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 bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
  80 {
  81         struct inode *inode;
  82         struct ceph_inode_info *ci;
  83         struct ceph_snap_context *snapc;
  84
  85         if (folio_test_dirty(folio)) {
  86                 dout("%p dirty_folio %p idx %lu -- already dirty\n",
  87                      mapping->host, folio, folio->index);
  88                 BUG_ON(!folio_get_private(folio));
  89                 return false;
  90         }
  91
  92         inode = mapping->host;
  93         ci = ceph_inode(inode);
  94
  95         /* dirty the head */
  96         spin_lock(&ci->i_ceph_lock);
  97         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  98         if (__ceph_have_pending_cap_snap(ci)) {
  99                 struct ceph_cap_snap *capsnap =
 100                                 list_last_entry(&ci->i_cap_snaps,
 101                                                 struct ceph_cap_snap,
 102                                                 ci_item);
 103                 snapc = ceph_get_snap_context(capsnap->context);
 104                 capsnap->dirty_pages++;
 105         } else {
 106                 BUG_ON(!ci->i_head_snapc);
 107                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 108                 ++ci->i_wrbuffer_ref_head;
 109         }
 110         if (ci->i_wrbuffer_ref == 0)
 111                 ihold(inode);
 112         ++ci->i_wrbuffer_ref;
 113         dout("%p dirty_folio %p idx %lu head %d/%d -> %d/%d "
 114              "snapc %p seq %lld (%d snaps)\n",
 115              mapping->host, folio, folio->index,
 116              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 117              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 118              snapc, snapc->seq, snapc->num_snaps);
 119         spin_unlock(&ci->i_ceph_lock);
 120
 121         /*
 122          * Reference snap context in folio->private.  Also set
 123          * PagePrivate so that we get invalidate_folio callback.
 124          */
 125         BUG_ON(folio_get_private(folio));
 126         folio_attach_private(folio, snapc);
 127
 128         return ceph_fscache_dirty_folio(mapping, folio);
 129 }
 130
 131 /*
 132  * If we are truncating the full folio (i.e. offset == 0), adjust the
 133  * dirty folio counters appropriately.  Only called if there is private
 134  * data on the folio.
 135  */
 136 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 137                                 size_t length)
 138 {
 139         struct inode *inode;
 140         struct ceph_inode_info *ci;
 141         struct ceph_snap_context *snapc;
 142
 143         inode = folio->mapping->host;
 144         ci = ceph_inode(inode);
 145
 146         if (offset != 0 || length != folio_size(folio)) {
 147                 dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
 148                      inode, folio->index, offset, length);
 149                 return;
 150         }
 151
 152         WARN_ON(!folio_test_locked(folio));
 153         if (folio_get_private(folio)) {
 154                 dout("%p invalidate_folio idx %lu full dirty page\n",
 155                      inode, folio->index);
 156
 157                 snapc = folio_detach_private(folio);
 158                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159                 ceph_put_snap_context(snapc);
 160         }
 161
 162         folio_wait_fscache(folio);
 163 }
 164
 165 static int ceph_releasepage(struct page *page, gfp_t gfp)
 166 {
 167         struct inode *inode = page->mapping->host;
 168
 169         dout("%llx:%llx releasepage %p idx %lu (%sdirty)\n",
 170              ceph_vinop(inode), page,
 171              page->index, PageDirty(page) ? "" : "not ");
 172
 173         if (PagePrivate(page))
 174                 return 0;
 175
 176         if (PageFsCache(page)) {
 177                 if (current_is_kswapd() || !(gfp & __GFP_FS))
 178                         return 0;
 179                 wait_on_page_fscache(page);
 180         }
 181         ceph_fscache_note_page_release(inode);
 182         return 1;
 183 }
 184
 185 static void ceph_netfs_expand_readahead(struct netfs_read_request *rreq)
 186 {
 187         struct inode *inode = rreq->mapping->host;
 188         struct ceph_inode_info *ci = ceph_inode(inode);
 189         struct ceph_file_layout *lo = &ci->i_layout;
 190         u32 blockoff;
 191         u64 blockno;
 192
 193         /* Expand the start downward */
 194         blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 195         rreq->start = blockno * lo->stripe_unit;
 196         rreq->len += blockoff;
 197
 198         /* Now, round up the length to the next block */
 199         rreq->len = roundup(rreq->len, lo->stripe_unit);
 200 }
 201
 202 static bool ceph_netfs_clamp_length(struct netfs_read_subrequest *subreq)
 203 {
 204         struct inode *inode = subreq->rreq->mapping->host;
 205         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 206         struct ceph_inode_info *ci = ceph_inode(inode);
 207         u64 objno, objoff;
 208         u32 xlen;
 209
 210         /* Truncate the extent at the end of the current block */
 211         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 212                                       &objno, &objoff, &xlen);
 213         subreq->len = min(xlen, fsc->mount_options->rsize);
 214         return true;
 215 }
 216
 217 static void finish_netfs_read(struct ceph_osd_request *req)
 218 {
 219         struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 220         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 221         struct netfs_read_subrequest *subreq = req->r_priv;
 222         int num_pages;
 223         int err = req->r_result;
 224
 225         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 226                                  req->r_end_latency, osd_data->length, err);
 227
 228         dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 229              subreq->len, i_size_read(req->r_inode));
 230
 231         /* no object means success but no data */
 232         if (err == -ENOENT)
 233                 err = 0;
 234         else if (err == -EBLOCKLISTED)
 235                 fsc->blocklisted = true;
 236
 237         if (err >= 0 && err < subreq->len)
 238                 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 239
 240         netfs_subreq_terminated(subreq, err, true);
 241
 242         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 243         ceph_put_page_vector(osd_data->pages, num_pages, false);
 244         iput(req->r_inode);
 245 }
 246
 247 static void ceph_netfs_issue_op(struct netfs_read_subrequest *subreq)
 248 {
 249         struct netfs_read_request *rreq = subreq->rreq;
 250         struct inode *inode = rreq->mapping->host;
 251         struct ceph_inode_info *ci = ceph_inode(inode);
 252         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 253         struct ceph_osd_request *req;
 254         struct ceph_vino vino = ceph_vino(inode);
 255         struct iov_iter iter;
 256         struct page **pages;
 257         size_t page_off;
 258         int err = 0;
 259         u64 len = subreq->len;
 260
 261         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 262                         0, 1, CEPH_OSD_OP_READ,
 263                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 264                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 265         if (IS_ERR(req)) {
 266                 err = PTR_ERR(req);
 267                 req = NULL;
 268                 goto out;
 269         }
 270
 271         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 272         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 273         err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 274         if (err < 0) {
 275                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 276                 goto out;
 277         }
 278
 279         /* should always give us a page-aligned read */
 280         WARN_ON_ONCE(page_off);
 281         len = err;
 282
 283         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 284         req->r_callback = finish_netfs_read;
 285         req->r_priv = subreq;
 286         req->r_inode = inode;
 287         ihold(inode);
 288
 289         err = ceph_osdc_start_request(req->r_osdc, req, false);
 290         if (err)
 291                 iput(inode);
 292 out:
 293         ceph_osdc_put_request(req);
 294         if (err)
 295                 netfs_subreq_terminated(subreq, err, false);
 296         dout("%s: result %d\n", __func__, err);
 297 }
 298
 299 static void ceph_readahead_cleanup(struct address_space *mapping, void *priv)
 300 {
 301         struct inode *inode = mapping->host;
 302         struct ceph_inode_info *ci = ceph_inode(inode);
 303         int got = (uintptr_t)priv;
 304
 305         if (got)
 306                 ceph_put_cap_refs(ci, got);
 307 }
 308
 309 static const struct netfs_read_request_ops ceph_netfs_read_ops = {
 310         .is_cache_enabled       = ceph_is_cache_enabled,
 311         .begin_cache_operation  = ceph_begin_cache_operation,
 312         .issue_op               = ceph_netfs_issue_op,
 313         .expand_readahead       = ceph_netfs_expand_readahead,
 314         .clamp_length           = ceph_netfs_clamp_length,
 315         .check_write_begin      = ceph_netfs_check_write_begin,
 316         .cleanup                = ceph_readahead_cleanup,
 317 };
 318
 319 /* read a single page, without unlocking it. */
 320 static int ceph_readpage(struct file *file, struct page *subpage)
 321 {
 322         struct folio *folio = page_folio(subpage);
 323         struct inode *inode = file_inode(file);
 324         struct ceph_inode_info *ci = ceph_inode(inode);
 325         struct ceph_vino vino = ceph_vino(inode);
 326         size_t len = folio_size(folio);
 327         u64 off = folio_file_pos(folio);
 328
 329         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 330                 /*
 331                  * Uptodate inline data should have been added
 332                  * into page cache while getting Fcr caps.
