return 0;
 343         }
 344
 345         dout("readpage ino %llx.%llx file %p off %llu len %zu folio %p index %lu\n",
 346              vino.ino, vino.snap, file, off, len, folio, folio_index(folio));
 347
 348         return netfs_readpage(file, folio, &ceph_netfs_read_ops, NULL);
 349 }
 350
 351 static void ceph_readahead(struct readahead_control *ractl)
 352 {
 353         struct inode *inode = file_inode(ractl->file);
 354         struct ceph_file_info *fi = ractl->file->private_data;
 355         struct ceph_rw_context *rw_ctx;
 356         int got = 0;
 357         int ret = 0;
 358
 359         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 360                 return;
 361
 362         rw_ctx = ceph_find_rw_context(fi);
 363         if (!rw_ctx) {
 364                 /*
 365                  * readahead callers do not necessarily hold Fcb caps
 366                  * (e.g. fadvise, madvise).
 367                  */
 368                 int want = CEPH_CAP_FILE_CACHE;
 369
 370                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 371                 if (ret < 0)
 372                         dout("start_read %p, error getting cap\n", inode);
 373                 else if (!(got & want))
 374                         dout("start_read %p, no cache cap\n", inode);
 375
 376                 if (ret <= 0)
 377                         return;
 378         }
 379         netfs_readahead(ractl, &ceph_netfs_read_ops, (void *)(uintptr_t)got);
 380 }
 381
 382 #ifdef CONFIG_CEPH_FSCACHE
 383 static void ceph_set_page_fscache(struct page *page)
 384 {
 385         set_page_fscache(page);
 386 }
 387
 388 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 389 {
 390         struct inode *inode = priv;
 391
 392         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 393                 ceph_fscache_invalidate(inode, false);
 394 }
 395
 396 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 397 {
 398         struct ceph_inode_info *ci = ceph_inode(inode);
 399         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 400
 401         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 402                                ceph_fscache_write_terminated, inode, caching);
 403 }
 404 #else
 405 static inline void ceph_set_page_fscache(struct page *page)
 406 {
 407 }
 408
 409 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 410 {
 411 }
 412 #endif /* CONFIG_CEPH_FSCACHE */
 413
 414 struct ceph_writeback_ctl
 415 {
 416         loff_t i_size;
 417         u64 truncate_size;
 418         u32 truncate_seq;
 419         bool size_stable;
 420         bool head_snapc;
 421 };
 422
 423 /*
 424  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 425  * only snap context we are allowed to write back.
 426  */
 427 static struct ceph_snap_context *
 428 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 429                    struct ceph_snap_context *page_snapc)
 430 {
 431         struct ceph_inode_info *ci = ceph_inode(inode);
 432         struct ceph_snap_context *snapc = NULL;
 433         struct ceph_cap_snap *capsnap = NULL;
 434
 435         spin_lock(&ci->i_ceph_lock);
 436         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 437                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 438                      capsnap->context, capsnap->dirty_pages);
 439                 if (!capsnap->dirty_pages)
 440                         continue;
 441
 442                 /* get i_size, truncate_{seq,size} for page_snapc? */
 443                 if (snapc && capsnap->context != page_snapc)
 444                         continue;
 445
 446                 if (ctl) {
 447                         if (capsnap->writing) {
 448                                 ctl->i_size = i_size_read(inode);
 449                                 ctl->size_stable = false;
 450                         } else {
 451                                 ctl->i_size = capsnap->size;
 452                                 ctl->size_stable = true;
 453                         }
 454                         ctl->truncate_size = capsnap->truncate_size;
 455                         ctl->truncate_seq = capsnap->truncate_seq;
 456                         ctl->head_snapc = false;
 457                 }
 458
 459                 if (snapc)
 460                         break;
 461
 462                 snapc = ceph_get_snap_context(capsnap->context);
 463                 if (!page_snapc ||
 464                     page_snapc == snapc ||
 465                     page_snapc->seq > snapc->seq)
 466                         break;
 467         }
 468         if (!snapc && ci->i_wrbuffer_ref_head) {
 469                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 470                 dout(" head snapc %p has %d dirty pages\n",
 471                      snapc, ci->i_wrbuffer_ref_head);
 472                 if (ctl) {
 473                         ctl->i_size = i_size_read(inode);
 474                         ctl->truncate_size = ci->i_truncate_size;
 475                         ctl->truncate_seq = ci->i_truncate_seq;
 476                         ctl->size_stable = false;
 477                         ctl->head_snapc = true;
 478                 }
 479         }
 480         spin_unlock(&ci->i_ceph_lock);
 481         return snapc;
 482 }
 483
 484 static u64 get_writepages_data_length(struct inode *inode,
 485                                       struct page *page, u64 start)
 486 {
 487         struct ceph_inode_info *ci = ceph_inode(inode);
 488         struct ceph_snap_context *snapc = page_snap_context(page);
 489         struct ceph_cap_snap *capsnap = NULL;
 490         u64 end = i_size_read(inode);
 491
 492         if (snapc != ci->i_head_snapc) {
 493                 bool found = false;
 494                 spin_lock(&ci->i_ceph_lock);
 495                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 496                         if (capsnap->context == snapc) {
 497                                 if (!capsnap->writing)
 498                                         end = capsnap->size;
 499                                 found = true;
 500                                 break;
 501                         }
 502                 }
 503                 spin_unlock(&ci->i_ceph_lock);
 504                 WARN_ON(!found);
 505         }
 506         if (end > page_offset(page) + thp_size(page))
 507                 end = page_offset(page) + thp_size(page);
 508         return end > start ? end - start : 0;
 509 }
 510
 511 /*
 512  * Write a single page, but leave the page locked.
 513  *
 514  * If we get a write error, mark the mapping for error, but still adjust the
 515  * dirty page accounting (i.e., page is no longer dirty).
