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
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "metric.h"
  20 #include <linux/ceph/osd_client.h>
  21 #include <linux/ceph/striper.h>
  22
  23 /*
  24  * Ceph address space ops.
  25  *
  26  * There are a few funny things going on here.
  27  *
  28  * The page->private field is used to reference a struct
  29  * ceph_snap_context for _every_ dirty page.  This indicates which
  30  * snapshot the page was logically dirtied in, and thus which snap
  31  * context needs to be associated with the osd write during writeback.
  32  *
  33  * Similarly, struct ceph_inode_info maintains a set of counters to
  34  * count dirty pages on the inode.  In the absence of snapshots,
  35  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  36  *
  37  * When a snapshot is taken (that is, when the client receives
  38  * notification that a snapshot was taken), each inode with caps and
  39  * with dirty pages (dirty pages implies there is a cap) gets a new
  40  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  41  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  42  * moved to capsnap->dirty. (Unless a sync write is currently in
  43  * progress.  In that case, the capsnap is said to be "pending", new
  44  * writes cannot start, and the capsnap isn't "finalized" until the
  45  * write completes (or fails) and a final size/mtime for the inode for
  46  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  47  *
  48  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  49  * we look for the first capsnap in i_cap_snaps and write out pages in
  50  * that snap context _only_.  Then we move on to the next capsnap,
  51  * eventually reaching the "live" or "head" context (i.e., pages that
  52  * are not yet snapped) and are writing the most recently dirtied
  53  * pages.
  54  *
  55  * Invalidate and so forth must take care to ensure the dirty page
  56  * accounting is preserved.
  57  */
  58
  59 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  60 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  61         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  62          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  63
  64 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  65 {
  66         if (PagePrivate(page))
  67                 return (void *)page->private;
  68         return NULL;
  69 }
  70
  71 /*
  72  * Dirty a page.  Optimistically adjust accounting, on the assumption
  73  * that we won't race with invalidate.  If we do, readjust.
  74  */
  75 static int ceph_set_page_dirty(struct page *page)
  76 {
  77         struct address_space *mapping = page->mapping;
  78         struct inode *inode;
  79         struct ceph_inode_info *ci;
  80         struct ceph_snap_context *snapc;
  81         int ret;
  82
  83         if (unlikely(!mapping))
  84                 return !TestSetPageDirty(page);
  85
  86         if (PageDirty(page)) {
  87                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  88                      mapping->host, page, page->index);
  89                 BUG_ON(!PagePrivate(page));
  90                 return 0;
  91         }
  92
  93         inode = mapping->host;
  94         ci = ceph_inode(inode);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  99         if (__ceph_have_pending_cap_snap(ci)) {
 100                 struct ceph_cap_snap *capsnap =
 101                                 list_last_entry(&ci->i_cap_snaps,
 102                                                 struct ceph_cap_snap,
 103                                                 ci_item);
 104                 snapc = ceph_get_snap_context(capsnap->context);
 105                 capsnap->dirty_pages++;
 106         } else {
 107                 BUG_ON(!ci->i_head_snapc);
 108                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 109                 ++ci->i_wrbuffer_ref_head;
 110         }
 111         if (ci->i_wrbuffer_ref == 0)
 112                 ihold(inode);
 113         ++ci->i_wrbuffer_ref;
 114         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 115              "snapc %p seq %lld (%d snaps)\n",
 116              mapping->host, page, page->index,
 117              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 118              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 119              snapc, snapc->seq, snapc->num_snaps);
 120         spin_unlock(&ci->i_ceph_lock);
 121
 122         /*
 123          * Reference snap context in page->private.  Also set
 124          * PagePrivate so that we get invalidatepage callback.
 125          */
 126         BUG_ON(PagePrivate(page));
 127         page->private = (unsigned long)snapc;
 128         SetPagePrivate(page);
 129
 130         ret = __set_page_dirty_nobuffers(page);
 131         WARN_ON(!PageLocked(page));
 132         WARN_ON(!page->mapping);
 133
 134         return ret;
 135 }
 136
 137 /*
 138  * If we are truncating the full page (i.e. offset == 0), adjust the
 139  * dirty page counters appropriately.  Only called if there is private
 140  * data on the page.
 141  */
 142 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 143                                 unsigned int length)
 144 {
 145         struct inode *inode;
 146         struct ceph_inode_info *ci;
 147         struct ceph_snap_context *snapc = page_snap_context(page);
 148
 149         inode = page->mapping->host;
 150         ci = ceph_inode(inode);
 151
 152         if (offset != 0 || length != PAGE_SIZE) {
 153                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 154                      inode, page, page->index, offset, length);
 155                 return;
 156         }
 157
 158         ceph_invalidate_fscache_page(inode, page);
 159
 160         WARN_ON(!PageLocked(page));
 161         if (!PagePrivate(page))
 162                 return;
 163
 164         dout("%p invalidatepage %p idx %lu full dirty page\n",
 165              inode, page, page->index);
 166
 167         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 168         ceph_put_snap_context(snapc);
 169         page->private = 0;
 170         ClearPagePrivate(page);
 171 }
 172
 173 static int ceph_releasepage(struct page *page, gfp_t g)
 174 {
 175         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 176              page, page->index, PageDirty(page) ? "" : "not ");
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /* read a single page, without unlocking it. */
 186 static int ceph_do_readpage(struct file *filp, struct page *page)
 187 {
 188         struct inode *inode = file_inode(filp);
 189         struct ceph_inode_info *ci = ceph_inode(inode);
 190         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 191         struct ceph_osd_client *osdc = &fsc->client->osdc;
 192         struct ceph_osd_request *req;
 193         struct ceph_vino vino = ceph_vino(inode);
 194         int err = 0;
 195         u64 off = page_offset(page);
 196         u64 len = PAGE_SIZE;
 197
 198         if (off >= i_size_read(inode)) {
 199                 zero_user_segment(page, 0, PAGE_SIZE);
 200                 SetPageUptodate(page);
 201                 return 0;
 202         }
 203
 204         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 205                 /*
 206                  * Uptodate inline data should have been added
 207                  * into page cache while getting Fcr caps.
 208                  */
 209                 if (off == 0)
 210                         return -EINVAL;
 211                 zero_user_segment(page, 0, PAGE_SIZE);
 212                 SetPageUptodate(page);
 213                 return 0;
 214         }
 215
 216         err = ceph_readpage_from_fscache(inode, page);
 217         if (err == 0)
 218                 return -EINPROGRESS;
 219
 220         dout("readpage ino %llx.%llx file %p off %llu len %llu page %p index %lu\n",
 221              vino.ino, vino.snap, filp, off, len, page, page->index);
 222         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, 0, 1,
 223                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, NULL,
 224                                     ci->i_truncate_seq, ci->i_truncate_size,
 225                                     false);
 226         if (IS_ERR(req))
 227                 return PTR_ERR(req);
 228
 229         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 230
 231         err = ceph_osdc_start_request(osdc, req, false);
 232         if (!err)
 233                 err = ceph_osdc_wait_request(osdc, req);
 234
 235         ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency,
 236                                  req->r_end_latency, err);
 237
 238         ceph_osdc_put_request(req);
 239         dout("readpage result %d\n", err);
 240
 241         if (err == -ENOENT)
 242                 err = 0;
 243         if (err < 0) {
 244                 ceph_fscache_readpage_cancel(inode, page);
 245                 if (err == -EBLOCKLISTED)
 246                         fsc->blocklisted = true;
 247                 goto out;
 248         }
 249         if (err < PAGE_SIZE)
 250                 /* zero fill remainder of page */
 251                 zero_user_segment(page, err, PAGE_SIZE);
 252         else
 253                 flush_dcache_page(page);
 254
 255         SetPageUptodate(page);
 256         ceph_readpage_to_fscache(inode, page);
 257
 258 out:
 259         return err < 0 ? err : 0;
 260 }
 261
 262 static int ceph_readpage(struct file *filp, struct page *page)
 263 {
 264         int r = ceph_do_readpage(filp, page);
 265         if (r != -EINPROGRESS)
 266                 unlock_page(page);
 267         else
 268                 r = 0;
 269         return r;
 270 }
 271
 272 /*
 273  * Finish an async read(ahead) op.
