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 = list_entry(page_list->prev, struct page, lru);
 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, &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 = list_entry(page_list->prev,
 337                                                   struct page, lru);
 338                                 list_del(&page->lru);
 339                                 put_page(page);
 340                         }
 341                         return ret;
 342                 }
 343         }
 344
 345         off = (u64) page_offset(page);
 346
 347         /* count pages */
 348         next_index = page->index;
 349         list_for_each_entry_reverse(page, page_list, lru) {
 350                 if (page->index != next_index)
 351                         break;
 352                 nr_pages++;
 353                 next_index++;
 354                 if (max && nr_pages == max)
 355                         break;
 356         }
 357         len = nr_pages << PAGE_SHIFT;
 358         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 359              off, len);
 360         vino = ceph_vino(inode);
 361         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 362                                     0, 1, CEPH_OSD_OP_READ,
 363                                     CEPH_OSD_FLAG_READ, NULL,
 364                                     ci->i_truncate_seq, ci->i_truncate_size,
 365                                     false);
 366         if (IS_ERR(req)) {
 367                 ret = PTR_ERR(req);
 368                 goto out;
 369         }
 370
 371         /* build page vector */
 372         nr_pages = calc_pages_for(0, len);
 373         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
 374         if (!pages) {
 375                 ret = -ENOMEM;
 376                 goto out_put;
 377         }
 378         for (i = 0; i < nr_pages; ++i) {
 379                 page = list_entry(page_list->prev, struct page, lru);
 380                 BUG_ON(PageLocked(page));
 381                 list_del(&page->lru);
 382
 383                 dout("start_read %p adding %p idx %lu\n", inode, page,
 384                      page->index);
 385                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 386                                           GFP_KERNEL)) {
 387                         ceph_fscache_uncache_page(inode, page);
 388                         put_page(page);
 389                         dout("start_read %p add_to_page_cache failed %p\n",
 390                              inode, page);
 391                         nr_pages = i;
 392                         if (nr_pages > 0) {
 393                                 len = nr_pages << PAGE_SHIFT;
 394                                 osd_req_op_extent_update(req, 0, len);
 395                                 break;
 396                         }
 397                         goto out_pages;
 398                 }
 399                 pages[i] = page;
 400         }
 401         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 402         req->r_callback = finish_read;
 403         req->r_inode = inode;
 404
 405         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 406         ret = ceph_osdc_start_request(osdc, req, false);
 407         if (ret < 0)
 408                 goto out_pages;
 409         ceph_osdc_put_request(req);
 410
 411         /* After adding locked pages to page cache, the inode holds cache cap.
 412          * So we can drop our cap refs. */
 413         if (got)
 414                 ceph_put_cap_refs(ci, got);
 415
 416         return nr_pages;
 417
 418 out_pages:
 419         for (i = 0; i < nr_pages; ++i) {
 420                 ceph_fscache_readpage_cancel(inode, pages[i]);
 421                 unlock_page(pages[i]);
 422         }
 423         ceph_put_page_vector(pages, nr_pages, false);
 424 out_put:
 425         ceph_osdc_put_request(req);
 426 out:
 427         if (got)
 428                 ceph_put_cap_refs(ci, got);
 429         return ret;
 430 }
 431
 432
 433 /*
 434  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 435  * the caller (VM) cleans them up.
 436  */
 437 static int ceph_readpages(struct file *file, struct address_space *mapping,
 438                           struct list_head *page_list, unsigned nr_pages)
 439 {
 440         struct inode *inode = file_inode(file);
 441         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 442         struct ceph_file_info *fi = file->private_data;
 443         struct ceph_rw_context *rw_ctx;
 444         int rc = 0;
 445         int max = 0;
 446
 447         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 448                 return -EINVAL;
 449
 450         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 451                                          &nr_pages);
 452
 453         if (rc == 0)
 454                 goto out;
 455
 456         rw_ctx = ceph_find_rw_context(fi);
 457         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 458         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 459              inode, file, rw_ctx, nr_pages, max);
 460         while (!list_empty(page_list)) {
 461                 rc = start_read(inode, rw_ctx, page_list, max);
 462                 if (rc < 0)
 463                         goto out;
 464         }
 465 out:
 466         ceph_fscache_readpages_cancel(inode, page_list);
 467
 468         dout("readpages %p file %p ret %d\n", inode, file, rc);
 469         return rc;
 470 }
 471
 472 struct ceph_writeback_ctl
 473 {
 474         loff_t i_size;
 475         u64 truncate_size;
 476         u32 truncate_seq;
 477         bool size_stable;
 478         bool head_snapc;
 479 };
 480
 481 /*
 482  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 483  * only snap context we are allowed to write back.
