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_array(nr_pages, sizeof(*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 timespec ts;
 578         struct inode *inode;
 579         struct ceph_inode_info *ci;
 580         struct ceph_fs_client *fsc;
 581         struct ceph_snap_context *snapc, *oldest;
 582         loff_t page_off = page_offset(page);
 583         int err, len = PAGE_SIZE;
 584         struct ceph_writeback_ctl ceph_wbc;
 585
 586         dout("writepage %p idx %lu\n", page, page->index);
 587
 588         inode = page->mapping->host;
 589         ci = ceph_inode(inode);
 590         fsc = ceph_inode_to_client(inode);
 591
 592         /* verify this is a writeable snap context */
 593         snapc = page_snap_context(page);
 594         if (!snapc) {
 595                 dout("writepage %p page %p not dirty?\n", inode, page);
 596                 return 0;
 597         }
 598         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 599         if (snapc->seq > oldest->seq) {
 600                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 601                      inode, page, snapc);
 602                 /* we should only noop if called by kswapd */
 603                 WARN_ON(!(current->flags & PF_MEMALLOC));
 604                 ceph_put_snap_context(oldest);
 605                 redirty_page_for_writepage(wbc, page);
 606                 return 0;
 607         }
 608         ceph_put_snap_context(oldest);
 609
 610         /* is this a partial page at end of file? */
 611         if (page_off >= ceph_wbc.i_size) {
 612                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 613                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 614                 return 0;
 615         }
 616
 617         if (ceph_wbc.i_size < page_off + len)
 618                 len = ceph_wbc.i_size - page_off;
 619
 620         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
 621              inode, page, page->index, page_off, len, snapc, snapc->seq);
 622
 623         if (atomic_long_inc_return(&fsc->writeback_count) >
 624             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 625                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 626
 627         set_page_writeback(page);
 628         ts = timespec64_to_timespec(inode->i_mtime);
 629         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
 630                                    &ci->i_layout, snapc, page_off, len,
 631                                    ceph_wbc.truncate_seq,
 632                                    ceph_wbc.truncate_size,
 633                                    &ts, &page, 1);
 634         if (err < 0) {
 635                 struct writeback_control tmp_wbc;
 636                 if (!wbc)
 637                         wbc = &tmp_wbc;
 638                 if (err == -ERESTARTSYS) {
 639                         /* killed by SIGKILL */
 640                         dout("writepage interrupted page %p\n", page);
 641                         redirty_page_for_writepage(wbc, page);
 642                         end_page_writeback(page);
 643                         return err;
 644                 }
 645                 dout("writepage setting page/mapping error %d %p\n",
 646                      err, page);
 647                 SetPageError(page);
 648                 mapping_set_error(&inode->i_data, err);
 649                 wbc->pages_skipped++;
 650         } else {
 651                 dout("writepage cleaned page %p\n", page);
 652                 err = 0;  /* vfs expects us to return 0 */
 653         }
 654         page->private = 0;
 655         ClearPagePrivate(page);
 656         end_page_writeback(page);
 657         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 658         ceph_put_snap_context(snapc);  /* page's reference */
 659
 660         if (atomic_long_dec_return(&fsc->writeback_count) <
 661             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 662                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 663
 664         return err;
 665 }
 666
 667 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 668 {
 669         int err;
 670         struct inode *inode = page->mapping->host;
 671         BUG_ON(!inode);
 672         ihold(inode);
 673         err = writepage_nounlock(page, wbc);
 674         if (err == -ERESTARTSYS) {
 675                 /* direct memory reclaimer was killed by SIGKILL. return 0
 676                  * to prevent caller from setting mapping/page error */
 677                 err = 0;
 678         }
 679         unlock_page(page);
 680         iput(inode);
 681         return err;
 682 }
 683
 684 /*
 685  * lame release_pages helper.  release_pages() isn't exported to
 686  * modules.
 687  */
 688 static void ceph_release_pages(struct page **pages, int num)
 689 {
 690         struct pagevec pvec;
 691         int i;
 692
 693         pagevec_init(&pvec);
 694         for (i = 0; i < num; i++) {
 695                 if (pagevec_add(&pvec, pages[i]) == 0)
 696                         pagevec_release(&pvec);
 697         }
 698         pagevec_release(&pvec);
 699 }
 700
 701 /*
 702  * async writeback completion handler.
 703  *
 704  * If we get an error, set the mapping error bit, but not the individual
 705  * page error bits.