 516  */
 517 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 518 {
 519         struct folio *folio = page_folio(page);
 520         struct inode *inode = page->mapping->host;
 521         struct ceph_inode_info *ci = ceph_inode(inode);
 522         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 523         struct ceph_snap_context *snapc, *oldest;
 524         loff_t page_off = page_offset(page);
 525         int err;
 526         loff_t len = thp_size(page);
 527         struct ceph_writeback_ctl ceph_wbc;
 528         struct ceph_osd_client *osdc = &fsc->client->osdc;
 529         struct ceph_osd_request *req;
 530         bool caching = ceph_is_cache_enabled(inode);
 531
 532         dout("writepage %p idx %lu\n", page, page->index);
 533
 534         /* verify this is a writeable snap context */
 535         snapc = page_snap_context(page);
 536         if (!snapc) {
 537                 dout("writepage %p page %p not dirty?\n", inode, page);
 538                 return 0;
 539         }
 540         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 541         if (snapc->seq > oldest->seq) {
 542                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 543                      inode, page, snapc);
 544                 /* we should only noop if called by kswapd */
 545                 WARN_ON(!(current->flags & PF_MEMALLOC));
 546                 ceph_put_snap_context(oldest);
 547                 redirty_page_for_writepage(wbc, page);
 548                 return 0;
 549         }
 550         ceph_put_snap_context(oldest);
 551
 552         /* is this a partial page at end of file? */
 553         if (page_off >= ceph_wbc.i_size) {
 554                 dout("folio at %lu beyond eof %llu\n", folio->index,
 555                                 ceph_wbc.i_size);
 556                 folio_invalidate(folio, 0, folio_size(folio));
 557                 return 0;
 558         }
 559
 560         if (ceph_wbc.i_size < page_off + len)
 561                 len = ceph_wbc.i_size - page_off;
 562
 563         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 564              inode, page, page->index, page_off, len, snapc, snapc->seq);
 565
 566         if (atomic_long_inc_return(&fsc->writeback_count) >
 567             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 568                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 569
 570         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 571                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 572                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 573                                     true);
 574         if (IS_ERR(req))
 575                 return PTR_ERR(req);
 576
 577         set_page_writeback(page);
 578         if (caching)
 579                 ceph_set_page_fscache(page);
 580         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 581
 582         /* it may be a short write due to an object boundary */
 583         WARN_ON_ONCE(len > thp_size(page));
 584         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 585         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 586
 587         req->r_mtime = inode->i_mtime;
 588         err = ceph_osdc_start_request(osdc, req, true);
 589         if (!err)
 590                 err = ceph_osdc_wait_request(osdc, req);
 591
 592         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 593                                   req->r_end_latency, len, err);
 594
 595         ceph_osdc_put_request(req);
 596         if (err == 0)
 597                 err = len;
 598
 599         if (err < 0) {
 600                 struct writeback_control tmp_wbc;
 601                 if (!wbc)
 602                         wbc = &tmp_wbc;
 603                 if (err == -ERESTARTSYS) {
 604                         /* killed by SIGKILL */
 605                         dout("writepage interrupted page %p\n", page);
 606                         redirty_page_for_writepage(wbc, page);
 607                         end_page_writeback(page);
 608                         return err;
 609                 }
 610                 if (err == -EBLOCKLISTED)
 611                         fsc->blocklisted = true;
 612                 dout("writepage setting page/mapping error %d %p\n",
 613                      err, page);
 614                 mapping_set_error(&inode->i_data, err);
 615                 wbc->pages_skipped++;
 616         } else {
 617                 dout("writepage cleaned page %p\n", page);
 618                 err = 0;  /* vfs expects us to return 0 */
 619         }
 620         oldest = detach_page_private(page);
 621         WARN_ON_ONCE(oldest != snapc);
 622         end_page_writeback(page);
 623         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 624         ceph_put_snap_context(snapc);  /* page's reference */
 625
 626         if (atomic_long_dec_return(&fsc->writeback_count) <
 627             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 628                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 629
 630         return err;
 631 }
 632
 633 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 634 {
 635         int err;
 636         struct inode *inode = page->mapping->host;
 637         BUG_ON(!inode);
 638         ihold(inode);
 639
 640         wait_on_page_fscache(page);
 641
 642         err = writepage_nounlock(page, wbc);
 643         if (err == -ERESTARTSYS) {
 644                 /* direct memory reclaimer was killed by SIGKILL. return 0
 645                  * to prevent caller from setting mapping/page error */
 646                 err = 0;
 647         }
 648         unlock_page(page);
 649         iput(inode);
 650         return err;
 651 }
 652
 653 /*
 654  * async writeback completion handler.
 655  *
 656  * If we get an error, set the mapping error bit, but not the individual
 657  * page error bits.
 658  */
 659 static void writepages_finish(struct ceph_osd_request *req)
 660 {
 661         struct inode *inode = req->r_inode;
 662         struct ceph_inode_info *ci = ceph_inode(inode);
 663         struct ceph_osd_data *osd_data;
 664         struct page *page;
 665         int num_pages, total_pages = 0;
 666         int i, j;
 667         int rc = req->r_result;
 668         struct ceph_snap_context *snapc = req->r_snapc;
 669         struct address_space *mapping = inode->i_mapping;
 670         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 671         unsigned int len = 0;
 672         bool remove_page;
 673
 674         dout("writepages_finish %p rc %d\n", inode, rc);
 675         if (rc < 0) {
 676                 mapping_set_error(mapping, rc);
 677                 ceph_set_error_write(ci);
 678                 if (rc == -EBLOCKLISTED)
 679                         fsc->blocklisted = true;
 680         } else {
 681                 ceph_clear_error_write(ci);
 682         }
 683
 684         /*
 685          * We lost the cache cap, need to truncate the page before
 686          * it is unlocked, otherwise we'd truncate it later in the
 687          * page truncation thread, possibly losing some data that
 688          * raced its way in
 689          */
 690         remove_page = !(ceph_caps_issued(ci) &
 691                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 692
 693         /* clean all pages */
 694         for (i = 0; i < req->r_num_ops; i++) {
 695                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 696                         break;
 697
 698                 osd_data = osd_req_op_extent_osd_data(req, i);
 699                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 700                 len += osd_data->length;
 701                 num_pages = calc_pages_for((u64)osd_data->alignment,
 702                                            (u64)osd_data->length);
 703                 total_pages += num_pages;
 704                 for (j = 0; j < num_pages; j++) {
 705                         page = osd_data->pages[j];
 706                         BUG_ON(!page);
 707                         WARN_ON(!PageUptodate(page));
 708
 709                         if (atomic_long_dec_return(&fsc->writeback_count) <
 710                              CONGESTION_OFF_THRESH(
 711                                         fsc->mount_options->congestion_kb))
 712                                 clear_bdi_congested(inode_to_bdi(inode),
 713                                                     BLK_RW_ASYNC);
 714
 715                         ceph_put_snap_context(detach_page_private(page));
 716                         end_page_writeback(page);
 717                         dout("unlocking %p\n", page);
 718
 719                         if (remove_page)
 720                                 generic_error_remove_page(inode->i_mapping,
 721                                                           page);
 722
 723                         unlock_page(page);
 724                 }
 725                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 726                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 727
 728                 release_pages(osd_data->pages, num_pages);
 729         }
 730
 731         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 732                                   req->r_end_latency, len, rc);
 733
 734         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 735
 736         osd_data = osd_req_op_extent_osd_data(req, 0);
 737         if (osd_data->pages_from_pool)
 738                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 739         else
 740                 kfree(osd_data->pages);
 741         ceph_osdc_put_request(req);
 742 }
 743
 744 /*
 745  * initiate async writeback
 746  */
 747 static int ceph_writepages_start(struct address_space *mapping,
 748                                  struct writeback_control *wbc)
 749 {
 750         struct inode *inode = mapping->host;
 751         struct ceph_inode_info *ci = ceph_inode(inode);
 752         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 753         struct ceph_vino vino = ceph_vino(inode);
 754         pgoff_t index, start_index, end = -1;
 755         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 756         struct pagevec pvec;
 757         int rc = 0;
 758         unsigned int wsize = i_blocksize(inode);
 759         struct ceph_osd_request *req = NULL;
 760         struct ceph_writeback_ctl ceph_wbc;
 761         bool should_loop, range_whole = false;
 762         bool done = false;
 763         bool caching = ceph_is_cache_enabled(inode);
 764
 765         dout("writepages_start %p (mode=%s)\n", inode,
 766              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 767              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 768
 769         if (ceph_inode_is_shutdown(inode)) {
 770                 if (ci->i_wrbuffer_ref > 0) {
 771                         pr_warn_ratelimited(
 772                                 "writepage_start %p %lld forced umount\n",
 773                                 inode, ceph_ino(inode));
 774                 }
 775                 mapping_set_error(mapping, -EIO);
 776                 return -EIO; /* we're in a forced umount, don't write! */
 777         }
 778         if (fsc->mount_options->wsize < wsize)
 779                 wsize = fsc->mount_options->wsize;
 780
 781         pagevec_init(&pvec);
 782
 783         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 784         index = start_index;
 785
 786 retry:
 787         /* find oldest snap context with dirty data */
 788         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 789         if (!snapc) {
 790                 /* hmm, why does writepages get called when there
 791                    is no dirty data? */
 792                 dout(" no snap context with dirty data?\n");
 793                 goto out;
 794         }
 795         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 796              snapc, snapc->seq, snapc->num_snaps);
 797
 798         should_loop = false;
 799         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 800                 /* where to start/end? */
 801                 if (wbc->range_cyclic) {
 802                         index = start_index;
 803                         end = -1;
 804                         if (index > 0)
 805                                 should_loop = true;
 806                         dout(" cyclic, start at %lu\n", index);
 807                 } else {
 808                         index = wbc->range_start >> PAGE_SHIFT;
 809                         end = wbc->range_end >> PAGE_SHIFT;
 810                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 811                                 range_whole = true;
 812                         dout(" not cyclic, %lu to %lu\n", index, end);
 813                 }
 814         } else if (!ceph_wbc.head_snapc) {
 815                 /* Do not respect wbc->range_{start,end}. Dirty pages
 816                  * in that range can be associated with newer snapc.