 274  */
 275 static void finish_read(struct ceph_osd_request *req)
 276 {
 277         struct inode *inode = req->r_inode;
 278         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 279         struct ceph_osd_data *osd_data;
 280         int rc = req->r_result <= 0 ? req->r_result : 0;
 281         int bytes = req->r_result >= 0 ? req->r_result : 0;
 282         int num_pages;
 283         int i;
 284
 285         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 286         if (rc == -EBLOCKLISTED)
 287                 ceph_inode_to_client(inode)->blocklisted = true;
 288
 289         /* unlock all pages, zeroing any data we didn't read */
 290         osd_data = osd_req_op_extent_osd_data(req, 0);
 291         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 292         num_pages = calc_pages_for((u64)osd_data->alignment,
 293                                         (u64)osd_data->length);
 294         for (i = 0; i < num_pages; i++) {
 295                 struct page *page = osd_data->pages[i];
 296
 297                 if (rc < 0 && rc != -ENOENT) {
 298                         ceph_fscache_readpage_cancel(inode, page);
 299                         goto unlock;
 300                 }
 301                 if (bytes < (int)PAGE_SIZE) {
 302                         /* zero (remainder of) page */
 303                         int s = bytes < 0 ? 0 : bytes;
 304                         zero_user_segment(page, s, PAGE_SIZE);
 305                 }
 306                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 307                      page->index);
 308                 flush_dcache_page(page);
 309                 SetPageUptodate(page);
 310                 ceph_readpage_to_fscache(inode, page);
 311 unlock:
 312                 unlock_page(page);
 313                 put_page(page);
 314                 bytes -= PAGE_SIZE;
 315         }
 316
 317         ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency,
 318                                  req->r_end_latency, rc);
 319
 320         kfree(osd_data->pages);
 321 }
 322
 323 /*
 324  * start an async read(ahead) operation.  return nr_pages we submitted
 325  * a read for on success, or negative error code.
 326  */
 327 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
 328                       struct list_head *page_list, int max)
 329 {
 330         struct ceph_osd_client *osdc =
 331                 &ceph_inode_to_client(inode)->client->osdc;
 332         struct ceph_inode_info *ci = ceph_inode(inode);
 333         struct page *page = lru_to_page(page_list);
 334         struct ceph_vino vino;
 335         struct ceph_osd_request *req;
 336         u64 off;
 337         u64 len;
 338         int i;
 339         struct page **pages;
 340         pgoff_t next_index;
 341         int nr_pages = 0;
 342         int got = 0;
 343         int ret = 0;
 344
 345         if (!rw_ctx) {
 346                 /* caller of readpages does not hold buffer and read caps
 347                  * (fadvise, madvise and readahead cases) */
 348                 int want = CEPH_CAP_FILE_CACHE;
 349                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want,
 350                                         true, &got);
 351                 if (ret < 0) {
 352                         dout("start_read %p, error getting cap\n", inode);
 353                 } else if (!(got & want)) {
 354                         dout("start_read %p, no cache cap\n", inode);
 355                         ret = 0;
 356                 }
 357                 if (ret <= 0) {
 358                         if (got)
 359                                 ceph_put_cap_refs(ci, got);
 360                         while (!list_empty(page_list)) {
 361                                 page = lru_to_page(page_list);
 362                                 list_del(&page->lru);
 363                                 put_page(page);
 364                         }
 365                         return ret;
 366                 }
 367         }
 368
 369         off = (u64) page_offset(page);
 370
 371         /* count pages */
 372         next_index = page->index;
 373         list_for_each_entry_reverse(page, page_list, lru) {
 374                 if (page->index != next_index)
 375                         break;
 376                 nr_pages++;
 377                 next_index++;
 378                 if (max && nr_pages == max)
 379                         break;
 380         }
 381         len = nr_pages << PAGE_SHIFT;
 382         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 383              off, len);
 384         vino = ceph_vino(inode);
 385         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 386                                     0, 1, CEPH_OSD_OP_READ,
 387                                     CEPH_OSD_FLAG_READ, NULL,
 388                                     ci->i_truncate_seq, ci->i_truncate_size,
 389                                     false);
 390         if (IS_ERR(req)) {
 391                 ret = PTR_ERR(req);
 392                 goto out;
 393         }
 394
 395         /* build page vector */
 396         nr_pages = calc_pages_for(0, len);
 397         pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
 398         if (!pages) {
 399                 ret = -ENOMEM;
 400                 goto out_put;
 401         }
 402         for (i = 0; i < nr_pages; ++i) {
 403                 page = list_entry(page_list->prev, struct page, lru);
 404                 BUG_ON(PageLocked(page));
 405                 list_del(&page->lru);
 406
 407                 dout("start_read %p adding %p idx %lu\n", inode, page,
 408                      page->index);
 409                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 410                                           GFP_KERNEL)) {
 411                         ceph_fscache_uncache_page(inode, page);
 412                         put_page(page);
 413                         dout("start_read %p add_to_page_cache failed %p\n",
 414                              inode, page);
 415                         nr_pages = i;
 416                         if (nr_pages > 0) {
 417                                 len = nr_pages << PAGE_SHIFT;
 418                                 osd_req_op_extent_update(req, 0, len);
 419                                 break;
 420                         }
 421                         goto out_pages;
 422                 }
 423                 pages[i] = page;
 424         }
 425         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 426         req->r_callback = finish_read;
 427         req->r_inode = inode;
 428
 429         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 430         ret = ceph_osdc_start_request(osdc, req, false);
 431         if (ret < 0)
 432                 goto out_pages;
 433         ceph_osdc_put_request(req);
 434
 435         /* After adding locked pages to page cache, the inode holds cache cap.
 436          * So we can drop our cap refs. */
 437         if (got)
 438                 ceph_put_cap_refs(ci, got);
 439
 440         return nr_pages;
 441
 442 out_pages:
 443         for (i = 0; i < nr_pages; ++i) {
 444                 ceph_fscache_readpage_cancel(inode, pages[i]);
 445                 unlock_page(pages[i]);
 446         }
 447         ceph_put_page_vector(pages, nr_pages, false);
 448 out_put:
 449         ceph_osdc_put_request(req);
 450 out:
 451         if (got)
 452                 ceph_put_cap_refs(ci, got);
 453         return ret;
 454 }
 455
 456
 457 /*
 458  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 459  * the caller (VM) cleans them up.
 460  */
 461 static int ceph_readpages(struct file *file, struct address_space *mapping,
 462                           struct list_head *page_list, unsigned nr_pages)
 463 {
 464         struct inode *inode = file_inode(file);
 465         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 466         struct ceph_file_info *fi = file->private_data;
 467         struct ceph_rw_context *rw_ctx;
 468         int rc = 0;
 469         int max = 0;
 470
 471         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 472                 return -EINVAL;
 473
 474         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 475                                          &nr_pages);
 476
 477         if (rc == 0)
 478                 goto out;
 479
 480         rw_ctx = ceph_find_rw_context(fi);
 481         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 482         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 483              inode, file, rw_ctx, nr_pages, max);
 484         while (!list_empty(page_list)) {
 485                 rc = start_read(inode, rw_ctx, page_list, max);
 486                 if (rc < 0)
 487                         goto out;
 488         }
 489 out:
 490         ceph_fscache_readpages_cancel(inode, page_list);
 491
 492         dout("readpages %p file %p ret %d\n", inode, file, rc);
 493         return rc;
 494 }
 495
 496 struct ceph_writeback_ctl
 497 {
 498         loff_t i_size;
 499         u64 truncate_size;
 500         u32 truncate_seq;
 501         bool size_stable;
 502         bool head_snapc;
 503 };
 504
 505 /*
 506  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 507  * only snap context we are allowed to write back.