 484  */
 485 static struct ceph_snap_context *
 486 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 487                    struct ceph_snap_context *page_snapc)
 488 {
 489         struct ceph_inode_info *ci = ceph_inode(inode);
 490         struct ceph_snap_context *snapc = NULL;
 491         struct ceph_cap_snap *capsnap = NULL;
 492
 493         spin_lock(&ci->i_ceph_lock);
 494         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 495                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 496                      capsnap->context, capsnap->dirty_pages);
 497                 if (!capsnap->dirty_pages)
 498                         continue;
 499
 500                 /* get i_size, truncate_{seq,size} for page_snapc? */
 501                 if (snapc && capsnap->context != page_snapc)
 502                         continue;
 503
 504                 if (ctl) {
 505                         if (capsnap->writing) {
 506                                 ctl->i_size = i_size_read(inode);
 507                                 ctl->size_stable = false;
 508                         } else {
 509                                 ctl->i_size = capsnap->size;
 510                                 ctl->size_stable = true;
 511                         }
 512                         ctl->truncate_size = capsnap->truncate_size;
 513                         ctl->truncate_seq = capsnap->truncate_seq;
 514                         ctl->head_snapc = false;
 515                 }
 516
 517                 if (snapc)
 518                         break;
 519
 520                 snapc = ceph_get_snap_context(capsnap->context);
 521                 if (!page_snapc ||
 522                     page_snapc == snapc ||
 523                     page_snapc->seq > snapc->seq)
 524                         break;
 525         }
 526         if (!snapc && ci->i_wrbuffer_ref_head) {
 527                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 528                 dout(" head snapc %p has %d dirty pages\n",
 529                      snapc, ci->i_wrbuffer_ref_head);
 530                 if (ctl) {
 531                         ctl->i_size = i_size_read(inode);
 532                         ctl->truncate_size = ci->i_truncate_size;
 533                         ctl->truncate_seq = ci->i_truncate_seq;
 534                         ctl->size_stable = false;
 535                         ctl->head_snapc = true;
 536                 }
 537         }
 538         spin_unlock(&ci->i_ceph_lock);
 539         return snapc;
 540 }
 541
 542 static u64 get_writepages_data_length(struct inode *inode,
 543                                       struct page *page, u64 start)
 544 {
 545         struct ceph_inode_info *ci = ceph_inode(inode);
 546         struct ceph_snap_context *snapc = page_snap_context(page);
 547         struct ceph_cap_snap *capsnap = NULL;
 548         u64 end = i_size_read(inode);
 549
 550         if (snapc != ci->i_head_snapc) {
 551                 bool found = false;
 552                 spin_lock(&ci->i_ceph_lock);
 553                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 554                         if (capsnap->context == snapc) {
 555                                 if (!capsnap->writing)
 556                                         end = capsnap->size;
 557                                 found = true;
 558                                 break;
 559                         }
 560                 }
 561                 spin_unlock(&ci->i_ceph_lock);
 562                 WARN_ON(!found);
 563         }
 564         if (end > page_offset(page) + PAGE_SIZE)
 565                 end = page_offset(page) + PAGE_SIZE;
 566         return end > start ? end - start : 0;
 567 }
 568
 569 /*
 570  * Write a single page, but leave the page locked.
 571  *
 572  * If we get a write error, set the page error bit, but still adjust the
 573  * dirty page accounting (i.e., page is no longer dirty).
 574  */
 575 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 576 {
 577         struct inode *inode;
 578         struct ceph_inode_info *ci;
 579         struct ceph_fs_client *fsc;
 580         struct ceph_snap_context *snapc, *oldest;
 581         loff_t page_off = page_offset(page);
 582         int err, len = PAGE_SIZE;
 583         struct ceph_writeback_ctl ceph_wbc;
 584
 585         dout("writepage %p idx %lu\n", page, page->index);
 586
 587         inode = page->mapping->host;
 588         ci = ceph_inode(inode);
 589         fsc = ceph_inode_to_client(inode);
 590
 591         /* verify this is a writeable snap context */
 592         snapc = page_snap_context(page);
 593         if (!snapc) {
 594                 dout("writepage %p page %p not dirty?\n", inode, page);
 595                 return 0;
 596         }
 597         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 598         if (snapc->seq > oldest->seq) {
 599                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 600                      inode, page, snapc);
 601                 /* we should only noop if called by kswapd */
 602                 WARN_ON(!(current->flags & PF_MEMALLOC));
 603                 ceph_put_snap_context(oldest);
 604                 redirty_page_for_writepage(wbc, page);
 605                 return 0;
 606         }
 607         ceph_put_snap_context(oldest);
 608
 609         /* is this a partial page at end of file? */
 610         if (page_off >= ceph_wbc.i_size) {
 611                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 612                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 613                 return 0;
 614         }
 615
 616         if (ceph_wbc.i_size < page_off + len)
 617                 len = ceph_wbc.i_size - page_off;
 618
 619         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 620              inode, page, page->index, page_off, len, snapc, snapc->seq);
 621
 622         if (atomic_long_inc_return(&fsc->writeback_count) >
 623             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 624                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 625
 626         set_page_writeback(page);
 627         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 628                                    &ci->i_layout, snapc, page_off, len,
 629                                    ceph_wbc.truncate_seq,
 630                                    ceph_wbc.truncate_size,