 817                  * They are not writeable until we write all dirty pages
 818                  * associated with 'snapc' get written */
 819                 if (index > 0)
 820                         should_loop = true;
 821                 dout(" non-head snapc, range whole\n");
 822         }
 823
 824         ceph_put_snap_context(last_snapc);
 825         last_snapc = snapc;
 826
 827         while (!done && index <= end) {
 828                 int num_ops = 0, op_idx;
 829                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 830                 struct page **pages = NULL, **data_pages;
 831                 struct page *page;
 832                 pgoff_t strip_unit_end = 0;
 833                 u64 offset = 0, len = 0;
 834                 bool from_pool = false;
 835
 836                 max_pages = wsize >> PAGE_SHIFT;
 837
 838 get_more_pages:
 839                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 840                                                 end, PAGECACHE_TAG_DIRTY);
 841                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 842                 if (!pvec_pages && !locked_pages)
 843                         break;
 844                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 845                         page = pvec.pages[i];
 846                         dout("? %p idx %lu\n", page, page->index);
 847                         if (locked_pages == 0)
 848                                 lock_page(page);  /* first page */
 849                         else if (!trylock_page(page))
 850                                 break;
 851
 852                         /* only dirty pages, or our accounting breaks */
 853                         if (unlikely(!PageDirty(page)) ||
 854                             unlikely(page->mapping != mapping)) {
 855                                 dout("!dirty or !mapping %p\n", page);
 856                                 unlock_page(page);
 857                                 continue;
 858                         }
 859                         /* only if matching snap context */
 860                         pgsnapc = page_snap_context(page);
 861                         if (pgsnapc != snapc) {
 862                                 dout("page snapc %p %lld != oldest %p %lld\n",
 863                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 864                                 if (!should_loop &&
 865                                     !ceph_wbc.head_snapc &&
 866                                     wbc->sync_mode != WB_SYNC_NONE)
 867                                         should_loop = true;
 868                                 unlock_page(page);
 869                                 continue;
 870                         }
 871                         if (page_offset(page) >= ceph_wbc.i_size) {
 872                                 struct folio *folio = page_folio(page);
 873
 874                                 dout("folio at %lu beyond eof %llu\n",
 875                                      folio->index, ceph_wbc.i_size);
 876                                 if ((ceph_wbc.size_stable ||
 877                                     folio_pos(folio) >= i_size_read(inode)) &&
 878                                     folio_clear_dirty_for_io(folio))
 879                                         folio_invalidate(folio, 0,
 880                                                         folio_size(folio));
 881                                 folio_unlock(folio);
 882                                 continue;
 883                         }
 884                         if (strip_unit_end && (page->index > strip_unit_end)) {
 885                                 dout("end of strip unit %p\n", page);
 886                                 unlock_page(page);
 887                                 break;
 888                         }
 889                         if (PageWriteback(page) || PageFsCache(page)) {
 890                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 891                                         dout("%p under writeback\n", page);
 892                                         unlock_page(page);
 893                                         continue;
 894                                 }
 895                                 dout("waiting on writeback %p\n", page);
 896                                 wait_on_page_writeback(page);
 897                                 wait_on_page_fscache(page);
 898                         }
 899
 900                         if (!clear_page_dirty_for_io(page)) {
 901                                 dout("%p !clear_page_dirty_for_io\n", page);
 902                                 unlock_page(page);
 903                                 continue;
 904                         }
 905
 906                         /*
 907                          * We have something to write.  If this is
 908                          * the first locked page this time through,
 909                          * calculate max possinle write size and
 910                          * allocate a page array
 911                          */
 912                         if (locked_pages == 0) {
 913                                 u64 objnum;
 914                                 u64 objoff;
 915                                 u32 xlen;
 916
 917                                 /* prepare async write request */
 918                                 offset = (u64)page_offset(page);
 919                                 ceph_calc_file_object_mapping(&ci->i_layout,
 920                                                               offset, wsize,
 921                                                               &objnum, &objoff,
 922                                                               &xlen);
 923                                 len = xlen;
 924
 925                                 num_ops = 1;
 926                                 strip_unit_end = page->index +
 927                                         ((len - 1) >> PAGE_SHIFT);
 928
 929                                 BUG_ON(pages);
 930                                 max_pages = calc_pages_for(0, (u64)len);
 931                                 pages = kmalloc_array(max_pages,
 932                                                       sizeof(*pages),
 933                                                       GFP_NOFS);
 934                                 if (!pages) {
 935                                         from_pool = true;
 936                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 937                                         BUG_ON(!pages);
 938                                 }
 939
 940                                 len = 0;
 941                         } else if (page->index !=
 942                                    (offset + len) >> PAGE_SHIFT) {
 943                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 944                                                              CEPH_OSD_MAX_OPS)) {
 945                                         redirty_page_for_writepage(wbc, page);
 946                                         unlock_page(page);
 947                                         break;
 948                                 }
 949
 950                                 num_ops++;
 951                                 offset = (u64)page_offset(page);
 952                                 len = 0;
 953                         }
 954
 955                         /* note position of first page in pvec */
 956                         dout("%p will write page %p idx %lu\n",
 957                              inode, page, page->index);
 958
 959                         if (atomic_long_inc_return(&fsc->writeback_count) >
 960                             CONGESTION_ON_THRESH(
 961                                     fsc->mount_options->congestion_kb)) {
 962                                 set_bdi_congested(inode_to_bdi(inode),
 963                                                   BLK_RW_ASYNC);
 964                         }
 965
 966
 967                         pages[locked_pages++] = page;
 968                         pvec.pages[i] = NULL;
 969
 970                         len += thp_size(page);
 971                 }
 972
 973                 /* did we get anything? */
 974                 if (!locked_pages)
 975                         goto release_pvec_pages;
 976                 if (i) {
 977                         unsigned j, n = 0;
 978                         /* shift unused page to beginning of pvec */
 979                         for (j = 0; j < pvec_pages; j++) {
 980                                 if (!pvec.pages[j])
 981                                         continue;
 982                                 if (n < j)
 983                                         pvec.pages[n] = pvec.pages[j];
 984                                 n++;
 985                         }
 986                         pvec.nr = n;
 987
 988                         if (pvec_pages && i == pvec_pages &&
 989                             locked_pages < max_pages) {
 990                                 dout("reached end pvec, trying for more\n");
 991                                 pagevec_release(&pvec);
 992                                 goto get_more_pages;