 706  */
 707 static void writepages_finish(struct ceph_osd_request *req)
 708 {
 709         struct inode *inode = req->r_inode;
 710         struct ceph_inode_info *ci = ceph_inode(inode);
 711         struct ceph_osd_data *osd_data;
 712         struct page *page;
 713         int num_pages, total_pages = 0;
 714         int i, j;
 715         int rc = req->r_result;
 716         struct ceph_snap_context *snapc = req->r_snapc;
 717         struct address_space *mapping = inode->i_mapping;
 718         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 719         bool remove_page;
 720
 721         dout("writepages_finish %p rc %d\n", inode, rc);
 722         if (rc < 0) {
 723                 mapping_set_error(mapping, rc);
 724                 ceph_set_error_write(ci);
 725         } else {
 726                 ceph_clear_error_write(ci);
 727         }
 728
 729         /*
 730          * We lost the cache cap, need to truncate the page before
 731          * it is unlocked, otherwise we'd truncate it later in the
 732          * page truncation thread, possibly losing some data that
 733          * raced its way in
 734          */
 735         remove_page = !(ceph_caps_issued(ci) &
 736                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 737
 738         /* clean all pages */
 739         for (i = 0; i < req->r_num_ops; i++) {
 740                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 741                         break;
 742
 743                 osd_data = osd_req_op_extent_osd_data(req, i);
 744                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 745                 num_pages = calc_pages_for((u64)osd_data->alignment,
 746                                            (u64)osd_data->length);
 747                 total_pages += num_pages;
 748                 for (j = 0; j < num_pages; j++) {
 749                         page = osd_data->pages[j];
 750                         BUG_ON(!page);
 751                         WARN_ON(!PageUptodate(page));
 752
 753                         if (atomic_long_dec_return(&fsc->writeback_count) <
 754                              CONGESTION_OFF_THRESH(
 755                                         fsc->mount_options->congestion_kb))
 756                                 clear_bdi_congested(inode_to_bdi(inode),
 757                                                     BLK_RW_ASYNC);
 758
 759                         ceph_put_snap_context(page_snap_context(page));
 760                         page->private = 0;
 761                         ClearPagePrivate(page);
 762                         dout("unlocking %p\n", page);
 763                         end_page_writeback(page);
 764
 765                         if (remove_page)
 766                                 generic_error_remove_page(inode->i_mapping,
 767                                                           page);
 768
 769                         unlock_page(page);
 770                 }
 771                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 772                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 773
 774                 ceph_release_pages(osd_data->pages, num_pages);
 775         }
 776
 777         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 778
 779         osd_data = osd_req_op_extent_osd_data(req, 0);
 780         if (osd_data->pages_from_pool)
 781                 mempool_free(osd_data->pages,
 782                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 783         else
 784                 kfree(osd_data->pages);
 785         ceph_osdc_put_request(req);
 786 }
 787
 788 /*
 789  * initiate async writeback
 790  */
 791 static int ceph_writepages_start(struct address_space *mapping,
 792                                  struct writeback_control *wbc)
 793 {
 794         struct inode *inode = mapping->host;
 795         struct ceph_inode_info *ci = ceph_inode(inode);
 796         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 797         struct ceph_vino vino = ceph_vino(inode);
 798         pgoff_t index, start_index, end = -1;
 799         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 800         struct pagevec pvec;
 801         int rc = 0;
 802         unsigned int wsize = i_blocksize(inode);
 803         struct ceph_osd_request *req = NULL;
 804         struct ceph_writeback_ctl ceph_wbc;
 805         bool should_loop, range_whole = false;
 806         bool done = false;
 807
 808         dout("writepages_start %p (mode=%s)\n", inode,
 809              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 810              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 811
 812         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 813                 if (ci->i_wrbuffer_ref > 0) {
 814                         pr_warn_ratelimited(
 815                                 "writepage_start %p %lld forced umount\n",
 816                                 inode, ceph_ino(inode));
 817                 }
 818                 mapping_set_error(mapping, -EIO);
 819                 return -EIO; /* we're in a forced umount, don't write! */
 820         }
 821         if (fsc->mount_options->wsize < wsize)
 822                 wsize = fsc->mount_options->wsize;
 823
 824         pagevec_init(&pvec);
 825
 826         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 827         index = start_index;
 828
 829 retry:
 830         /* find oldest snap context with dirty data */
 831         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 832         if (!snapc) {
 833                 /* hmm, why does writepages get called when there
 834                    is no dirty data? */
 835                 dout(" no snap context with dirty data?\n");
 836                 goto out;
 837         }
 838         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 839              snapc, snapc->seq, snapc->num_snaps);
 840
 841         should_loop = false;
 842         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 843                 /* where to start/end? */
 844                 if (wbc->range_cyclic) {
 845                         index = start_index;
 846                         end = -1;
 847                         if (index > 0)
 848                                 should_loop = true;
 849                         dout(" cyclic, start at %lu\n", index);
 850                 } else {
 851                         index = wbc->range_start >> PAGE_SHIFT;
 852                         end = wbc->range_end >> PAGE_SHIFT;
 853                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 854                                 range_whole = true;
 855                         dout(" not cyclic, %lu to %lu\n", index, end);
 856                 }
 857         } else if (!ceph_wbc.head_snapc) {
 858                 /* Do not respect wbc->range_{start,end}. Dirty pages
 859                  * in that range can be associated with newer snapc.
 860                  * They are not writeable until we write all dirty pages
 861                  * associated with 'snapc' get written */
 862                 if (index > 0)
 863                         should_loop = true;
 864                 dout(" non-head snapc, range whole\n");
 865         }
 866
 867         ceph_put_snap_context(last_snapc);
 868         last_snapc = snapc;
 869
 870         while (!done && index <= end) {
 871                 int num_ops = 0, op_idx;
 872                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 873                 struct page **pages = NULL, **data_pages;
 874                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 875                 struct page *page;
 876                 pgoff_t strip_unit_end = 0;
 877                 u64 offset = 0, len = 0;
 878
 879                 max_pages = wsize >> PAGE_SHIFT;
 880
 881 get_more_pages:
 882                 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
 883                                                 end, PAGECACHE_TAG_DIRTY,
 884                                                 max_pages - locked_pages);
 885                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 886                 if (!pvec_pages && !locked_pages)
 887                         break;
 888                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 889                         page = pvec.pages[i];
 890                         dout("? %p idx %lu\n", page, page->index);
 891                         if (locked_pages == 0)
 892                                 lock_page(page);  /* first page */
 893                         else if (!trylock_page(page))
 894                                 break;
 895
 896                         /* only dirty pages, or our accounting breaks */
 897                         if (unlikely(!PageDirty(page)) ||