 508  */
 509 static struct ceph_snap_context *
 510 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 511                    struct ceph_snap_context *page_snapc)
 512 {
 513         struct ceph_inode_info *ci = ceph_inode(inode);
 514         struct ceph_snap_context *snapc = NULL;
 515         struct ceph_cap_snap *capsnap = NULL;
 516
 517         spin_lock(&ci->i_ceph_lock);
 518         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 519                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 520                      capsnap->context, capsnap->dirty_pages);
 521                 if (!capsnap->dirty_pages)
 522                         continue;
 523
 524                 /* get i_size, truncate_{seq,size} for page_snapc? */
 525                 if (snapc && capsnap->context != page_snapc)
 526                         continue;
 527
 528                 if (ctl) {
 529                         if (capsnap->writing) {
 530                                 ctl->i_size = i_size_read(inode);
 531                                 ctl->size_stable = false;
 532                         } else {
 533                                 ctl->i_size = capsnap->size;
 534                                 ctl->size_stable = true;
 535                         }
 536                         ctl->truncate_size = capsnap->truncate_size;
 537                         ctl->truncate_seq = capsnap->truncate_seq;
 538                         ctl->head_snapc = false;
 539                 }
 540
 541                 if (snapc)
 542                         break;
 543
 544                 snapc = ceph_get_snap_context(capsnap->context);
 545                 if (!page_snapc ||
 546                     page_snapc == snapc ||
 547                     page_snapc->seq > snapc->seq)
 548                         break;
 549         }
 550         if (!snapc && ci->i_wrbuffer_ref_head) {
 551                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 552                 dout(" head snapc %p has %d dirty pages\n",
 553                      snapc, ci->i_wrbuffer_ref_head);
 554                 if (ctl) {
 555                         ctl->i_size = i_size_read(inode);
 556                         ctl->truncate_size = ci->i_truncate_size;
 557                         ctl->truncate_seq = ci->i_truncate_seq;
 558                         ctl->size_stable = false;
 559                         ctl->head_snapc = true;
 560                 }
 561         }
 562         spin_unlock(&ci->i_ceph_lock);
 563         return snapc;
 564 }
 565
 566 static u64 get_writepages_data_length(struct inode *inode,
 567                                       struct page *page, u64 start)
 568 {
 569         struct ceph_inode_info *ci = ceph_inode(inode);
 570         struct ceph_snap_context *snapc = page_snap_context(page);
 571         struct ceph_cap_snap *capsnap = NULL;
 572         u64 end = i_size_read(inode);
 573
 574         if (snapc != ci->i_head_snapc) {
 575                 bool found = false;
 576                 spin_lock(&ci->i_ceph_lock);
 577                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 578                         if (capsnap->context == snapc) {
 579                                 if (!capsnap->writing)
 580                                         end = capsnap->size;
 581                                 found = true;
 582                                 break;
 583                         }
 584                 }
 585                 spin_unlock(&ci->i_ceph_lock);
 586                 WARN_ON(!found);
 587         }
 588         if (end > page_offset(page) + PAGE_SIZE)
 589                 end = page_offset(page) + PAGE_SIZE;
 590         return end > start ? end - start : 0;
 591 }
 592
 593 /*
 594  * Write a single page, but leave the page locked.
 595  *
 596  * If we get a write error, mark the mapping for error, but still adjust the
 597  * dirty page accounting (i.e., page is no longer dirty).
 598  */
 599 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 600 {
 601         struct inode *inode = page->mapping->host;
 602         struct ceph_inode_info *ci = ceph_inode(inode);
 603         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 604         struct ceph_snap_context *snapc, *oldest;
 605         loff_t page_off = page_offset(page);
 606         int err;
 607         loff_t len = PAGE_SIZE;
 608         struct ceph_writeback_ctl ceph_wbc;
 609         struct ceph_osd_client *osdc = &fsc->client->osdc;
 610         struct ceph_osd_request *req;
 611
 612         dout("writepage %p idx %lu\n", page, page->index);
 613
 614         /* verify this is a writeable snap context */
 615         snapc = page_snap_context(page);
 616         if (!snapc) {
 617                 dout("writepage %p page %p not dirty?\n", inode, page);
 618                 return 0;
 619         }
 620         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 621         if (snapc->seq > oldest->seq) {
 622                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 623                      inode, page, snapc);
 624                 /* we should only noop if called by kswapd */
 625                 WARN_ON(!(current->flags & PF_MEMALLOC));
 626                 ceph_put_snap_context(oldest);
 627                 redirty_page_for_writepage(wbc, page);
 628                 return 0;
 629         }
 630         ceph_put_snap_context(oldest);
 631
 632         /* is this a partial page at end of file? */
 633         if (page_off >= ceph_wbc.i_size) {
 634                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 635                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 636                 return 0;
 637         }
 638
 639         if (ceph_wbc.i_size < page_off + len)
 640                 len = ceph_wbc.i_size - page_off;
 641
 642         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 643              inode, page, page->index, page_off, len, snapc, snapc->seq);
 644
 645         if (atomic_long_inc_return(&fsc->writeback_count) >
 646             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 647                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 648
 649         set_page_writeback(page);
 650         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 651                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 652                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 653                                     true);
 654         if (IS_ERR(req)) {
 655                 redirty_page_for_writepage(wbc, page);
 656                 end_page_writeback(page);
 657                 return PTR_ERR(req);
 658         }
 659
 660         /* it may be a short write due to an object boundary */
 661         WARN_ON_ONCE(len > PAGE_SIZE);
 662         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 663         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 664
 665         req->r_mtime = inode->i_mtime;
 666         err = ceph_osdc_start_request(osdc, req, true);
 667         if (!err)
 668                 err = ceph_osdc_wait_request(osdc, req);
 669
 670         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
 671                                   req->r_end_latency, err);
 672
 673         ceph_osdc_put_request(req);
 674         if (err == 0)
 675                 err = len;
 676
 677         if (err < 0) {
 678                 struct writeback_control tmp_wbc;
 679                 if (!wbc)
 680                         wbc = &tmp_wbc;
 681                 if (err == -ERESTARTSYS) {
 682                         /* killed by SIGKILL */
 683                         dout("writepage interrupted page %p\n", page);
 684                         redirty_page_for_writepage(wbc, page);
 685                         end_page_writeback(page);
 686                         return err;
 687                 }
 688                 if (err == -EBLOCKLISTED)
 689                         fsc->blocklisted = true;
 690                 dout("writepage setting page/mapping error %d %p\n",
 691                      err, page);
 692                 mapping_set_error(&inode->i_data, err);
 693                 wbc->pages_skipped++;
 694         } else {
 695                 dout("writepage cleaned page %p\n", page);
 696                 err = 0;  /* vfs expects us to return 0 */
 697         }
 698         page->private = 0;
 699         ClearPagePrivate(page);
 700         end_page_writeback(page);
 701         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 702         ceph_put_snap_context(snapc);  /* page's reference */
 703
 704         if (atomic_long_dec_return(&fsc->writeback_count) <
 705             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 706                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 707
 708         return err;
 709 }
 710
 711 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 712 {
 713         int err;
 714         struct inode *inode = page->mapping->host;
 715         BUG_ON(!inode);
 716         ihold(inode);
 717         err = writepage_nounlock(page, wbc);
 718         if (err == -ERESTARTSYS) {
 719                 /* direct memory reclaimer was killed by SIGKILL. return 0
 720                  * to prevent caller from setting mapping/page error */
 721                 err = 0;
 722         }
 723         unlock_page(page);
 724         iput(inode);
 725         return err;
 726 }
 727
 728 /*
 729  * async writeback completion handler.
 730  *
 731  * If we get an error, set the mapping error bit, but not the individual
 732  * page error bits.
