fs/ceph/addr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/backing-dev.h>
   5 #include <linux/fs.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>    /* generic_writepages */
   9 #include <linux/slab.h>
  10 #include <linux/pagevec.h>
  11 #include <linux/task_io_accounting_ops.h>
  12 #include <linux/signal.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15 #include <linux/netfs.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include "metric.h"
  21 #include <linux/ceph/osd_client.h>
  22 #include <linux/ceph/striper.h>
  23
  24 /*
  25  * Ceph address space ops.
  26  *
  27  * There are a few funny things going on here.
  28  *
  29  * The page->private field is used to reference a struct
  30  * ceph_snap_context for _every_ dirty page.  This indicates which
  31  * snapshot the page was logically dirtied in, and thus which snap
  32  * context needs to be associated with the osd write during writeback.
  33  *
  34  * Similarly, struct ceph_inode_info maintains a set of counters to
  35  * count dirty pages on the inode.  In the absence of snapshots,
  36  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  37  *
  38  * When a snapshot is taken (that is, when the client receives
  39  * notification that a snapshot was taken), each inode with caps and
  40  * with dirty pages (dirty pages implies there is a cap) gets a new
  41  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  42  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  43  * moved to capsnap->dirty. (Unless a sync write is currently in
  44  * progress.  In that case, the capsnap is said to be "pending", new
  45  * writes cannot start, and the capsnap isn't "finalized" until the
  46  * write completes (or fails) and a final size/mtime for the inode for
  47  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  48  *
  49  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  50  * we look for the first capsnap in i_cap_snaps and write out pages in
  51  * that snap context _only_.  Then we move on to the next capsnap,
  52  * eventually reaching the "live" or "head" context (i.e., pages that
  53  * are not yet snapped) and are writing the most recently dirtied
  54  * pages.
  55  *
  56  * Invalidate and so forth must take care to ensure the dirty page
  57  * accounting is preserved.
  58  */
  59
  60 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  61 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  62         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  63          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  64
  65 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  66                                         struct page *page, void **_fsdata);
  67
  68 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  69 {
  70         if (PagePrivate(page))
  71                 return (void *)page->private;
  72         return NULL;
  73 }
  74
  75 /*
  76  * Dirty a page.  Optimistically adjust accounting, on the assumption
  77  * that we won't race with invalidate.  If we do, readjust.
  78  */
  79 static int ceph_set_page_dirty(struct page *page)
  80 {
  81         struct address_space *mapping = page->mapping;
  82         struct inode *inode;
  83         struct ceph_inode_info *ci;
  84         struct ceph_snap_context *snapc;
  85         int ret;
  86
  87         if (unlikely(!mapping))
  88                 return !TestSetPageDirty(page);
  89
  90         if (PageDirty(page)) {
  91                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  92                      mapping->host, page, page->index);
  93                 BUG_ON(!PagePrivate(page));
  94                 return 0;
  95         }
  96
  97         inode = mapping->host;
  98         ci = ceph_inode(inode);
  99
 100         /* dirty the head */
 101         spin_lock(&ci->i_ceph_lock);
 102         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
 103         if (__ceph_have_pending_cap_snap(ci)) {
 104                 struct ceph_cap_snap *capsnap =
 105                                 list_last_entry(&ci->i_cap_snaps,
 106                                                 struct ceph_cap_snap,
 107                                                 ci_item);
 108                 snapc = ceph_get_snap_context(capsnap->context);
 109                 capsnap->dirty_pages++;
 110         } else {
 111                 BUG_ON(!ci->i_head_snapc);
 112                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 113                 ++ci->i_wrbuffer_ref_head;
 114         }
 115         if (ci->i_wrbuffer_ref == 0)
 116                 ihold(inode);
 117         ++ci->i_wrbuffer_ref;
 118         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 119              "snapc %p seq %lld (%d snaps)\n",
 120              mapping->host, page, page->index,
 121              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 122              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 123              snapc, snapc->seq, snapc->num_snaps);
 124         spin_unlock(&ci->i_ceph_lock);
 125
 126         /*
 127          * Reference snap context in page->private.  Also set
 128          * PagePrivate so that we get invalidatepage callback.
 129          */
 130         BUG_ON(PagePrivate(page));
 131         attach_page_private(page, snapc);
 132
 133         ret = __set_page_dirty_nobuffers(page);
 134         WARN_ON(!PageLocked(page));
 135         WARN_ON(!page->mapping);
 136
 137         return ret;
 138 }
 139
 140 /*
 141  * If we are truncating the full page (i.e. offset == 0), adjust the
 142  * dirty page counters appropriately.  Only called if there is private
 143  * data on the page.
 144  */
 145 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 146                                 unsigned int length)
 147 {
 148         struct inode *inode;
 149         struct ceph_inode_info *ci;
 150         struct ceph_snap_context *snapc;
 151
 152         wait_on_page_fscache(page);
 153
 154         inode = page->mapping->host;
 155         ci = ceph_inode(inode);
 156
 157         if (offset != 0 || length != thp_size(page)) {
 158                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 159                      inode, page, page->index, offset, length);
 160                 return;
 161         }
 162
 163         WARN_ON(!PageLocked(page));
 164         if (!PagePrivate(page))
 165                 return;
 166
 167         dout("%p invalidatepage %p idx %lu full dirty page\n",
 168              inode, page, page->index);
 169
 170         snapc = detach_page_private(page);
 171         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 172         ceph_put_snap_context(snapc);
 173 }
 174
 175 static int ceph_releasepage(struct page *page, gfp_t gfp)
 176 {
 177         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 178              page, page->index, PageDirty(page) ? "" : "not ");
 179
 180         if (PageFsCache(page)) {
 181                 if (!(gfp & __GFP_DIRECT_RECLAIM) || !(gfp & __GFP_FS))
 182                         return 0;
 183                 wait_on_page_fscache(page);
 184         }
 185         return !PagePrivate(page);
 186 }
 187
 188 static void ceph_netfs_expand_readahead(struct netfs_read_request *rreq)
 189 {
 190         struct inode *inode = rreq->mapping->host;
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_file_layout *lo = &ci->i_layout;
 193         u32 blockoff;
 194         u64 blockno;
 195
 196         /* Expand the start downward */
 197         blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 198         rreq->start = blockno * lo->stripe_unit;
 199         rreq->len += blockoff;
 200
 201         /* Now, round up the length to the next block */
 202         rreq->len = roundup(rreq->len, lo->stripe_unit);
 203 }
 204
 205 static bool ceph_netfs_clamp_length(struct netfs_read_subrequest *subreq)
 206 {
 207         struct inode *inode = subreq->rreq->mapping->host;
 208         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 209         struct ceph_inode_info *ci = ceph_inode(inode);
 210         u64 objno, objoff;
 211         u32 xlen;
 212
 213         /* Truncate the extent at the end of the current block */
 214         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 215                                       &objno, &objoff, &xlen);
 216         subreq->len = min(xlen, fsc->mount_options->rsize);
 217         return true;
 218 }
 219
 220 static void finish_netfs_read(struct ceph_osd_request *req)
 221 {
 222         struct ceph_fs_client *fsc = ceph_inode_to_client(req->r_inode);
 223         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 224         struct netfs_read_subrequest *subreq = req->r_priv;
 225         int num_pages;
 226         int err = req->r_result;
 227
 228         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 229                                  req->r_end_latency, err);
 230
 231         dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, req->r_result,
 232              subreq->len, i_size_read(req->r_inode));
 233
 234         /* no object means success but no data */
 235         if (err == -ENOENT)
 236                 err = 0;
 237         else if (err == -EBLOCKLISTED)
 238                 fsc->blocklisted = true;
 239
 240         if (err >= 0 && err < subreq->len)
 241                 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 242
 243         netfs_subreq_terminated(subreq, err, true);
 244
 245         num_pages = calc_pages_for(osd_data->alignment, osd_data->length);
 246         ceph_put_page_vector(osd_data->pages, num_pages, false);
 247         iput(req->r_inode);
 248 }
 249
 250 static void ceph_netfs_issue_op(struct netfs_read_subrequest *subreq)
 251 {