 631                                    &inode->i_mtime, &page, 1);
 632         if (err < 0) {
 633                 struct writeback_control tmp_wbc;
 634                 if (!wbc)
 635                         wbc = &tmp_wbc;
 636                 if (err == -ERESTARTSYS) {
 637                         /* killed by SIGKILL */
 638                         dout("writepage interrupted page %p\n", page);
 639                         redirty_page_for_writepage(wbc, page);
 640                         end_page_writeback(page);
 641                         return err;
 642                 }
 643                 dout("writepage setting page/mapping error %d %p\n",
 644                      err, page);
 645                 SetPageError(page);
 646                 mapping_set_error(&inode->i_data, err);
 647                 wbc->pages_skipped++;
 648         } else {
 649                 dout("writepage cleaned page %p\n", page);
 650                 err = 0;  /* vfs expects us to return 0 */
 651         }
 652         page->private = 0;
 653         ClearPagePrivate(page);
 654         end_page_writeback(page);
 655         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 656         ceph_put_snap_context(snapc);  /* page's reference */
 657
 658         if (atomic_long_dec_return(&fsc->writeback_count) <
 659             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 660                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 661
 662         return err;
 663 }
 664
 665 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 666 {
 667         int err;
 668         struct inode *inode = page->mapping->host;
 669         BUG_ON(!inode);
 670         ihold(inode);
 671         err = writepage_nounlock(page, wbc);
 672         if (err == -ERESTARTSYS) {
 673                 /* direct memory reclaimer was killed by SIGKILL. return 0
 674                  * to prevent caller from setting mapping/page error */
 675                 err = 0;
 676         }
 677         unlock_page(page);
 678         iput(inode);
 679         return err;
 680 }
 681
 682 /*
 683  * lame release_pages helper.  release_pages() isn't exported to
 684  * modules.
 685  */
 686 static void ceph_release_pages(struct page **pages, int num)
 687 {
 688         struct pagevec pvec;
 689         int i;
 690
 691         pagevec_init(&pvec);
 692         for (i = 0; i < num; i++) {
 693                 if (pagevec_add(&pvec, pages[i]) == 0)
 694                         pagevec_release(&pvec);
 695         }
 696         pagevec_release(&pvec);
 697 }
 698
 699 /*
 700  * async writeback completion handler.
 701  *
 702  * If we get an error, set the mapping error bit, but not the individual
 703  * page error bits.
 704  */
 705 static void writepages_finish(struct ceph_osd_request *req)
 706 {
 707         struct inode *inode = req->r_inode;
 708         struct ceph_inode_info *ci = ceph_inode(inode);
 709         struct ceph_osd_data *osd_data;
 710         struct page *page;
 711         int num_pages, total_pages = 0;
 712         int i, j;
 713         int rc = req->r_result;
 714         struct ceph_snap_context *snapc = req->r_snapc;
 715         struct address_space *mapping = inode->i_mapping;
 716         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 717         bool remove_page;
 718
 719         dout("writepages_finish %p rc %d\n", inode, rc);
 720         if (rc < 0) {
 721                 mapping_set_error(mapping, rc);
 722                 ceph_set_error_write(ci);
 723         } else {
 724                 ceph_clear_error_write(ci);
 725         }
 726
 727         /*
 728          * We lost the cache cap, need to truncate the page before
 729          * it is unlocked, otherwise we'd truncate it later in the
 730          * page truncation thread, possibly losing some data that
 731          * raced its way in
 732          */
 733         remove_page = !(ceph_caps_issued(ci) &
 734                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 735
 736         /* clean all pages */
 737         for (i = 0; i < req->r_num_ops; i++) {
 738                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 739                         break;
 740
 741                 osd_data = osd_req_op_extent_osd_data(req, i);
 742                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 743                 num_pages = calc_pages_for((u64)osd_data->alignment,
 744                                            (u64)osd_data->length);
 745                 total_pages += num_pages;
 746                 for (j = 0; j < num_pages; j++) {
 747                         page = osd_data->pages[j];
 748                         BUG_ON(!page);
 749                         WARN_ON(!PageUptodate(page));
 750
 751                         if (atomic_long_dec_return(&fsc->writeback_count) <
 752                              CONGESTION_OFF_THRESH(
 753                                         fsc->mount_options->congestion_kb))
 754                                 clear_bdi_congested(inode_to_bdi(inode),
 755                                                     BLK_RW_ASYNC);
 756
 757                         ceph_put_snap_context(page_snap_context(page));
 758                         page->private = 0;
 759                         ClearPagePrivate(page);
 760                         dout("unlocking %p\n", page);
 761                         end_page_writeback(page);
 762
 763                         if (remove_page)
 764                                 generic_error_remove_page(inode->i_mapping,
 765                                                           page);
 766
 767                         unlock_page(page);
 768                 }
 769                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 770                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 771
 772                 ceph_release_pages(osd_data->pages, num_pages);
 773         }
 774
 775         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 776
 777         osd_data = osd_req_op_extent_osd_data(req, 0);
 778         if (osd_data->pages_from_pool)
 779                 mempool_free(osd_data->pages,
 780                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 781         else
 782                 kfree(osd_data->pages);
 783         ceph_osdc_put_request(req);
 784 }
 785
 786 /*
 787  * initiate async writeback
 788  */
 789 static int ceph_writepages_start(struct address_space *mapping,