 733  */
 734 static void writepages_finish(struct ceph_osd_request *req)
 735 {
 736         struct inode *inode = req->r_inode;
 737         struct ceph_inode_info *ci = ceph_inode(inode);
 738         struct ceph_osd_data *osd_data;
 739         struct page *page;
 740         int num_pages, total_pages = 0;
 741         int i, j;
 742         int rc = req->r_result;
 743         struct ceph_snap_context *snapc = req->r_snapc;
 744         struct address_space *mapping = inode->i_mapping;
 745         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 746         bool remove_page;
 747
 748         dout("writepages_finish %p rc %d\n", inode, rc);
 749         if (rc < 0) {
 750                 mapping_set_error(mapping, rc);
 751                 ceph_set_error_write(ci);
 752                 if (rc == -EBLOCKLISTED)
 753                         fsc->blocklisted = true;
 754         } else {
 755                 ceph_clear_error_write(ci);
 756         }
 757
 758         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
 759                                   req->r_end_latency, rc);
 760
 761         /*
 762          * We lost the cache cap, need to truncate the page before
 763          * it is unlocked, otherwise we'd truncate it later in the
 764          * page truncation thread, possibly losing some data that
 765          * raced its way in
 766          */
 767         remove_page = !(ceph_caps_issued(ci) &
 768                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 769
 770         /* clean all pages */
 771         for (i = 0; i < req->r_num_ops; i++) {
 772                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 773                         break;
 774
 775                 osd_data = osd_req_op_extent_osd_data(req, i);
 776                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 777                 num_pages = calc_pages_for((u64)osd_data->alignment,
 778                                            (u64)osd_data->length);
 779                 total_pages += num_pages;
 780                 for (j = 0; j < num_pages; j++) {
 781                         page = osd_data->pages[j];
 782                         BUG_ON(!page);
 783                         WARN_ON(!PageUptodate(page));
 784
 785                         if (atomic_long_dec_return(&fsc->writeback_count) <
 786                              CONGESTION_OFF_THRESH(
 787                                         fsc->mount_options->congestion_kb))
 788                                 clear_bdi_congested(inode_to_bdi(inode),
 789                                                     BLK_RW_ASYNC);
 790
 791                         ceph_put_snap_context(page_snap_context(page));
 792                         page->private = 0;
 793                         ClearPagePrivate(page);
 794                         dout("unlocking %p\n", page);
 795                         end_page_writeback(page);
 796
 797                         if (remove_page)
 798                                 generic_error_remove_page(inode->i_mapping,
 799                                                           page);
 800
 801                         unlock_page(page);
 802                 }
 803                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 804                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 805
 806                 release_pages(osd_data->pages, num_pages);
 807         }
 808
 809         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 810
 811         osd_data = osd_req_op_extent_osd_data(req, 0);
 812         if (osd_data->pages_from_pool)
 813                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 814         else
 815                 kfree(osd_data->pages);
 816         ceph_osdc_put_request(req);
 817 }
 818
 819 /*
 820  * initiate async writeback
 821  */
 822 static int ceph_writepages_start(struct address_space *mapping,
 823                                  struct writeback_control *wbc)
 824 {
 825         struct inode *inode = mapping->host;
 826         struct ceph_inode_info *ci = ceph_inode(inode);
 827         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 828         struct ceph_vino vino = ceph_vino(inode);
 829         pgoff_t index, start_index, end = -1;
 830         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 831         struct pagevec pvec;
 832         int rc = 0;
 833         unsigned int wsize = i_blocksize(inode);
 834         struct ceph_osd_request *req = NULL;
 835         struct ceph_writeback_ctl ceph_wbc;
 836         bool should_loop, range_whole = false;
 837         bool done = false;
 838
 839         dout("writepages_start %p (mode=%s)\n", inode,
 840              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 841              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 842
 843         if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
 844                 if (ci->i_wrbuffer_ref > 0) {
 845                         pr_warn_ratelimited(
 846                                 "writepage_start %p %lld forced umount\n",
 847                                 inode, ceph_ino(inode));
 848                 }
 849                 mapping_set_error(mapping, -EIO);
 850                 return -EIO; /* we're in a forced umount, don't write! */
 851         }
 852         if (fsc->mount_options->wsize < wsize)
 853                 wsize = fsc->mount_options->wsize;
 854
 855         pagevec_init(&pvec);
 856
 857         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 858         index = start_index;
 859
 860 retry:
 861         /* find oldest snap context with dirty data */
 862         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 863         if (!snapc) {
 864                 /* hmm, why does writepages get called when there
 865                    is no dirty data? */
 866                 dout(" no snap context with dirty data?\n");
 867                 goto out;
 868         }
 869         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 870              snapc, snapc->seq, snapc->num_snaps);
 871
 872         should_loop = false;
 873         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 874                 /* where to start/end? */
 875                 if (wbc->range_cyclic) {
 876                         index = start_index;
 877                         end = -1;
 878                         if (index > 0)
 879                                 should_loop = true;
 880                         dout(" cyclic, start at %lu\n", index);
 881                 } else {
 882                         index = wbc->range_start >> PAGE_SHIFT;
 883                         end = wbc->range_end >> PAGE_SHIFT;
 884                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 885                                 range_whole = true;
 886                         dout(" not cyclic, %lu to %lu\n", index, end);
 887                 }
 888         } else if (!ceph_wbc.head_snapc) {
 889                 /* Do not respect wbc->range_{start,end}. Dirty pages
 890                  * in that range can be associated with newer snapc.
 891                  * They are not writeable until we write all dirty pages
 892                  * associated with 'snapc' get written */
 893                 if (index > 0)
 894                         should_loop = true;
 895                 dout(" non-head snapc, range whole\n");
 896         }
 897
 898         ceph_put_snap_context(last_snapc);
 899         last_snapc = snapc;
 900
 901         while (!done && index <= end) {
 902                 int num_ops = 0, op_idx;
 903                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 904                 struct page **pages = NULL, **data_pages;
 905                 struct page *page;
 906                 pgoff_t strip_unit_end = 0;
 907                 u64 offset = 0, len = 0;
 908                 bool from_pool = false;
 909
 910                 max_pages = wsize >> PAGE_SHIFT;
 911
 912 get_more_pages:
 913                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 914                                                 end, PAGECACHE_TAG_DIRTY);
 915                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 916                 if (!pvec_pages && !locked_pages)
 917                         break;
 918                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 919                         page = pvec.pages[i];
 920                         dout("? %p idx %lu\n", page, page->index);
 921                         if (locked_pages == 0)
 922                                 lock_page(page);  /* first page */
 923                         else if (!trylock_page(page))
 924                                 break;
 925
 926                         /* only dirty pages, or our accounting breaks */
 927                         if (unlikely(!PageDirty(page)) ||
 928                             unlikely(page->mapping != mapping)) {
 929                                 dout("!dirty or !mapping %p\n", page);
 930                                 unlock_page(page);
 931                                 continue;
 932                         }
 933                         /* only if matching snap context */
 934                         pgsnapc = page_snap_context(page);
 935                         if (pgsnapc != snapc) {
 936                                 dout("page snapc %p %lld != oldest %p %lld\n",
 937                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 938                                 if (!should_loop &&
 939                                     !ceph_wbc.head_snapc &&
 940                                     wbc->sync_mode != WB_SYNC_NONE)
 941                                         should_loop = true;
 942                                 unlock_page(page);
 943                                 continue;
 944                         }
 945                         if (page_offset(page) >= ceph_wbc.i_size) {
 946                                 dout("%p page eof %llu\n",
 947                                      page, ceph_wbc.i_size);
 948                                 if ((ceph_wbc.size_stable ||
 949                                     page_offset(page) >= i_size_read(inode)) &&
 950                                     clear_page_dirty_for_io(page))
 951                                         mapping->a_ops->invalidatepage(page,
 952                                                                 0, PAGE_SIZE);
 953                                 unlock_page(page);
 954                                 continue;
 955                         }
 956                         if (strip_unit_end && (page->index > strip_unit_end)) {
 957                                 dout("end of strip unit %p\n", page);
 958                                 unlock_page(page);
 959                                 break;
 960                         }
 961                         if (PageWriteback(page)) {
 962                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 963                                         dout("%p under writeback\n", page);
 964                                         unlock_page(page);
 965                                         continue;
 966                                 }
 967                                 dout("waiting on writeback %p\n", page);
 968                                 wait_on_page_writeback(page);
 969                         }
 970
 971                         if (!clear_page_dirty_for_io(page)) {
 972                                 dout("%p !clear_page_dirty_for_io\n", page);
 973                                 unlock_page(page);
 974                                 continue;
 975                         }
 976
 977                         /*
 978                          * We have something to write.  If this is
 979                          * the first locked page this time through,
 980                          * calculate max possinle write size and
 981                          * allocate a page array
 982                          */
 983                         if (locked_pages == 0) {
 984                                 u64 objnum;
 985                                 u64 objoff;
 986                                 u32 xlen;
 987
 988                                 /* prepare async write request */
 989                                 offset = (u64)page_offset(page);
 990                                 ceph_calc_file_object_mapping(&ci->i_layout,
 991                                                               offset, wsize,
 992                                                               &objnum, &objoff,
 993                                                               &xlen);
 994                                 len = xlen;
 995
 996                                 num_ops = 1;
 997                                 strip_unit_end = page->index +
 998                                         ((len - 1) >> PAGE_SHIFT);
 999
1000                                 BUG_ON(pages);
1001                                 max_pages = calc_pages_for(0, (u64)len);
1002                                 pages = kmalloc_array(max_pages,
1003                                                       sizeof(*pages),