 898                             unlikely(page->mapping != mapping)) {
 899                                 dout("!dirty or !mapping %p\n", page);
 900                                 unlock_page(page);
 901                                 continue;
 902                         }
 903                         /* only if matching snap context */
 904                         pgsnapc = page_snap_context(page);
 905                         if (pgsnapc != snapc) {
 906                                 dout("page snapc %p %lld != oldest %p %lld\n",
 907                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 908                                 if (!should_loop &&
 909                                     !ceph_wbc.head_snapc &&
 910                                     wbc->sync_mode != WB_SYNC_NONE)
 911                                         should_loop = true;
 912                                 unlock_page(page);
 913                                 continue;
 914                         }
 915                         if (page_offset(page) >= ceph_wbc.i_size) {
 916                                 dout("%p page eof %llu\n",
 917                                      page, ceph_wbc.i_size);
 918                                 if (ceph_wbc.size_stable ||
 919                                     page_offset(page) >= i_size_read(inode))
 920                                         mapping->a_ops->invalidatepage(page,
 921                                                                 0, PAGE_SIZE);
 922                                 unlock_page(page);
 923                                 continue;
 924                         }
 925                         if (strip_unit_end && (page->index > strip_unit_end)) {
 926                                 dout("end of strip unit %p\n", page);
 927                                 unlock_page(page);
 928                                 break;
 929                         }
 930                         if (PageWriteback(page)) {
 931                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 932                                         dout("%p under writeback\n", page);
 933                                         unlock_page(page);
 934                                         continue;
 935                                 }
 936                                 dout("waiting on writeback %p\n", page);
 937                                 wait_on_page_writeback(page);
 938                         }
 939
 940                         if (!clear_page_dirty_for_io(page)) {
 941                                 dout("%p !clear_page_dirty_for_io\n", page);
 942                                 unlock_page(page);
 943                                 continue;
 944                         }
 945
 946                         /*
 947                          * We have something to write.  If this is
 948                          * the first locked page this time through,
 949                          * calculate max possinle write size and
 950                          * allocate a page array
 951                          */
 952                         if (locked_pages == 0) {
 953                                 u64 objnum;
 954                                 u64 objoff;
 955                                 u32 xlen;
 956
 957                                 /* prepare async write request */
 958                                 offset = (u64)page_offset(page);
 959                                 ceph_calc_file_object_mapping(&ci->i_layout,
 960                                                               offset, wsize,
 961                                                               &objnum, &objoff,
 962                                                               &xlen);
 963                                 len = xlen;
 964
 965                                 num_ops = 1;
 966                                 strip_unit_end = page->index +
 967                                         ((len - 1) >> PAGE_SHIFT);
 968
 969                                 BUG_ON(pages);
 970                                 max_pages = calc_pages_for(0, (u64)len);
 971                                 pages = kmalloc_array(max_pages,
 972                                                       sizeof(*pages),
 973                                                       GFP_NOFS);
 974                                 if (!pages) {
 975                                         pool = fsc->wb_pagevec_pool;
 976                                         pages = mempool_alloc(pool, GFP_NOFS);
 977                                         BUG_ON(!pages);
 978                                 }
 979
 980                                 len = 0;
 981                         } else if (page->index !=
 982                                    (offset + len) >> PAGE_SHIFT) {
 983                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
 984                                                         CEPH_OSD_MAX_OPS)) {
 985                                         redirty_page_for_writepage(wbc, page);
 986                                         unlock_page(page);
 987                                         break;
 988                                 }
 989
 990                                 num_ops++;
 991                                 offset = (u64)page_offset(page);
 992                                 len = 0;
 993                         }
 994
 995                         /* note position of first page in pvec */
 996                         dout("%p will write page %p idx %lu\n",
 997                              inode, page, page->index);
 998
 999                         if (atomic_long_inc_return(&fsc->writeback_count) >
1000                             CONGESTION_ON_THRESH(
1001                                     fsc->mount_options->congestion_kb)) {
1002                                 set_bdi_congested(inode_to_bdi(inode),
1003                                                   BLK_RW_ASYNC);
1004                         }
1005
1006
1007                         pages[locked_pages++] = page;
1008                         pvec.pages[i] = NULL;
1009
1010                         len += PAGE_SIZE;
1011                 }
1012
1013                 /* did we get anything? */
1014                 if (!locked_pages)
1015                         goto release_pvec_pages;
1016                 if (i) {
1017                         unsigned j, n = 0;
1018                         /* shift unused page to beginning of pvec */
1019                         for (j = 0; j < pvec_pages; j++) {
1020                                 if (!pvec.pages[j])
1021                                         continue;
1022                                 if (n < j)
1023                                         pvec.pages[n] = pvec.pages[j];
1024                                 n++;
1025                         }
1026                         pvec.nr = n;
1027
1028                         if (pvec_pages && i == pvec_pages &&
1029                             locked_pages < max_pages) {
1030                                 dout("reached end pvec, trying for more\n");
1031                                 pagevec_release(&pvec);
1032                                 goto get_more_pages;
1033                         }
1034                 }
1035
1036 new_request:
1037                 offset = page_offset(pages[0]);
1038                 len = wsize;
1039
1040                 req = ceph_osdc_new_request(&fsc->client->osdc,
1041                                         &ci->i_layout, vino,
1042                                         offset, &len, 0, num_ops,
1043                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1044                                         snapc, ceph_wbc.truncate_seq,
1045                                         ceph_wbc.truncate_size, false);
1046                 if (IS_ERR(req)) {
1047                         req = ceph_osdc_new_request(&fsc->client->osdc,
1048                                                 &ci->i_layout, vino,
1049                                                 offset, &len, 0,
1050                                                 min(num_ops,
1051                                                     CEPH_OSD_SLAB_OPS),
1052                                                 CEPH_OSD_OP_WRITE,
1053                                                 CEPH_OSD_FLAG_WRITE,
1054                                                 snapc, ceph_wbc.truncate_seq,
1055                                                 ceph_wbc.truncate_size, true);
1056                         BUG_ON(IS_ERR(req));
1057                 }
1058                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1059                              PAGE_SIZE - offset);
1060
1061                 req->r_callback = writepages_finish;
1062                 req->r_inode = inode;