 333                  */
 334                 if (off == 0) {
 335                         folio_unlock(folio);
 336                         return -EINVAL;
 337                 }
 338                 zero_user_segment(&folio->page, 0, folio_size(folio));
 339                 folio_mark_uptodate(folio);
 340                 folio_unlock(folio);
 341                 return 0;
 342         }
 343
 344         dout("readpage ino %llx.%llx file %p off %llu len %zu folio %p index %lu\n",
 345              vino.ino, vino.snap, file, off, len, folio, folio_index(folio));
 346
 347         return netfs_readpage(file, folio, &ceph_netfs_read_ops, NULL);
 348 }
 349
 350 static void ceph_readahead(struct readahead_control *ractl)
 351 {
 352         struct inode *inode = file_inode(ractl->file);
 353         struct ceph_file_info *fi = ractl->file->private_data;
 354         struct ceph_rw_context *rw_ctx;
 355         int got = 0;
 356         int ret = 0;
 357
 358         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 359                 return;
 360
 361         rw_ctx = ceph_find_rw_context(fi);
 362         if (!rw_ctx) {
 363                 /*
 364                  * readahead callers do not necessarily hold Fcb caps
 365                  * (e.g. fadvise, madvise).
 366                  */
 367                 int want = CEPH_CAP_FILE_CACHE;
 368
 369                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 370                 if (ret < 0)
 371                         dout("start_read %p, error getting cap\n", inode);
 372                 else if (!(got & want))
 373                         dout("start_read %p, no cache cap\n", inode);
 374
 375                 if (ret <= 0)
 376                         return;
 377         }
 378         netfs_readahead(ractl, &ceph_netfs_read_ops, (void *)(uintptr_t)got);
 379 }
 380
 381 #ifdef CONFIG_CEPH_FSCACHE
 382 static void ceph_set_page_fscache(struct page *page)
 383 {
 384         set_page_fscache(page);
 385 }
 386
 387 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 388 {
 389         struct inode *inode = priv;
 390
 391         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 392                 ceph_fscache_invalidate(inode, false);
 393 }
 394
 395 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 396 {
 397         struct ceph_inode_info *ci = ceph_inode(inode);
 398         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 399
 400         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 401                                ceph_fscache_write_terminated, inode, caching);
 402 }
 403 #else
 404 static inline void ceph_set_page_fscache(struct page *page)
 405 {
 406 }
 407
 408 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 409 {
 410 }
 411 #endif /* CONFIG_CEPH_FSCACHE */
 412
 413 struct ceph_writeback_ctl
 414 {
 415         loff_t i_size;
 416         u64 truncate_size;
 417         u32 truncate_seq;
 418         bool size_stable;
 419         bool head_snapc;
 420 };
 421
 422 /*
 423  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 424  * only snap context we are allowed to write back.
 425  */
 426 static struct ceph_snap_context *
 427 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 428                    struct ceph_snap_context *page_snapc)
 429 {
 430         struct ceph_inode_info *ci = ceph_inode(inode);
 431         struct ceph_snap_context *snapc = NULL;
 432         struct ceph_cap_snap *capsnap = NULL;
 433
 434         spin_lock(&ci->i_ceph_lock);
 435         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 436                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 437                      capsnap->context, capsnap->dirty_pages);
 438                 if (!capsnap->dirty_pages)
 439                         continue;
 440
 441                 /* get i_size, truncate_{seq,size} for page_snapc? */
 442                 if (snapc && capsnap->context != page_snapc)
 443                         continue;
 444
 445                 if (ctl) {
 446                         if (capsnap->writing) {
 447                                 ctl->i_size = i_size_read(inode);
 448                                 ctl->size_stable = false;
 449                         } else {
 450                                 ctl->i_size = capsnap->size;
 451                                 ctl->size_stable = true;
 452                         }
 453                         ctl->truncate_size = capsnap->truncate_size;
 454                         ctl->truncate_seq = capsnap->truncate_seq;
 455                         ctl->head_snapc = false;
 456                 }
 457
 458                 if (snapc)
 459                         break;
 460
 461                 snapc = ceph_get_snap_context(capsnap->context);
 462                 if (!page_snapc ||
 463                     page_snapc == snapc ||
 464                     page_snapc->seq > snapc->seq)
 465                         break;
 466         }
 467         if (!snapc && ci->i_wrbuffer_ref_head) {
 468                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 469                 dout(" head snapc %p has %d dirty pages\n",
 470                      snapc, ci->i_wrbuffer_ref_head);
 471                 if (ctl) {
 472                         ctl->i_size = i_size_read(inode);
 473                         ctl->truncate_size = ci->i_truncate_size;
 474                         ctl->truncate_seq = ci->i_truncate_seq;
 475                         ctl->size_stable = false;
 476                         ctl->head_snapc = true;
 477                 }
 478         }
 479         spin_unlock(&ci->i_ceph_lock);
 480         return snapc;
 481 }
 482
 483 static u64 get_writepages_data_length(struct inode *inode,
 484                                       struct page *page, u64 start)
 485 {
 486         struct ceph_inode_info *ci = ceph_inode(inode);
 487         struct ceph_snap_context *snapc = page_snap_context(page);
 488         struct ceph_cap_snap *capsnap = NULL;
 489         u64 end = i_size_read(inode);
 490
 491         if (snapc != ci->i_head_snapc) {
 492                 bool found = false;
 493                 spin_lock(&ci->i_ceph_lock);
 494                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 495                         if (capsnap->context == snapc) {
 496                                 if (!capsnap->writing)
 497                                         end = capsnap->size;
 498                                 found = true;
 499                                 break;
 500                         }
 501                 }
 502                 spin_unlock(&ci->i_ceph_lock);
 503                 WARN_ON(!found);
 504         }
 505         if (end > page_offset(page) + thp_size(page))
 506                 end = page_offset(page) + thp_size(page);
 507         return end > start ? end - start : 0;
 508 }
 509
 510 /*
 511  * Write a single page, but leave the page locked.
 512  *
 513  * If we get a write error, mark the mapping for error, but still adjust the
 514  * dirty page accounting (i.e., page is no longer dirty).