 993                         }
 994                 }
 995
 996 new_request:
 997                 offset = page_offset(pages[0]);
 998                 len = wsize;
 999
1000                 req = ceph_osdc_new_request(&fsc->client->osdc,
1001                                         &ci->i_layout, vino,
1002                                         offset, &len, 0, num_ops,
1003                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1004                                         snapc, ceph_wbc.truncate_seq,
1005                                         ceph_wbc.truncate_size, false);
1006                 if (IS_ERR(req)) {
1007                         req = ceph_osdc_new_request(&fsc->client->osdc,
1008                                                 &ci->i_layout, vino,
1009                                                 offset, &len, 0,
1010                                                 min(num_ops,
1011                                                     CEPH_OSD_SLAB_OPS),
1012                                                 CEPH_OSD_OP_WRITE,
1013                                                 CEPH_OSD_FLAG_WRITE,
1014                                                 snapc, ceph_wbc.truncate_seq,
1015                                                 ceph_wbc.truncate_size, true);
1016                         BUG_ON(IS_ERR(req));
1017                 }
1018                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1019                              thp_size(page) - offset);
1020
1021                 req->r_callback = writepages_finish;
1022                 req->r_inode = inode;
1023
1024                 /* Format the osd request message and submit the write */
1025                 len = 0;
1026                 data_pages = pages;
1027                 op_idx = 0;
1028                 for (i = 0; i < locked_pages; i++) {
1029                         u64 cur_offset = page_offset(pages[i]);
1030                         /*
1031                          * Discontinuity in page range? Ceph can handle that by just passing
1032                          * multiple extents in the write op.
1033                          */
1034                         if (offset + len != cur_offset) {
1035                                 /* If it's full, stop here */
1036                                 if (op_idx + 1 == req->r_num_ops)
1037                                         break;
1038
1039                                 /* Kick off an fscache write with what we have so far. */
1040                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1041
1042                                 /* Start a new extent */
1043                                 osd_req_op_extent_dup_last(req, op_idx,
1044                                                            cur_offset - offset);
1045                                 dout("writepages got pages at %llu~%llu\n",
1046                                      offset, len);
1047                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1048                                                         data_pages, len, 0,
1049                                                         from_pool, false);
1050                                 osd_req_op_extent_update(req, op_idx, len);
1051
1052                                 len = 0;
1053                                 offset = cur_offset;
1054                                 data_pages = pages + i;
1055                                 op_idx++;
1056                         }
1057
1058                         set_page_writeback(pages[i]);
1059                         if (caching)
1060                                 ceph_set_page_fscache(pages[i]);
1061                         len += thp_size(page);
1062                 }
1063                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1064
1065                 if (ceph_wbc.size_stable) {
1066                         len = min(len, ceph_wbc.i_size - offset);
1067                 } else if (i == locked_pages) {
1068                         /* writepages_finish() clears writeback pages
1069                          * according to the data length, so make sure
1070                          * data length covers all locked pages */
1071                         u64 min_len = len + 1 - thp_size(page);
1072                         len = get_writepages_data_length(inode, pages[i - 1],
1073                                                          offset);
1074                         len = max(len, min_len);
1075                 }
1076                 dout("writepages got pages at %llu~%llu\n", offset, len);
1077
1078                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1079                                                  0, from_pool, false);
1080                 osd_req_op_extent_update(req, op_idx, len);
1081
1082                 BUG_ON(op_idx + 1 != req->r_num_ops);
1083
1084                 from_pool = false;
1085                 if (i < locked_pages) {
1086                         BUG_ON(num_ops <= req->r_num_ops);
1087                         num_ops -= req->r_num_ops;
1088                         locked_pages -= i;
1089
1090                         /* allocate new pages array for next request */
1091                         data_pages = pages;
1092                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1093                                               GFP_NOFS);
1094                         if (!pages) {
1095                                 from_pool = true;
1096                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1097                                 BUG_ON(!pages);
1098                         }
1099                         memcpy(pages, data_pages + i,
1100                                locked_pages * sizeof(*pages));
1101                         memset(data_pages + i, 0,
1102                                locked_pages * sizeof(*pages));
1103                 } else {
1104                         BUG_ON(num_ops != req->r_num_ops);
1105                         index = pages[i - 1]->index + 1;
1106                         /* request message now owns the pages array */
1107                         pages = NULL;
1108                 }
1109
1110                 req->r_mtime = inode->i_mtime;
1111                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1112                 BUG_ON(rc);
1113                 req = NULL;
1114
1115                 wbc->nr_to_write -= i;
1116                 if (pages)
1117                         goto new_request;
1118
1119                 /*
1120                  * We stop writing back only if we are not doing
1121                  * integrity sync. In case of integrity sync we have to
1122                  * keep going until we have written all the pages
1123                  * we tagged for writeback prior to entering this loop.
1124                  */
1125                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1126                         done = true;
1127
1128 release_pvec_pages:
1129                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1130                      pvec.nr ? pvec.pages[0] : NULL);
1131                 pagevec_release(&pvec);
1132         }
1133
1134         if (should_loop && !done) {
1135                 /* more to do; loop back to beginning of file */
1136                 dout("writepages looping back to beginning of file\n");
1137                 end = start_index - 1; /* OK even when start_index == 0 */
1138
1139                 /* to write dirty pages associated with next snapc,
1140                  * we need to wait until current writes complete */
1141                 if (wbc->sync_mode != WB_SYNC_NONE &&
1142                     start_index == 0 && /* all dirty pages were checked */
1143                     !ceph_wbc.head_snapc) {
1144                         struct page *page;
1145                         unsigned i, nr;
1146                         index = 0;
1147                         while ((index <= end) &&
1148                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1149                                                 PAGECACHE_TAG_WRITEBACK))) {
1150                                 for (i = 0; i < nr; i++) {
1151                                         page = pvec.pages[i];
1152                                         if (page_snap_context(page) != snapc)
1153                                                 continue;
1154                                         wait_on_page_writeback(page);
1155                                 }
1156                                 pagevec_release(&pvec);
1157                                 cond_resched();
1158                         }
1159                 }
1160
1161                 start_index = 0;
1162                 index = 0;
1163                 goto retry;
1164         }
1165
1166         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1167                 mapping->writeback_index = index;
1168
1169 out:
1170         ceph_osdc_put_request(req);
1171         ceph_put_snap_context(last_snapc);
1172         dout("writepages dend - startone, rc = %d\n", rc);
1173         return rc;
1174 }
1175
1176
1177
1178 /*
1179  * See if a given @snapc is either writeable, or already written.
