fs/ceph/addr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/backing-dev.h>
   5 #include <linux/fs.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>    /* generic_writepages */
   9 #include <linux/slab.h>
  10 #include <linux/pagevec.h>
  11 #include <linux/task_io_accounting_ops.h>
  12 #include <linux/signal.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "metric.h"
  20 #include <linux/ceph/osd_client.h>
  21 #include <linux/ceph/striper.h>
  22
  23 /*
  24  * Ceph address space ops.
  25  *
  26  * There are a few funny things going on here.
  27  *
  28  * The page->private field is used to reference a struct
  29  * ceph_snap_context for _every_ dirty page.  This indicates which
  30  * snapshot the page was logically dirtied in, and thus which snap
  31  * context needs to be associated with the osd write during writeback.
  32  *
  33  * Similarly, struct ceph_inode_info maintains a set of counters to
  34  * count dirty pages on the inode.  In the absence of snapshots,
  35  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  36  *
  37  * When a snapshot is taken (that is, when the client receives
  38  * notification that a snapshot was taken), each inode with caps and
  39  * with dirty pages (dirty pages implies there is a cap) gets a new
  40  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  41  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  42  * moved to capsnap->dirty. (Unless a sync write is currently in
  43  * progress.  In that case, the capsnap is said to be "pending", new
  44  * writes cannot start, and the capsnap isn't "finalized" until the
  45  * write completes (or fails) and a final size/mtime for the inode for
  46  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  47  *
  48  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  49  * we look for the first capsnap in i_cap_snaps and write out pages in
  50  * that snap context _only_.  Then we move on to the next capsnap,
  51  * eventually reaching the "live" or "head" context (i.e., pages that
  52  * are not yet snapped) and are writing the most recently dirtied
  53  * pages.
  54  *
  55  * Invalidate and so forth must take care to ensure the dirty page
  56  * accounting is preserved.
  57  */
  58
  59 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  60 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  61         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  62          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  63
  64 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  65 {
  66         if (PagePrivate(page))
  67                 return (void *)page->private;
  68         return NULL;
  69 }
  70
  71 /*
  72  * Dirty a page.  Optimistically adjust accounting, on the assumption
  73  * that we won't race with invalidate.  If we do, readjust.
  74  */
  75 static int ceph_set_page_dirty(struct page *page)
  76 {
  77         struct address_space *mapping = page->mapping;
  78         struct inode *inode;
  79         struct ceph_inode_info *ci;
  80         struct ceph_snap_context *snapc;
  81
  82         if (PageDirty(page)) {
  83                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  84                      mapping->host, page, page->index);
  85                 BUG_ON(!PagePrivate(page));
  86                 return 0;
  87         }
  88
  89         inode = mapping->host;
  90         ci = ceph_inode(inode);
  91
  92         /* dirty the head */
  93         spin_lock(&ci->i_ceph_lock);
  94         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  95         if (__ceph_have_pending_cap_snap(ci)) {
  96                 struct ceph_cap_snap *capsnap =
  97                                 list_last_entry(&ci->i_cap_snaps,
  98                                                 struct ceph_cap_snap,
  99                                                 ci_item);
 100                 snapc = ceph_get_snap_context(capsnap->context);
 101                 capsnap->dirty_pages++;
 102         } else {
 103                 BUG_ON(!ci->i_head_snapc);
 104                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 105                 ++ci->i_wrbuffer_ref_head;
 106         }
 107         if (ci->i_wrbuffer_ref == 0)
 108                 ihold(inode);
 109         ++ci->i_wrbuffer_ref;
 110         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 111              "snapc %p seq %lld (%d snaps)\n",
 112              mapping->host, page, page->index,
 113              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 114              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 115              snapc, snapc->seq, snapc->num_snaps);
 116         spin_unlock(&ci->i_ceph_lock);
 117
 118         /*
 119          * Reference snap context in page->private.  Also set
 120          * PagePrivate so that we get invalidatepage callback.
 121          */
 122         BUG_ON(PagePrivate(page));
 123         page->private = (unsigned long)snapc;
 124         SetPagePrivate(page);
 125
 126         return __set_page_dirty_nobuffers(page);
 127 }
 128
 129 /*
 130  * If we are truncating the full page (i.e. offset == 0), adjust the
 131  * dirty page counters appropriately.  Only called if there is private
 132  * data on the page.