 252         struct netfs_read_request *rreq = subreq->rreq;
 253         struct inode *inode = rreq->mapping->host;
 254         struct ceph_inode_info *ci = ceph_inode(inode);
 255         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 256         struct ceph_osd_request *req;
 257         struct ceph_vino vino = ceph_vino(inode);
 258         struct iov_iter iter;
 259         struct page **pages;
 260         size_t page_off;
 261         int err = 0;
 262         u64 len = subreq->len;
 263
 264         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino, subreq->start, &len,
 265                         0, 1, CEPH_OSD_OP_READ,
 266                         CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
 267                         NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
 268         if (IS_ERR(req)) {
 269                 err = PTR_ERR(req);
 270                 req = NULL;
 271                 goto out;
 272         }
 273
 274         dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
 275         iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages, subreq->start, len);
 276         err = iov_iter_get_pages_alloc(&iter, &pages, len, &page_off);
 277         if (err < 0) {
 278                 dout("%s: iov_ter_get_pages_alloc returned %d\n", __func__, err);
 279                 goto out;
 280         }
 281
 282         /* should always give us a page-aligned read */
 283         WARN_ON_ONCE(page_off);
 284         len = err;
 285
 286         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 287         req->r_callback = finish_netfs_read;
 288         req->r_priv = subreq;
 289         req->r_inode = inode;
 290         ihold(inode);
 291
 292         err = ceph_osdc_start_request(req->r_osdc, req, false);
 293         if (err)
 294                 iput(inode);
 295 out:
 296         ceph_osdc_put_request(req);
 297         if (err)
 298                 netfs_subreq_terminated(subreq, err, false);
 299         dout("%s: result %d\n", __func__, err);
 300 }
 301
 302 static void ceph_init_rreq(struct netfs_read_request *rreq, struct file *file)
 303 {
 304 }
 305
 306 static void ceph_readahead_cleanup(struct address_space *mapping, void *priv)
 307 {
 308         struct inode *inode = mapping->host;
 309         struct ceph_inode_info *ci = ceph_inode(inode);
 310         int got = (uintptr_t)priv;
 311
 312         if (got)
 313                 ceph_put_cap_refs(ci, got);
 314 }
 315
 316 const struct netfs_read_request_ops ceph_netfs_read_ops = {
 317         .init_rreq              = ceph_init_rreq,
 318         .is_cache_enabled       = ceph_is_cache_enabled,
 319         .begin_cache_operation  = ceph_begin_cache_operation,
 320         .issue_op               = ceph_netfs_issue_op,
 321         .expand_readahead       = ceph_netfs_expand_readahead,
 322         .clamp_length           = ceph_netfs_clamp_length,
 323         .check_write_begin      = ceph_netfs_check_write_begin,
 324         .cleanup                = ceph_readahead_cleanup,
 325 };
 326
 327 /* read a single page, without unlocking it. */
 328 static int ceph_readpage(struct file *file, struct page *page)
 329 {
 330         struct inode *inode = file_inode(file);
 331         struct ceph_inode_info *ci = ceph_inode(inode);
 332         struct ceph_vino vino = ceph_vino(inode);
 333         u64 off = page_offset(page);
 334         u64 len = thp_size(page);
 335
 336         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 337                 /*
 338                  * Uptodate inline data should have been added
 339                  * into page cache while getting Fcr caps.
 340                  */
 341                 if (off == 0) {
 342                         unlock_page(page);
 343                         return -EINVAL;
 344                 }
 345                 zero_user_segment(page, 0, thp_size(page));
 346                 SetPageUptodate(page);
 347                 unlock_page(page);
 348                 return 0;
 349         }
 350
 351         dout("readpage ino %llx.%llx file %p off %llu len %llu page %p index %lu\n",
 352              vino.ino, vino.snap, file, off, len, page, page->index);
 353
 354         return netfs_readpage(file, page, &ceph_netfs_read_ops, NULL);
 355 }
 356
 357 static void ceph_readahead(struct readahead_control *ractl)
 358 {
 359         struct inode *inode = file_inode(ractl->file);
 360         struct ceph_file_info *fi = ractl->file->private_data;
 361         struct ceph_rw_context *rw_ctx;
 362         int got = 0;
 363         int ret = 0;
 364
 365         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 366                 return;
 367
 368         rw_ctx = ceph_find_rw_context(fi);
 369         if (!rw_ctx) {
 370                 /*
 371                  * readahead callers do not necessarily hold Fcb caps
 372                  * (e.g. fadvise, madvise).
 373                  */
 374                 int want = CEPH_CAP_FILE_CACHE;
 375
 376                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 377                 if (ret < 0)
 378                         dout("start_read %p, error getting cap\n", inode);
 379                 else if (!(got & want))
 380                         dout("start_read %p, no cache cap\n", inode);
 381
 382                 if (ret <= 0)
 383                         return;
 384         }
 385         netfs_readahead(ractl, &ceph_netfs_read_ops, (void *)(uintptr_t)got);
 386 }
 387
 388 struct ceph_writeback_ctl
 389 {
 390         loff_t i_size;
 391         u64 truncate_size;
 392         u32 truncate_seq;
 393         bool size_stable;
 394         bool head_snapc;
 395 };
 396
 397 /*
 398  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 399  * only snap context we are allowed to write back.
 400  */
 401 static struct ceph_snap_context *
 402 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 403                    struct ceph_snap_context *page_snapc)
 404 {
 405         struct ceph_inode_info *ci = ceph_inode(inode);
 406         struct ceph_snap_context *snapc = NULL;
 407         struct ceph_cap_snap *capsnap = NULL;
 408
 409         spin_lock(&ci->i_ceph_lock);
 410         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 411                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 412                      capsnap->context, capsnap->dirty_pages);
 413                 if (!capsnap->dirty_pages)
 414                         continue;
 415
 416                 /* get i_size, truncate_{seq,size} for page_snapc? */
 417                 if (snapc && capsnap->context != page_snapc)
 418                         continue;
 419
 420                 if (ctl) {
 421                         if (capsnap->writing) {
 422                                 ctl->i_size = i_size_read(inode);
 423                                 ctl->size_stable = false;
 424                         } else {
 425                                 ctl->i_size = capsnap->size;
 426                                 ctl->size_stable = true;
 427                         }
 428                         ctl->truncate_size = capsnap->truncate_size;
 429                         ctl->truncate_seq = capsnap->truncate_seq;
 430                         ctl->head_snapc = false;
 431                 }
 432
 433                 if (snapc)
 434                         break;
 435
 436                 snapc = ceph_get_snap_context(capsnap->context);
 437                 if (!page_snapc ||
 438                     page_snapc == snapc ||
 439                     page_snapc->seq > snapc->seq)
 440                         break;
 441         }
 442         if (!snapc && ci->i_wrbuffer_ref_head) {
 443                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 444                 dout(" head snapc %p has %d dirty pages\n",
 445                      snapc, ci->i_wrbuffer_ref_head);
 446                 if (ctl) {
 447                         ctl->i_size = i_size_read(inode);
 448                         ctl->truncate_size = ci->i_truncate_size;
 449                         ctl->truncate_seq = ci->i_truncate_seq;
 450                         ctl->size_stable = false;
 451                         ctl->head_snapc = true;
 452                 }
 453         }
 454         spin_unlock(&ci->i_ceph_lock);
 455         return snapc;
 456 }
 457
 458 static u64 get_writepages_data_length(struct inode *inode,
 459                                       struct page *page, u64 start)
 460 {
 461         struct ceph_inode_info *ci = ceph_inode(inode);
 462         struct ceph_snap_context *snapc = page_snap_context(page);
 463         struct ceph_cap_snap *capsnap = NULL;
 464         u64 end = i_size_read(inode);
 465
 466         if (snapc != ci->i_head_snapc) {
 467                 bool found = false;
 468                 spin_lock(&ci->i_ceph_lock);
 469                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 470                         if (capsnap->context == snapc) {
 471                                 if (!capsnap->writing)
 472                                         end = capsnap->size;
 473                                 found = true;
 474                                 break;
 475                         }
 476                 }
 477                 spin_unlock(&ci->i_ceph_lock);
 478                 WARN_ON(!found);
 479         }
 480         if (end > page_offset(page) + thp_size(page))
 481                 end = page_offset(page) + thp_size(page);
 482         return end > start ? end - start : 0;
 483 }
 484
 485 /*
 486  * Write a single page, but leave the page locked.
 487  *
 488  * If we get a write error, mark the mapping for error, but still adjust the
 489  * dirty page accounting (i.e., page is no longer dirty).