 515  */
 516 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 517 {
 518         struct folio *folio = page_folio(page);
 519         struct inode *inode = page->mapping->host;
 520         struct ceph_inode_info *ci = ceph_inode(inode);
 521         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 522         struct ceph_snap_context *snapc, *oldest;
 523         loff_t page_off = page_offset(page);
 524         int err;
 525         loff_t len = thp_size(page);
 526         struct ceph_writeback_ctl ceph_wbc;
 527         struct ceph_osd_client *osdc = &fsc->client->osdc;
 528         struct ceph_osd_request *req;
 529         bool caching = ceph_is_cache_enabled(inode);
 530
 531         dout("writepage %p idx %lu\n", page, page->index);
 532
 533         /* verify this is a writeable snap context */
 534         snapc = page_snap_context(page);
 535         if (!snapc) {
 536                 dout("writepage %p page %p not dirty?\n", inode, page);
 537                 return 0;
 538         }
 539         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 540         if (snapc->seq > oldest->seq) {
 541                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 542                      inode, page, snapc);
 543                 /* we should only noop if called by kswapd */
 544                 WARN_ON(!(current->flags & PF_MEMALLOC));
 545                 ceph_put_snap_context(oldest);
 546                 redirty_page_for_writepage(wbc, page);
 547                 return 0;
 548         }
 549         ceph_put_snap_context(oldest);
 550
 551         /* is this a partial page at end of file? */
 552         if (page_off >= ceph_wbc.i_size) {
 553                 dout("folio at %lu beyond eof %llu\n", folio->index,
 554                                 ceph_wbc.i_size);
 555                 folio_invalidate(folio, 0, folio_size(folio));
 556                 return 0;
 557         }
 558
 559         if (ceph_wbc.i_size < page_off + len)
 560                 len = ceph_wbc.i_size - page_off;
 561
 562         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 563              inode, page, page->index, page_off, len, snapc, snapc->seq);
 564
 565         if (atomic_long_inc_return(&fsc->writeback_count) >
 566             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 567                 fsc->write_congested = true;
 568
 569         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 570                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 571                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 572                                     true);
 573         if (IS_ERR(req))
 574                 return PTR_ERR(req);
 575
 576         set_page_writeback(page);
 577         if (caching)
 578                 ceph_set_page_fscache(page);
 579         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 580
 581         /* it may be a short write due to an object boundary */
 582         WARN_ON_ONCE(len > thp_size(page));
 583         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 584         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 585
 586         req->r_mtime = inode->i_mtime;
 587         err = ceph_osdc_start_request(osdc, req, true);
 588         if (!err)
 589                 err = ceph_osdc_wait_request(osdc, req);
 590
 591         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 592                                   req->r_end_latency, len, err);
 593
 594         ceph_osdc_put_request(req);
 595         if (err == 0)
 596                 err = len;
 597
 598         if (err < 0) {
 599                 struct writeback_control tmp_wbc;
 600                 if (!wbc)
 601                         wbc = &tmp_wbc;
 602                 if (err == -ERESTARTSYS) {
 603                         /* killed by SIGKILL */
 604                         dout("writepage interrupted page %p\n", page);
 605                         redirty_page_for_writepage(wbc, page);
 606                         end_page_writeback(page);
 607                         return err;
 608                 }
 609                 if (err == -EBLOCKLISTED)
 610                         fsc->blocklisted = true;
 611                 dout("writepage setting page/mapping error %d %p\n",
 612                      err, page);
 613                 mapping_set_error(&inode->i_data, err);
 614                 wbc->pages_skipped++;
 615         } else {
 616                 dout("writepage cleaned page %p\n", page);
 617                 err = 0;  /* vfs expects us to return 0 */
 618         }
 619         oldest = detach_page_private(page);
 620         WARN_ON_ONCE(oldest != snapc);
 621         end_page_writeback(page);
 622         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 623         ceph_put_snap_context(snapc);  /* page's reference */
 624
 625         if (atomic_long_dec_return(&fsc->writeback_count) <
 626             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 627                 fsc->write_congested = false;
 628
 629         return err;
 630 }
 631
 632 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 633 {
 634         int err;
 635         struct inode *inode = page->mapping->host;
 636         BUG_ON(!inode);
 637         ihold(inode);
 638
 639         if (wbc->sync_mode == WB_SYNC_NONE &&
 640             ceph_inode_to_client(inode)->write_congested)
 641                 return AOP_WRITEPAGE_ACTIVATE;
 642
 643         wait_on_page_fscache(page);
 644
 645         err = writepage_nounlock(page, wbc);
 646         if (err == -ERESTARTSYS) {
 647                 /* direct memory reclaimer was killed by SIGKILL. return 0
 648                  * to prevent caller from setting mapping/page error */
 649                 err = 0;
 650         }
 651         unlock_page(page);
 652         iput(inode);
 653         return err;
 654 }
 655
 656 /*
 657  * async writeback completion handler.
 658  *
 659  * If we get an error, set the mapping error bit, but not the individual
 660  * page error bits.
 661  */
 662 static void writepages_finish(struct ceph_osd_request *req)
 663 {
 664         struct inode *inode = req->r_inode;
 665         struct ceph_inode_info *ci = ceph_inode(inode);
 666         struct ceph_osd_data *osd_data;
 667         struct page *page;
 668         int num_pages, total_pages = 0;
 669         int i, j;
 670         int rc = req->r_result;
 671         struct ceph_snap_context *snapc = req->r_snapc;
 672         struct address_space *mapping = inode->i_mapping;
 673         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 674         unsigned int len = 0;
 675         bool remove_page;
 676
 677         dout("writepages_finish %p rc %d\n", inode, rc);
 678         if (rc < 0) {
 679                 mapping_set_error(mapping, rc);
 680                 ceph_set_error_write(ci);
 681                 if (rc == -EBLOCKLISTED)
 682                         fsc->blocklisted = true;
 683         } else {
 684                 ceph_clear_error_write(ci);
 685         }
 686
 687         /*
 688          * We lost the cache cap, need to truncate the page before
 689          * it is unlocked, otherwise we'd truncate it later in the
 690          * page truncation thread, possibly losing some data that
 691          * raced its way in
 692          */
 693         remove_page = !(ceph_caps_issued(ci) &
 694                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 695
 696         /* clean all pages */
 697         for (i = 0; i < req->r_num_ops; i++) {
 698                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 699                         break;
 700
 701                 osd_data = osd_req_op_extent_osd_data(req, i);
 702                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 703                 len += osd_data->length;
 704                 num_pages = calc_pages_for((u64)osd_data->alignment,
 705                                            (u64)osd_data->length);
 706                 total_pages += num_pages;
 707                 for (j = 0; j < num_pages; j++) {
 708                         page = osd_data->pages[j];
 709                         BUG_ON(!page);
 710                         WARN_ON(!PageUptodate(page));
 711
 712                         if (atomic_long_dec_return(&fsc->writeback_count) <
 713                              CONGESTION_OFF_THRESH(
 714                                         fsc->mount_options->congestion_kb))
 715                                 fsc->write_congested = false;
 716
 717                         ceph_put_snap_context(detach_page_private(page));
 718                         end_page_writeback(page);
 719                         dout("unlocking %p\n", page);
 720
 721                         if (remove_page)
 722                                 generic_error_remove_page(inode->i_mapping,
 723                                                           page);
 724
 725                         unlock_page(page);
 726                 }
 727                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 728                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 729
 730                 release_pages(osd_data->pages, num_pages);
 731         }
 732
 733         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 734                                   req->r_end_latency, len, rc);
 735
 736         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 737
 738         osd_data = osd_req_op_extent_osd_data(req, 0);
 739         if (osd_data->pages_from_pool)
 740                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 741         else
 742                 kfree(osd_data->pages);
 743         ceph_osdc_put_request(req);
 744 }
 745
 746 /*
 747  * initiate async writeback
 748  */
 749 static int ceph_writepages_start(struct address_space *mapping,
 750                                  struct writeback_control *wbc)
 751 {
 752         struct inode *inode = mapping->host;
 753         struct ceph_inode_info *ci = ceph_inode(inode);
 754         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 755         struct ceph_vino vino = ceph_vino(inode);
 756         pgoff_t index, start_index, end = -1;
 757         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 758         struct pagevec pvec;
 759         int rc = 0;
 760         unsigned int wsize = i_blocksize(inode);
 761         struct ceph_osd_request *req = NULL;
 762         struct ceph_writeback_ctl ceph_wbc;
 763         bool should_loop, range_whole = false;
 764         bool done = false;
 765         bool caching = ceph_is_cache_enabled(inode);
 766
 767         if (wbc->sync_mode == WB_SYNC_NONE &&
 768             fsc->write_congested)
 769                 return 0;
 770
 771         dout("writepages_start %p (mode=%s)\n", inode,
 772              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 773              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 774
 775         if (ceph_inode_is_shutdown(inode)) {
 776                 if (ci->i_wrbuffer_ref > 0) {
 777                         pr_warn_ratelimited(
 778                                 "writepage_start %p %lld forced umount\n",
 779                                 inode, ceph_ino(inode));
 780                 }
 781                 mapping_set_error(mapping, -EIO);
 782                 return -EIO; /* we're in a forced umount, don't write! */
 783         }
 784         if (fsc->mount_options->wsize < wsize)
 785                 wsize = fsc->mount_options->wsize;
 786
 787         pagevec_init(&pvec);
 788
 789         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 790         index = start_index;
 791
 792 retry:
 793         /* find oldest snap context with dirty data */
 794         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 795         if (!snapc) {
 796                 /* hmm, why does writepages get called when there
 797                    is no dirty data? */
 798                 dout(" no snap context with dirty data?\n");
 799                 goto out;
 800         }
 801         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 802              snapc, snapc->seq, snapc->num_snaps);
 803
 804         should_loop = false;
 805         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 806                 /* where to start/end? */
 807                 if (wbc->range_cyclic) {
 808                         index = start_index;
 809                         end = -1;
 810                         if (index > 0)
 811                                 should_loop = true;
 812                         dout(" cyclic, start at %lu\n", index);
 813                 } else {
 814                         index = wbc->range_start >> PAGE_SHIFT;
 815                         end = wbc->range_end >> PAGE_SHIFT;
 816                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 817                                 range_whole = true;
 818                         dout(" not cyclic, %lu to %lu\n", index, end);
 819                 }
 820         } else if (!ceph_wbc.head_snapc) {
 821                 /* Do not respect wbc->range_{start,end}. Dirty pages
 822                  * in that range can be associated with newer snapc.