1180  */
1181 static int context_is_writeable_or_written(struct inode *inode,
1182                                            struct ceph_snap_context *snapc)
1183 {
1184         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1185         int ret = !oldest || snapc->seq <= oldest->seq;
1186
1187         ceph_put_snap_context(oldest);
1188         return ret;
1189 }
1190
1191 /**
1192  * ceph_find_incompatible - find an incompatible context and return it
1193  * @page: page being dirtied
1194  *
1195  * We are only allowed to write into/dirty a page if the page is
1196  * clean, or already dirty within the same snap context. Returns a
1197  * conflicting context if there is one, NULL if there isn't, or a
1198  * negative error code on other errors.
1199  *
1200  * Must be called with page lock held.
1201  */
1202 static struct ceph_snap_context *
1203 ceph_find_incompatible(struct page *page)
1204 {
1205         struct inode *inode = page->mapping->host;
1206         struct ceph_inode_info *ci = ceph_inode(inode);
1207
1208         if (ceph_inode_is_shutdown(inode)) {
1209                 dout(" page %p %llx:%llx is shutdown\n", page,
1210                      ceph_vinop(inode));
1211                 return ERR_PTR(-ESTALE);
1212         }
1213
1214         for (;;) {
1215                 struct ceph_snap_context *snapc, *oldest;
1216
1217                 wait_on_page_writeback(page);
1218
1219                 snapc = page_snap_context(page);
1220                 if (!snapc || snapc == ci->i_head_snapc)
1221                         break;
1222
1223                 /*
1224                  * this page is already dirty in another (older) snap
1225                  * context!  is it writeable now?
1226                  */
1227                 oldest = get_oldest_context(inode, NULL, NULL);
1228                 if (snapc->seq > oldest->seq) {
1229                         /* not writeable -- return it for the caller to deal with */
1230                         ceph_put_snap_context(oldest);
1231                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1232                         return ceph_get_snap_context(snapc);
1233                 }
1234                 ceph_put_snap_context(oldest);
1235
1236                 /* yay, writeable, do it now (without dropping page lock) */
1237                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1238                 if (clear_page_dirty_for_io(page)) {
1239                         int r = writepage_nounlock(page, NULL);
1240                         if (r < 0)
1241                                 return ERR_PTR(r);
1242                 }
1243         }
1244         return NULL;
1245 }
1246
1247 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1248                                         struct folio *folio, void **_fsdata)
1249 {
1250         struct inode *inode = file_inode(file);
1251         struct ceph_inode_info *ci = ceph_inode(inode);
1252         struct ceph_snap_context *snapc;
1253
1254         snapc = ceph_find_incompatible(folio_page(folio, 0));
1255         if (snapc) {
1256                 int r;
1257
1258                 folio_unlock(folio);
1259                 folio_put(folio);
1260                 if (IS_ERR(snapc))
1261                         return PTR_ERR(snapc);
1262
1263                 ceph_queue_writeback(inode);
1264                 r = wait_event_killable(ci->i_cap_wq,
1265                                         context_is_writeable_or_written(inode, snapc));
1266                 ceph_put_snap_context(snapc);
1267                 return r == 0 ? -EAGAIN : r;
1268         }
1269         return 0;
1270 }
1271
1272 /*
1273  * We are only allowed to write into/dirty the page if the page is
1274  * clean, or already dirty within the same snap context.
1275  */
1276 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1277                             loff_t pos, unsigned len, unsigned aop_flags,
1278                             struct page **pagep, void **fsdata)
1279 {
1280         struct inode *inode = file_inode(file);
1281         struct ceph_inode_info *ci = ceph_inode(inode);
1282         struct folio *folio = NULL;
1283         pgoff_t index = pos >> PAGE_SHIFT;
1284         int r;
1285
1286         /*
1287          * Uninlining should have already been done and everything updated, EXCEPT
1288          * for inline_version sent to the MDS.
1289          */
1290         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1291                 unsigned int fgp_flags = FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE;
1292                 if (aop_flags & AOP_FLAG_NOFS)
1293                         fgp_flags |= FGP_NOFS;
1294                 folio = __filemap_get_folio(mapping, index, fgp_flags,
1295                                             mapping_gfp_mask(mapping));
1296                 if (!folio)
1297                         return -ENOMEM;
1298
1299                 /*
1300                  * The inline_version on a new inode is set to 1. If that's the
1301                  * case, then the folio is brand new and isn't yet Uptodate.
1302                  */
1303                 r = 0;
1304                 if (index == 0 && ci->i_inline_version != 1) {
1305                         if (!folio_test_uptodate(folio)) {
1306                                 WARN_ONCE(1, "ceph: write_begin called on still-inlined inode (inline_version %llu)!\n",
1307                                           ci->i_inline_version);
1308                                 r = -EINVAL;
1309                         }
1310                         goto out;
1311                 }
1312                 zero_user_segment(&folio->page, 0, folio_size(folio));
1313                 folio_mark_uptodate(folio);
1314                 goto out;
1315         }
1316
1317         r = netfs_write_begin(file, inode->i_mapping, pos, len, 0, &folio, NULL,
1318                               &ceph_netfs_read_ops, NULL);
1319 out:
1320         if (r == 0)
1321                 folio_wait_fscache(folio);
1322         if (r < 0) {
1323                 if (folio)
1324                         folio_put(folio);
1325         } else {
1326                 WARN_ON_ONCE(!folio_test_locked(folio));
1327                 *pagep = &folio->page;
1328         }
1329         return r;
1330 }
1331
1332 /*
1333  * we don't do anything in here that simple_write_end doesn't do
1334  * except adjust dirty page accounting
1335  */
1336 static int ceph_write_end(struct file *file, struct address_space *mapping,
1337                           loff_t pos, unsigned len, unsigned copied,
1338                           struct page *subpage, void *fsdata)
1339 {
1340         struct folio *folio = page_folio(subpage);
1341         struct inode *inode = file_inode(file);
1342         bool check_cap = false;
1343
1344         dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1345              inode, folio, (int)pos, (int)copied, (int)len);
1346
1347         if (!folio_test_uptodate(folio)) {
1348                 /* just return that nothing was copied on a short copy */
1349                 if (copied < len) {
1350                         copied = 0;
1351                         goto out;
1352                 }
1353                 folio_mark_uptodate(folio);
1354         }
1355
1356         /* did file size increase? */
1357         if (pos+copied > i_size_read(inode))
1358                 check_cap = ceph_inode_set_size(inode, pos+copied);
1359
1360         folio_mark_dirty(folio);
1361
1362 out:
1363         folio_unlock(folio);
1364         folio_put(folio);
1365
1366         if (check_cap)
1367                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1368
1369         return copied;
1370 }
1371
1372 const struct address_space_operations ceph_aops = {
1373         .readpage = ceph_readpage,
1374         .readahead = ceph_readahead,
1375         .writepage = ceph_writepage,
1376         .writepages = ceph_writepages_start,
1377         .write_begin = ceph_write_begin,
1378         .write_end = ceph_write_end,
1379         .set_page_dirty = ceph_set_page_dirty,
1380         .invalidate_folio = ceph_invalidate_folio,
1381         .releasepage = ceph_releasepage,
1382         .direct_IO = noop_direct_IO,
1383 };
1384
1385 static void ceph_block_sigs(sigset_t *oldset)
1386 {
1387         sigset_t mask;
1388         siginitsetinv(&mask, sigmask(SIGKILL));
1389         sigprocmask(SIG_BLOCK, &mask, oldset);
1390 }
1391
1392 static void ceph_restore_sigs(sigset_t *oldset)
1393 {
1394         sigprocmask(SIG_SETMASK, oldset, NULL);
1395 }
1396
1397 /*
1398  * vm ops
1399  */
1400 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1401 {