 790                                  struct writeback_control *wbc)
 791 {
 792         struct inode *inode = mapping->host;
 793         struct ceph_inode_info *ci = ceph_inode(inode);
 794         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 795         struct ceph_vino vino = ceph_vino(inode);
 796         pgoff_t index, start_index, end = -1;
 797         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 798         struct pagevec pvec;
 799         int rc = 0;
 800         unsigned int wsize = i_blocksize(inode);
 801         struct ceph_osd_request *req = NULL;
 802         struct ceph_writeback_ctl ceph_wbc;
 803         bool should_loop, range_whole = false;
 804         bool done = false;
 805
 806         dout("writepages_start %p (mode=%s)\n", inode,
 807              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 808              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 809
 810         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 811                 if (ci->i_wrbuffer_ref > 0) {
 812                         pr_warn_ratelimited(
 813                                 "writepage_start %p %lld forced umount\n",
 814                                 inode, ceph_ino(inode));
 815                 }
 816                 mapping_set_error(mapping, -EIO);
 817                 return -EIO; /* we're in a forced umount, don't write! */
 818         }
 819         if (fsc->mount_options->wsize < wsize)
 820                 wsize = fsc->mount_options->wsize;
 821
 822         pagevec_init(&pvec);
 823
 824         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 825         index = start_index;
 826
 827 retry:
 828         /* find oldest snap context with dirty data */
 829         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 830         if (!snapc) {
 831                 /* hmm, why does writepages get called when there
 832                    is no dirty data? */
 833                 dout(" no snap context with dirty data?\n");
 834                 goto out;
 835         }
 836         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 837              snapc, snapc->seq, snapc->num_snaps);
 838
 839         should_loop = false;
 840         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 841                 /* where to start/end? */
 842                 if (wbc->range_cyclic) {
 843                         index = start_index;
 844                         end = -1;
 845                         if (index > 0)
 846                                 should_loop = true;
 847                         dout(" cyclic, start at %lu\n", index);
 848                 } else {
 849                         index = wbc->range_start >> PAGE_SHIFT;
 850                         end = wbc->range_end >> PAGE_SHIFT;
 851                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 852                                 range_whole = true;
 853                         dout(" not cyclic, %lu to %lu\n", index, end);
 854                 }
 855         } else if (!ceph_wbc.head_snapc) {
 856                 /* Do not respect wbc->range_{start,end}. Dirty pages
 857                  * in that range can be associated with newer snapc.
 858                  * They are not writeable until we write all dirty pages
 859                  * associated with 'snapc' get written */
 860                 if (index > 0)
 861                         should_loop = true;
 862                 dout(" non-head snapc, range whole\n");
 863         }
 864
 865         ceph_put_snap_context(last_snapc);
 866         last_snapc = snapc;
 867
 868         while (!done && index <= end) {
 869                 int num_ops = 0, op_idx;
 870                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 871                 struct page **pages = NULL, **data_pages;
 872                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 873                 struct page *page;
 874                 pgoff_t strip_unit_end = 0;
 875                 u64 offset = 0, len = 0;
 876
 877                 max_pages = wsize >> PAGE_SHIFT;
 878
 879 get_more_pages:
 880                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 881                                                 end, PAGECACHE_TAG_DIRTY,
 882                                                 max_pages - locked_pages);
 883                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 884                 if (!pvec_pages && !locked_pages)
 885                         break;
 886                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 887                         page = pvec.pages[i];
 888                         dout("? %p idx %lu\n", page, page->index);
 889                         if (locked_pages == 0)
 890                                 lock_page(page);  /* first page */
 891                         else if (!trylock_page(page))
 892                                 break;
 893
 894                         /* only dirty pages, or our accounting breaks */
 895                         if (unlikely(!PageDirty(page)) ||
 896                             unlikely(page->mapping != mapping)) {
 897                                 dout("!dirty or !mapping %p\n", page);
 898                                 unlock_page(page);
 899                                 continue;
 900                         }
 901                         /* only if matching snap context */
 902                         pgsnapc = page_snap_context(page);
 903                         if (pgsnapc != snapc) {
 904                                 dout("page snapc %p %lld != oldest %p %lld\n",
 905                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 906                                 if (!should_loop &&
 907                                     !ceph_wbc.head_snapc &&
 908                                     wbc->sync_mode != WB_SYNC_NONE)
 909                                         should_loop = true;
 910                                 unlock_page(page);
 911                                 continue;
 912                         }
 913                         if (page_offset(page) >= ceph_wbc.i_size) {
 914                                 dout("%p page eof %llu\n",
 915                                      page, ceph_wbc.i_size);
 916                                 if (ceph_wbc.size_stable ||
 917                                     page_offset(page) >= i_size_read(inode))