 823                  * They are not writeable until we write all dirty pages
 824                  * associated with 'snapc' get written */
 825                 if (index > 0)
 826                         should_loop = true;
 827                 dout(" non-head snapc, range whole\n");
 828         }
 829
 830         ceph_put_snap_context(last_snapc);
 831         last_snapc = snapc;
 832
 833         while (!done && index <= end) {
 834                 int num_ops = 0, op_idx;
 835                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 836                 struct page **pages = NULL, **data_pages;
 837                 struct page *page;
 838                 pgoff_t strip_unit_end = 0;
 839                 u64 offset = 0, len = 0;
 840                 bool from_pool = false;
 841
 842                 max_pages = wsize >> PAGE_SHIFT;
 843
 844 get_more_pages:
 845                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 846                                                 end, PAGECACHE_TAG_DIRTY);
 847                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 848                 if (!pvec_pages && !locked_pages)
 849                         break;
 850                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 851                         page = pvec.pages[i];
 852                         dout("? %p idx %lu\n", page, page->index);
 853                         if (locked_pages == 0)
 854                                 lock_page(page);  /* first page */
 855                         else if (!trylock_page(page))
 856                                 break;
 857
 858                         /* only dirty pages, or our accounting breaks */
 859                         if (unlikely(!PageDirty(page)) ||
 860                             unlikely(page->mapping != mapping)) {
 861                                 dout("!dirty or !mapping %p\n", page);
 862                                 unlock_page(page);
 863                                 continue;
 864                         }
 865                         /* only if matching snap context */
 866                         pgsnapc = page_snap_context(page);
 867                         if (pgsnapc != snapc) {
 868                                 dout("page snapc %p %lld != oldest %p %lld\n",
 869                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 870                                 if (!should_loop &&
 871                                     !ceph_wbc.head_snapc &&
 872                                     wbc->sync_mode != WB_SYNC_NONE)
 873                                         should_loop = true;
 874                                 unlock_page(page);
 875                                 continue;
 876                         }
 877                         if (page_offset(page) >= ceph_wbc.i_size) {
 878                                 struct folio *folio = page_folio(page);
 879
 880                                 dout("folio at %lu beyond eof %llu\n",
 881                                      folio->index, ceph_wbc.i_size);
 882                                 if ((ceph_wbc.size_stable ||
 883                                     folio_pos(folio) >= i_size_read(inode)) &&
 884                                     folio_clear_dirty_for_io(folio))
 885                                         folio_invalidate(folio, 0,
 886                                                         folio_size(folio));
 887                                 folio_unlock(folio);
 888                                 continue;
 889                         }
 890                         if (strip_unit_end && (page->index > strip_unit_end)) {
 891                                 dout("end of strip unit %p\n", page);
 892                                 unlock_page(page);
 893                                 break;
 894                         }
 895                         if (PageWriteback(page) || PageFsCache(page)) {
 896                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 897                                         dout("%p under writeback\n", page);
 898                                         unlock_page(page);
 899                                         continue;
 900                                 }
 901                                 dout("waiting on writeback %p\n", page);
 902                                 wait_on_page_writeback(page);
 903                                 wait_on_page_fscache(page);
 904                         }
 905
 906                         if (!clear_page_dirty_for_io(page)) {
 907                                 dout("%p !clear_page_dirty_for_io\n", page);
 908                                 unlock_page(page);
 909                                 continue;
 910                         }
 911
 912                         /*
 913                          * We have something to write.  If this is
 914                          * the first locked page this time through,
 915                          * calculate max possinle write size and
 916                          * allocate a page array
 917                          */
 918                         if (locked_pages == 0) {
 919                                 u64 objnum;
 920                                 u64 objoff;
 921                                 u32 xlen;
 922
 923                                 /* prepare async write request */
 924                                 offset = (u64)page_offset(page);
 925                                 ceph_calc_file_object_mapping(&ci->i_layout,
 926                                                               offset, wsize,
 927                                                               &objnum, &objoff,
 928                                                               &xlen);
 929                                 len = xlen;
 930
 931                                 num_ops = 1;
 932                                 strip_unit_end = page->index +
 933                                         ((len - 1) >> PAGE_SHIFT);
 934
 935                                 BUG_ON(pages);
 936                                 max_pages = calc_pages_for(0, (u64)len);
 937                                 pages = kmalloc_array(max_pages,
 938                                                       sizeof(*pages),
 939                                                       GFP_NOFS);
 940                                 if (!pages) {
 941                                         from_pool = true;
 942                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 943                                         BUG_ON(!pages);
 944                                 }
 945
 946                                 len = 0;
 947                         } else if (page->index !=
 948                                    (offset + len) >> PAGE_SHIFT) {
 949                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 950                                                              CEPH_OSD_MAX_OPS)) {
 951                                         redirty_page_for_writepage(wbc, page);
 952                                         unlock_page(page);
 953                                         break;
 954                                 }
 955
 956                                 num_ops++;
 957                                 offset = (u64)page_offset(page);
 958                                 len = 0;
 959                         }
 960
 961                         /* note position of first page in pvec */
 962                         dout("%p will write page %p idx %lu\n",
 963                              inode, page, page->index);
 964
 965                         if (atomic_long_inc_return(&fsc->writeback_count) >
 966                             CONGESTION_ON_THRESH(
 967                                     fsc->mount_options->congestion_kb))
 968                                 fsc->write_congested = true;
 969
 970                         pages[locked_pages++] = page;
 971                         pvec.pages[i] = NULL;
 972
 973                         len += thp_size(page);
 974                 }
 975
 976                 /* did we get anything? */
 977                 if (!locked_pages)
 978                         goto release_pvec_pages;
 979                 if (i) {
 980                         unsigned j, n = 0;
 981                         /* shift unused page to beginning of pvec */
 982                         for (j = 0; j < pvec_pages; j++) {
 983                                 if (!pvec.pages[j])
 984                                         continue;
 985                                 if (n < j)
 986                                         pvec.pages[n] = pvec.pages[j];
 987                                 n++;
 988                         }
 989                         pvec.nr = n;
 990
 991                         if (pvec_pages && i == pvec_pages &&
 992                             locked_pages < max_pages) {
 993                                 dout("reached end pvec, trying for more\n");
 994                                 pagevec_release(&pvec);
 995                                 goto get_more_pages;
 996                         }
 997                 }
 998
 999 new_request:
1000                 offset = page_offset(pages[0]);
1001                 len = wsize;
1002
1003                 req = ceph_osdc_new_request(&fsc->client->osdc,
1004                                         &ci->i_layout, vino,
1005                                         offset, &len, 0, num_ops,
1006                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1007                                         snapc, ceph_wbc.truncate_seq,
1008                                         ceph_wbc.truncate_size, false);
1009                 if (IS_ERR(req)) {
1010                         req = ceph_osdc_new_request(&fsc->client->osdc,
1011                                                 &ci->i_layout, vino,
1012                                                 offset, &len, 0,
1013                                                 min(num_ops,
1014                                                     CEPH_OSD_SLAB_OPS),
1015                                                 CEPH_OSD_OP_WRITE,
1016                                                 CEPH_OSD_FLAG_WRITE,
1017                                                 snapc, ceph_wbc.truncate_seq,
1018                                                 ceph_wbc.truncate_size, true);
1019                         BUG_ON(IS_ERR(req));
1020                 }
1021                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1022                              thp_size(page) - offset);
1023
1024                 req->r_callback = writepages_finish;
1025                 req->r_inode = inode;
1026
1027                 /* Format the osd request message and submit the write */
1028                 len = 0;
1029                 data_pages = pages;
1030                 op_idx = 0;
1031                 for (i = 0; i < locked_pages; i++) {
1032                         u64 cur_offset = page_offset(pages[i]);
1033                         /*
1034                          * Discontinuity in page range? Ceph can handle that by just passing
1035                          * multiple extents in the write op.