1402         struct vm_area_struct *vma = vmf->vma;
1403         struct inode *inode = file_inode(vma->vm_file);
1404         struct ceph_inode_info *ci = ceph_inode(inode);
1405         struct ceph_file_info *fi = vma->vm_file->private_data;
1406         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1407         int want, got, err;
1408         sigset_t oldset;
1409         vm_fault_t ret = VM_FAULT_SIGBUS;
1410
1411         if (ceph_inode_is_shutdown(inode))
1412                 return ret;
1413
1414         ceph_block_sigs(&oldset);
1415
1416         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1417              inode, ceph_vinop(inode), off);
1418         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1419                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1420         else
1421                 want = CEPH_CAP_FILE_CACHE;
1422
1423         got = 0;
1424         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1425         if (err < 0)
1426                 goto out_restore;
1427
1428         dout("filemap_fault %p %llu got cap refs on %s\n",
1429              inode, off, ceph_cap_string(got));
1430
1431         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1432             ci->i_inline_version == CEPH_INLINE_NONE) {
1433                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1434                 ceph_add_rw_context(fi, &rw_ctx);
1435                 ret = filemap_fault(vmf);
1436                 ceph_del_rw_context(fi, &rw_ctx);
1437                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1438                      inode, off, ceph_cap_string(got), ret);
1439         } else
1440                 err = -EAGAIN;
1441
1442         ceph_put_cap_refs(ci, got);
1443
1444         if (err != -EAGAIN)
1445                 goto out_restore;
1446
1447         /* read inline data */
1448         if (off >= PAGE_SIZE) {
1449                 /* does not support inline data > PAGE_SIZE */
1450                 ret = VM_FAULT_SIGBUS;
1451         } else {
1452                 struct address_space *mapping = inode->i_mapping;
1453                 struct page *page;
1454
1455                 filemap_invalidate_lock_shared(mapping);
1456                 page = find_or_create_page(mapping, 0,
1457                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1458                 if (!page) {
1459                         ret = VM_FAULT_OOM;
1460                         goto out_inline;
1461                 }
1462                 err = __ceph_do_getattr(inode, page,
1463                                          CEPH_STAT_CAP_INLINE_DATA, true);
1464                 if (err < 0 || off >= i_size_read(inode)) {
1465                         unlock_page(page);
1466                         put_page(page);
1467                         ret = vmf_error(err);
1468                         goto out_inline;
1469                 }
1470                 if (err < PAGE_SIZE)
1471                         zero_user_segment(page, err, PAGE_SIZE);
1472                 else
1473                         flush_dcache_page(page);
1474                 SetPageUptodate(page);
1475                 vmf->page = page;
1476                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1477 out_inline:
1478                 filemap_invalidate_unlock_shared(mapping);
1479                 dout("filemap_fault %p %llu read inline data ret %x\n",
1480                      inode, off, ret);
1481         }
1482 out_restore:
1483         ceph_restore_sigs(&oldset);
1484         if (err < 0)
1485                 ret = vmf_error(err);
1486
1487         return ret;
1488 }
1489
1490 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1491 {
1492         struct vm_area_struct *vma = vmf->vma;
1493         struct inode *inode = file_inode(vma->vm_file);
1494         struct ceph_inode_info *ci = ceph_inode(inode);
1495         struct ceph_file_info *fi = vma->vm_file->private_data;
1496         struct ceph_cap_flush *prealloc_cf;
1497         struct page *page = vmf->page;
1498         loff_t off = page_offset(page);
1499         loff_t size = i_size_read(inode);
1500         size_t len;
1501         int want, got, err;
1502         sigset_t oldset;
1503         vm_fault_t ret = VM_FAULT_SIGBUS;
1504
1505         if (ceph_inode_is_shutdown(inode))
1506                 return ret;
1507
1508         prealloc_cf = ceph_alloc_cap_flush();
1509         if (!prealloc_cf)
1510                 return VM_FAULT_OOM;
1511
1512         sb_start_pagefault(inode->i_sb);
1513         ceph_block_sigs(&oldset);
1514
1515         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1516                 struct page *locked_page = NULL;
1517                 if (off == 0) {
1518                         lock_page(page);
1519                         locked_page = page;
1520                 }
1521                 err = ceph_uninline_data(vma->vm_file, locked_page);
1522                 if (locked_page)
1523                         unlock_page(locked_page);
1524                 if (err < 0)
1525                         goto out_free;
1526         }
1527
1528         if (off + thp_size(page) <= size)
1529                 len = thp_size(page);
1530         else
1531                 len = offset_in_thp(page, size);
1532
1533         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1534              inode, ceph_vinop(inode), off, len, size);
1535         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1536                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1537         else
1538                 want = CEPH_CAP_FILE_BUFFER;
1539
1540         got = 0;
1541         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1542         if (err < 0)
1543                 goto out_free;
1544
1545         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1546              inode, off, len, ceph_cap_string(got));
1547
1548         /* Update time before taking page lock */
1549         file_update_time(vma->vm_file);
1550         inode_inc_iversion_raw(inode);
1551
1552         do {
1553                 struct ceph_snap_context *snapc;
1554
1555                 lock_page(page);
1556
1557                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1558                         unlock_page(page);
1559                         ret = VM_FAULT_NOPAGE;
1560                         break;
1561                 }
1562
1563                 snapc = ceph_find_incompatible(page);
1564                 if (!snapc) {
1565                         /* success.  we'll keep the page locked. */
1566                         set_page_dirty(page);
1567                         ret = VM_FAULT_LOCKED;
1568                         break;
1569                 }
1570
1571                 unlock_page(page);
1572
1573                 if (IS_ERR(snapc)) {
1574                         ret = VM_FAULT_SIGBUS;
1575                         break;
1576                 }
1577
1578                 ceph_queue_writeback(inode);
1579                 err = wait_event_killable(ci->i_cap_wq,
1580                                 context_is_writeable_or_written(inode, snapc));
1581                 ceph_put_snap_context(snapc);
1582         } while (err == 0);
1583
1584         if (ret == VM_FAULT_LOCKED ||
1585             ci->i_inline_version != CEPH_INLINE_NONE) {
1586                 int dirty;
1587                 spin_lock(&ci->i_ceph_lock);
1588                 ci->i_inline_version = CEPH_INLINE_NONE;
1589                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1590                                                &prealloc_cf);
1591                 spin_unlock(&ci->i_ceph_lock);
1592                 if (dirty)
1593                         __mark_inode_dirty(inode, dirty);
1594         }
1595
1596         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1597              inode, off, len, ceph_cap_string(got), ret);
1598         ceph_put_cap_refs_async(ci, got);
1599 out_free:
1600         ceph_restore_sigs(&oldset);
1601         sb_end_pagefault(inode->i_sb);
1602         ceph_free_cap_flush(prealloc_cf);
1603         if (err < 0)
1604                 ret = vmf_error(err);
1605         return ret;
1606 }
1607
1608 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1609                            char *data, size_t len)
1610 {
1611         struct address_space *mapping = inode->i_mapping;
1612         struct page *page;
1613
1614         if (locked_page) {
1615                 page = locked_page;
1616         } else {
1617                 if (i_size_read(inode) == 0)
1618                         return;
1619                 page = find_or_create_page(mapping, 0,