1004                                                       GFP_NOFS);
1005                                 if (!pages) {
1006                                         from_pool = true;
1007                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1008                                         BUG_ON(!pages);
1009                                 }
1010
1011                                 len = 0;
1012                         } else if (page->index !=
1013                                    (offset + len) >> PAGE_SHIFT) {
1014                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1015                                                              CEPH_OSD_MAX_OPS)) {
1016                                         redirty_page_for_writepage(wbc, page);
1017                                         unlock_page(page);
1018                                         break;
1019                                 }
1020
1021                                 num_ops++;
1022                                 offset = (u64)page_offset(page);
1023                                 len = 0;
1024                         }
1025
1026                         /* note position of first page in pvec */
1027                         dout("%p will write page %p idx %lu\n",
1028                              inode, page, page->index);
1029
1030                         if (atomic_long_inc_return(&fsc->writeback_count) >
1031                             CONGESTION_ON_THRESH(
1032                                     fsc->mount_options->congestion_kb)) {
1033                                 set_bdi_congested(inode_to_bdi(inode),
1034                                                   BLK_RW_ASYNC);
1035                         }
1036
1037
1038                         pages[locked_pages++] = page;
1039                         pvec.pages[i] = NULL;
1040
1041                         len += PAGE_SIZE;
1042                 }
1043
1044                 /* did we get anything? */
1045                 if (!locked_pages)
1046                         goto release_pvec_pages;
1047                 if (i) {
1048                         unsigned j, n = 0;
1049                         /* shift unused page to beginning of pvec */
1050                         for (j = 0; j < pvec_pages; j++) {
1051                                 if (!pvec.pages[j])
1052                                         continue;
1053                                 if (n < j)
1054                                         pvec.pages[n] = pvec.pages[j];
1055                                 n++;
1056                         }
1057                         pvec.nr = n;
1058
1059                         if (pvec_pages && i == pvec_pages &&
1060                             locked_pages < max_pages) {
1061                                 dout("reached end pvec, trying for more\n");
1062                                 pagevec_release(&pvec);
1063                                 goto get_more_pages;
1064                         }
1065                 }
1066
1067 new_request:
1068                 offset = page_offset(pages[0]);
1069                 len = wsize;
1070
1071                 req = ceph_osdc_new_request(&fsc->client->osdc,
1072                                         &ci->i_layout, vino,
1073                                         offset, &len, 0, num_ops,
1074                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1075                                         snapc, ceph_wbc.truncate_seq,
1076                                         ceph_wbc.truncate_size, false);
1077                 if (IS_ERR(req)) {
1078                         req = ceph_osdc_new_request(&fsc->client->osdc,
1079                                                 &ci->i_layout, vino,
1080                                                 offset, &len, 0,
1081                                                 min(num_ops,
1082                                                     CEPH_OSD_SLAB_OPS),
1083                                                 CEPH_OSD_OP_WRITE,
1084                                                 CEPH_OSD_FLAG_WRITE,
1085                                                 snapc, ceph_wbc.truncate_seq,
1086                                                 ceph_wbc.truncate_size, true);
1087                         BUG_ON(IS_ERR(req));
1088                 }
1089                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1090                              PAGE_SIZE - offset);
1091
1092                 req->r_callback = writepages_finish;
1093                 req->r_inode = inode;
1094
1095                 /* Format the osd request message and submit the write */
1096                 len = 0;
1097                 data_pages = pages;
1098                 op_idx = 0;
1099                 for (i = 0; i < locked_pages; i++) {
1100                         u64 cur_offset = page_offset(pages[i]);
1101                         if (offset + len != cur_offset) {
1102                                 if (op_idx + 1 == req->r_num_ops)
1103                                         break;
1104                                 osd_req_op_extent_dup_last(req, op_idx,
1105                                                            cur_offset - offset);
1106                                 dout("writepages got pages at %llu~%llu\n",
1107                                      offset, len);
1108                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1109                                                         data_pages, len, 0,
1110                                                         from_pool, false);
1111                                 osd_req_op_extent_update(req, op_idx, len);
1112
1113                                 len = 0;
1114                                 offset = cur_offset;
1115                                 data_pages = pages + i;
1116                                 op_idx++;
1117                         }
1118
1119                         set_page_writeback(pages[i]);
1120                         len += PAGE_SIZE;
1121                 }
1122
1123                 if (ceph_wbc.size_stable) {
1124                         len = min(len, ceph_wbc.i_size - offset);
1125                 } else if (i == locked_pages) {
1126                         /* writepages_finish() clears writeback pages
1127                          * according to the data length, so make sure
1128                          * data length covers all locked pages */
1129                         u64 min_len = len + 1 - PAGE_SIZE;
1130                         len = get_writepages_data_length(inode, pages[i - 1],
1131                                                          offset);
1132                         len = max(len, min_len);
1133                 }
1134                 dout("writepages got pages at %llu~%llu\n", offset, len);
1135
1136                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1137                                                  0, from_pool, false);
1138                 osd_req_op_extent_update(req, op_idx, len);
1139
1140                 BUG_ON(op_idx + 1 != req->r_num_ops);
1141
1142                 from_pool = false;
1143                 if (i < locked_pages) {
1144                         BUG_ON(num_ops <= req->r_num_ops);
1145                         num_ops -= req->r_num_ops;
1146                         locked_pages -= i;
1147
1148                         /* allocate new pages array for next request */
1149                         data_pages = pages;
1150                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1151                                               GFP_NOFS);
1152                         if (!pages) {
1153                                 from_pool = true;
1154                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1155                                 BUG_ON(!pages);
1156                         }
1157                         memcpy(pages, data_pages + i,
1158                                locked_pages * sizeof(*pages));
1159                         memset(data_pages + i, 0,
1160                                locked_pages * sizeof(*pages));
1161                 } else {
1162                         BUG_ON(num_ops != req->r_num_ops);
1163                         index = pages[i - 1]->index + 1;
1164                         /* request message now owns the pages array */
1165                         pages = NULL;
1166                 }
1167
1168                 req->r_mtime = inode->i_mtime;
1169                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1170                 BUG_ON(rc);
1171                 req = NULL;
1172
1173                 wbc->nr_to_write -= i;
1174                 if (pages)
1175                         goto new_request;
1176
1177                 /*
1178                  * We stop writing back only if we are not doing
1179                  * integrity sync. In case of integrity sync we have to
1180                  * keep going until we have written all the pages
1181                  * we tagged for writeback prior to entering this loop.
1182                  */
1183                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1184                         done = true;
1185
1186 release_pvec_pages:
1187                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1188                      pvec.nr ? pvec.pages[0] : NULL);
1189                 pagevec_release(&pvec);
1190         }
1191
1192         if (should_loop && !done) {
1193                 /* more to do; loop back to beginning of file */
1194                 dout("writepages looping back to beginning of file\n");
1195                 end = start_index - 1; /* OK even when start_index == 0 */
1196
1197                 /* to write dirty pages associated with next snapc,
1198                  * we need to wait until current writes complete */
1199                 if (wbc->sync_mode != WB_SYNC_NONE &&
1200                     start_index == 0 && /* all dirty pages were checked */
1201                     !ceph_wbc.head_snapc) {
1202                         struct page *page;
1203                         unsigned i, nr;
1204                         index = 0;
1205                         while ((index <= end) &&
1206                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1207                                                 PAGECACHE_TAG_WRITEBACK))) {
1208                                 for (i = 0; i < nr; i++) {
1209                                         page = pvec.pages[i];
1210                                         if (page_snap_context(page) != snapc)
1211                                                 continue;
1212                                         wait_on_page_writeback(page);
1213                                 }
1214                                 pagevec_release(&pvec);
1215                                 cond_resched();
1216                         }
1217                 }
1218
1219                 start_index = 0;
1220                 index = 0;
1221                 goto retry;
1222         }
1223
1224         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1225                 mapping->writeback_index = index;
1226
1227 out:
1228         ceph_osdc_put_request(req);
1229         ceph_put_snap_context(last_snapc);
1230         dout("writepages dend - startone, rc = %d\n", rc);
1231         return rc;
1232 }
1233
1234
1235
1236 /*
1237  * See if a given @snapc is either writeable, or already written.
1238  */
1239 static int context_is_writeable_or_written(struct inode *inode,
1240                                            struct ceph_snap_context *snapc)
1241 {
1242         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1243         int ret = !oldest || snapc->seq <= oldest->seq;
1244
1245         ceph_put_snap_context(oldest);
1246         return ret;
1247 }
1248
1249 /**
1250  * ceph_find_incompatible - find an incompatible context and return it
1251  * @page: page being dirtied
1252  *
1253  * We are only allowed to write into/dirty a page if the page is
1254  * clean, or already dirty within the same snap context. Returns a
1255  * conflicting context if there is one, NULL if there isn't, or a
1256  * negative error code on other errors.
1257  *
1258  * Must be called with page lock held.
1259  */
1260 static struct ceph_snap_context *
1261 ceph_find_incompatible(struct page *page)
1262 {
1263         struct inode *inode = page->mapping->host;
1264         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1265         struct ceph_inode_info *ci = ceph_inode(inode);
1266
1267         if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1268                 dout(" page %p forced umount\n", page);
1269                 return ERR_PTR(-EIO);
1270         }
1271
1272         for (;;) {
1273                 struct ceph_snap_context *snapc, *oldest;
1274
1275                 wait_on_page_writeback(page);
1276
1277                 snapc = page_snap_context(page);
1278                 if (!snapc || snapc == ci->i_head_snapc)
1279                         break;
1280
1281                 /*
1282                  * this page is already dirty in another (older) snap
1283                  * context!  is it writeable now?