1063
1064                 /* Format the osd request message and submit the write */
1065                 len = 0;
1066                 data_pages = pages;
1067                 op_idx = 0;
1068                 for (i = 0; i < locked_pages; i++) {
1069                         u64 cur_offset = page_offset(pages[i]);
1070                         if (offset + len != cur_offset) {
1071                                 if (op_idx + 1 == req->r_num_ops)
1072                                         break;
1073                                 osd_req_op_extent_dup_last(req, op_idx,
1074                                                            cur_offset - offset);
1075                                 dout("writepages got pages at %llu~%llu\n",
1076                                      offset, len);
1077                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1078                                                         data_pages, len, 0,
1079                                                         !!pool, false);
1080                                 osd_req_op_extent_update(req, op_idx, len);
1081
1082                                 len = 0;
1083                                 offset = cur_offset;
1084                                 data_pages = pages + i;
1085                                 op_idx++;
1086                         }
1087
1088                         set_page_writeback(pages[i]);
1089                         len += PAGE_SIZE;
1090                 }
1091
1092                 if (ceph_wbc.size_stable) {
1093                         len = min(len, ceph_wbc.i_size - offset);
1094                 } else if (i == locked_pages) {
1095                         /* writepages_finish() clears writeback pages
1096                          * according to the data length, so make sure
1097                          * data length covers all locked pages */
1098                         u64 min_len = len + 1 - PAGE_SIZE;
1099                         len = get_writepages_data_length(inode, pages[i - 1],
1100                                                          offset);
1101                         len = max(len, min_len);
1102                 }
1103                 dout("writepages got pages at %llu~%llu\n", offset, len);
1104
1105                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1106                                                  0, !!pool, false);
1107                 osd_req_op_extent_update(req, op_idx, len);
1108
1109                 BUG_ON(op_idx + 1 != req->r_num_ops);
1110
1111                 pool = NULL;
1112                 if (i < locked_pages) {
1113                         BUG_ON(num_ops <= req->r_num_ops);
1114                         num_ops -= req->r_num_ops;
1115                         locked_pages -= i;
1116
1117                         /* allocate new pages array for next request */
1118                         data_pages = pages;
1119                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1120                                               GFP_NOFS);
1121                         if (!pages) {
1122                                 pool = fsc->wb_pagevec_pool;
1123                                 pages = mempool_alloc(pool, GFP_NOFS);
1124                                 BUG_ON(!pages);
1125                         }
1126                         memcpy(pages, data_pages + i,
1127                                locked_pages * sizeof(*pages));
1128                         memset(data_pages + i, 0,
1129                                locked_pages * sizeof(*pages));
1130                 } else {
1131                         BUG_ON(num_ops != req->r_num_ops);
1132                         index = pages[i - 1]->index + 1;
1133                         /* request message now owns the pages array */
1134                         pages = NULL;
1135                 }
1136
1137                 req->r_mtime = timespec64_to_timespec(inode->i_mtime);
1138                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1139                 BUG_ON(rc);
1140                 req = NULL;
1141
1142                 wbc->nr_to_write -= i;
1143                 if (pages)
1144                         goto new_request;
1145
1146                 /*
1147                  * We stop writing back only if we are not doing
1148                  * integrity sync. In case of integrity sync we have to
1149                  * keep going until we have written all the pages
1150                  * we tagged for writeback prior to entering this loop.
1151                  */
1152                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1153                         done = true;
1154
1155 release_pvec_pages:
1156                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1157                      pvec.nr ? pvec.pages[0] : NULL);
1158                 pagevec_release(&pvec);
1159         }
1160
1161         if (should_loop && !done) {
1162                 /* more to do; loop back to beginning of file */
1163                 dout("writepages looping back to beginning of file\n");
1164                 end = start_index - 1; /* OK even when start_index == 0 */
1165
1166                 /* to write dirty pages associated with next snapc,
1167                  * we need to wait until current writes complete */
1168                 if (wbc->sync_mode != WB_SYNC_NONE &&
1169                     start_index == 0 && /* all dirty pages were checked */
1170                     !ceph_wbc.head_snapc) {
1171                         struct page *page;
1172                         unsigned i, nr;
1173                         index = 0;
1174                         while ((index <= end) &&
1175                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1176                                                 PAGECACHE_TAG_WRITEBACK))) {
1177                                 for (i = 0; i < nr; i++) {
1178                                         page = pvec.pages[i];
1179                                         if (page_snap_context(page) != snapc)
1180                                                 continue;
1181                                         wait_on_page_writeback(page);
1182                                 }
1183                                 pagevec_release(&pvec);
1184                                 cond_resched();
1185                         }
1186                 }
1187
1188                 start_index = 0;
1189                 index = 0;
1190                 goto retry;
1191         }
1192
1193         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1194                 mapping->writeback_index = index;
1195
1196 out:
1197         ceph_osdc_put_request(req);
1198         ceph_put_snap_context(last_snapc);
1199         dout("writepages dend - startone, rc = %d\n", rc);
1200         return rc;
1201 }
1202
1203
1204
1205 /*
1206  * See if a given @snapc is either writeable, or already written.
1207  */
1208 static int context_is_writeable_or_written(struct inode *inode,
1209                                            struct ceph_snap_context *snapc)
1210 {
1211         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1212         int ret = !oldest || snapc->seq <= oldest->seq;
1213
1214         ceph_put_snap_context(oldest);
1215         return ret;
1216 }
1217
1218 /*
1219  * We are only allowed to write into/dirty the page if the page is
1220  * clean, or already dirty within the same snap context.
1221  *
1222  * called with page locked.
1223  * return success with page locked,
1224  * or any failure (incl -EAGAIN) with page unlocked.
1225  */
1226 static int ceph_update_writeable_page(struct file *file,
1227                             loff_t pos, unsigned len,
1228                             struct page *page)
1229 {
1230         struct inode *inode = file_inode(file);
1231         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1232         struct ceph_inode_info *ci = ceph_inode(inode);
1233         loff_t page_off = pos & PAGE_MASK;
1234         int pos_in_page = pos & ~PAGE_MASK;
1235         int end_in_page = pos_in_page + len;
1236         loff_t i_size;
1237         int r;
1238         struct ceph_snap_context *snapc, *oldest;
1239
1240         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1241                 dout(" page %p forced umount\n", page);
1242                 unlock_page(page);
1243                 return -EIO;
1244         }
1245
1246 retry_locked:
1247         /* writepages currently holds page lock, but if we change that later, */
1248         wait_on_page_writeback(page);
1249
1250         snapc = page_snap_context(page);
1251         if (snapc && snapc != ci->i_head_snapc) {
1252                 /*
1253                  * this page is already dirty in another (older) snap
1254                  * context!  is it writeable now?