1620                                            mapping_gfp_constraint(mapping,
1621                                            ~__GFP_FS));
1622                 if (!page)
1623                         return;
1624                 if (PageUptodate(page)) {
1625                         unlock_page(page);
1626                         put_page(page);
1627                         return;
1628                 }
1629         }
1630
1631         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1632              inode, ceph_vinop(inode), len, locked_page);
1633
1634         if (len > 0) {
1635                 void *kaddr = kmap_atomic(page);
1636                 memcpy(kaddr, data, len);
1637                 kunmap_atomic(kaddr);
1638         }
1639
1640         if (page != locked_page) {
1641                 if (len < PAGE_SIZE)
1642                         zero_user_segment(page, len, PAGE_SIZE);
1643                 else
1644                         flush_dcache_page(page);
1645
1646                 SetPageUptodate(page);
1647                 unlock_page(page);
1648                 put_page(page);
1649         }
1650 }
1651
1652 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1653 {
1654         struct inode *inode = file_inode(filp);
1655         struct ceph_inode_info *ci = ceph_inode(inode);
1656         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1657         struct ceph_osd_request *req;
1658         struct page *page = NULL;
1659         u64 len, inline_version;
1660         int err = 0;
1661         bool from_pagecache = false;
1662
1663         spin_lock(&ci->i_ceph_lock);
1664         inline_version = ci->i_inline_version;
1665         spin_unlock(&ci->i_ceph_lock);
1666
1667         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1668              inode, ceph_vinop(inode), inline_version);
1669
1670         if (inline_version == 1 || /* initial version, no data */
1671             inline_version == CEPH_INLINE_NONE)
1672                 goto out;
1673
1674         if (locked_page) {
1675                 page = locked_page;
1676                 WARN_ON(!PageUptodate(page));
1677         } else if (ceph_caps_issued(ci) &
1678                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1679                 page = find_get_page(inode->i_mapping, 0);
1680                 if (page) {
1681                         if (PageUptodate(page)) {
1682                                 from_pagecache = true;
1683                                 lock_page(page);
1684                         } else {
1685                                 put_page(page);
1686                                 page = NULL;
1687                         }
1688                 }
1689         }
1690
1691         if (page) {
1692                 len = i_size_read(inode);
1693                 if (len > PAGE_SIZE)
1694                         len = PAGE_SIZE;
1695         } else {
1696                 page = __page_cache_alloc(GFP_NOFS);
1697                 if (!page) {
1698                         err = -ENOMEM;
1699                         goto out;
1700                 }
1701                 err = __ceph_do_getattr(inode, page,
1702                                         CEPH_STAT_CAP_INLINE_DATA, true);
1703                 if (err < 0) {
1704                         /* no inline data */
1705                         if (err == -ENODATA)
1706                                 err = 0;
1707                         goto out;
1708                 }
1709                 len = err;
1710         }
1711
1712         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1713                                     ceph_vino(inode), 0, &len, 0, 1,
1714                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1715                                     NULL, 0, 0, false);
1716         if (IS_ERR(req)) {
1717                 err = PTR_ERR(req);
1718                 goto out;
1719         }
1720
1721         req->r_mtime = inode->i_mtime;
1722         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1723         if (!err)
1724                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1725         ceph_osdc_put_request(req);
1726         if (err < 0)
1727                 goto out;
1728
1729         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1730                                     ceph_vino(inode), 0, &len, 1, 3,
1731                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1732                                     NULL, ci->i_truncate_seq,
1733                                     ci->i_truncate_size, false);
1734         if (IS_ERR(req)) {
1735                 err = PTR_ERR(req);
1736                 goto out;
1737         }
1738
1739         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1740
1741         {
1742                 __le64 xattr_buf = cpu_to_le64(inline_version);
1743                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1744                                             "inline_version", &xattr_buf,
1745                                             sizeof(xattr_buf),
1746                                             CEPH_OSD_CMPXATTR_OP_GT,
1747                                             CEPH_OSD_CMPXATTR_MODE_U64);
1748                 if (err)
1749                         goto out_put;
1750         }
1751
1752         {
1753                 char xattr_buf[32];
1754                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1755                                          "%llu", inline_version);
1756                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1757                                             "inline_version",
1758                                             xattr_buf, xattr_len, 0, 0);
1759                 if (err)
1760                         goto out_put;
1761         }
1762
1763         req->r_mtime = inode->i_mtime;
1764         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1765         if (!err)
1766                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1767
1768         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1769                                   req->r_end_latency, len, err);
1770
1771 out_put:
1772         ceph_osdc_put_request(req);
1773         if (err == -ECANCELED)
1774                 err = 0;
1775 out:
1776         if (page && page != locked_page) {
1777                 if (from_pagecache) {
1778                         unlock_page(page);
1779                         put_page(page);
1780                 } else
1781                         __free_pages(page, 0);
1782         }
1783
1784         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1785              inode, ceph_vinop(inode), inline_version, err);
1786         return err;
1787 }
1788
1789 static const struct vm_operations_struct ceph_vmops = {
1790         .fault          = ceph_filemap_fault,
1791         .page_mkwrite   = ceph_page_mkwrite,
1792 };
1793
1794 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1795 {
1796         struct address_space *mapping = file->f_mapping;
1797
1798         if (!mapping->a_ops->readpage)
1799                 return -ENOEXEC;
1800         file_accessed(file);
1801         vma->vm_ops = &ceph_vmops;
1802         return 0;
1803 }
1804
1805 enum {
1806         POOL_READ       = 1,
1807         POOL_WRITE      = 2,
1808 };
1809
1810 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1811                                 s64 pool, struct ceph_string *pool_ns)
1812 {
1813         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1814         struct ceph_mds_client *mdsc = fsc->mdsc;
1815         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1816         struct rb_node **p, *parent;
1817         struct ceph_pool_perm *perm;
1818         struct page **pages;
1819         size_t pool_ns_len;
1820         int err = 0, err2 = 0, have = 0;
1821
1822         down_read(&mdsc->pool_perm_rwsem);
1823         p = &mdsc->pool_perm_tree.rb_node;
1824         while (*p) {
1825                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1826                 if (pool < perm->pool)
1827                         p = &(*p)->rb_left;
1828                 else if (pool > perm->pool)
1829                         p = &(*p)->rb_right;
1830                 else {
1831                         int ret = ceph_compare_string(pool_ns,
1832                                                 perm->pool_ns,
1833                                                 perm->pool_ns_len);
1834                         if (ret < 0)
1835                                 p = &(*p)->rb_left;
1836                         else if (ret > 0)
1837                                 p = &(*p)->rb_right;
1838                         else {