1284                  */
1285                 oldest = get_oldest_context(inode, NULL, NULL);
1286                 if (snapc->seq > oldest->seq) {
1287                         /* not writeable -- return it for the caller to deal with */
1288                         ceph_put_snap_context(oldest);
1289                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1290                         return ceph_get_snap_context(snapc);
1291                 }
1292                 ceph_put_snap_context(oldest);
1293
1294                 /* yay, writeable, do it now (without dropping page lock) */
1295                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1296                 if (clear_page_dirty_for_io(page)) {
1297                         int r = writepage_nounlock(page, NULL);
1298                         if (r < 0)
1299                                 return ERR_PTR(r);
1300                 }
1301         }
1302         return NULL;
1303 }
1304
1305 /*
1306  * We are only allowed to write into/dirty the page if the page is
1307  * clean, or already dirty within the same snap context.
1308  */
1309 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1310                             loff_t pos, unsigned len, unsigned flags,
1311                             struct page **pagep, void **fsdata)
1312 {
1313         struct inode *inode = file_inode(file);
1314         struct ceph_inode_info *ci = ceph_inode(inode);
1315         struct ceph_snap_context *snapc;
1316         struct page *page = NULL;
1317         pgoff_t index = pos >> PAGE_SHIFT;
1318         int pos_in_page = pos & ~PAGE_MASK;
1319         int r = 0;
1320
1321         dout("write_begin file %p inode %p page %p %d~%d\n", file, inode, page, (int)pos, (int)len);
1322
1323         for (;;) {
1324                 page = grab_cache_page_write_begin(mapping, index, flags);
1325                 if (!page) {
1326                         r = -ENOMEM;
1327                         break;
1328                 }
1329
1330                 snapc = ceph_find_incompatible(page);
1331                 if (snapc) {
1332                         if (IS_ERR(snapc)) {
1333                                 r = PTR_ERR(snapc);
1334                                 break;
1335                         }
1336                         unlock_page(page);
1337                         put_page(page);
1338                         page = NULL;
1339                         ceph_queue_writeback(inode);
1340                         r = wait_event_killable(ci->i_cap_wq,
1341                                                 context_is_writeable_or_written(inode, snapc));
1342                         ceph_put_snap_context(snapc);
1343                         if (r != 0)
1344                                 break;
1345                         continue;
1346                 }
1347
1348                 if (PageUptodate(page)) {
1349                         dout(" page %p already uptodate\n", page);
1350                         break;
1351                 }
1352
1353                 /*
1354                  * In some cases we don't need to read at all:
1355                  * - full page write
1356                  * - write that lies completely beyond EOF
1357                  * - write that covers the the page from start to EOF or beyond it
1358                  */
1359                 if ((pos_in_page == 0 && len == PAGE_SIZE) ||
1360                     (pos >= i_size_read(inode)) ||
1361                     (pos_in_page == 0 && (pos + len) >= i_size_read(inode))) {
1362                         zero_user_segments(page, 0, pos_in_page,
1363                                            pos_in_page + len, PAGE_SIZE);
1364                         break;
1365                 }
1366
1367                 /*
1368                  * We need to read it. If we get back -EINPROGRESS, then the page was
1369                  * handed off to fscache and it will be unlocked when the read completes.
1370                  * Refind the page in that case so we can reacquire the page lock. Otherwise
1371                  * we got a hard error or the read was completed synchronously.
1372                  */
1373                 r = ceph_do_readpage(file, page);
1374                 if (r != -EINPROGRESS)
1375                         break;
1376         }
1377
1378         if (r < 0) {
1379                 if (page) {
1380                         unlock_page(page);
1381                         put_page(page);
1382                 }
1383         } else {
1384                 *pagep = page;
1385         }
1386         return r;
1387 }
1388
1389 /*
1390  * we don't do anything in here that simple_write_end doesn't do
1391  * except adjust dirty page accounting
1392  */
1393 static int ceph_write_end(struct file *file, struct address_space *mapping,
1394                           loff_t pos, unsigned len, unsigned copied,
1395                           struct page *page, void *fsdata)
1396 {
1397         struct inode *inode = file_inode(file);
1398         bool check_cap = false;
1399
1400         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1401              inode, page, (int)pos, (int)copied, (int)len);
1402
1403         /* zero the stale part of the page if we did a short copy */
1404         if (!PageUptodate(page)) {
1405                 if (copied < len) {
1406                         copied = 0;
1407                         goto out;
1408                 }
1409                 SetPageUptodate(page);
1410         }
1411
1412         /* did file size increase? */
1413         if (pos+copied > i_size_read(inode))
1414                 check_cap = ceph_inode_set_size(inode, pos+copied);
1415
1416         set_page_dirty(page);
1417
1418 out:
1419         unlock_page(page);
1420         put_page(page);
1421
1422         if (check_cap)
1423                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1424
1425         return copied;
1426 }
1427
1428 /*
1429  * we set .direct_IO to indicate direct io is supported, but since we
1430  * intercept O_DIRECT reads and writes early, this function should
1431  * never get called.
1432  */
1433 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1434 {
1435         WARN_ON(1);
1436         return -EINVAL;
1437 }
1438
1439 const struct address_space_operations ceph_aops = {
1440         .readpage = ceph_readpage,
1441         .readpages = ceph_readpages,
1442         .writepage = ceph_writepage,
1443         .writepages = ceph_writepages_start,
1444         .write_begin = ceph_write_begin,
1445         .write_end = ceph_write_end,
1446         .set_page_dirty = ceph_set_page_dirty,
1447         .invalidatepage = ceph_invalidatepage,
1448         .releasepage = ceph_releasepage,
1449         .direct_IO = ceph_direct_io,
1450 };
1451
1452 static void ceph_block_sigs(sigset_t *oldset)
1453 {
1454         sigset_t mask;
1455         siginitsetinv(&mask, sigmask(SIGKILL));
1456         sigprocmask(SIG_BLOCK, &mask, oldset);
1457 }
1458
1459 static void ceph_restore_sigs(sigset_t *oldset)
1460 {
1461         sigprocmask(SIG_SETMASK, oldset, NULL);
1462 }
1463
1464 /*
1465  * vm ops
1466  */
1467 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1468 {
1469         struct vm_area_struct *vma = vmf->vma;
1470         struct inode *inode = file_inode(vma->vm_file);
1471         struct ceph_inode_info *ci = ceph_inode(inode);
1472         struct ceph_file_info *fi = vma->vm_file->private_data;
1473         struct page *pinned_page = NULL;
1474         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1475         int want, got, err;
1476         sigset_t oldset;
1477         vm_fault_t ret = VM_FAULT_SIGBUS;
1478
1479         ceph_block_sigs(&oldset);
1480
1481         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1482              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1483         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1484                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1485         else
1486                 want = CEPH_CAP_FILE_CACHE;
1487
1488         got = 0;
1489         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1,
1490                             &got, &pinned_page);
1491         if (err < 0)
1492                 goto out_restore;
1493
1494         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1495              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1496
1497         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1498             ci->i_inline_version == CEPH_INLINE_NONE) {
1499                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1500                 ceph_add_rw_context(fi, &rw_ctx);
1501                 ret = filemap_fault(vmf);
1502                 ceph_del_rw_context(fi, &rw_ctx);
1503                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1504                         inode, off, (size_t)PAGE_SIZE,
1505                                 ceph_cap_string(got), ret);
1506         } else
1507                 err = -EAGAIN;
1508
1509         if (pinned_page)
1510                 put_page(pinned_page);
1511         ceph_put_cap_refs(ci, got);
1512
1513         if (err != -EAGAIN)
1514                 goto out_restore;
1515
1516         /* read inline data */
1517         if (off >= PAGE_SIZE) {
1518                 /* does not support inline data > PAGE_SIZE */
1519                 ret = VM_FAULT_SIGBUS;
1520         } else {
1521                 struct address_space *mapping = inode->i_mapping;
1522                 struct page *page = find_or_create_page(mapping, 0,
1523                                                 mapping_gfp_constraint(mapping,
1524                                                 ~__GFP_FS));
1525                 if (!page) {
1526                         ret = VM_FAULT_OOM;
1527                         goto out_inline;
1528                 }
1529                 err = __ceph_do_getattr(inode, page,
1530                                          CEPH_STAT_CAP_INLINE_DATA, true);
1531                 if (err < 0 || off >= i_size_read(inode)) {
1532                         unlock_page(page);
1533                         put_page(page);
1534                         ret = vmf_error(err);
1535                         goto out_inline;
1536                 }
1537                 if (err < PAGE_SIZE)
1538                         zero_user_segment(page, err, PAGE_SIZE);
1539                 else
1540                         flush_dcache_page(page);
1541                 SetPageUptodate(page);
1542                 vmf->page = page;
1543                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1544 out_inline:
1545                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1546                      inode, off, (size_t)PAGE_SIZE, ret);
1547         }
1548 out_restore:
1549         ceph_restore_sigs(&oldset);
1550         if (err < 0)
1551                 ret = vmf_error(err);
1552
1553         return ret;
1554 }
1555
1556 /*
1557  * Reuse write_begin here for simplicity.