 490  */
 491 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 492 {
 493         struct inode *inode = page->mapping->host;
 494         struct ceph_inode_info *ci = ceph_inode(inode);
 495         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 496         struct ceph_snap_context *snapc, *oldest;
 497         loff_t page_off = page_offset(page);
 498         int err;
 499         loff_t len = thp_size(page);
 500         struct ceph_writeback_ctl ceph_wbc;
 501         struct ceph_osd_client *osdc = &fsc->client->osdc;
 502         struct ceph_osd_request *req;
 503
 504         dout("writepage %p idx %lu\n", page, page->index);
 505
 506         /* verify this is a writeable snap context */
 507         snapc = page_snap_context(page);
 508         if (!snapc) {
 509                 dout("writepage %p page %p not dirty?\n", inode, page);
 510                 return 0;
 511         }
 512         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 513         if (snapc->seq > oldest->seq) {
 514                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 515                      inode, page, snapc);
 516                 /* we should only noop if called by kswapd */
 517                 WARN_ON(!(current->flags & PF_MEMALLOC));
 518                 ceph_put_snap_context(oldest);
 519                 redirty_page_for_writepage(wbc, page);
 520                 return 0;
 521         }
 522         ceph_put_snap_context(oldest);
 523
 524         /* is this a partial page at end of file? */
 525         if (page_off >= ceph_wbc.i_size) {
 526                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 527                 page->mapping->a_ops->invalidatepage(page, 0, thp_size(page));
 528                 return 0;
 529         }
 530
 531         if (ceph_wbc.i_size < page_off + len)
 532                 len = ceph_wbc.i_size - page_off;
 533
 534         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 535              inode, page, page->index, page_off, len, snapc, snapc->seq);
 536
 537         if (atomic_long_inc_return(&fsc->writeback_count) >
 538             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 539                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 540
 541         set_page_writeback(page);
 542         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 543                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 544                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 545                                     true);
 546         if (IS_ERR(req)) {
 547                 redirty_page_for_writepage(wbc, page);
 548                 end_page_writeback(page);
 549                 return PTR_ERR(req);
 550         }
 551
 552         /* it may be a short write due to an object boundary */
 553         WARN_ON_ONCE(len > thp_size(page));
 554         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 555         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 556
 557         req->r_mtime = inode->i_mtime;
 558         err = ceph_osdc_start_request(osdc, req, true);
 559         if (!err)
 560                 err = ceph_osdc_wait_request(osdc, req);
 561
 562         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 563                                   req->r_end_latency, err);
 564
 565         ceph_osdc_put_request(req);
 566         if (err == 0)
 567                 err = len;
 568
 569         if (err < 0) {
 570                 struct writeback_control tmp_wbc;
 571                 if (!wbc)
 572                         wbc = &tmp_wbc;
 573                 if (err == -ERESTARTSYS) {
 574                         /* killed by SIGKILL */
 575                         dout("writepage interrupted page %p\n", page);
 576                         redirty_page_for_writepage(wbc, page);
 577                         end_page_writeback(page);
 578                         return err;
 579                 }
 580                 if (err == -EBLOCKLISTED)
 581                         fsc->blocklisted = true;
 582                 dout("writepage setting page/mapping error %d %p\n",
 583                      err, page);
 584                 mapping_set_error(&inode->i_data, err);
 585                 wbc->pages_skipped++;
 586         } else {
 587                 dout("writepage cleaned page %p\n", page);
 588                 err = 0;  /* vfs expects us to return 0 */
 589         }
 590         oldest = detach_page_private(page);
 591         WARN_ON_ONCE(oldest != snapc);
 592         end_page_writeback(page);
 593         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 594         ceph_put_snap_context(snapc);  /* page's reference */
 595
 596         if (atomic_long_dec_return(&fsc->writeback_count) <
 597             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 598                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 599
 600         return err;
 601 }
 602
 603 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 604 {
 605         int err;
 606         struct inode *inode = page->mapping->host;
 607         BUG_ON(!inode);
 608         ihold(inode);
 609         err = writepage_nounlock(page, wbc);
 610         if (err == -ERESTARTSYS) {
 611                 /* direct memory reclaimer was killed by SIGKILL. return 0
 612                  * to prevent caller from setting mapping/page error */
 613                 err = 0;
 614         }
 615         unlock_page(page);
 616         iput(inode);
 617         return err;
 618 }
 619
 620 /*
 621  * async writeback completion handler.
 622  *
 623  * If we get an error, set the mapping error bit, but not the individual
 624  * page error bits.
 625  */
 626 static void writepages_finish(struct ceph_osd_request *req)
 627 {
 628         struct inode *inode = req->r_inode;
 629         struct ceph_inode_info *ci = ceph_inode(inode);
 630         struct ceph_osd_data *osd_data;
 631         struct page *page;
 632         int num_pages, total_pages = 0;
 633         int i, j;
 634         int rc = req->r_result;
 635         struct ceph_snap_context *snapc = req->r_snapc;
 636         struct address_space *mapping = inode->i_mapping;
 637         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 638         bool remove_page;
 639
 640         dout("writepages_finish %p rc %d\n", inode, rc);
 641         if (rc < 0) {
 642                 mapping_set_error(mapping, rc);
 643                 ceph_set_error_write(ci);
 644                 if (rc == -EBLOCKLISTED)
 645                         fsc->blocklisted = true;
 646         } else {
 647                 ceph_clear_error_write(ci);
 648         }
 649
 650         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 651                                   req->r_end_latency, rc);
 652
 653         /*
 654          * We lost the cache cap, need to truncate the page before
 655          * it is unlocked, otherwise we'd truncate it later in the
 656          * page truncation thread, possibly losing some data that
 657          * raced its way in
 658          */
 659         remove_page = !(ceph_caps_issued(ci) &
 660                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 661
 662         /* clean all pages */
 663         for (i = 0; i < req->r_num_ops; i++) {
 664                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 665                         break;
 666
 667                 osd_data = osd_req_op_extent_osd_data(req, i);
 668                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 669                 num_pages = calc_pages_for((u64)osd_data->alignment,
 670                                            (u64)osd_data->length);
 671                 total_pages += num_pages;
 672                 for (j = 0; j < num_pages; j++) {
 673                         page = osd_data->pages[j];
 674                         BUG_ON(!page);
 675                         WARN_ON(!PageUptodate(page));
 676
 677                         if (atomic_long_dec_return(&fsc->writeback_count) <
 678                              CONGESTION_OFF_THRESH(
 679                                         fsc->mount_options->congestion_kb))
 680                                 clear_bdi_congested(inode_to_bdi(inode),
 681                                                     BLK_RW_ASYNC);
 682
 683                         ceph_put_snap_context(detach_page_private(page));
 684                         end_page_writeback(page);
 685                         dout("unlocking %p\n", page);
 686
 687                         if (remove_page)
 688                                 generic_error_remove_page(inode->i_mapping,
 689                                                           page);
 690
 691                         unlock_page(page);
 692                 }
 693                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 694                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 695
 696                 release_pages(osd_data->pages, num_pages);
 697         }
 698
 699         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 700
 701         osd_data = osd_req_op_extent_osd_data(req, 0);
 702         if (osd_data->pages_from_pool)
 703                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 704         else
 705                 kfree(osd_data->pages);
 706         ceph_osdc_put_request(req);
 707 }
 708
 709 /*
 710  * initiate async writeback
 711  */
 712 static int ceph_writepages_start(struct address_space *mapping,
 713                                  struct writeback_control *wbc)
 714 {
 715         struct inode *inode = mapping->host;
 716         struct ceph_inode_info *ci = ceph_inode(inode);
 717         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 718         struct ceph_vino vino = ceph_vino(inode);
 719         pgoff_t index, start_index, end = -1;
 720         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 721         struct pagevec pvec;
 722         int rc = 0;
 723         unsigned int wsize = i_blocksize(inode);
 724         struct ceph_osd_request *req = NULL;
 725         struct ceph_writeback_ctl ceph_wbc;
 726         bool should_loop, range_whole = false;
 727         bool done = false;
 728
 729         dout("writepages_start %p (mode=%s)\n", inode,
 730              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 731              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 732
 733         if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
 734                 if (ci->i_wrbuffer_ref > 0) {
 735                         pr_warn_ratelimited(
 736                                 "writepage_start %p %lld forced umount\n",
 737                                 inode, ceph_ino(inode));
 738                 }
 739                 mapping_set_error(mapping, -EIO);
 740                 return -EIO; /* we're in a forced umount, don't write! */
 741         }
 742         if (fsc->mount_options->wsize < wsize)
 743                 wsize = fsc->mount_options->wsize;
 744
 745         pagevec_init(&pvec);
 746
 747         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 748         index = start_index;
 749
 750 retry:
 751         /* find oldest snap context with dirty data */
 752         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 753         if (!snapc) {
 754                 /* hmm, why does writepages get called when there
 755                    is no dirty data? */
 756                 dout(" no snap context with dirty data?\n");
 757                 goto out;
 758         }
 759         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 760              snapc, snapc->seq, snapc->num_snaps);
 761
 762         should_loop = false;
 763         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 764                 /* where to start/end? */
 765                 if (wbc->range_cyclic) {
 766                         index = start_index;
 767                         end = -1;
 768                         if (index > 0)
 769                                 should_loop = true;
 770                         dout(" cyclic, start at %lu\n", index);
 771                 } else {
 772                         index = wbc->range_start >> PAGE_SHIFT;
 773                         end = wbc->range_end >> PAGE_SHIFT;
 774                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 775                                 range_whole = true;
 776                         dout(" not cyclic, %lu to %lu\n", index, end);
 777                 }
 778         } else if (!ceph_wbc.head_snapc) {
 779                 /* Do not respect wbc->range_{start,end}. Dirty pages
 780                  * in that range can be associated with newer snapc.