 918                                         mapping->a_ops->invalidatepage(page,
 919                                                                 0, PAGE_SIZE);
 920                                 unlock_page(page);
 921                                 continue;
 922                         }
 923                         if (strip_unit_end && (page->index > strip_unit_end)) {
 924                                 dout("end of strip unit %p\n", page);
 925                                 unlock_page(page);
 926                                 break;
 927                         }
 928                         if (PageWriteback(page)) {
 929                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 930                                         dout("%p under writeback\n", page);
 931                                         unlock_page(page);
 932                                         continue;
 933                                 }
 934                                 dout("waiting on writeback %p\n", page);
 935                                 wait_on_page_writeback(page);
 936                         }
 937
 938                         if (!clear_page_dirty_for_io(page)) {
 939                                 dout("%p !clear_page_dirty_for_io\n", page);
 940                                 unlock_page(page);
 941                                 continue;
 942                         }
 943
 944                         /*
 945                          * We have something to write.  If this is
 946                          * the first locked page this time through,
 947                          * calculate max possinle write size and
 948                          * allocate a page array
 949                          */
 950                         if (locked_pages == 0) {
 951                                 u64 objnum;
 952                                 u64 objoff;
 953                                 u32 xlen;
 954
 955                                 /* prepare async write request */
 956                                 offset = (u64)page_offset(page);
 957                                 ceph_calc_file_object_mapping(&ci->i_layout,
 958                                                               offset, wsize,
 959                                                               &objnum, &objoff,
 960                                                               &xlen);
 961                                 len = xlen;
 962
 963                                 num_ops = 1;
 964                                 strip_unit_end = page->index +
 965                                         ((len - 1) >> PAGE_SHIFT);
 966
 967                                 BUG_ON(pages);
 968                                 max_pages = calc_pages_for(0, (u64)len);
 969                                 pages = kmalloc(max_pages * 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(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 int 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, ret;
1440         sigset_t oldset;
1441
1442         ceph_block_sigs(&oldset);
1443
1444         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1445              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1446         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1447                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1448         else
1449                 want = CEPH_CAP_FILE_CACHE;
1450
1451         got = 0;
1452         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1453         if (ret < 0)
1454                 goto out_restore;
1455
1456         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1457              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1458
1459         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1460             ci->i_inline_version == CEPH_INLINE_NONE) {
1461                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1462                 ceph_add_rw_context(fi, &rw_ctx);
1463                 ret = filemap_fault(vmf);
1464                 ceph_del_rw_context(fi, &rw_ctx);
1465         } else
1466                 ret = -EAGAIN;
1467
1468         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1469              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1470         if (pinned_page)
1471                 put_page(pinned_page);
1472         ceph_put_cap_refs(ci, got);
1473
1474         if (ret != -EAGAIN)
1475                 goto out_restore;
1476
1477         /* read inline data */
1478         if (off >= PAGE_SIZE) {
1479                 /* does not support inline data > PAGE_SIZE */
1480                 ret = VM_FAULT_SIGBUS;
1481         } else {
1482                 int ret1;
1483                 struct address_space *mapping = inode->i_mapping;
1484                 struct page *page = find_or_create_page(mapping, 0,
1485                                                 mapping_gfp_constraint(mapping,
1486                                                 ~__GFP_FS));
1487                 if (!page) {
1488                         ret = VM_FAULT_OOM;
1489                         goto out_inline;
1490                 }
1491                 ret1 = __ceph_do_getattr(inode, page,
1492                                          CEPH_STAT_CAP_INLINE_DATA, true);
1493                 if (ret1 < 0 || off >= i_size_read(inode)) {
1494                         unlock_page(page);
1495                         put_page(page);
1496                         if (ret1 < 0)
1497                                 ret = ret1;
1498                         else
1499                                 ret = VM_FAULT_SIGBUS;
1500                         goto out_inline;
1501                 }
1502                 if (ret1 < PAGE_SIZE)
1503                         zero_user_segment(page, ret1, PAGE_SIZE);
1504                 else
1505                         flush_dcache_page(page);
1506                 SetPageUptodate(page);
1507                 vmf->page = page;
1508                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1509 out_inline:
1510                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1511                      inode, off, (size_t)PAGE_SIZE, ret);
1512         }
1513 out_restore:
1514         ceph_restore_sigs(&oldset);
1515         if (ret < 0)
1516                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1517
1518         return ret;
1519 }
1520
1521 /*
1522  * Reuse write_begin here for simplicity.