1036                          */
1037                         if (offset + len != cur_offset) {
1038                                 /* If it's full, stop here */
1039                                 if (op_idx + 1 == req->r_num_ops)
1040                                         break;
1041
1042                                 /* Kick off an fscache write with what we have so far. */
1043                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1044
1045                                 /* Start a new extent */
1046                                 osd_req_op_extent_dup_last(req, op_idx,
1047                                                            cur_offset - offset);
1048                                 dout("writepages got pages at %llu~%llu\n",
1049                                      offset, len);
1050                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1051                                                         data_pages, len, 0,
1052                                                         from_pool, false);
1053                                 osd_req_op_extent_update(req, op_idx, len);
1054
1055                                 len = 0;
1056                                 offset = cur_offset;
1057                                 data_pages = pages + i;
1058                                 op_idx++;
1059                         }
1060
1061                         set_page_writeback(pages[i]);
1062                         if (caching)
1063                                 ceph_set_page_fscache(pages[i]);
1064                         len += thp_size(page);
1065                 }
1066                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1067
1068                 if (ceph_wbc.size_stable) {
1069                         len = min(len, ceph_wbc.i_size - offset);
1070                 } else if (i == locked_pages) {
1071                         /* writepages_finish() clears writeback pages
1072                          * according to the data length, so make sure
1073                          * data length covers all locked pages */
1074                         u64 min_len = len + 1 - thp_size(page);
1075                         len = get_writepages_data_length(inode, pages[i - 1],
1076                                                          offset);
1077                         len = max(len, min_len);
1078                 }
1079                 dout("writepages got pages at %llu~%llu\n", offset, len);
1080
1081                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1082                                                  0, from_pool, false);
1083                 osd_req_op_extent_update(req, op_idx, len);
1084
1085                 BUG_ON(op_idx + 1 != req->r_num_ops);
1086
1087                 from_pool = false;
1088                 if (i < locked_pages) {
1089                         BUG_ON(num_ops <= req->r_num_ops);
1090                         num_ops -= req->r_num_ops;
1091                         locked_pages -= i;
1092
1093                         /* allocate new pages array for next request */
1094                         data_pages = pages;
1095                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1096                                               GFP_NOFS);
1097                         if (!pages) {
1098                                 from_pool = true;
1099                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1100                                 BUG_ON(!pages);
1101                         }
1102                         memcpy(pages, data_pages + i,
1103                                locked_pages * sizeof(*pages));
1104                         memset(data_pages + i, 0,
1105                                locked_pages * sizeof(*pages));
1106                 } else {
1107                         BUG_ON(num_ops != req->r_num_ops);
1108                         index = pages[i - 1]->index + 1;
1109                         /* request message now owns the pages array */
1110                         pages = NULL;
1111                 }
1112
1113                 req->r_mtime = inode->i_mtime;
1114                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1115                 BUG_ON(rc);
1116                 req = NULL;
1117
1118                 wbc->nr_to_write -= i;
1119                 if (pages)
1120                         goto new_request;
1121
1122                 /*
1123                  * We stop writing back only if we are not doing
1124                  * integrity sync. In case of integrity sync we have to
1125                  * keep going until we have written all the pages
1126                  * we tagged for writeback prior to entering this loop.
1127                  */
1128                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1129                         done = true;
1130
1131 release_pvec_pages:
1132                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1133                      pvec.nr ? pvec.pages[0] : NULL);
1134                 pagevec_release(&pvec);
1135         }
1136
1137         if (should_loop && !done) {
1138                 /* more to do; loop back to beginning of file */
1139                 dout("writepages looping back to beginning of file\n");
1140                 end = start_index - 1; /* OK even when start_index == 0 */
1141
1142                 /* to write dirty pages associated with next snapc,
1143                  * we need to wait until current writes complete */
1144                 if (wbc->sync_mode != WB_SYNC_NONE &&
1145                     start_index == 0 && /* all dirty pages were checked */
1146                     !ceph_wbc.head_snapc) {
1147                         struct page *page;
1148                         unsigned i, nr;
1149                         index = 0;
1150                         while ((index <= end) &&
1151                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1152                                                 PAGECACHE_TAG_WRITEBACK))) {
1153                                 for (i = 0; i < nr; i++) {
1154                                         page = pvec.pages[i];
1155                                         if (page_snap_context(page) != snapc)
1156                                                 continue;
1157                                         wait_on_page_writeback(page);
1158                                 }
1159                                 pagevec_release(&pvec);
1160                                 cond_resched();
1161                         }
1162                 }
1163
1164                 start_index = 0;
1165                 index = 0;
1166                 goto retry;
1167         }
1168
1169         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1170                 mapping->writeback_index = index;
1171
1172 out:
1173         ceph_osdc_put_request(req);
1174         ceph_put_snap_context(last_snapc);
1175         dout("writepages dend - startone, rc = %d\n", rc);
1176         return rc;
1177 }
1178
1179
1180
1181 /*
1182  * See if a given @snapc is either writeable, or already written.
1183  */
1184 static int context_is_writeable_or_written(struct inode *inode,
1185                                            struct ceph_snap_context *snapc)
1186 {
1187         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1188         int ret = !oldest || snapc->seq <= oldest->seq;
1189
1190         ceph_put_snap_context(oldest);
1191         return ret;
1192 }
1193
1194 /**
1195  * ceph_find_incompatible - find an incompatible context and return it
1196  * @page: page being dirtied
1197  *
1198  * We are only allowed to write into/dirty a page if the page is
1199  * clean, or already dirty within the same snap context. Returns a
1200  * conflicting context if there is one, NULL if there isn't, or a
1201  * negative error code on other errors.
1202  *
1203  * Must be called with page lock held.
1204  */
1205 static struct ceph_snap_context *
1206 ceph_find_incompatible(struct page *page)
1207 {
1208         struct inode *inode = page->mapping->host;
1209         struct ceph_inode_info *ci = ceph_inode(inode);
1210
1211         if (ceph_inode_is_shutdown(inode)) {
1212                 dout(" page %p %llx:%llx is shutdown\n", page,
1213                      ceph_vinop(inode));
1214                 return ERR_PTR(-ESTALE);
1215         }
1216
1217         for (;;) {
1218                 struct ceph_snap_context *snapc, *oldest;
1219
1220                 wait_on_page_writeback(page);
1221
1222                 snapc = page_snap_context(page);
1223                 if (!snapc || snapc == ci->i_head_snapc)
1224                         break;
1225
1226                 /*
1227                  * this page is already dirty in another (older) snap
1228                  * context!  is it writeable now?
1229                  */
1230                 oldest = get_oldest_context(inode, NULL, NULL);
1231                 if (snapc->seq > oldest->seq) {
1232                         /* not writeable -- return it for the caller to deal with */
1233                         ceph_put_snap_context(oldest);
1234                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1235                         return ceph_get_snap_context(snapc);
1236                 }
1237                 ceph_put_snap_context(oldest);
1238
1239                 /* yay, writeable, do it now (without dropping page lock) */
1240                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1241                 if (clear_page_dirty_for_io(page)) {
1242                         int r = writepage_nounlock(page, NULL);
1243                         if (r < 0)
1244                                 return ERR_PTR(r);
1245                 }
1246         }
1247         return NULL;
1248 }
1249
1250 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1251                                         struct folio *folio, void **_fsdata)
1252 {
1253         struct inode *inode = file_inode(file);
1254         struct ceph_inode_info *ci = ceph_inode(inode);
1255         struct ceph_snap_context *snapc;
1256
1257         snapc = ceph_find_incompatible(folio_page(folio, 0));
1258         if (snapc) {
1259                 int r;
1260
1261                 folio_unlock(folio);
1262                 folio_put(folio);
1263                 if (IS_ERR(snapc))
1264                         return PTR_ERR(snapc);
1265
1266                 ceph_queue_writeback(inode);
1267                 r = wait_event_killable(ci->i_cap_wq,
1268                                         context_is_writeable_or_written(inode, snapc));
1269                 ceph_put_snap_context(snapc);
1270                 return r == 0 ? -EAGAIN : r;
1271         }
1272         return 0;
1273 }
1274
1275 /*
1276  * We are only allowed to write into/dirty the page if the page is
1277  * clean, or already dirty within the same snap context.