 781                  * They are not writeable until we write all dirty pages
 782                  * associated with 'snapc' get written */
 783                 if (index > 0)
 784                         should_loop = true;
 785                 dout(" non-head snapc, range whole\n");
 786         }
 787
 788         ceph_put_snap_context(last_snapc);
 789         last_snapc = snapc;
 790
 791         while (!done && index <= end) {
 792                 int num_ops = 0, op_idx;
 793                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 794                 struct page **pages = NULL, **data_pages;
 795                 struct page *page;
 796                 pgoff_t strip_unit_end = 0;
 797                 u64 offset = 0, len = 0;
 798                 bool from_pool = false;
 799
 800                 max_pages = wsize >> PAGE_SHIFT;
 801
 802 get_more_pages:
 803                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 804                                                 end, PAGECACHE_TAG_DIRTY);
 805                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 806                 if (!pvec_pages && !locked_pages)
 807                         break;
 808                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 809                         page = pvec.pages[i];
 810                         dout("? %p idx %lu\n", page, page->index);
 811                         if (locked_pages == 0)
 812                                 lock_page(page);  /* first page */
 813                         else if (!trylock_page(page))
 814                                 break;
 815
 816                         /* only dirty pages, or our accounting breaks */
 817                         if (unlikely(!PageDirty(page)) ||
 818                             unlikely(page->mapping != mapping)) {
 819                                 dout("!dirty or !mapping %p\n", page);
 820                                 unlock_page(page);
 821                                 continue;
 822                         }
 823                         /* only if matching snap context */
 824                         pgsnapc = page_snap_context(page);
 825                         if (pgsnapc != snapc) {
 826                                 dout("page snapc %p %lld != oldest %p %lld\n",
 827                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 828                                 if (!should_loop &&
 829                                     !ceph_wbc.head_snapc &&
 830                                     wbc->sync_mode != WB_SYNC_NONE)
 831                                         should_loop = true;
 832                                 unlock_page(page);
 833                                 continue;
 834                         }
 835                         if (page_offset(page) >= ceph_wbc.i_size) {
 836                                 dout("%p page eof %llu\n",
 837                                      page, ceph_wbc.i_size);
 838                                 if ((ceph_wbc.size_stable ||
 839                                     page_offset(page) >= i_size_read(inode)) &&
 840                                     clear_page_dirty_for_io(page))
 841                                         mapping->a_ops->invalidatepage(page,
 842                                                                 0, thp_size(page));
 843                                 unlock_page(page);
 844                                 continue;
 845                         }
 846                         if (strip_unit_end && (page->index > strip_unit_end)) {
 847                                 dout("end of strip unit %p\n", page);
 848                                 unlock_page(page);
 849                                 break;
 850                         }
 851                         if (PageWriteback(page)) {
 852                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 853                                         dout("%p under writeback\n", page);
 854                                         unlock_page(page);
 855                                         continue;
 856                                 }
 857                                 dout("waiting on writeback %p\n", page);
 858                                 wait_on_page_writeback(page);
 859                         }
 860
 861                         if (!clear_page_dirty_for_io(page)) {
 862                                 dout("%p !clear_page_dirty_for_io\n", page);
 863                                 unlock_page(page);
 864                                 continue;
 865                         }
 866
 867                         /*
 868                          * We have something to write.  If this is
 869                          * the first locked page this time through,
 870                          * calculate max possinle write size and
 871                          * allocate a page array
 872                          */
 873                         if (locked_pages == 0) {
 874                                 u64 objnum;
 875                                 u64 objoff;
 876                                 u32 xlen;
 877
 878                                 /* prepare async write request */
 879                                 offset = (u64)page_offset(page);
 880                                 ceph_calc_file_object_mapping(&ci->i_layout,
 881                                                               offset, wsize,
 882                                                               &objnum, &objoff,
 883                                                               &xlen);
 884                                 len = xlen;
 885
 886                                 num_ops = 1;
 887                                 strip_unit_end = page->index +
 888                                         ((len - 1) >> PAGE_SHIFT);
 889
 890                                 BUG_ON(pages);
 891                                 max_pages = calc_pages_for(0, (u64)len);
 892                                 pages = kmalloc_array(max_pages,
 893                                                       sizeof(*pages),
 894                                                       GFP_NOFS);
 895                                 if (!pages) {
 896                                         from_pool = true;
 897                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
 898                                         BUG_ON(!pages);
 899                                 }
 900
 901                                 len = 0;
 902                         } else if (page->index !=
 903                                    (offset + len) >> PAGE_SHIFT) {
 904                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
 905                                                              CEPH_OSD_MAX_OPS)) {
 906                                         redirty_page_for_writepage(wbc, page);
 907                                         unlock_page(page);
 908                                         break;
 909                                 }
 910
 911                                 num_ops++;
 912                                 offset = (u64)page_offset(page);
 913                                 len = 0;
 914                         }
 915
 916                         /* note position of first page in pvec */
 917                         dout("%p will write page %p idx %lu\n",
 918                              inode, page, page->index);
 919
 920                         if (atomic_long_inc_return(&fsc->writeback_count) >
 921                             CONGESTION_ON_THRESH(
 922                                     fsc->mount_options->congestion_kb)) {
 923                                 set_bdi_congested(inode_to_bdi(inode),
 924                                                   BLK_RW_ASYNC);
 925                         }
 926
 927
 928                         pages[locked_pages++] = page;
 929                         pvec.pages[i] = NULL;
 930
 931                         len += thp_size(page);
 932                 }
 933
 934                 /* did we get anything? */
 935                 if (!locked_pages)
 936                         goto release_pvec_pages;
 937                 if (i) {
 938                         unsigned j, n = 0;
 939                         /* shift unused page to beginning of pvec */
 940                         for (j = 0; j < pvec_pages; j++) {
 941                                 if (!pvec.pages[j])
 942                                         continue;
 943                                 if (n < j)
 944                                         pvec.pages[n] = pvec.pages[j];
 945                                 n++;
 946                         }
 947                         pvec.nr = n;
 948
 949                         if (pvec_pages && i == pvec_pages &&
 950                             locked_pages < max_pages) {
 951                                 dout("reached end pvec, trying for more\n");
 952                                 pagevec_release(&pvec);
 953                                 goto get_more_pages;
 954                         }
 955                 }
 956
 957 new_request:
 958                 offset = page_offset(pages[0]);
 959                 len = wsize;
 960
 961                 req = ceph_osdc_new_request(&fsc->client->osdc,
 962                                         &ci->i_layout, vino,
 963                                         offset, &len, 0, num_ops,
 964                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
 965                                         snapc, ceph_wbc.truncate_seq,
 966                                         ceph_wbc.truncate_size, false);
 967                 if (IS_ERR(req)) {
 968                         req = ceph_osdc_new_request(&fsc->client->osdc,
 969                                                 &ci->i_layout, vino,
 970                                                 offset, &len, 0,
 971                                                 min(num_ops,
 972                                                     CEPH_OSD_SLAB_OPS),
 973                                                 CEPH_OSD_OP_WRITE,
 974                                                 CEPH_OSD_FLAG_WRITE,
 975                                                 snapc, ceph_wbc.truncate_seq,
 976                                                 ceph_wbc.truncate_size, true);
 977                         BUG_ON(IS_ERR(req));
 978                 }
 979                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
 980                              thp_size(page) - offset);
 981
 982                 req->r_callback = writepages_finish;
 983                 req->r_inode = inode;