1523  */
1524 static int ceph_page_mkwrite(struct vm_fault *vmf)
1525 {
1526         struct vm_area_struct *vma = vmf->vma;
1527         struct inode *inode = file_inode(vma->vm_file);
1528         struct ceph_inode_info *ci = ceph_inode(inode);
1529         struct ceph_file_info *fi = vma->vm_file->private_data;
1530         struct ceph_cap_flush *prealloc_cf;
1531         struct page *page = vmf->page;
1532         loff_t off = page_offset(page);
1533         loff_t size = i_size_read(inode);
1534         size_t len;
1535         int want, got, ret;
1536         sigset_t oldset;
1537
1538         prealloc_cf = ceph_alloc_cap_flush();
1539         if (!prealloc_cf)
1540                 return VM_FAULT_OOM;
1541
1542         ceph_block_sigs(&oldset);
1543
1544         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1545                 struct page *locked_page = NULL;
1546                 if (off == 0) {
1547                         lock_page(page);
1548                         locked_page = page;
1549                 }
1550                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1551                 if (locked_page)
1552                         unlock_page(locked_page);
1553                 if (ret < 0)
1554                         goto out_free;
1555         }
1556
1557         if (off + PAGE_SIZE <= size)
1558                 len = PAGE_SIZE;
1559         else
1560                 len = size & ~PAGE_MASK;
1561
1562         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1563              inode, ceph_vinop(inode), off, len, size);
1564         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1565                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1566         else
1567                 want = CEPH_CAP_FILE_BUFFER;
1568
1569         got = 0;
1570         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1571                             &got, NULL);
1572         if (ret < 0)
1573                 goto out_free;
1574
1575         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1576              inode, off, len, ceph_cap_string(got));
1577
1578         /* Update time before taking page lock */
1579         file_update_time(vma->vm_file);
1580
1581         do {
1582                 lock_page(page);
1583
1584                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1585                         unlock_page(page);
1586                         ret = VM_FAULT_NOPAGE;
1587                         break;
1588                 }
1589
1590                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1591                 if (ret >= 0) {
1592                         /* success.  we'll keep the page locked. */
1593                         set_page_dirty(page);
1594                         ret = VM_FAULT_LOCKED;
1595                 }
1596         } while (ret == -EAGAIN);
1597
1598         if (ret == VM_FAULT_LOCKED ||
1599             ci->i_inline_version != CEPH_INLINE_NONE) {
1600                 int dirty;
1601                 spin_lock(&ci->i_ceph_lock);
1602                 ci->i_inline_version = CEPH_INLINE_NONE;
1603                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1604                                                &prealloc_cf);
1605                 spin_unlock(&ci->i_ceph_lock);
1606                 if (dirty)
1607                         __mark_inode_dirty(inode, dirty);
1608         }
1609
1610         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1611              inode, off, len, ceph_cap_string(got), ret);
1612         ceph_put_cap_refs(ci, got);
1613 out_free:
1614         ceph_restore_sigs(&oldset);
1615         ceph_free_cap_flush(prealloc_cf);
1616         if (ret < 0)
1617                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1618         return ret;
1619 }
1620
1621 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1622                            char *data, size_t len)
1623 {
1624         struct address_space *mapping = inode->i_mapping;
1625         struct page *page;
1626
1627         if (locked_page) {
1628                 page = locked_page;
1629         } else {
1630                 if (i_size_read(inode) == 0)
1631                         return;
1632                 page = find_or_create_page(mapping, 0,
1633                                            mapping_gfp_constraint(mapping,
1634                                            ~__GFP_FS));
1635                 if (!page)
1636                         return;
1637                 if (PageUptodate(page)) {
1638                         unlock_page(page);
1639                         put_page(page);
1640                         return;
1641                 }
1642         }
1643
1644         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1645              inode, ceph_vinop(inode), len, locked_page);
1646
1647         if (len > 0) {
1648                 void *kaddr = kmap_atomic(page);
1649                 memcpy(kaddr, data, len);
1650                 kunmap_atomic(kaddr);
1651         }
1652
1653         if (page != locked_page) {
1654                 if (len < PAGE_SIZE)
1655                         zero_user_segment(page, len, PAGE_SIZE);
1656                 else
1657                         flush_dcache_page(page);
1658
1659                 SetPageUptodate(page);
1660                 unlock_page(page);
1661                 put_page(page);
1662         }
1663 }
1664
1665 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1666 {
1667         struct inode *inode = file_inode(filp);
1668         struct ceph_inode_info *ci = ceph_inode(inode);
1669         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1670         struct ceph_osd_request *req;
1671         struct page *page = NULL;
1672         u64 len, inline_version;
1673         int err = 0;
1674         bool from_pagecache = false;
1675
1676         spin_lock(&ci->i_ceph_lock);
1677         inline_version = ci->i_inline_version;
1678         spin_unlock(&ci->i_ceph_lock);
1679
1680         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1681              inode, ceph_vinop(inode), inline_version);
1682
1683         if (inline_version == 1 || /* initial version, no data */
1684             inline_version == CEPH_INLINE_NONE)
1685                 goto out;
1686
1687         if (locked_page) {
1688                 page = locked_page;
1689                 WARN_ON(!PageUptodate(page));
1690         } else if (ceph_caps_issued(ci) &
1691                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1692                 page = find_get_page(inode->i_mapping, 0);
1693                 if (page) {
1694                         if (PageUptodate(page)) {
1695                                 from_pagecache = true;
1696                                 lock_page(page);
1697                         } else {
1698                                 put_page(page);
1699                                 page = NULL;
1700                         }
1701                 }
1702         }
1703
1704         if (page) {
1705                 len = i_size_read(inode);