1278  */
1279 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1280                             loff_t pos, unsigned len, unsigned aop_flags,
1281                             struct page **pagep, void **fsdata)
1282 {
1283         struct inode *inode = file_inode(file);
1284         struct ceph_inode_info *ci = ceph_inode(inode);
1285         struct folio *folio = NULL;
1286         pgoff_t index = pos >> PAGE_SHIFT;
1287         int r;
1288
1289         /*
1290          * Uninlining should have already been done and everything updated, EXCEPT
1291          * for inline_version sent to the MDS.
1292          */
1293         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1294                 unsigned int fgp_flags = FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE;
1295                 if (aop_flags & AOP_FLAG_NOFS)
1296                         fgp_flags |= FGP_NOFS;
1297                 folio = __filemap_get_folio(mapping, index, fgp_flags,
1298                                             mapping_gfp_mask(mapping));
1299                 if (!folio)
1300                         return -ENOMEM;
1301
1302                 /*
1303                  * The inline_version on a new inode is set to 1. If that's the
1304                  * case, then the folio is brand new and isn't yet Uptodate.
1305                  */
1306                 r = 0;
1307                 if (index == 0 && ci->i_inline_version != 1) {
1308                         if (!folio_test_uptodate(folio)) {
1309                                 WARN_ONCE(1, "ceph: write_begin called on still-inlined inode (inline_version %llu)!\n",
1310                                           ci->i_inline_version);
1311                                 r = -EINVAL;
1312                         }
1313                         goto out;
1314                 }
1315                 zero_user_segment(&folio->page, 0, folio_size(folio));
1316                 folio_mark_uptodate(folio);
1317                 goto out;
1318         }
1319
1320         r = netfs_write_begin(file, inode->i_mapping, pos, len, 0, &folio, NULL,
1321                               &ceph_netfs_read_ops, NULL);
1322 out:
1323         if (r == 0)
1324                 folio_wait_fscache(folio);
1325         if (r < 0) {
1326                 if (folio)
1327                         folio_put(folio);
1328         } else {
1329                 WARN_ON_ONCE(!folio_test_locked(folio));
1330                 *pagep = &folio->page;
1331         }
1332         return r;
1333 }
1334
1335 /*
1336  * we don't do anything in here that simple_write_end doesn't do
1337  * except adjust dirty page accounting
1338  */
1339 static int ceph_write_end(struct file *file, struct address_space *mapping,
1340                           loff_t pos, unsigned len, unsigned copied,
1341                           struct page *subpage, void *fsdata)
1342 {
1343         struct folio *folio = page_folio(subpage);
1344         struct inode *inode = file_inode(file);
1345         bool check_cap = false;
1346
1347         dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1348              inode, folio, (int)pos, (int)copied, (int)len);
1349
1350         if (!folio_test_uptodate(folio)) {
1351                 /* just return that nothing was copied on a short copy */
1352                 if (copied < len) {
1353                         copied = 0;
1354                         goto out;
1355                 }
1356                 folio_mark_uptodate(folio);
1357         }
1358
1359         /* did file size increase? */
1360         if (pos+copied > i_size_read(inode))
1361                 check_cap = ceph_inode_set_size(inode, pos+copied);
1362
1363         folio_mark_dirty(folio);
1364
1365 out:
1366         folio_unlock(folio);
1367         folio_put(folio);
1368
1369         if (check_cap)
1370                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1371
1372         return copied;
1373 }
1374
1375 const struct address_space_operations ceph_aops = {
1376         .readpage = ceph_readpage,
1377         .readahead = ceph_readahead,
1378         .writepage = ceph_writepage,
1379         .writepages = ceph_writepages_start,
1380         .write_begin = ceph_write_begin,
1381         .write_end = ceph_write_end,
1382         .dirty_folio = ceph_dirty_folio,
1383         .invalidate_folio = ceph_invalidate_folio,
1384         .releasepage = ceph_releasepage,
1385         .direct_IO = noop_direct_IO,
1386 };
1387
1388 static void ceph_block_sigs(sigset_t *oldset)
1389 {
1390         sigset_t mask;
1391         siginitsetinv(&mask, sigmask(SIGKILL));
1392         sigprocmask(SIG_BLOCK, &mask, oldset);
1393 }
1394
1395 static void ceph_restore_sigs(sigset_t *oldset)
1396 {
1397         sigprocmask(SIG_SETMASK, oldset, NULL);
1398 }
1399
1400 /*
1401  * vm ops
1402  */
1403 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1404 {
1405         struct vm_area_struct *vma = vmf->vma;
1406         struct inode *inode = file_inode(vma->vm_file);
1407         struct ceph_inode_info *ci = ceph_inode(inode);
1408         struct ceph_file_info *fi = vma->vm_file->private_data;
1409         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1410         int want, got, err;
1411         sigset_t oldset;
1412         vm_fault_t ret = VM_FAULT_SIGBUS;
1413
1414         if (ceph_inode_is_shutdown(inode))
1415                 return ret;
1416
1417         ceph_block_sigs(&oldset);
1418
1419         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1420              inode, ceph_vinop(inode), off);
1421         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1422                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1423         else
1424                 want = CEPH_CAP_FILE_CACHE;
1425
1426         got = 0;
1427         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1428         if (err < 0)
1429                 goto out_restore;
1430
1431         dout("filemap_fault %p %llu got cap refs on %s\n",
1432              inode, off, ceph_cap_string(got));
1433
1434         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1435             ci->i_inline_version == CEPH_INLINE_NONE) {
1436                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1437                 ceph_add_rw_context(fi, &rw_ctx);
1438                 ret = filemap_fault(vmf);
1439                 ceph_del_rw_context(fi, &rw_ctx);
1440                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1441                      inode, off, ceph_cap_string(got), ret);
1442         } else
1443                 err = -EAGAIN;
1444
1445         ceph_put_cap_refs(ci, got);
1446
1447         if (err != -EAGAIN)
1448                 goto out_restore;
1449
1450         /* read inline data */
1451         if (off >= PAGE_SIZE) {
1452                 /* does not support inline data > PAGE_SIZE */
1453                 ret = VM_FAULT_SIGBUS;
1454         } else {
1455                 struct address_space *mapping = inode->i_mapping;
1456                 struct page *page;
1457
1458                 filemap_invalidate_lock_shared(mapping);
1459                 page = find_or_create_page(mapping, 0,
1460                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1461                 if (!page) {
1462                         ret = VM_FAULT_OOM;
1463                         goto out_inline;
1464                 }
1465                 err = __ceph_do_getattr(inode, page,
1466                                          CEPH_STAT_CAP_INLINE_DATA, true);
1467                 if (err < 0 || off >= i_size_read(inode)) {
1468                         unlock_page(page);
1469                         put_page(page);
1470                         ret = vmf_error(err);
1471                         goto out_inline;
1472                 }
1473                 if (err < PAGE_SIZE)
1474                         zero_user_segment(page, err, PAGE_SIZE);
1475                 else
1476                         flush_dcache_page(page);
1477                 SetPageUptodate(page);
1478                 vmf->page = page;
1479                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1480 out_inline:
1481                 filemap_invalidate_unlock_shared(mapping);
1482                 dout("filemap_fault %p %llu read inline data ret %x\n",
1483                      inode, off, ret);
1484         }
1485 out_restore:
1486         ceph_restore_sigs(&oldset);
1487         if (err < 0)
1488                 ret = vmf_error(err);
1489
1490         return ret;
1491 }
1492
1493 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1494 {
1495         struct vm_area_struct *vma = vmf->vma;
1496         struct inode *inode = file_inode(vma->vm_file);
1497         struct ceph_inode_info *ci = ceph_inode(inode);
1498         struct ceph_file_info *fi = vma->vm_file->private_data;
1499         struct ceph_cap_flush *prealloc_cf;
1500         struct page *page = vmf->page;
1501         loff_t off = page_offset(page);
1502         loff_t size = i_size_read(inode);
1503         size_t len;
1504         int want, got, err;
1505         sigset_t oldset;
1506         vm_fault_t ret = VM_FAULT_SIGBUS;
1507
1508         if (ceph_inode_is_shutdown(inode))
1509                 return ret;
1510
1511         prealloc_cf = ceph_alloc_cap_flush();
1512         if (!prealloc_cf)
1513                 return VM_FAULT_OOM;
1514
1515         sb_start_pagefault(inode->i_sb);
1516         ceph_block_sigs(&oldset);
1517
1518         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1519                 struct page *locked_page = NULL;
1520                 if (off == 0) {
1521                         lock_page(page);
1522                         locked_page = page;
1523                 }
1524                 err = ceph_uninline_data(vma->vm_file, locked_page);
1525                 if (locked_page)
1526                         unlock_page(locked_page);
1527                 if (err < 0)
1528                         goto out_free;
1529         }
1530
1531         if (off + thp_size(page) <= size)
1532                 len = thp_size(page);
1533         else
1534                 len = offset_in_thp(page, size);
1535
1536         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1537              inode, ceph_vinop(inode), off, len, size);