 984
 985                 /* Format the osd request message and submit the write */
 986                 len = 0;
 987                 data_pages = pages;
 988                 op_idx = 0;
 989                 for (i = 0; i < locked_pages; i++) {
 990                         u64 cur_offset = page_offset(pages[i]);
 991                         if (offset + len != cur_offset) {
 992                                 if (op_idx + 1 == req->r_num_ops)
 993                                         break;
 994                                 osd_req_op_extent_dup_last(req, op_idx,
 995                                                            cur_offset - offset);
 996                                 dout("writepages got pages at %llu~%llu\n",
 997                                      offset, len);
 998                                 osd_req_op_extent_osd_data_pages(req, op_idx,
 999                                                         data_pages, len, 0,
1000                                                         from_pool, false);
1001                                 osd_req_op_extent_update(req, op_idx, len);
1002
1003                                 len = 0;
1004                                 offset = cur_offset;
1005                                 data_pages = pages + i;
1006                                 op_idx++;
1007                         }
1008
1009                         set_page_writeback(pages[i]);
1010                         len += thp_size(page);
1011                 }
1012
1013                 if (ceph_wbc.size_stable) {
1014                         len = min(len, ceph_wbc.i_size - offset);
1015                 } else if (i == locked_pages) {
1016                         /* writepages_finish() clears writeback pages
1017                          * according to the data length, so make sure
1018                          * data length covers all locked pages */
1019                         u64 min_len = len + 1 - thp_size(page);
1020                         len = get_writepages_data_length(inode, pages[i - 1],
1021                                                          offset);
1022                         len = max(len, min_len);
1023                 }
1024                 dout("writepages got pages at %llu~%llu\n", offset, len);
1025
1026                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1027                                                  0, from_pool, false);
1028                 osd_req_op_extent_update(req, op_idx, len);
1029
1030                 BUG_ON(op_idx + 1 != req->r_num_ops);
1031
1032                 from_pool = false;
1033                 if (i < locked_pages) {
1034                         BUG_ON(num_ops <= req->r_num_ops);
1035                         num_ops -= req->r_num_ops;
1036                         locked_pages -= i;
1037
1038                         /* allocate new pages array for next request */
1039                         data_pages = pages;
1040                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1041                                               GFP_NOFS);
1042                         if (!pages) {
1043                                 from_pool = true;
1044                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1045                                 BUG_ON(!pages);
1046                         }
1047                         memcpy(pages, data_pages + i,
1048                                locked_pages * sizeof(*pages));
1049                         memset(data_pages + i, 0,
1050                                locked_pages * sizeof(*pages));
1051                 } else {
1052                         BUG_ON(num_ops != req->r_num_ops);
1053                         index = pages[i - 1]->index + 1;
1054                         /* request message now owns the pages array */
1055                         pages = NULL;
1056                 }
1057
1058                 req->r_mtime = inode->i_mtime;
1059                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1060                 BUG_ON(rc);
1061                 req = NULL;
1062
1063                 wbc->nr_to_write -= i;
1064                 if (pages)
1065                         goto new_request;
1066
1067                 /*
1068                  * We stop writing back only if we are not doing
1069                  * integrity sync. In case of integrity sync we have to
1070                  * keep going until we have written all the pages
1071                  * we tagged for writeback prior to entering this loop.
1072                  */
1073                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1074                         done = true;
1075
1076 release_pvec_pages:
1077                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1078                      pvec.nr ? pvec.pages[0] : NULL);
1079                 pagevec_release(&pvec);
1080         }
1081
1082         if (should_loop && !done) {
1083                 /* more to do; loop back to beginning of file */
1084                 dout("writepages looping back to beginning of file\n");
1085                 end = start_index - 1; /* OK even when start_index == 0 */
1086
1087                 /* to write dirty pages associated with next snapc,
1088                  * we need to wait until current writes complete */
1089                 if (wbc->sync_mode != WB_SYNC_NONE &&
1090                     start_index == 0 && /* all dirty pages were checked */
1091                     !ceph_wbc.head_snapc) {
1092                         struct page *page;
1093                         unsigned i, nr;
1094                         index = 0;
1095                         while ((index <= end) &&
1096                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1097                                                 PAGECACHE_TAG_WRITEBACK))) {
1098                                 for (i = 0; i < nr; i++) {
1099                                         page = pvec.pages[i];
1100                                         if (page_snap_context(page) != snapc)
1101                                                 continue;
1102                                         wait_on_page_writeback(page);
1103                                 }
1104                                 pagevec_release(&pvec);
1105                                 cond_resched();
1106                         }
1107                 }
1108
1109                 start_index = 0;
1110                 index = 0;
1111                 goto retry;
1112         }
1113
1114         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1115                 mapping->writeback_index = index;
1116
1117 out:
1118         ceph_osdc_put_request(req);
1119         ceph_put_snap_context(last_snapc);
1120         dout("writepages dend - startone, rc = %d\n", rc);
1121         return rc;
1122 }
1123
1124
1125
1126 /*
1127  * See if a given @snapc is either writeable, or already written.
1128  */
1129 static int context_is_writeable_or_written(struct inode *inode,
1130                                            struct ceph_snap_context *snapc)
1131 {
1132         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1133         int ret = !oldest || snapc->seq <= oldest->seq;
1134
1135         ceph_put_snap_context(oldest);
1136         return ret;
1137 }
1138
1139 /**
1140  * ceph_find_incompatible - find an incompatible context and return it
1141  * @page: page being dirtied
1142  *
1143  * We are only allowed to write into/dirty a page if the page is
1144  * clean, or already dirty within the same snap context. Returns a
1145  * conflicting context if there is one, NULL if there isn't, or a
1146  * negative error code on other errors.
1147  *
1148  * Must be called with page lock held.
1149  */
1150 static struct ceph_snap_context *
1151 ceph_find_incompatible(struct page *page)
1152 {
1153         struct inode *inode = page->mapping->host;
1154         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1155         struct ceph_inode_info *ci = ceph_inode(inode);
1156
1157         if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1158                 dout(" page %p forced umount\n", page);
1159                 return ERR_PTR(-EIO);
1160         }
1161
1162         for (;;) {
1163                 struct ceph_snap_context *snapc, *oldest;
1164
1165                 wait_on_page_writeback(page);
1166
1167                 snapc = page_snap_context(page);
1168                 if (!snapc || snapc == ci->i_head_snapc)
1169                         break;
1170
1171                 /*
1172                  * this page is already dirty in another (older) snap
1173                  * context!  is it writeable now?
1174                  */
1175                 oldest = get_oldest_context(inode, NULL, NULL);
1176                 if (snapc->seq > oldest->seq) {
1177                         /* not writeable -- return it for the caller to deal with */
1178                         ceph_put_snap_context(oldest);
1179                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1180                         return ceph_get_snap_context(snapc);
1181                 }
1182                 ceph_put_snap_context(oldest);
1183
1184                 /* yay, writeable, do it now (without dropping page lock) */
1185                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1186                 if (clear_page_dirty_for_io(page)) {
1187                         int r = writepage_nounlock(page, NULL);
1188                         if (r < 0)
1189                                 return ERR_PTR(r);
1190                 }
1191         }
1192         return NULL;
1193 }
1194
1195 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1196                                         struct page *page, void **_fsdata)
1197 {
1198         struct inode *inode = file_inode(file);
1199         struct ceph_inode_info *ci = ceph_inode(inode);
1200         struct ceph_snap_context *snapc;
1201
1202         snapc = ceph_find_incompatible(page);
1203         if (snapc) {
1204                 int r;
1205
1206                 unlock_page(page);
1207                 put_page(page);
1208                 if (IS_ERR(snapc))
1209                         return PTR_ERR(snapc);
1210
1211                 ceph_queue_writeback(inode);
1212                 r = wait_event_killable(ci->i_cap_wq,
1213                                         context_is_writeable_or_written(inode, snapc));
1214                 ceph_put_snap_context(snapc);
1215                 return r == 0 ? -EAGAIN : r;
1216         }
1217         return 0;
1218 }
1219
1220 /*
1221  * We are only allowed to write into/dirty the page if the page is
1222  * clean, or already dirty within the same snap context.
