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