1255                  */
1256                 oldest = get_oldest_context(inode, NULL, NULL);
1257                 if (snapc->seq > oldest->seq) {
1258                         ceph_put_snap_context(oldest);
1259                         dout(" page %p snapc %p not current or oldest\n",
1260                              page, snapc);
1261                         /*
1262                          * queue for writeback, and wait for snapc to
1263                          * be writeable or written
1264                          */
1265                         snapc = ceph_get_snap_context(snapc);
1266                         unlock_page(page);
1267                         ceph_queue_writeback(inode);
1268                         r = wait_event_killable(ci->i_cap_wq,
1269                                context_is_writeable_or_written(inode, snapc));
1270                         ceph_put_snap_context(snapc);
1271                         if (r == -ERESTARTSYS)
1272                                 return r;
1273                         return -EAGAIN;
1274                 }
1275                 ceph_put_snap_context(oldest);
1276
1277                 /* yay, writeable, do it now (without dropping page lock) */
1278                 dout(" page %p snapc %p not current, but oldest\n",
1279                      page, snapc);
1280                 if (!clear_page_dirty_for_io(page))
1281                         goto retry_locked;
1282                 r = writepage_nounlock(page, NULL);
1283                 if (r < 0)
1284                         goto fail_unlock;
1285                 goto retry_locked;
1286         }
1287
1288         if (PageUptodate(page)) {
1289                 dout(" page %p already uptodate\n", page);
1290                 return 0;
1291         }
1292
1293         /* full page? */
1294         if (pos_in_page == 0 && len == PAGE_SIZE)
1295                 return 0;
1296
1297         /* past end of file? */
1298         i_size = i_size_read(inode);
1299
1300         if (page_off >= i_size ||
1301             (pos_in_page == 0 && (pos+len) >= i_size &&
1302              end_in_page - pos_in_page != PAGE_SIZE)) {
1303                 dout(" zeroing %p 0 - %d and %d - %d\n",
1304                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1305                 zero_user_segments(page,
1306                                    0, pos_in_page,
1307                                    end_in_page, PAGE_SIZE);
1308                 return 0;
1309         }
1310
1311         /* we need to read it. */
1312         r = ceph_do_readpage(file, page);
1313         if (r < 0) {
1314                 if (r == -EINPROGRESS)
1315                         return -EAGAIN;
1316                 goto fail_unlock;
1317         }
1318         goto retry_locked;
1319 fail_unlock:
1320         unlock_page(page);
1321         return r;
1322 }
1323
1324 /*
1325  * We are only allowed to write into/dirty the page if the page is
1326  * clean, or already dirty within the same snap context.
1327  */
1328 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1329                             loff_t pos, unsigned len, unsigned flags,
1330                             struct page **pagep, void **fsdata)
1331 {
1332         struct inode *inode = file_inode(file);
1333         struct page *page;
1334         pgoff_t index = pos >> PAGE_SHIFT;
1335         int r;
1336
1337         do {
1338                 /* get a page */
1339                 page = grab_cache_page_write_begin(mapping, index, 0);
1340                 if (!page)
1341                         return -ENOMEM;
1342
1343                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1344                      inode, page, (int)pos, (int)len);
1345
1346                 r = ceph_update_writeable_page(file, pos, len, page);
1347                 if (r < 0)
1348                         put_page(page);
1349                 else
1350                         *pagep = page;
1351         } while (r == -EAGAIN);
1352
1353         return r;
1354 }
1355
1356 /*
1357  * we don't do anything in here that simple_write_end doesn't do
1358  * except adjust dirty page accounting
1359  */
1360 static int ceph_write_end(struct file *file, struct address_space *mapping,
1361                           loff_t pos, unsigned len, unsigned copied,
1362                           struct page *page, void *fsdata)
1363 {
1364         struct inode *inode = file_inode(file);
1365         bool check_cap = false;
1366
1367         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1368              inode, page, (int)pos, (int)copied, (int)len);
1369
1370         /* zero the stale part of the page if we did a short copy */
1371         if (!PageUptodate(page)) {
1372                 if (copied < len) {
1373                         copied = 0;
1374                         goto out;
1375                 }
1376                 SetPageUptodate(page);
1377         }
1378
1379         /* did file size increase? */
1380         if (pos+copied > i_size_read(inode))
1381                 check_cap = ceph_inode_set_size(inode, pos+copied);
1382
1383         set_page_dirty(page);
1384
1385 out:
1386         unlock_page(page);
1387         put_page(page);
1388
1389         if (check_cap)
1390                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1391
1392         return copied;
1393 }
1394
1395 /*
1396  * we set .direct_IO to indicate direct io is supported, but since we
1397  * intercept O_DIRECT reads and writes early, this function should
1398  * never get called.
1399  */
1400 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1401 {
1402         WARN_ON(1);
1403         return -EINVAL;
1404 }
1405
1406 const struct address_space_operations ceph_aops = {
1407         .readpage = ceph_readpage,
1408         .readpages = ceph_readpages,
1409         .writepage = ceph_writepage,
1410         .writepages = ceph_writepages_start,
1411         .write_begin = ceph_write_begin,
1412         .write_end = ceph_write_end,
1413         .set_page_dirty = ceph_set_page_dirty,
1414         .invalidatepage = ceph_invalidatepage,
1415         .releasepage = ceph_releasepage,
1416         .direct_IO = ceph_direct_io,
1417 };
1418
1419 static void ceph_block_sigs(sigset_t *oldset)
1420 {
1421         sigset_t mask;
1422         siginitsetinv(&mask, sigmask(SIGKILL));
1423         sigprocmask(SIG_BLOCK, &mask, oldset);
1424 }
1425
1426 static void ceph_restore_sigs(sigset_t *oldset)
1427 {
1428         sigprocmask(SIG_SETMASK, oldset, NULL);
1429 }
1430
1431 /*
1432  * vm ops
1433  */
1434 static int ceph_filemap_fault(struct vm_fault *vmf)
1435 {
1436         struct vm_area_struct *vma = vmf->vma;
1437         struct inode *inode = file_inode(vma->vm_file);
1438         struct ceph_inode_info *ci = ceph_inode(inode);
1439         struct ceph_file_info *fi = vma->vm_file->private_data;
1440         struct page *pinned_page = NULL;
1441         loff_t off = vmf->pgoff << PAGE_SHIFT;
1442         int want, got, ret;
1443         sigset_t oldset;
1444
1445         ceph_block_sigs(&oldset);
1446
1447         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1448              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1449         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1450                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1451         else
1452                 want = CEPH_CAP_FILE_CACHE;
1453
1454         got = 0;
1455         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1456         if (ret < 0)
1457                 goto out_restore;
1458
1459         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1460              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1461
1462         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1463             ci->i_inline_version == CEPH_INLINE_NONE) {
1464                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1465                 ceph_add_rw_context(fi, &rw_ctx);