1538         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1539                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1540         else
1541                 want = CEPH_CAP_FILE_BUFFER;
1542
1543         got = 0;
1544         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1545         if (err < 0)
1546                 goto out_free;
1547
1548         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1549              inode, off, len, ceph_cap_string(got));
1550
1551         /* Update time before taking page lock */
1552         file_update_time(vma->vm_file);
1553         inode_inc_iversion_raw(inode);
1554
1555         do {
1556                 struct ceph_snap_context *snapc;
1557
1558                 lock_page(page);
1559
1560                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1561                         unlock_page(page);
1562                         ret = VM_FAULT_NOPAGE;
1563                         break;
1564                 }
1565
1566                 snapc = ceph_find_incompatible(page);
1567                 if (!snapc) {
1568                         /* success.  we'll keep the page locked. */
1569                         set_page_dirty(page);
1570                         ret = VM_FAULT_LOCKED;
1571                         break;
1572                 }
1573
1574                 unlock_page(page);
1575
1576                 if (IS_ERR(snapc)) {
1577                         ret = VM_FAULT_SIGBUS;
1578                         break;
1579                 }
1580
1581                 ceph_queue_writeback(inode);
1582                 err = wait_event_killable(ci->i_cap_wq,
1583                                 context_is_writeable_or_written(inode, snapc));
1584                 ceph_put_snap_context(snapc);
1585         } while (err == 0);
1586
1587         if (ret == VM_FAULT_LOCKED ||
1588             ci->i_inline_version != CEPH_INLINE_NONE) {
1589                 int dirty;
1590                 spin_lock(&ci->i_ceph_lock);
1591                 ci->i_inline_version = CEPH_INLINE_NONE;
1592                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1593                                                &prealloc_cf);
1594                 spin_unlock(&ci->i_ceph_lock);
1595                 if (dirty)
1596                         __mark_inode_dirty(inode, dirty);
1597         }
1598
1599         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1600              inode, off, len, ceph_cap_string(got), ret);
1601         ceph_put_cap_refs_async(ci, got);
1602 out_free:
1603         ceph_restore_sigs(&oldset);
1604         sb_end_pagefault(inode->i_sb);
1605         ceph_free_cap_flush(prealloc_cf);
1606         if (err < 0)
1607                 ret = vmf_error(err);
1608         return ret;
1609 }
1610
1611 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1612                            char *data, size_t len)
1613 {
1614         struct address_space *mapping = inode->i_mapping;
1615         struct page *page;
1616
1617         if (locked_page) {
1618                 page = locked_page;
1619         } else {
1620                 if (i_size_read(inode) == 0)
1621                         return;
1622                 page = find_or_create_page(mapping, 0,
1623                                            mapping_gfp_constraint(mapping,
1624                                            ~__GFP_FS));
1625                 if (!page)
1626                         return;
1627                 if (PageUptodate(page)) {
1628                         unlock_page(page);
1629                         put_page(page);
1630                         return;
1631                 }
1632         }
1633
1634         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1635              inode, ceph_vinop(inode), len, locked_page);
1636
1637         if (len > 0) {
1638                 void *kaddr = kmap_atomic(page);
1639                 memcpy(kaddr, data, len);
1640                 kunmap_atomic(kaddr);
1641         }
1642
1643         if (page != locked_page) {
1644                 if (len < PAGE_SIZE)
1645                         zero_user_segment(page, len, PAGE_SIZE);
1646                 else
1647                         flush_dcache_page(page);
1648
1649                 SetPageUptodate(page);
1650                 unlock_page(page);
1651                 put_page(page);
1652         }
1653 }
1654
1655 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1656 {
1657         struct inode *inode = file_inode(filp);
1658         struct ceph_inode_info *ci = ceph_inode(inode);
1659         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1660         struct ceph_osd_request *req;
1661         struct page *page = NULL;
1662         u64 len, inline_version;
1663         int err = 0;
1664         bool from_pagecache = false;
1665
1666         spin_lock(&ci->i_ceph_lock);
1667         inline_version = ci->i_inline_version;
1668         spin_unlock(&ci->i_ceph_lock);
1669
1670         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1671              inode, ceph_vinop(inode), inline_version);
1672
1673         if (inline_version == 1 || /* initial version, no data */
1674             inline_version == CEPH_INLINE_NONE)
1675                 goto out;
1676
1677         if (locked_page) {
1678                 page = locked_page;
1679                 WARN_ON(!PageUptodate(page));
1680         } else if (ceph_caps_issued(ci) &
1681                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1682                 page = find_get_page(inode->i_mapping, 0);
1683                 if (page) {
1684                         if (PageUptodate(page)) {
1685                                 from_pagecache = true;
1686                                 lock_page(page);
1687                         } else {
1688                                 put_page(page);
1689                                 page = NULL;
1690                         }
1691                 }
1692         }
1693
1694         if (page) {
1695                 len = i_size_read(inode);
1696                 if (len > PAGE_SIZE)
1697                         len = PAGE_SIZE;
1698         } else {
1699                 page = __page_cache_alloc(GFP_NOFS);
1700                 if (!page) {
1701                         err = -ENOMEM;
1702                         goto out;
1703                 }
1704                 err = __ceph_do_getattr(inode, page,
1705                                         CEPH_STAT_CAP_INLINE_DATA, true);
1706                 if (err < 0) {
1707                         /* no inline data */
1708                         if (err == -ENODATA)
1709                                 err = 0;
1710                         goto out;
1711                 }
1712                 len = err;
1713         }
1714
1715         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1716                                     ceph_vino(inode), 0, &len, 0, 1,
1717                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1718                                     NULL, 0, 0, false);
1719         if (IS_ERR(req)) {
1720                 err = PTR_ERR(req);
1721                 goto out;
1722         }
1723
1724         req->r_mtime = inode->i_mtime;
1725         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1726         if (!err)
1727                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1728         ceph_osdc_put_request(req);
1729         if (err < 0)
1730                 goto out;
1731
1732         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1733                                     ceph_vino(inode), 0, &len, 1, 3,
1734                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1735                                     NULL, ci->i_truncate_seq,
1736                                     ci->i_truncate_size, false);
1737         if (IS_ERR(req)) {
1738                 err = PTR_ERR(req);
1739                 goto out;
1740         }
1741
1742         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1743
1744         {
1745                 __le64 xattr_buf = cpu_to_le64(inline_version);
1746                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1747                                             "inline_version", &xattr_buf,
1748                                             sizeof(xattr_buf),
1749                                             CEPH_OSD_CMPXATTR_OP_GT,
1750                                             CEPH_OSD_CMPXATTR_MODE_U64);
1751                 if (err)
1752                         goto out_put;
1753         }
1754
1755         {
1756                 char xattr_buf[32];
1757                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1758                                          "%llu", inline_version);
1759                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1760                                             "inline_version",
1761                                             xattr_buf, xattr_len, 0, 0);
1762                 if (err)
1763                         goto out_put;
1764         }
1765
1766         req->r_mtime = inode->i_mtime;
1767         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1768         if (!err)
1769                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1770
1771         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1772                                   req->r_end_latency, len, err);
1773
1774 out_put:
1775         ceph_osdc_put_request(req);
1776         if (err == -ECANCELED)
1777                 err = 0;
1778 out:
1779         if (page && page != locked_page) {
1780                 if (from_pagecache) {
1781                         unlock_page(page);
1782                         put_page(page);
1783                 } else
1784                         __free_pages(page, 0);
1785         }
1786
1787         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1788              inode, ceph_vinop(inode), inline_version, err);
1789         return err;
1790 }
1791
1792 static const struct vm_operations_struct ceph_vmops = {
1793         .fault          = ceph_filemap_fault,
1794         .page_mkwrite   = ceph_page_mkwrite,
1795 };
1796
1797 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1798 {
1799         struct address_space *mapping = file->f_mapping;
1800
1801         if (!mapping->a_ops->readpage)