1223  */
1224 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1225                             loff_t pos, unsigned len, unsigned flags,
1226                             struct page **pagep, void **fsdata)
1227 {
1228         struct inode *inode = file_inode(file);
1229         struct ceph_inode_info *ci = ceph_inode(inode);
1230         struct page *page = NULL;
1231         pgoff_t index = pos >> PAGE_SHIFT;
1232         int r;
1233
1234         /*
1235          * Uninlining should have already been done and everything updated, EXCEPT
1236          * for inline_version sent to the MDS.
1237          */
1238         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1239                 page = grab_cache_page_write_begin(mapping, index, flags);
1240                 if (!page)
1241                         return -ENOMEM;
1242
1243                 /*
1244                  * The inline_version on a new inode is set to 1. If that's the
1245                  * case, then the page is brand new and isn't yet Uptodate.
1246                  */
1247                 r = 0;
1248                 if (index == 0 && ci->i_inline_version != 1) {
1249                         if (!PageUptodate(page)) {
1250                                 WARN_ONCE(1, "ceph: write_begin called on still-inlined inode (inline_version %llu)!\n",
1251                                           ci->i_inline_version);
1252                                 r = -EINVAL;
1253                         }
1254                         goto out;
1255                 }
1256                 zero_user_segment(page, 0, thp_size(page));
1257                 SetPageUptodate(page);
1258                 goto out;
1259         }
1260
1261         r = netfs_write_begin(file, inode->i_mapping, pos, len, 0, &page, NULL,
1262                               &ceph_netfs_read_ops, NULL);
1263 out:
1264         if (r == 0)
1265                 wait_on_page_fscache(page);
1266         if (r < 0) {
1267                 if (page)
1268                         put_page(page);
1269         } else {
1270                 WARN_ON_ONCE(!PageLocked(page));
1271                 *pagep = page;
1272         }
1273         return r;
1274 }
1275
1276 /*
1277  * we don't do anything in here that simple_write_end doesn't do
1278  * except adjust dirty page accounting
1279  */
1280 static int ceph_write_end(struct file *file, struct address_space *mapping,
1281                           loff_t pos, unsigned len, unsigned copied,
1282                           struct page *page, void *fsdata)
1283 {
1284         struct inode *inode = file_inode(file);
1285         bool check_cap = false;
1286
1287         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1288              inode, page, (int)pos, (int)copied, (int)len);
1289
1290         /* zero the stale part of the page if we did a short copy */
1291         if (!PageUptodate(page)) {
1292                 if (copied < len) {
1293                         copied = 0;
1294                         goto out;
1295                 }
1296                 SetPageUptodate(page);
1297         }
1298
1299         /* did file size increase? */
1300         if (pos+copied > i_size_read(inode))
1301                 check_cap = ceph_inode_set_size(inode, pos+copied);
1302
1303         set_page_dirty(page);
1304
1305 out:
1306         unlock_page(page);
1307         put_page(page);
1308
1309         if (check_cap)
1310                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1311
1312         return copied;
1313 }
1314
1315 /*
1316  * we set .direct_IO to indicate direct io is supported, but since we
1317  * intercept O_DIRECT reads and writes early, this function should
1318  * never get called.
1319  */
1320 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1321 {
1322         WARN_ON(1);
1323         return -EINVAL;
1324 }
1325
1326 const struct address_space_operations ceph_aops = {
1327         .readpage = ceph_readpage,
1328         .readahead = ceph_readahead,
1329         .writepage = ceph_writepage,
1330         .writepages = ceph_writepages_start,
1331         .write_begin = ceph_write_begin,
1332         .write_end = ceph_write_end,
1333         .set_page_dirty = ceph_set_page_dirty,
1334         .invalidatepage = ceph_invalidatepage,
1335         .releasepage = ceph_releasepage,
1336         .direct_IO = ceph_direct_io,
1337 };
1338
1339 static void ceph_block_sigs(sigset_t *oldset)
1340 {
1341         sigset_t mask;
1342         siginitsetinv(&mask, sigmask(SIGKILL));
1343         sigprocmask(SIG_BLOCK, &mask, oldset);
1344 }
1345
1346 static void ceph_restore_sigs(sigset_t *oldset)
1347 {
1348         sigprocmask(SIG_SETMASK, oldset, NULL);
1349 }
1350
1351 /*
1352  * vm ops
1353  */
1354 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1355 {
1356         struct vm_area_struct *vma = vmf->vma;
1357         struct inode *inode = file_inode(vma->vm_file);
1358         struct ceph_inode_info *ci = ceph_inode(inode);
1359         struct ceph_file_info *fi = vma->vm_file->private_data;
1360         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1361         int want, got, err;
1362         sigset_t oldset;
1363         vm_fault_t ret = VM_FAULT_SIGBUS;
1364
1365         ceph_block_sigs(&oldset);
1366
1367         dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1368              inode, ceph_vinop(inode), off);
1369         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1370                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1371         else
1372                 want = CEPH_CAP_FILE_CACHE;
1373
1374         got = 0;
1375         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1376         if (err < 0)
1377                 goto out_restore;
1378
1379         dout("filemap_fault %p %llu got cap refs on %s\n",
1380              inode, off, ceph_cap_string(got));
1381
1382         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1383             ci->i_inline_version == CEPH_INLINE_NONE) {
1384                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1385                 ceph_add_rw_context(fi, &rw_ctx);
1386                 ret = filemap_fault(vmf);
1387                 ceph_del_rw_context(fi, &rw_ctx);
1388                 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1389                      inode, off, ceph_cap_string(got), ret);
1390         } else
1391                 err = -EAGAIN;
1392
1393         ceph_put_cap_refs(ci, got);
1394
1395         if (err != -EAGAIN)
1396                 goto out_restore;
1397
1398         /* read inline data */
1399         if (off >= PAGE_SIZE) {
1400                 /* does not support inline data > PAGE_SIZE */
1401                 ret = VM_FAULT_SIGBUS;
1402         } else {
1403                 struct address_space *mapping = inode->i_mapping;
1404                 struct page *page = find_or_create_page(mapping, 0,
1405                                                 mapping_gfp_constraint(mapping,
1406                                                 ~__GFP_FS));
1407                 if (!page) {
1408                         ret = VM_FAULT_OOM;
1409                         goto out_inline;
1410                 }
1411                 err = __ceph_do_getattr(inode, page,
1412                                          CEPH_STAT_CAP_INLINE_DATA, true);
1413                 if (err < 0 || off >= i_size_read(inode)) {
1414                         unlock_page(page);
1415                         put_page(page);
1416                         ret = vmf_error(err);
1417                         goto out_inline;
1418                 }
1419                 if (err < PAGE_SIZE)
1420                         zero_user_segment(page, err, PAGE_SIZE);
1421                 else
1422                         flush_dcache_page(page);
1423                 SetPageUptodate(page);
1424                 vmf->page = page;
1425                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1426 out_inline:
1427                 dout("filemap_fault %p %llu read inline data ret %x\n",
1428                      inode, off, ret);
1429         }
1430 out_restore:
1431         ceph_restore_sigs(&oldset);
1432         if (err < 0)
1433                 ret = vmf_error(err);
1434
1435         return ret;
1436 }
1437
1438 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1439 {
1440         struct vm_area_struct *vma = vmf->vma;
1441         struct inode *inode = file_inode(vma->vm_file);
1442         struct ceph_inode_info *ci = ceph_inode(inode);
1443         struct ceph_file_info *fi = vma->vm_file->private_data;
1444         struct ceph_cap_flush *prealloc_cf;
1445         struct page *page = vmf->page;
1446         loff_t off = page_offset(page);
1447         loff_t size = i_size_read(inode);
1448         size_t len;
1449         int want, got, err;
1450         sigset_t oldset;
1451         vm_fault_t ret = VM_FAULT_SIGBUS;
1452
1453         prealloc_cf = ceph_alloc_cap_flush();
1454         if (!prealloc_cf)
1455                 return VM_FAULT_OOM;
1456
1457         sb_start_pagefault(inode->i_sb);
1458         ceph_block_sigs(&oldset);
1459
1460         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1461                 struct page *locked_page = NULL;
1462                 if (off == 0) {
1463                         lock_page(page);
1464                         locked_page = page;
1465                 }
1466                 err = ceph_uninline_data(vma->vm_file, locked_page);
1467                 if (locked_page)
1468                         unlock_page(locked_page);
1469                 if (err < 0)
1470                         goto out_free;
1471         }
1472
1473         if (off + thp_size(page) <= size)
1474                 len = thp_size(page);
1475         else
1476                 len = offset_in_thp(page, size);
1477
1478         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1479              inode, ceph_vinop(inode), off, len, size);
1480         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1481                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1482         else
1483                 want = CEPH_CAP_FILE_BUFFER;
1484
1485         got = 0;