1706                 if (len > PAGE_SIZE)
1707                         len = PAGE_SIZE;
1708         } else {
1709                 page = __page_cache_alloc(GFP_NOFS);
1710                 if (!page) {
1711                         err = -ENOMEM;
1712                         goto out;
1713                 }
1714                 err = __ceph_do_getattr(inode, page,
1715                                         CEPH_STAT_CAP_INLINE_DATA, true);
1716                 if (err < 0) {
1717                         /* no inline data */
1718                         if (err == -ENODATA)
1719                                 err = 0;
1720                         goto out;
1721                 }
1722                 len = err;
1723         }
1724
1725         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1726                                     ceph_vino(inode), 0, &len, 0, 1,
1727                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1728                                     NULL, 0, 0, false);
1729         if (IS_ERR(req)) {
1730                 err = PTR_ERR(req);
1731                 goto out;
1732         }
1733
1734         req->r_mtime = inode->i_mtime;
1735         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1736         if (!err)
1737                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1738         ceph_osdc_put_request(req);
1739         if (err < 0)
1740                 goto out;
1741
1742         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1743                                     ceph_vino(inode), 0, &len, 1, 3,
1744                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1745                                     NULL, ci->i_truncate_seq,
1746                                     ci->i_truncate_size, false);
1747         if (IS_ERR(req)) {
1748                 err = PTR_ERR(req);
1749                 goto out;
1750         }
1751
1752         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1753
1754         {
1755                 __le64 xattr_buf = cpu_to_le64(inline_version);
1756                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1757                                             "inline_version", &xattr_buf,
1758                                             sizeof(xattr_buf),
1759                                             CEPH_OSD_CMPXATTR_OP_GT,
1760                                             CEPH_OSD_CMPXATTR_MODE_U64);
1761                 if (err)
1762                         goto out_put;
1763         }
1764
1765         {
1766                 char xattr_buf[32];
1767                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1768                                          "%llu", inline_version);
1769                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1770                                             "inline_version",
1771                                             xattr_buf, xattr_len, 0, 0);
1772                 if (err)
1773                         goto out_put;
1774         }
1775
1776         req->r_mtime = inode->i_mtime;
1777         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1778         if (!err)
1779                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1780 out_put:
1781         ceph_osdc_put_request(req);
1782         if (err == -ECANCELED)
1783                 err = 0;
1784 out:
1785         if (page && page != locked_page) {
1786                 if (from_pagecache) {
1787                         unlock_page(page);
1788                         put_page(page);
1789                 } else
1790                         __free_pages(page, 0);
1791         }
1792
1793         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1794              inode, ceph_vinop(inode), inline_version, err);
1795         return err;
1796 }
1797
1798 static const struct vm_operations_struct ceph_vmops = {
1799         .fault          = ceph_filemap_fault,
1800         .page_mkwrite   = ceph_page_mkwrite,
1801 };
1802
1803 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1804 {
1805         struct address_space *mapping = file->f_mapping;
1806
1807         if (!mapping->a_ops->readpage)
1808                 return -ENOEXEC;
1809         file_accessed(file);
1810         vma->vm_ops = &ceph_vmops;
1811         return 0;
1812 }
1813
1814 enum {
1815         POOL_READ       = 1,
1816         POOL_WRITE      = 2,
1817 };
1818
1819 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1820                                 s64 pool, struct ceph_string *pool_ns)
1821 {
1822         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1823         struct ceph_mds_client *mdsc = fsc->mdsc;
1824         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1825         struct rb_node **p, *parent;
1826         struct ceph_pool_perm *perm;
1827         struct page **pages;
1828         size_t pool_ns_len;
1829         int err = 0, err2 = 0, have = 0;
1830
1831         down_read(&mdsc->pool_perm_rwsem);
1832         p = &mdsc->pool_perm_tree.rb_node;
1833         while (*p) {
1834                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1835                 if (pool < perm->pool)
1836                         p = &(*p)->rb_left;
1837                 else if (pool > perm->pool)
1838                         p = &(*p)->rb_right;
1839                 else {
1840                         int ret = ceph_compare_string(pool_ns,
1841                                                 perm->pool_ns,
1842                                                 perm->pool_ns_len);
1843                         if (ret < 0)
1844                                 p = &(*p)->rb_left;
1845                         else if (ret > 0)
1846                                 p = &(*p)->rb_right;
1847                         else {
1848                                 have = perm->perm;
1849                                 break;
1850                         }
1851                 }
1852         }
1853         up_read(&mdsc->pool_perm_rwsem);
1854         if (*p)
1855                 goto out;
1856
1857         if (pool_ns)
1858                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1859                      pool, (int)pool_ns->len, pool_ns->str);
1860         else
1861                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1862
1863         down_write(&mdsc->pool_perm_rwsem);
1864         p = &mdsc->pool_perm_tree.rb_node;
1865         parent = NULL;
1866         while (*p) {
1867                 parent = *p;
1868                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1869                 if (pool < perm->pool)
1870                         p = &(*p)->rb_left;
1871                 else if (pool > perm->pool)
1872                         p = &(*p)->rb_right;
1873                 else {
1874                         int ret = ceph_compare_string(pool_ns,
1875                                                 perm->pool_ns,
1876                                                 perm->pool_ns_len);
1877                         if (ret < 0)