1558  */
1559 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1560 {
1561         struct vm_area_struct *vma = vmf->vma;
1562         struct inode *inode = file_inode(vma->vm_file);
1563         struct ceph_inode_info *ci = ceph_inode(inode);
1564         struct ceph_file_info *fi = vma->vm_file->private_data;
1565         struct ceph_cap_flush *prealloc_cf;
1566         struct page *page = vmf->page;
1567         loff_t off = page_offset(page);
1568         loff_t size = i_size_read(inode);
1569         size_t len;
1570         int want, got, err;
1571         sigset_t oldset;
1572         vm_fault_t ret = VM_FAULT_SIGBUS;
1573
1574         prealloc_cf = ceph_alloc_cap_flush();
1575         if (!prealloc_cf)
1576                 return VM_FAULT_OOM;
1577
1578         sb_start_pagefault(inode->i_sb);
1579         ceph_block_sigs(&oldset);
1580
1581         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1582                 struct page *locked_page = NULL;
1583                 if (off == 0) {
1584                         lock_page(page);
1585                         locked_page = page;
1586                 }
1587                 err = ceph_uninline_data(vma->vm_file, locked_page);
1588                 if (locked_page)
1589                         unlock_page(locked_page);
1590                 if (err < 0)
1591                         goto out_free;
1592         }
1593
1594         if (off + PAGE_SIZE <= size)
1595                 len = PAGE_SIZE;
1596         else
1597                 len = size & ~PAGE_MASK;
1598
1599         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1600              inode, ceph_vinop(inode), off, len, size);
1601         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1602                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1603         else
1604                 want = CEPH_CAP_FILE_BUFFER;
1605
1606         got = 0;
1607         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len,
1608                             &got, NULL);
1609         if (err < 0)
1610                 goto out_free;
1611
1612         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1613              inode, off, len, ceph_cap_string(got));
1614
1615         /* Update time before taking page lock */
1616         file_update_time(vma->vm_file);
1617         inode_inc_iversion_raw(inode);
1618
1619         do {
1620                 struct ceph_snap_context *snapc;
1621
1622                 lock_page(page);
1623
1624                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1625                         unlock_page(page);
1626                         ret = VM_FAULT_NOPAGE;
1627                         break;
1628                 }
1629
1630                 snapc = ceph_find_incompatible(page);
1631                 if (!snapc) {
1632                         /* success.  we'll keep the page locked. */
1633                         set_page_dirty(page);
1634                         ret = VM_FAULT_LOCKED;
1635                         break;
1636                 }
1637
1638                 unlock_page(page);
1639
1640                 if (IS_ERR(snapc)) {
1641                         ret = VM_FAULT_SIGBUS;
1642                         break;
1643                 }
1644
1645                 ceph_queue_writeback(inode);
1646                 err = wait_event_killable(ci->i_cap_wq,
1647                                 context_is_writeable_or_written(inode, snapc));
1648                 ceph_put_snap_context(snapc);
1649         } while (err == 0);
1650
1651         if (ret == VM_FAULT_LOCKED ||
1652             ci->i_inline_version != CEPH_INLINE_NONE) {
1653                 int dirty;
1654                 spin_lock(&ci->i_ceph_lock);
1655                 ci->i_inline_version = CEPH_INLINE_NONE;
1656                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1657                                                &prealloc_cf);
1658                 spin_unlock(&ci->i_ceph_lock);
1659                 if (dirty)
1660                         __mark_inode_dirty(inode, dirty);
1661         }
1662
1663         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1664              inode, off, len, ceph_cap_string(got), ret);
1665         ceph_put_cap_refs(ci, got);
1666 out_free:
1667         ceph_restore_sigs(&oldset);
1668         sb_end_pagefault(inode->i_sb);
1669         ceph_free_cap_flush(prealloc_cf);
1670         if (err < 0)
1671                 ret = vmf_error(err);
1672         return ret;
1673 }
1674
1675 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1676                            char *data, size_t len)
1677 {
1678         struct address_space *mapping = inode->i_mapping;
1679         struct page *page;
1680
1681         if (locked_page) {
1682                 page = locked_page;
1683         } else {
1684                 if (i_size_read(inode) == 0)
1685                         return;
1686                 page = find_or_create_page(mapping, 0,
1687                                            mapping_gfp_constraint(mapping,
1688                                            ~__GFP_FS));
1689                 if (!page)
1690                         return;
1691                 if (PageUptodate(page)) {
1692                         unlock_page(page);
1693                         put_page(page);
1694                         return;
1695                 }
1696         }
1697
1698         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1699              inode, ceph_vinop(inode), len, locked_page);
1700
1701         if (len > 0) {
1702                 void *kaddr = kmap_atomic(page);
1703                 memcpy(kaddr, data, len);
1704                 kunmap_atomic(kaddr);
1705         }
1706
1707         if (page != locked_page) {
1708                 if (len < PAGE_SIZE)
1709                         zero_user_segment(page, len, PAGE_SIZE);
1710                 else
1711                         flush_dcache_page(page);
1712
1713                 SetPageUptodate(page);
1714                 unlock_page(page);
1715                 put_page(page);
1716         }
1717 }
1718
1719 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1720 {
1721         struct inode *inode = file_inode(filp);
1722         struct ceph_inode_info *ci = ceph_inode(inode);
1723         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1724         struct ceph_osd_request *req;
1725         struct page *page = NULL;
1726         u64 len, inline_version;
1727         int err = 0;
1728         bool from_pagecache = false;
1729
1730         spin_lock(&ci->i_ceph_lock);
1731         inline_version = ci->i_inline_version;
1732         spin_unlock(&ci->i_ceph_lock);
1733
1734         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1735              inode, ceph_vinop(inode), inline_version);
1736
1737         if (inline_version == 1 || /* initial version, no data */
1738             inline_version == CEPH_INLINE_NONE)
1739                 goto out;
1740
1741         if (locked_page) {
1742                 page = locked_page;
1743                 WARN_ON(!PageUptodate(page));
1744         } else if (ceph_caps_issued(ci) &
1745                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1746                 page = find_get_page(inode->i_mapping, 0);
1747                 if (page) {
1748                         if (PageUptodate(page)) {
1749                                 from_pagecache = true;
1750                                 lock_page(page);
1751                         } else {
1752                                 put_page(page);
1753                                 page = NULL;
1754                         }
1755                 }
1756         }
1757
1758         if (page) {
1759                 len = i_size_read(inode);
1760                 if (len > PAGE_SIZE)
1761                         len = PAGE_SIZE;
1762         } else {
1763                 page = __page_cache_alloc(GFP_NOFS);
1764                 if (!page) {
1765                         err = -ENOMEM;
1766                         goto out;
1767                 }
1768                 err = __ceph_do_getattr(inode, page,
1769                                         CEPH_STAT_CAP_INLINE_DATA, true);
1770                 if (err < 0) {
1771                         /* no inline data */
1772                         if (err == -ENODATA)
1773                                 err = 0;
1774                         goto out;
1775                 }
1776                 len = err;
1777         }
1778
1779         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1780                                     ceph_vino(inode), 0, &len, 0, 1,
1781                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1782                                     NULL, 0, 0, false);
1783         if (IS_ERR(req)) {
1784                 err = PTR_ERR(req);
1785                 goto out;
1786         }
1787
1788         req->r_mtime = inode->i_mtime;
1789         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1790         if (!err)
1791                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1792         ceph_osdc_put_request(req);
1793         if (err < 0)
1794                 goto out;
1795
1796         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1797                                     ceph_vino(inode), 0, &len, 1, 3,
1798                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1799                                     NULL, ci->i_truncate_seq,
1800                                     ci->i_truncate_size, false);
1801         if (IS_ERR(req)) {
1802                 err = PTR_ERR(req);
1803                 goto out;
1804         }
1805
1806         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1807
1808         {
1809                 __le64 xattr_buf = cpu_to_le64(inline_version);
1810                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1811                                             "inline_version", &xattr_buf,
1812                                             sizeof(xattr_buf),
1813                                             CEPH_OSD_CMPXATTR_OP_GT,
1814                                             CEPH_OSD_CMPXATTR_MODE_U64);
1815                 if (err)
1816                         goto out_put;
1817         }
1818
1819         {
1820                 char xattr_buf[32];
1821                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1822                                          "%llu", inline_version);
1823                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1824                                             "inline_version",
1825                                             xattr_buf, xattr_len, 0, 0);
1826                 if (err)
1827                         goto out_put;
1828         }
1829
1830         req->r_mtime = inode->i_mtime;
1831         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1832         if (!err)
1833                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1834
1835         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
1836                                   req->r_end_latency, err);