1486         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1487         if (err < 0)
1488                 goto out_free;
1489
1490         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1491              inode, off, len, ceph_cap_string(got));
1492
1493         /* Update time before taking page lock */
1494         file_update_time(vma->vm_file);
1495         inode_inc_iversion_raw(inode);
1496
1497         do {
1498                 struct ceph_snap_context *snapc;
1499
1500                 lock_page(page);
1501
1502                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1503                         unlock_page(page);
1504                         ret = VM_FAULT_NOPAGE;
1505                         break;
1506                 }
1507
1508                 snapc = ceph_find_incompatible(page);
1509                 if (!snapc) {
1510                         /* success.  we'll keep the page locked. */
1511                         set_page_dirty(page);
1512                         ret = VM_FAULT_LOCKED;
1513                         break;
1514                 }
1515
1516                 unlock_page(page);
1517
1518                 if (IS_ERR(snapc)) {
1519                         ret = VM_FAULT_SIGBUS;
1520                         break;
1521                 }
1522
1523                 ceph_queue_writeback(inode);
1524                 err = wait_event_killable(ci->i_cap_wq,
1525                                 context_is_writeable_or_written(inode, snapc));
1526                 ceph_put_snap_context(snapc);
1527         } while (err == 0);
1528
1529         if (ret == VM_FAULT_LOCKED ||
1530             ci->i_inline_version != CEPH_INLINE_NONE) {
1531                 int dirty;
1532                 spin_lock(&ci->i_ceph_lock);
1533                 ci->i_inline_version = CEPH_INLINE_NONE;
1534                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1535                                                &prealloc_cf);
1536                 spin_unlock(&ci->i_ceph_lock);
1537                 if (dirty)
1538                         __mark_inode_dirty(inode, dirty);
1539         }
1540
1541         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1542              inode, off, len, ceph_cap_string(got), ret);
1543         ceph_put_cap_refs_async(ci, got);
1544 out_free:
1545         ceph_restore_sigs(&oldset);
1546         sb_end_pagefault(inode->i_sb);
1547         ceph_free_cap_flush(prealloc_cf);
1548         if (err < 0)
1549                 ret = vmf_error(err);
1550         return ret;
1551 }
1552
1553 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1554                            char *data, size_t len)
1555 {
1556         struct address_space *mapping = inode->i_mapping;
1557         struct page *page;
1558
1559         if (locked_page) {
1560                 page = locked_page;
1561         } else {
1562                 if (i_size_read(inode) == 0)
1563                         return;
1564                 page = find_or_create_page(mapping, 0,
1565                                            mapping_gfp_constraint(mapping,
1566                                            ~__GFP_FS));
1567                 if (!page)
1568                         return;
1569                 if (PageUptodate(page)) {
1570                         unlock_page(page);
1571                         put_page(page);
1572                         return;
1573                 }
1574         }
1575
1576         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1577              inode, ceph_vinop(inode), len, locked_page);
1578
1579         if (len > 0) {
1580                 void *kaddr = kmap_atomic(page);
1581                 memcpy(kaddr, data, len);
1582                 kunmap_atomic(kaddr);
1583         }
1584
1585         if (page != locked_page) {
1586                 if (len < PAGE_SIZE)
1587                         zero_user_segment(page, len, PAGE_SIZE);
1588                 else
1589                         flush_dcache_page(page);
1590
1591                 SetPageUptodate(page);
1592                 unlock_page(page);
1593                 put_page(page);
1594         }
1595 }
1596
1597 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1598 {
1599         struct inode *inode = file_inode(filp);
1600         struct ceph_inode_info *ci = ceph_inode(inode);
1601         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1602         struct ceph_osd_request *req;
1603         struct page *page = NULL;
1604         u64 len, inline_version;
1605         int err = 0;
1606         bool from_pagecache = false;
1607
1608         spin_lock(&ci->i_ceph_lock);
1609         inline_version = ci->i_inline_version;
1610         spin_unlock(&ci->i_ceph_lock);
1611
1612         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1613              inode, ceph_vinop(inode), inline_version);
1614
1615         if (inline_version == 1 || /* initial version, no data */
1616             inline_version == CEPH_INLINE_NONE)
1617                 goto out;
1618
1619         if (locked_page) {
1620                 page = locked_page;
1621                 WARN_ON(!PageUptodate(page));
1622         } else if (ceph_caps_issued(ci) &
1623                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1624                 page = find_get_page(inode->i_mapping, 0);
1625                 if (page) {
1626                         if (PageUptodate(page)) {
1627                                 from_pagecache = true;
1628                                 lock_page(page);
1629                         } else {
1630                                 put_page(page);
1631                                 page = NULL;
1632                         }
1633                 }
1634         }
1635
1636         if (page) {
1637                 len = i_size_read(inode);
1638                 if (len > PAGE_SIZE)
1639                         len = PAGE_SIZE;
1640         } else {
1641                 page = __page_cache_alloc(GFP_NOFS);
1642                 if (!page) {
1643                         err = -ENOMEM;
1644                         goto out;
1645                 }
1646                 err = __ceph_do_getattr(inode, page,
1647                                         CEPH_STAT_CAP_INLINE_DATA, true);
1648                 if (err < 0) {
1649                         /* no inline data */
1650                         if (err == -ENODATA)
1651                                 err = 0;
1652                         goto out;
1653                 }
1654                 len = err;
1655         }
1656
1657         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1658                                     ceph_vino(inode), 0, &len, 0, 1,
1659                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1660                                     NULL, 0, 0, false);
1661         if (IS_ERR(req)) {
1662                 err = PTR_ERR(req);
1663                 goto out;
1664         }
1665
1666         req->r_mtime = inode->i_mtime;
1667         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1668         if (!err)
1669                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1670         ceph_osdc_put_request(req);
1671         if (err < 0)
1672                 goto out;
1673
1674         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1675                                     ceph_vino(inode), 0, &len, 1, 3,
1676                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1677                                     NULL, ci->i_truncate_seq,
1678                                     ci->i_truncate_size, false);
1679         if (IS_ERR(req)) {
1680                 err = PTR_ERR(req);
1681                 goto out;
1682         }
1683
1684         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1685
1686         {
1687                 __le64 xattr_buf = cpu_to_le64(inline_version);
1688                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1689                                             "inline_version", &xattr_buf,
1690                                             sizeof(xattr_buf),
1691                                             CEPH_OSD_CMPXATTR_OP_GT,
1692                                             CEPH_OSD_CMPXATTR_MODE_U64);
1693                 if (err)
1694                         goto out_put;
1695         }
1696
1697         {
1698                 char xattr_buf[32];
1699                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1700                                          "%llu", inline_version);
1701                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1702                                             "inline_version",
1703                                             xattr_buf, xattr_len, 0, 0);
1704                 if (err)
1705                         goto out_put;
1706         }
1707
1708         req->r_mtime = inode->i_mtime;
1709         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1710         if (!err)
1711                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1712
1713         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1714                                   req->r_end_latency, err);
1715
1716 out_put:
1717         ceph_osdc_put_request(req);
1718         if (err == -ECANCELED)
1719                 err = 0;
1720 out:
1721         if (page && page != locked_page) {
1722                 if (from_pagecache) {
1723                         unlock_page(page);
1724                         put_page(page);
1725                 } else
1726                         __free_pages(page, 0);
1727         }
1728
1729         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1730              inode, ceph_vinop(inode), inline_version, err);
1731         return err;
1732 }
1733
1734 static const struct vm_operations_struct ceph_vmops = {
1735         .fault          = ceph_filemap_fault,
1736         .page_mkwrite   = ceph_page_mkwrite,
1737 };
1738
1739 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1740 {
1741         struct address_space *mapping = file->f_mapping;
1742
1743         if (!mapping->a_ops->readpage)
1744                 return -ENOEXEC;
1745         file_accessed(file);
1746         vma->vm_ops = &ceph_vmops;