1466                 ret = filemap_fault(vmf);
1467                 ceph_del_rw_context(fi, &rw_ctx);
1468         } else
1469                 ret = -EAGAIN;
1470
1471         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1472              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1473         if (pinned_page)
1474                 put_page(pinned_page);
1475         ceph_put_cap_refs(ci, got);
1476
1477         if (ret != -EAGAIN)
1478                 goto out_restore;
1479
1480         /* read inline data */
1481         if (off >= PAGE_SIZE) {
1482                 /* does not support inline data > PAGE_SIZE */
1483                 ret = VM_FAULT_SIGBUS;
1484         } else {
1485                 int ret1;
1486                 struct address_space *mapping = inode->i_mapping;
1487                 struct page *page = find_or_create_page(mapping, 0,
1488                                                 mapping_gfp_constraint(mapping,
1489                                                 ~__GFP_FS));
1490                 if (!page) {
1491                         ret = VM_FAULT_OOM;
1492                         goto out_inline;
1493                 }
1494                 ret1 = __ceph_do_getattr(inode, page,
1495                                          CEPH_STAT_CAP_INLINE_DATA, true);
1496                 if (ret1 < 0 || off >= i_size_read(inode)) {
1497                         unlock_page(page);
1498                         put_page(page);
1499                         if (ret1 < 0)
1500                                 ret = ret1;
1501                         else
1502                                 ret = VM_FAULT_SIGBUS;
1503                         goto out_inline;
1504                 }
1505                 if (ret1 < PAGE_SIZE)
1506                         zero_user_segment(page, ret1, PAGE_SIZE);
1507                 else
1508                         flush_dcache_page(page);
1509                 SetPageUptodate(page);
1510                 vmf->page = page;
1511                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1512 out_inline:
1513                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1514                      inode, off, (size_t)PAGE_SIZE, ret);
1515         }
1516 out_restore:
1517         ceph_restore_sigs(&oldset);
1518         if (ret < 0)
1519                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1520
1521         return ret;
1522 }
1523
1524 /*
1525  * Reuse write_begin here for simplicity.
1526  */
1527 static int ceph_page_mkwrite(struct vm_fault *vmf)
1528 {
1529         struct vm_area_struct *vma = vmf->vma;
1530         struct inode *inode = file_inode(vma->vm_file);
1531         struct ceph_inode_info *ci = ceph_inode(inode);
1532         struct ceph_file_info *fi = vma->vm_file->private_data;
1533         struct ceph_cap_flush *prealloc_cf;
1534         struct page *page = vmf->page;
1535         loff_t off = page_offset(page);
1536         loff_t size = i_size_read(inode);
1537         size_t len;
1538         int want, got, ret;
1539         sigset_t oldset;
1540
1541         prealloc_cf = ceph_alloc_cap_flush();
1542         if (!prealloc_cf)
1543                 return VM_FAULT_OOM;
1544
1545         ceph_block_sigs(&oldset);
1546
1547         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1548                 struct page *locked_page = NULL;
1549                 if (off == 0) {
1550                         lock_page(page);
1551                         locked_page = page;
1552                 }
1553                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1554                 if (locked_page)
1555                         unlock_page(locked_page);
1556                 if (ret < 0)
1557                         goto out_free;
1558         }
1559
1560         if (off + PAGE_SIZE <= size)
1561                 len = PAGE_SIZE;
1562         else
1563                 len = size & ~PAGE_MASK;
1564
1565         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1566              inode, ceph_vinop(inode), off, len, size);
1567         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1568                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1569         else
1570                 want = CEPH_CAP_FILE_BUFFER;
1571
1572         got = 0;
1573         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1574                             &got, NULL);
1575         if (ret < 0)
1576                 goto out_free;
1577
1578         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1579              inode, off, len, ceph_cap_string(got));
1580
1581         /* Update time before taking page lock */
1582         file_update_time(vma->vm_file);
1583
1584         do {
1585                 lock_page(page);
1586
1587                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1588                         unlock_page(page);
1589                         ret = VM_FAULT_NOPAGE;
1590                         break;
1591                 }
1592
1593                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1594                 if (ret >= 0) {
1595                         /* success.  we'll keep the page locked. */
1596                         set_page_dirty(page);
1597                         ret = VM_FAULT_LOCKED;
1598                 }
1599         } while (ret == -EAGAIN);
1600
1601         if (ret == VM_FAULT_LOCKED ||
1602             ci->i_inline_version != CEPH_INLINE_NONE) {
1603                 int dirty;
1604                 spin_lock(&ci->i_ceph_lock);
1605                 ci->i_inline_version = CEPH_INLINE_NONE;
1606                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1607                                                &prealloc_cf);
1608                 spin_unlock(&ci->i_ceph_lock);
1609                 if (dirty)
1610                         __mark_inode_dirty(inode, dirty);
1611         }
1612
1613         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1614              inode, off, len, ceph_cap_string(got), ret);
1615         ceph_put_cap_refs(ci, got);
1616 out_free:
1617         ceph_restore_sigs(&oldset);
1618         ceph_free_cap_flush(prealloc_cf);
1619         if (ret < 0)
1620                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1621         return ret;
1622 }
1623
1624 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1625                            char *data, size_t len)
1626 {
1627         struct address_space *mapping = inode->i_mapping;
1628         struct page *page;
1629
1630         if (locked_page) {
1631                 page = locked_page;
1632         } else {
1633                 if (i_size_read(inode) == 0)
1634                         return;
1635                 page = find_or_create_page(mapping, 0,
1636                                            mapping_gfp_constraint(mapping,
1637                                            ~__GFP_FS));
1638                 if (!page)
1639                         return;
1640                 if (PageUptodate(page)) {
1641                         unlock_page(page);
1642                         put_page(page);
1643                         return;
1644                 }
1645         }
1646
1647         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1648              inode, ceph_vinop(inode), len, locked_page);
1649
1650         if (len > 0) {
1651                 void *kaddr = kmap_atomic(page);
1652                 memcpy(kaddr, data, len);
1653                 kunmap_atomic(kaddr);
1654         }
1655
1656         if (page != locked_page) {
1657                 if (len < PAGE_SIZE)
1658                         zero_user_segment(page, len, PAGE_SIZE);
1659                 else
1660                         flush_dcache_page(page);
1661
1662                 SetPageUptodate(page);
1663                 unlock_page(page);