1802                 return -ENOEXEC;
1803         file_accessed(file);
1804         vma->vm_ops = &ceph_vmops;
1805         return 0;
1806 }
1807
1808 enum {
1809         POOL_READ       = 1,
1810         POOL_WRITE      = 2,
1811 };
1812
1813 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1814                                 s64 pool, struct ceph_string *pool_ns)
1815 {
1816         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1817         struct ceph_mds_client *mdsc = fsc->mdsc;
1818         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1819         struct rb_node **p, *parent;
1820         struct ceph_pool_perm *perm;
1821         struct page **pages;
1822         size_t pool_ns_len;
1823         int err = 0, err2 = 0, have = 0;
1824
1825         down_read(&mdsc->pool_perm_rwsem);
1826         p = &mdsc->pool_perm_tree.rb_node;
1827         while (*p) {
1828                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1829                 if (pool < perm->pool)
1830                         p = &(*p)->rb_left;
1831                 else if (pool > perm->pool)
1832                         p = &(*p)->rb_right;
1833                 else {
1834                         int ret = ceph_compare_string(pool_ns,
1835                                                 perm->pool_ns,
1836                                                 perm->pool_ns_len);
1837                         if (ret < 0)
1838                                 p = &(*p)->rb_left;
1839                         else if (ret > 0)
1840                                 p = &(*p)->rb_right;
1841                         else {
1842                                 have = perm->perm;
1843                                 break;
1844                         }
1845                 }
1846         }
1847         up_read(&mdsc->pool_perm_rwsem);
1848         if (*p)
1849                 goto out;
1850
1851         if (pool_ns)
1852                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1853                      pool, (int)pool_ns->len, pool_ns->str);
1854         else
1855                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1856
1857         down_write(&mdsc->pool_perm_rwsem);
1858         p = &mdsc->pool_perm_tree.rb_node;
1859         parent = NULL;
1860         while (*p) {
1861                 parent = *p;
1862                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1863                 if (pool < perm->pool)
1864                         p = &(*p)->rb_left;
1865                 else if (pool > perm->pool)
1866                         p = &(*p)->rb_right;
1867                 else {
1868                         int ret = ceph_compare_string(pool_ns,
1869                                                 perm->pool_ns,
1870                                                 perm->pool_ns_len);
1871                         if (ret < 0)
1872                                 p = &(*p)->rb_left;
1873                         else if (ret > 0)
1874                                 p = &(*p)->rb_right;
1875                         else {
1876                                 have = perm->perm;
1877                                 break;
1878                         }
1879                 }
1880         }
1881         if (*p) {
1882                 up_write(&mdsc->pool_perm_rwsem);
1883                 goto out;
1884         }
1885
1886         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1887                                          1, false, GFP_NOFS);
1888         if (!rd_req) {
1889                 err = -ENOMEM;
1890                 goto out_unlock;
1891         }
1892
1893         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1894         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1895         rd_req->r_base_oloc.pool = pool;
1896         if (pool_ns)
1897                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1898         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1899
1900         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1901         if (err)
1902                 goto out_unlock;
1903
1904         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1905                                          1, false, GFP_NOFS);
1906         if (!wr_req) {
1907                 err = -ENOMEM;
1908                 goto out_unlock;
1909         }
1910
1911         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1912         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1913         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1914         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1915
1916         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1917         if (err)
1918                 goto out_unlock;
1919
1920         /* one page should be large enough for STAT data */
1921         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1922         if (IS_ERR(pages)) {
1923                 err = PTR_ERR(pages);
1924                 goto out_unlock;
1925         }
1926
1927         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1928                                      0, false, true);
1929         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1930
1931         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1932         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1933
1934         if (!err)
1935                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1936         if (!err2)
1937                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1938
1939         if (err >= 0 || err == -ENOENT)
1940                 have |= POOL_READ;
1941         else if (err != -EPERM) {
1942                 if (err == -EBLOCKLISTED)
1943                         fsc->blocklisted = true;
1944                 goto out_unlock;
1945         }
1946
1947         if (err2 == 0 || err2 == -EEXIST)
1948                 have |= POOL_WRITE;
1949         else if (err2 != -EPERM) {
1950                 if (err2 == -EBLOCKLISTED)
1951                         fsc->blocklisted = true;
1952                 err = err2;
1953                 goto out_unlock;
1954         }
1955
1956         pool_ns_len = pool_ns ? pool_ns->len : 0;
1957         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1958         if (!perm) {
1959                 err = -ENOMEM;
1960                 goto out_unlock;
1961         }
1962
1963         perm->pool = pool;
1964         perm->perm = have;
1965         perm->pool_ns_len = pool_ns_len;
1966         if (pool_ns_len > 0)
1967                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1968         perm->pool_ns[pool_ns_len] = 0;
1969
1970         rb_link_node(&perm->node, parent, p);
1971         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1972         err = 0;
1973 out_unlock:
1974         up_write(&mdsc->pool_perm_rwsem);
1975
1976         ceph_osdc_put_request(rd_req);
1977         ceph_osdc_put_request(wr_req);
1978 out:
1979         if (!err)
1980                 err = have;
1981         if (pool_ns)
1982                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1983                      pool, (int)pool_ns->len, pool_ns->str, err);
1984         else
1985                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1986         return err;
1987 }
1988
1989 int ceph_pool_perm_check(struct inode *inode, int need)
1990 {
1991         struct ceph_inode_info *ci = ceph_inode(inode);
1992         struct ceph_string *pool_ns;
1993         s64 pool;
1994         int ret, flags;
1995
1996         /* Only need to do this for regular files */
1997         if (!S_ISREG(inode->i_mode))
1998                 return 0;
1999
2000         if (ci->i_vino.snap != CEPH_NOSNAP) {
2001                 /*
2002                  * Pool permission check needs to write to the first object.
2003                  * But for snapshot, head of the first object may have alread
2004                  * been deleted. Skip check to avoid creating orphan object.
2005                  */
2006                 return 0;
2007         }
2008
2009         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2010                                 NOPOOLPERM))
2011                 return 0;
2012
2013         spin_lock(&ci->i_ceph_lock);
2014         flags = ci->i_ceph_flags;
2015         pool = ci->i_layout.pool_id;
2016         spin_unlock(&ci->i_ceph_lock);
2017 check:
2018         if (flags & CEPH_I_POOL_PERM) {
2019                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2020                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2021                              pool);
2022                         return -EPERM;
2023                 }
2024                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2025                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2026                              pool);
2027                         return -EPERM;
2028                 }
2029                 return 0;
2030         }
2031
2032         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2033         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2034         ceph_put_string(pool_ns);
2035         if (ret < 0)
2036                 return ret;
2037
2038         flags = CEPH_I_POOL_PERM;
2039         if (ret & POOL_READ)
2040                 flags |= CEPH_I_POOL_RD;
2041         if (ret & POOL_WRITE)
2042                 flags |= CEPH_I_POOL_WR;
2043
2044         spin_lock(&ci->i_ceph_lock);
2045         if (pool == ci->i_layout.pool_id &&
2046             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2047                 ci->i_ceph_flags |= flags;
2048         } else {
2049                 pool = ci->i_layout.pool_id;
2050                 flags = ci->i_ceph_flags;
2051         }
2052         spin_unlock(&ci->i_ceph_lock);
2053         goto check;
2054 }
2055
2056 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2057 {
2058         struct ceph_pool_perm *perm;
2059         struct rb_node *n;
2060
2061         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2062                 n = rb_first(&mdsc->pool_perm_tree);
2063                 perm = rb_entry(n, struct ceph_pool_perm, node);
2064                 rb_erase(n, &mdsc->pool_perm_tree);
2065                 kfree(perm);
2066         }
2067 }