1747         return 0;
1748 }
1749
1750 enum {
1751         POOL_READ       = 1,
1752         POOL_WRITE      = 2,
1753 };
1754
1755 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1756                                 s64 pool, struct ceph_string *pool_ns)
1757 {
1758         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1759         struct ceph_mds_client *mdsc = fsc->mdsc;
1760         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1761         struct rb_node **p, *parent;
1762         struct ceph_pool_perm *perm;
1763         struct page **pages;
1764         size_t pool_ns_len;
1765         int err = 0, err2 = 0, have = 0;
1766
1767         down_read(&mdsc->pool_perm_rwsem);
1768         p = &mdsc->pool_perm_tree.rb_node;
1769         while (*p) {
1770                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1771                 if (pool < perm->pool)
1772                         p = &(*p)->rb_left;
1773                 else if (pool > perm->pool)
1774                         p = &(*p)->rb_right;
1775                 else {
1776                         int ret = ceph_compare_string(pool_ns,
1777                                                 perm->pool_ns,
1778                                                 perm->pool_ns_len);
1779                         if (ret < 0)
1780                                 p = &(*p)->rb_left;
1781                         else if (ret > 0)
1782                                 p = &(*p)->rb_right;
1783                         else {
1784                                 have = perm->perm;
1785                                 break;
1786                         }
1787                 }
1788         }
1789         up_read(&mdsc->pool_perm_rwsem);
1790         if (*p)
1791                 goto out;
1792
1793         if (pool_ns)
1794                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1795                      pool, (int)pool_ns->len, pool_ns->str);
1796         else
1797                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1798
1799         down_write(&mdsc->pool_perm_rwsem);
1800         p = &mdsc->pool_perm_tree.rb_node;
1801         parent = NULL;
1802         while (*p) {
1803                 parent = *p;
1804                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1805                 if (pool < perm->pool)
1806                         p = &(*p)->rb_left;
1807                 else if (pool > perm->pool)
1808                         p = &(*p)->rb_right;
1809                 else {
1810                         int ret = ceph_compare_string(pool_ns,
1811                                                 perm->pool_ns,
1812                                                 perm->pool_ns_len);
1813                         if (ret < 0)
1814                                 p = &(*p)->rb_left;
1815                         else if (ret > 0)
1816                                 p = &(*p)->rb_right;
1817                         else {
1818                                 have = perm->perm;
1819                                 break;
1820                         }
1821                 }
1822         }
1823         if (*p) {
1824                 up_write(&mdsc->pool_perm_rwsem);
1825                 goto out;
1826         }
1827
1828         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1829                                          1, false, GFP_NOFS);
1830         if (!rd_req) {
1831                 err = -ENOMEM;
1832                 goto out_unlock;
1833         }
1834
1835         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1836         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1837         rd_req->r_base_oloc.pool = pool;
1838         if (pool_ns)
1839                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1840         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1841
1842         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1843         if (err)
1844                 goto out_unlock;
1845
1846         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1847                                          1, false, GFP_NOFS);
1848         if (!wr_req) {
1849                 err = -ENOMEM;
1850                 goto out_unlock;
1851         }
1852
1853         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1854         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1855         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1856         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1857
1858         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1859         if (err)
1860                 goto out_unlock;
1861
1862         /* one page should be large enough for STAT data */
1863         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1864         if (IS_ERR(pages)) {
1865                 err = PTR_ERR(pages);
1866                 goto out_unlock;
1867         }
1868
1869         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1870                                      0, false, true);
1871         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1872
1873         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1874         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1875
1876         if (!err)
1877                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1878         if (!err2)
1879                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1880
1881         if (err >= 0 || err == -ENOENT)
1882                 have |= POOL_READ;
1883         else if (err != -EPERM) {
1884                 if (err == -EBLOCKLISTED)
1885                         fsc->blocklisted = true;
1886                 goto out_unlock;
1887         }
1888
1889         if (err2 == 0 || err2 == -EEXIST)
1890                 have |= POOL_WRITE;
1891         else if (err2 != -EPERM) {
1892                 if (err2 == -EBLOCKLISTED)
1893                         fsc->blocklisted = true;
1894                 err = err2;
1895                 goto out_unlock;
1896         }
1897
1898         pool_ns_len = pool_ns ? pool_ns->len : 0;
1899         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1900         if (!perm) {
1901                 err = -ENOMEM;
1902                 goto out_unlock;
1903         }
1904
1905         perm->pool = pool;
1906         perm->perm = have;
1907         perm->pool_ns_len = pool_ns_len;
1908         if (pool_ns_len > 0)
1909                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1910         perm->pool_ns[pool_ns_len] = 0;
1911
1912         rb_link_node(&perm->node, parent, p);
1913         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1914         err = 0;
1915 out_unlock:
1916         up_write(&mdsc->pool_perm_rwsem);
1917
1918         ceph_osdc_put_request(rd_req);
1919         ceph_osdc_put_request(wr_req);
1920 out:
1921         if (!err)
1922                 err = have;
1923         if (pool_ns)
1924                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1925                      pool, (int)pool_ns->len, pool_ns->str, err);
1926         else
1927                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1928         return err;
1929 }
1930
1931 int ceph_pool_perm_check(struct inode *inode, int need)
1932 {
1933         struct ceph_inode_info *ci = ceph_inode(inode);
1934         struct ceph_string *pool_ns;
1935         s64 pool;
1936         int ret, flags;
1937
1938         /* Only need to do this for regular files */
1939         if (!S_ISREG(inode->i_mode))
1940                 return 0;
1941
1942         if (ci->i_vino.snap != CEPH_NOSNAP) {
1943                 /*
1944                  * Pool permission check needs to write to the first object.
1945                  * But for snapshot, head of the first object may have alread
1946                  * been deleted. Skip check to avoid creating orphan object.
1947                  */
1948                 return 0;
1949         }
1950
1951         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
1952                                 NOPOOLPERM))
1953                 return 0;
1954
1955         spin_lock(&ci->i_ceph_lock);
1956         flags = ci->i_ceph_flags;
1957         pool = ci->i_layout.pool_id;
1958         spin_unlock(&ci->i_ceph_lock);
1959 check:
1960         if (flags & CEPH_I_POOL_PERM) {
1961                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1962                         dout("ceph_pool_perm_check pool %lld no read perm\n",
1963                              pool);
1964                         return -EPERM;
1965                 }
1966                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1967                         dout("ceph_pool_perm_check pool %lld no write perm\n",
1968                              pool);
1969                         return -EPERM;
1970                 }
1971                 return 0;
1972         }
1973
1974         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1975         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1976         ceph_put_string(pool_ns);
1977         if (ret < 0)
1978                 return ret;
1979
1980         flags = CEPH_I_POOL_PERM;
1981         if (ret & POOL_READ)
1982                 flags |= CEPH_I_POOL_RD;
1983         if (ret & POOL_WRITE)
1984                 flags |= CEPH_I_POOL_WR;
1985
1986         spin_lock(&ci->i_ceph_lock);
1987         if (pool == ci->i_layout.pool_id &&
1988             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
1989                 ci->i_ceph_flags |= flags;
1990         } else {
1991                 pool = ci->i_layout.pool_id;
1992                 flags = ci->i_ceph_flags;
1993         }
1994         spin_unlock(&ci->i_ceph_lock);
1995         goto check;
1996 }
1997
1998 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1999 {
2000         struct ceph_pool_perm *perm;
2001         struct rb_node *n;
2002
2003         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2004                 n = rb_first(&mdsc->pool_perm_tree);
2005                 perm = rb_entry(n, struct ceph_pool_perm, node);
2006                 rb_erase(n, &mdsc->pool_perm_tree);
2007                 kfree(perm);
2008         }
2009 }