1839                                 have = perm->perm;
1840                                 break;
1841                         }
1842                 }
1843         }
1844         up_read(&mdsc->pool_perm_rwsem);
1845         if (*p)
1846                 goto out;
1847
1848         if (pool_ns)
1849                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1850                      pool, (int)pool_ns->len, pool_ns->str);
1851         else
1852                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1853
1854         down_write(&mdsc->pool_perm_rwsem);
1855         p = &mdsc->pool_perm_tree.rb_node;
1856         parent = NULL;
1857         while (*p) {
1858                 parent = *p;
1859                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1860                 if (pool < perm->pool)
1861                         p = &(*p)->rb_left;
1862                 else if (pool > perm->pool)
1863                         p = &(*p)->rb_right;
1864                 else {
1865                         int ret = ceph_compare_string(pool_ns,
1866                                                 perm->pool_ns,
1867                                                 perm->pool_ns_len);
1868                         if (ret < 0)
1869                                 p = &(*p)->rb_left;
1870                         else if (ret > 0)
1871                                 p = &(*p)->rb_right;
1872                         else {
1873                                 have = perm->perm;
1874                                 break;
1875                         }
1876                 }
1877         }
1878         if (*p) {
1879                 up_write(&mdsc->pool_perm_rwsem);
1880                 goto out;
1881         }
1882
1883         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1884                                          1, false, GFP_NOFS);
1885         if (!rd_req) {
1886                 err = -ENOMEM;
1887                 goto out_unlock;
1888         }
1889
1890         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1891         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1892         rd_req->r_base_oloc.pool = pool;
1893         if (pool_ns)
1894                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1895         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1896
1897         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1898         if (err)
1899                 goto out_unlock;
1900
1901         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1902                                          1, false, GFP_NOFS);
1903         if (!wr_req) {
1904                 err = -ENOMEM;
1905                 goto out_unlock;
1906         }
1907
1908         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1909         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1910         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1911         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1912
1913         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1914         if (err)
1915                 goto out_unlock;
1916
1917         /* one page should be large enough for STAT data */
1918         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1919         if (IS_ERR(pages)) {
1920                 err = PTR_ERR(pages);
1921                 goto out_unlock;
1922         }
1923
1924         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1925                                      0, false, true);
1926         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1927
1928         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1929         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1930
1931         if (!err)
1932                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1933         if (!err2)
1934                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1935
1936         if (err >= 0 || err == -ENOENT)
1937                 have |= POOL_READ;
1938         else if (err != -EPERM) {
1939                 if (err == -EBLOCKLISTED)
1940                         fsc->blocklisted = true;
1941                 goto out_unlock;
1942         }
1943
1944         if (err2 == 0 || err2 == -EEXIST)
1945                 have |= POOL_WRITE;
1946         else if (err2 != -EPERM) {
1947                 if (err2 == -EBLOCKLISTED)
1948                         fsc->blocklisted = true;
1949                 err = err2;
1950                 goto out_unlock;
1951         }
1952
1953         pool_ns_len = pool_ns ? pool_ns->len : 0;
1954         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1955         if (!perm) {
1956                 err = -ENOMEM;
1957                 goto out_unlock;
1958         }
1959
1960         perm->pool = pool;
1961         perm->perm = have;
1962         perm->pool_ns_len = pool_ns_len;
1963         if (pool_ns_len > 0)
1964                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1965         perm->pool_ns[pool_ns_len] = 0;
1966
1967         rb_link_node(&perm->node, parent, p);
1968         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1969         err = 0;
1970 out_unlock:
1971         up_write(&mdsc->pool_perm_rwsem);
1972
1973         ceph_osdc_put_request(rd_req);
1974         ceph_osdc_put_request(wr_req);
1975 out:
1976         if (!err)
1977                 err = have;
1978         if (pool_ns)
1979                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1980                      pool, (int)pool_ns->len, pool_ns->str, err);
1981         else
1982                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1983         return err;
1984 }
1985
1986 int ceph_pool_perm_check(struct inode *inode, int need)
1987 {
1988         struct ceph_inode_info *ci = ceph_inode(inode);
1989         struct ceph_string *pool_ns;
1990         s64 pool;
1991         int ret, flags;
1992
1993         /* Only need to do this for regular files */
1994         if (!S_ISREG(inode->i_mode))
1995                 return 0;
1996
1997         if (ci->i_vino.snap != CEPH_NOSNAP) {
1998                 /*
1999                  * Pool permission check needs to write to the first object.
2000                  * But for snapshot, head of the first object may have alread
2001                  * been deleted. Skip check to avoid creating orphan object.
2002                  */
2003                 return 0;
2004         }
2005
2006         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2007                                 NOPOOLPERM))
2008                 return 0;
2009
2010         spin_lock(&ci->i_ceph_lock);
2011         flags = ci->i_ceph_flags;
2012         pool = ci->i_layout.pool_id;
2013         spin_unlock(&ci->i_ceph_lock);
2014 check:
2015         if (flags & CEPH_I_POOL_PERM) {
2016                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2017                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2018                              pool);
2019                         return -EPERM;
2020                 }
2021                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2022                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2023                              pool);
2024                         return -EPERM;
2025                 }
2026                 return 0;
2027         }
2028
2029         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2030         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2031         ceph_put_string(pool_ns);
2032         if (ret < 0)
2033                 return ret;
2034
2035         flags = CEPH_I_POOL_PERM;
2036         if (ret & POOL_READ)
2037                 flags |= CEPH_I_POOL_RD;
2038         if (ret & POOL_WRITE)
2039                 flags |= CEPH_I_POOL_WR;
2040
2041         spin_lock(&ci->i_ceph_lock);
2042         if (pool == ci->i_layout.pool_id &&
2043             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2044                 ci->i_ceph_flags |= flags;
2045         } else {
2046                 pool = ci->i_layout.pool_id;
2047                 flags = ci->i_ceph_flags;
2048         }
2049         spin_unlock(&ci->i_ceph_lock);
2050         goto check;
2051 }
2052
2053 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2054 {
2055         struct ceph_pool_perm *perm;
2056         struct rb_node *n;
2057
2058         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2059                 n = rb_first(&mdsc->pool_perm_tree);
2060                 perm = rb_entry(n, struct ceph_pool_perm, node);
2061                 rb_erase(n, &mdsc->pool_perm_tree);
2062                 kfree(perm);
2063         }
2064 }