1664                 put_page(page);
1665         }
1666 }
1667
1668 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1669 {
1670         struct inode *inode = file_inode(filp);
1671         struct ceph_inode_info *ci = ceph_inode(inode);
1672         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1673         struct ceph_osd_request *req;
1674         struct page *page = NULL;
1675         u64 len, inline_version;
1676         int err = 0;
1677         bool from_pagecache = false;
1678
1679         spin_lock(&ci->i_ceph_lock);
1680         inline_version = ci->i_inline_version;
1681         spin_unlock(&ci->i_ceph_lock);
1682
1683         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1684              inode, ceph_vinop(inode), inline_version);
1685
1686         if (inline_version == 1 || /* initial version, no data */
1687             inline_version == CEPH_INLINE_NONE)
1688                 goto out;
1689
1690         if (locked_page) {
1691                 page = locked_page;
1692                 WARN_ON(!PageUptodate(page));
1693         } else if (ceph_caps_issued(ci) &
1694                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1695                 page = find_get_page(inode->i_mapping, 0);
1696                 if (page) {
1697                         if (PageUptodate(page)) {
1698                                 from_pagecache = true;
1699                                 lock_page(page);
1700                         } else {
1701                                 put_page(page);
1702                                 page = NULL;
1703                         }
1704                 }
1705         }
1706
1707         if (page) {
1708                 len = i_size_read(inode);
1709                 if (len > PAGE_SIZE)
1710                         len = PAGE_SIZE;
1711         } else {
1712                 page = __page_cache_alloc(GFP_NOFS);
1713                 if (!page) {
1714                         err = -ENOMEM;
1715                         goto out;
1716                 }
1717                 err = __ceph_do_getattr(inode, page,
1718                                         CEPH_STAT_CAP_INLINE_DATA, true);
1719                 if (err < 0) {
1720                         /* no inline data */
1721                         if (err == -ENODATA)
1722                                 err = 0;
1723                         goto out;
1724                 }
1725                 len = err;
1726         }
1727
1728         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1729                                     ceph_vino(inode), 0, &len, 0, 1,
1730                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1731                                     NULL, 0, 0, false);
1732         if (IS_ERR(req)) {
1733                 err = PTR_ERR(req);
1734                 goto out;
1735         }
1736
1737         req->r_mtime = timespec64_to_timespec(inode->i_mtime);
1738         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1739         if (!err)
1740                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1741         ceph_osdc_put_request(req);
1742         if (err < 0)
1743                 goto out;
1744
1745         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1746                                     ceph_vino(inode), 0, &len, 1, 3,
1747                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1748                                     NULL, ci->i_truncate_seq,
1749                                     ci->i_truncate_size, false);
1750         if (IS_ERR(req)) {
1751                 err = PTR_ERR(req);
1752                 goto out;
1753         }
1754
1755         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1756
1757         {
1758                 __le64 xattr_buf = cpu_to_le64(inline_version);
1759                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1760                                             "inline_version", &xattr_buf,
1761                                             sizeof(xattr_buf),
1762                                             CEPH_OSD_CMPXATTR_OP_GT,
1763                                             CEPH_OSD_CMPXATTR_MODE_U64);
1764                 if (err)
1765                         goto out_put;
1766         }
1767
1768         {
1769                 char xattr_buf[32];
1770                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1771                                          "%llu", inline_version);
1772                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1773                                             "inline_version",
1774                                             xattr_buf, xattr_len, 0, 0);
1775                 if (err)
1776                         goto out_put;
1777         }
1778
1779         req->r_mtime = timespec64_to_timespec(inode->i_mtime);
1780         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1781         if (!err)
1782                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1783 out_put:
1784         ceph_osdc_put_request(req);
1785         if (err == -ECANCELED)
1786                 err = 0;
1787 out:
1788         if (page && page != locked_page) {
1789                 if (from_pagecache) {
1790                         unlock_page(page);
1791                         put_page(page);
1792                 } else
1793                         __free_pages(page, 0);
1794         }
1795
1796         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1797              inode, ceph_vinop(inode), inline_version, err);
1798         return err;
1799 }
1800
1801 static const struct vm_operations_struct ceph_vmops = {
1802         .fault          = ceph_filemap_fault,
1803         .page_mkwrite   = ceph_page_mkwrite,
1804 };
1805
1806 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1807 {
1808         struct address_space *mapping = file->f_mapping;
1809
1810         if (!mapping->a_ops->readpage)
1811                 return -ENOEXEC;
1812         file_accessed(file);
1813         vma->vm_ops = &ceph_vmops;
1814         return 0;
1815 }
1816
1817 enum {
1818         POOL_READ       = 1,
1819         POOL_WRITE      = 2,
1820 };
1821
1822 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1823                                 s64 pool, struct ceph_string *pool_ns)
1824 {
1825         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1826         struct ceph_mds_client *mdsc = fsc->mdsc;
1827         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1828         struct rb_node **p, *parent;
1829         struct ceph_pool_perm *perm;
1830         struct page **pages;
1831         size_t pool_ns_len;
1832         int err = 0, err2 = 0, have = 0;
1833
1834         down_read(&mdsc->pool_perm_rwsem);
1835         p = &mdsc->pool_perm_tree.rb_node;
1836         while (*p) {
1837                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1838                 if (pool < perm->pool)
1839                         p = &(*p)->rb_left;
1840                 else if (pool > perm->pool)
1841                         p = &(*p)->rb_right;
1842                 else {
1843                         int ret = ceph_compare_string(pool_ns,
1844                                                 perm->pool_ns,
1845                                                 perm->pool_ns_len);
1846                         if (ret < 0)
1847                                 p = &(*p)->rb_left;
1848                         else if (ret > 0)
1849                                 p = &(*p)->rb_right;
1850                         else {
1851                                 have = perm->perm;
1852                                 break;
1853                         }
1854                 }
1855         }
1856         up_read(&mdsc->pool_perm_rwsem);
1857         if (*p)
1858                 goto out;
1859
1860         if (pool_ns)