1878                                 p = &(*p)->rb_left;
1879                         else if (ret > 0)
1880                                 p = &(*p)->rb_right;
1881                         else {
1882                                 have = perm->perm;
1883                                 break;
1884                         }
1885                 }
1886         }
1887         if (*p) {
1888                 up_write(&mdsc->pool_perm_rwsem);
1889                 goto out;
1890         }
1891
1892         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1893                                          1, false, GFP_NOFS);
1894         if (!rd_req) {
1895                 err = -ENOMEM;
1896                 goto out_unlock;
1897         }
1898
1899         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1900         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1901         rd_req->r_base_oloc.pool = pool;
1902         if (pool_ns)
1903                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1904         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1905
1906         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1907         if (err)
1908                 goto out_unlock;
1909
1910         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1911                                          1, false, GFP_NOFS);
1912         if (!wr_req) {
1913                 err = -ENOMEM;
1914                 goto out_unlock;
1915         }
1916
1917         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1918         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1919         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1920         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1921
1922         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1923         if (err)
1924                 goto out_unlock;
1925
1926         /* one page should be large enough for STAT data */
1927         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1928         if (IS_ERR(pages)) {
1929                 err = PTR_ERR(pages);
1930                 goto out_unlock;
1931         }
1932
1933         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1934                                      0, false, true);
1935         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1936
1937         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1938         wr_req->r_abort_on_full = true;
1939         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1940
1941         if (!err)
1942                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1943         if (!err2)
1944                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1945
1946         if (err >= 0 || err == -ENOENT)
1947                 have |= POOL_READ;
1948         else if (err != -EPERM)
1949                 goto out_unlock;
1950
1951         if (err2 == 0 || err2 == -EEXIST)
1952                 have |= POOL_WRITE;
1953         else if (err2 != -EPERM) {
1954                 err = err2;
1955                 goto out_unlock;
1956         }
1957
1958         pool_ns_len = pool_ns ? pool_ns->len : 0;
1959         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1960         if (!perm) {
1961                 err = -ENOMEM;
1962                 goto out_unlock;
1963         }
1964
1965         perm->pool = pool;
1966         perm->perm = have;
1967         perm->pool_ns_len = pool_ns_len;
1968         if (pool_ns_len > 0)
1969                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1970         perm->pool_ns[pool_ns_len] = 0;
1971
1972         rb_link_node(&perm->node, parent, p);
1973         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1974         err = 0;
1975 out_unlock:
1976         up_write(&mdsc->pool_perm_rwsem);
1977
1978         ceph_osdc_put_request(rd_req);
1979         ceph_osdc_put_request(wr_req);
1980 out:
1981         if (!err)
1982                 err = have;
1983         if (pool_ns)
1984                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1985                      pool, (int)pool_ns->len, pool_ns->str, err);
1986         else
1987                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1988         return err;
1989 }
1990
1991 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1992 {
1993         s64 pool;
1994         struct ceph_string *pool_ns;
1995         int ret, flags;
1996
1997         if (ci->i_vino.snap != CEPH_NOSNAP) {
1998                 /*
1999                  * Pool permission check needs to write to the first object.
2000                  * But for snapshot, head of the first object may have alread
2001                  * been deleted. Skip check to avoid creating orphan object.
2002                  */
2003                 return 0;
2004         }
2005
2006         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2007                                 NOPOOLPERM))
2008                 return 0;
2009
2010         spin_lock(&ci->i_ceph_lock);
2011         flags = ci->i_ceph_flags;
2012         pool = ci->i_layout.pool_id;
2013         spin_unlock(&ci->i_ceph_lock);
2014 check:
2015         if (flags & CEPH_I_POOL_PERM) {
2016                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2017                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2018                              pool);
2019                         return -EPERM;
2020                 }
2021                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2022                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2023                              pool);
2024                         return -EPERM;
2025                 }
2026                 return 0;
2027         }
2028
2029         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2030         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2031         ceph_put_string(pool_ns);
2032         if (ret < 0)
2033                 return ret;
2034
2035         flags = CEPH_I_POOL_PERM;
2036         if (ret & POOL_READ)
2037                 flags |= CEPH_I_POOL_RD;
2038         if (ret & POOL_WRITE)
2039                 flags |= CEPH_I_POOL_WR;
2040
2041         spin_lock(&ci->i_ceph_lock);
2042         if (pool == ci->i_layout.pool_id &&
2043             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2044                 ci->i_ceph_flags |= flags;
2045         } else {
2046                 pool = ci->i_layout.pool_id;
2047                 flags = ci->i_ceph_flags;
2048         }
2049         spin_unlock(&ci->i_ceph_lock);
2050         goto check;
2051 }
2052
2053 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2054 {
2055         struct ceph_pool_perm *perm;
2056         struct rb_node *n;
2057
2058         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2059                 n = rb_first(&mdsc->pool_perm_tree);
2060                 perm = rb_entry(n, struct ceph_pool_perm, node);
2061                 rb_erase(n, &mdsc->pool_perm_tree);
2062                 kfree(perm);
2063         }
2064 }