1837
1838 out_put:
1839         ceph_osdc_put_request(req);
1840         if (err == -ECANCELED)
1841                 err = 0;
1842 out:
1843         if (page && page != locked_page) {
1844                 if (from_pagecache) {
1845                         unlock_page(page);
1846                         put_page(page);
1847                 } else
1848                         __free_pages(page, 0);
1849         }
1850
1851         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1852              inode, ceph_vinop(inode), inline_version, err);
1853         return err;
1854 }
1855
1856 static const struct vm_operations_struct ceph_vmops = {
1857         .fault          = ceph_filemap_fault,
1858         .page_mkwrite   = ceph_page_mkwrite,
1859 };
1860
1861 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1862 {
1863         struct address_space *mapping = file->f_mapping;
1864
1865         if (!mapping->a_ops->readpage)
1866                 return -ENOEXEC;
1867         file_accessed(file);
1868         vma->vm_ops = &ceph_vmops;
1869         return 0;
1870 }
1871
1872 enum {
1873         POOL_READ       = 1,
1874         POOL_WRITE      = 2,
1875 };
1876
1877 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1878                                 s64 pool, struct ceph_string *pool_ns)
1879 {
1880         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1881         struct ceph_mds_client *mdsc = fsc->mdsc;
1882         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1883         struct rb_node **p, *parent;
1884         struct ceph_pool_perm *perm;
1885         struct page **pages;
1886         size_t pool_ns_len;
1887         int err = 0, err2 = 0, have = 0;
1888
1889         down_read(&mdsc->pool_perm_rwsem);
1890         p = &mdsc->pool_perm_tree.rb_node;
1891         while (*p) {
1892                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1893                 if (pool < perm->pool)
1894                         p = &(*p)->rb_left;
1895                 else if (pool > perm->pool)
1896                         p = &(*p)->rb_right;
1897                 else {
1898                         int ret = ceph_compare_string(pool_ns,
1899                                                 perm->pool_ns,
1900                                                 perm->pool_ns_len);
1901                         if (ret < 0)
1902                                 p = &(*p)->rb_left;
1903                         else if (ret > 0)
1904                                 p = &(*p)->rb_right;
1905                         else {
1906                                 have = perm->perm;
1907                                 break;
1908                         }
1909                 }
1910         }
1911         up_read(&mdsc->pool_perm_rwsem);
1912         if (*p)
1913                 goto out;
1914
1915         if (pool_ns)
1916                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1917                      pool, (int)pool_ns->len, pool_ns->str);
1918         else
1919                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1920
1921         down_write(&mdsc->pool_perm_rwsem);
1922         p = &mdsc->pool_perm_tree.rb_node;
1923         parent = NULL;
1924         while (*p) {
1925                 parent = *p;
1926                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1927                 if (pool < perm->pool)
1928                         p = &(*p)->rb_left;
1929                 else if (pool > perm->pool)
1930                         p = &(*p)->rb_right;
1931                 else {
1932                         int ret = ceph_compare_string(pool_ns,
1933                                                 perm->pool_ns,
1934                                                 perm->pool_ns_len);
1935                         if (ret < 0)
1936                                 p = &(*p)->rb_left;
1937                         else if (ret > 0)
1938                                 p = &(*p)->rb_right;
1939                         else {
1940                                 have = perm->perm;
1941                                 break;
1942                         }
1943                 }
1944         }
1945         if (*p) {
1946                 up_write(&mdsc->pool_perm_rwsem);
1947                 goto out;
1948         }
1949
1950         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1951                                          1, false, GFP_NOFS);
1952         if (!rd_req) {
1953                 err = -ENOMEM;
1954                 goto out_unlock;
1955         }
1956
1957         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1958         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1959         rd_req->r_base_oloc.pool = pool;
1960         if (pool_ns)
1961                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1962         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1963
1964         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1965         if (err)
1966                 goto out_unlock;
1967
1968         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1969                                          1, false, GFP_NOFS);
1970         if (!wr_req) {
1971                 err = -ENOMEM;
1972                 goto out_unlock;
1973         }
1974
1975         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1976         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1977         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1978         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1979
1980         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1981         if (err)
1982                 goto out_unlock;
1983
1984         /* one page should be large enough for STAT data */
1985         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1986         if (IS_ERR(pages)) {
1987                 err = PTR_ERR(pages);
1988                 goto out_unlock;
1989         }
1990
1991         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1992                                      0, false, true);
1993         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1994
1995         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1996         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1997
1998         if (!err)
1999                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2000         if (!err2)
2001                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2002
2003         if (err >= 0 || err == -ENOENT)
2004                 have |= POOL_READ;
2005         else if (err != -EPERM) {
2006                 if (err == -EBLOCKLISTED)
2007                         fsc->blocklisted = true;
2008                 goto out_unlock;
2009         }
2010
2011         if (err2 == 0 || err2 == -EEXIST)
2012                 have |= POOL_WRITE;
2013         else if (err2 != -EPERM) {
2014                 if (err2 == -EBLOCKLISTED)
2015                         fsc->blocklisted = true;
2016                 err = err2;
2017                 goto out_unlock;
2018         }
2019
2020         pool_ns_len = pool_ns ? pool_ns->len : 0;
2021         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2022         if (!perm) {
2023                 err = -ENOMEM;
2024                 goto out_unlock;
2025         }
2026
2027         perm->pool = pool;
2028         perm->perm = have;
2029         perm->pool_ns_len = pool_ns_len;
2030         if (pool_ns_len > 0)
2031                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2032         perm->pool_ns[pool_ns_len] = 0;
2033
2034         rb_link_node(&perm->node, parent, p);
2035         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2036         err = 0;
2037 out_unlock:
2038         up_write(&mdsc->pool_perm_rwsem);
2039
2040         ceph_osdc_put_request(rd_req);
2041         ceph_osdc_put_request(wr_req);
2042 out:
2043         if (!err)
2044                 err = have;
2045         if (pool_ns)
2046                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2047                      pool, (int)pool_ns->len, pool_ns->str, err);
2048         else
2049                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2050         return err;
2051 }
2052
2053 int ceph_pool_perm_check(struct inode *inode, int need)
2054 {
2055         struct ceph_inode_info *ci = ceph_inode(inode);
2056         struct ceph_string *pool_ns;
2057         s64 pool;
2058         int ret, flags;
2059
2060         if (ci->i_vino.snap != CEPH_NOSNAP) {
2061                 /*
2062                  * Pool permission check needs to write to the first object.
2063                  * But for snapshot, head of the first object may have alread
2064                  * been deleted. Skip check to avoid creating orphan object.
2065                  */
2066                 return 0;
2067         }
2068
2069         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2070                                 NOPOOLPERM))
2071                 return 0;
2072
2073         spin_lock(&ci->i_ceph_lock);
2074         flags = ci->i_ceph_flags;
2075         pool = ci->i_layout.pool_id;
2076         spin_unlock(&ci->i_ceph_lock);
2077 check:
2078         if (flags & CEPH_I_POOL_PERM) {
2079                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2080                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2081                              pool);
2082                         return -EPERM;
2083                 }
2084                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2085                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2086                              pool);
2087                         return -EPERM;
2088                 }
2089                 return 0;
2090         }
2091
2092         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2093         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2094         ceph_put_string(pool_ns);
2095         if (ret < 0)
2096                 return ret;
2097
2098         flags = CEPH_I_POOL_PERM;
2099         if (ret & POOL_READ)
2100                 flags |= CEPH_I_POOL_RD;
2101         if (ret & POOL_WRITE)
2102                 flags |= CEPH_I_POOL_WR;
2103
2104         spin_lock(&ci->i_ceph_lock);
2105         if (pool == ci->i_layout.pool_id &&
2106             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2107                 ci->i_ceph_flags |= flags;
2108         } else {
2109                 pool = ci->i_layout.pool_id;
2110                 flags = ci->i_ceph_flags;
2111         }
2112         spin_unlock(&ci->i_ceph_lock);
2113         goto check;
2114 }
2115
2116 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2117 {
2118         struct ceph_pool_perm *perm;
2119         struct rb_node *n;
2120
2121         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2122                 n = rb_first(&mdsc->pool_perm_tree);
2123                 perm = rb_entry(n, struct ceph_pool_perm, node);
2124                 rb_erase(n, &mdsc->pool_perm_tree);
2125                 kfree(perm);
2126         }
2127 }