1861                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1862                      pool, (int)pool_ns->len, pool_ns->str);
1863         else
1864                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1865
1866         down_write(&mdsc->pool_perm_rwsem);
1867         p = &mdsc->pool_perm_tree.rb_node;
1868         parent = NULL;
1869         while (*p) {
1870                 parent = *p;
1871                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1872                 if (pool < perm->pool)
1873                         p = &(*p)->rb_left;
1874                 else if (pool > perm->pool)
1875                         p = &(*p)->rb_right;
1876                 else {
1877                         int ret = ceph_compare_string(pool_ns,
1878                                                 perm->pool_ns,
1879                                                 perm->pool_ns_len);
1880                         if (ret < 0)
1881                                 p = &(*p)->rb_left;
1882                         else if (ret > 0)
1883                                 p = &(*p)->rb_right;
1884                         else {
1885                                 have = perm->perm;
1886                                 break;
1887                         }
1888                 }
1889         }
1890         if (*p) {
1891                 up_write(&mdsc->pool_perm_rwsem);
1892                 goto out;
1893         }
1894
1895         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1896                                          1, false, GFP_NOFS);
1897         if (!rd_req) {
1898                 err = -ENOMEM;
1899                 goto out_unlock;
1900         }
1901
1902         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1903         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1904         rd_req->r_base_oloc.pool = pool;
1905         if (pool_ns)
1906                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1907         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1908
1909         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1910         if (err)
1911                 goto out_unlock;
1912
1913         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1914                                          1, false, GFP_NOFS);
1915         if (!wr_req) {
1916                 err = -ENOMEM;
1917                 goto out_unlock;
1918         }
1919
1920         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1921         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1922         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1923         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1924
1925         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1926         if (err)
1927                 goto out_unlock;
1928
1929         /* one page should be large enough for STAT data */
1930         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1931         if (IS_ERR(pages)) {
1932                 err = PTR_ERR(pages);
1933                 goto out_unlock;
1934         }
1935
1936         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1937                                      0, false, true);
1938         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1939
1940         wr_req->r_mtime = timespec64_to_timespec(ci->vfs_inode.i_mtime);
1941         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1942
1943         if (!err)
1944                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1945         if (!err2)
1946                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1947
1948         if (err >= 0 || err == -ENOENT)
1949                 have |= POOL_READ;
1950         else if (err != -EPERM)
1951                 goto out_unlock;
1952
1953         if (err2 == 0 || err2 == -EEXIST)
1954                 have |= POOL_WRITE;
1955         else if (err2 != -EPERM) {
1956                 err = err2;
1957                 goto out_unlock;
1958         }
1959
1960         pool_ns_len = pool_ns ? pool_ns->len : 0;
1961         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1962         if (!perm) {
1963                 err = -ENOMEM;
1964                 goto out_unlock;
1965         }
1966
1967         perm->pool = pool;
1968         perm->perm = have;
1969         perm->pool_ns_len = pool_ns_len;
1970         if (pool_ns_len > 0)
1971                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1972         perm->pool_ns[pool_ns_len] = 0;
1973
1974         rb_link_node(&perm->node, parent, p);
1975         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1976         err = 0;
1977 out_unlock:
1978         up_write(&mdsc->pool_perm_rwsem);
1979
1980         ceph_osdc_put_request(rd_req);
1981         ceph_osdc_put_request(wr_req);
1982 out:
1983         if (!err)
1984                 err = have;
1985         if (pool_ns)
1986                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1987                      pool, (int)pool_ns->len, pool_ns->str, err);
1988         else
1989                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1990         return err;
1991 }
1992
1993 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1994 {
1995         s64 pool;
1996         struct ceph_string *pool_ns;
1997         int ret, flags;
1998
1999         if (ci->i_vino.snap != CEPH_NOSNAP) {
2000                 /*
2001                  * Pool permission check needs to write to the first object.
2002                  * But for snapshot, head of the first object may have alread
2003                  * been deleted. Skip check to avoid creating orphan object.
2004                  */
2005                 return 0;
2006         }
2007
2008         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2009                                 NOPOOLPERM))
2010                 return 0;
2011
2012         spin_lock(&ci->i_ceph_lock);
2013         flags = ci->i_ceph_flags;
2014         pool = ci->i_layout.pool_id;
2015         spin_unlock(&ci->i_ceph_lock);
2016 check:
2017         if (flags & CEPH_I_POOL_PERM) {
2018                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2019                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2020                              pool);
2021                         return -EPERM;
2022                 }
2023                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2024                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2025                              pool);
2026                         return -EPERM;
2027                 }
2028                 return 0;
2029         }
2030
2031         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2032         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2033         ceph_put_string(pool_ns);
2034         if (ret < 0)
2035                 return ret;
2036
2037         flags = CEPH_I_POOL_PERM;
2038         if (ret & POOL_READ)
2039                 flags |= CEPH_I_POOL_RD;
2040         if (ret & POOL_WRITE)
2041                 flags |= CEPH_I_POOL_WR;
2042
2043         spin_lock(&ci->i_ceph_lock);
2044         if (pool == ci->i_layout.pool_id &&
2045             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2046                 ci->i_ceph_flags |= flags;
2047         } else {
2048                 pool = ci->i_layout.pool_id;
2049                 flags = ci->i_ceph_flags;
2050         }
2051         spin_unlock(&ci->i_ceph_lock);
2052         goto check;
2053 }
2054
2055 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2056 {
2057         struct ceph_pool_perm *perm;
2058         struct rb_node *n;
2059
2060         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2061                 n = rb_first(&mdsc->pool_perm_tree);
2062                 perm = rb_entry(n, struct ceph_pool_perm, node);
2063                 rb_erase(n, &mdsc->pool_perm_tree);
2064                 kfree(perm);
2065         }
2066 }