 133  */
 134 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 135                                 unsigned int length)
 136 {
 137         struct inode *inode;
 138         struct ceph_inode_info *ci;
 139         struct ceph_snap_context *snapc = page_snap_context(page);
 140
 141         inode = page->mapping->host;
 142         ci = ceph_inode(inode);
 143
 144         if (offset != 0 || length != PAGE_SIZE) {
 145                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 146                      inode, page, page->index, offset, length);
 147                 return;
 148         }
 149
 150         ceph_invalidate_fscache_page(inode, page);
 151
 152         WARN_ON(!PageLocked(page));
 153         if (!PagePrivate(page))
 154                 return;
 155
 156         dout("%p invalidatepage %p idx %lu full dirty page\n",
 157              inode, page, page->index);
 158
 159         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 160         ceph_put_snap_context(snapc);
 161         page->private = 0;
 162         ClearPagePrivate(page);
 163 }
 164
 165 static int ceph_releasepage(struct page *page, gfp_t g)
 166 {
 167         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
 168              page, page->index, PageDirty(page) ? "" : "not ");
 169
 170         /* Can we release the page from the cache? */
 171         if (!ceph_release_fscache_page(page, g))
 172                 return 0;
 173
 174         return !PagePrivate(page);
 175 }
 176
 177 /* read a single page, without unlocking it. */
 178 static int ceph_do_readpage(struct file *filp, struct page *page)
 179 {
 180         struct inode *inode = file_inode(filp);
 181         struct ceph_inode_info *ci = ceph_inode(inode);
 182         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 183         struct ceph_osd_client *osdc = &fsc->client->osdc;
 184         struct ceph_osd_request *req;
 185         struct ceph_vino vino = ceph_vino(inode);
 186         int err = 0;
 187         u64 off = page_offset(page);
 188         u64 len = PAGE_SIZE;
 189
 190         if (off >= i_size_read(inode)) {
 191                 zero_user_segment(page, 0, PAGE_SIZE);
 192                 SetPageUptodate(page);
 193                 return 0;
 194         }
 195
 196         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 197                 /*
 198                  * Uptodate inline data should have been added
 199                  * into page cache while getting Fcr caps.
 200                  */
 201                 if (off == 0)
 202                         return -EINVAL;
 203                 zero_user_segment(page, 0, PAGE_SIZE);
 204                 SetPageUptodate(page);
 205                 return 0;
 206         }
 207
 208         err = ceph_readpage_from_fscache(inode, page);
 209         if (err == 0)
 210                 return -EINPROGRESS;
 211
 212         dout("readpage ino %llx.%llx file %p off %llu len %llu page %p index %lu\n",
 213              vino.ino, vino.snap, filp, off, len, page, page->index);
 214         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, 0, 1,
 215                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, NULL,
 216                                     ci->i_truncate_seq, ci->i_truncate_size,
 217                                     false);
 218         if (IS_ERR(req))
 219                 return PTR_ERR(req);
 220
 221         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 222
 223         err = ceph_osdc_start_request(osdc, req, false);
 224         if (!err)
 225                 err = ceph_osdc_wait_request(osdc, req);
 226
 227         ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency,
 228                                  req->r_end_latency, err);
 229
 230         ceph_osdc_put_request(req);
 231         dout("readpage result %d\n", err);
 232
 233         if (err == -ENOENT)
 234                 err = 0;
 235         if (err < 0) {
 236                 ceph_fscache_readpage_cancel(inode, page);
 237                 if (err == -EBLOCKLISTED)
 238                         fsc->blocklisted = true;
 239                 goto out;
 240         }
 241         if (err < PAGE_SIZE)
 242                 /* zero fill remainder of page */
 243                 zero_user_segment(page, err, PAGE_SIZE);
 244         else
 245                 flush_dcache_page(page);
 246
 247         SetPageUptodate(page);
 248         ceph_readpage_to_fscache(inode, page);
 249
 250 out:
 251         return err < 0 ? err : 0;
 252 }
 253
 254 static int ceph_readpage(struct file *filp, struct page *page)
 255 {
 256         int r = ceph_do_readpage(filp, page);
 257         if (r != -EINPROGRESS)
 258                 unlock_page(page);
 259         else
 260                 r = 0;
 261         return r;
 262 }
 263
 264 /*
 265  * Finish an async read(ahead) op.
 266  */
 267 static void finish_read(struct ceph_osd_request *req)
 268 {
 269         struct inode *inode = req->r_inode;
 270         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 271         struct ceph_osd_data *osd_data;
 272         int rc = req->r_result <= 0 ? req->r_result : 0;
 273         int bytes = req->r_result >= 0 ? req->r_result : 0;
 274         int num_pages;
 275         int i;
 276
 277         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 278         if (rc == -EBLOCKLISTED)
 279                 ceph_inode_to_client(inode)->blocklisted = true;
 280
 281         /* unlock all pages, zeroing any data we didn't read */
 282         osd_data = osd_req_op_extent_osd_data(req, 0);
 283         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 284         num_pages = calc_pages_for((u64)osd_data->alignment,
 285                                         (u64)osd_data->length);
 286         for (i = 0; i < num_pages; i++) {
 287                 struct page *page = osd_data->pages[i];
 288
 289                 if (rc < 0 && rc != -ENOENT) {
 290                         ceph_fscache_readpage_cancel(inode, page);
 291                         goto unlock;
 292                 }
 293                 if (bytes < (int)PAGE_SIZE) {
 294                         /* zero (remainder of) page */
 295                         int s = bytes < 0 ? 0 : bytes;
 296                         zero_user_segment(page, s, PAGE_SIZE);
 297                 }
 298                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 299                      page->index);
 300                 flush_dcache_page(page);
 301                 SetPageUptodate(page);
 302                 ceph_readpage_to_fscache(inode, page);
 303 unlock:
 304                 unlock_page(page);
 305                 put_page(page);
 306                 bytes -= PAGE_SIZE;
 307         }
 308
 309         ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency,
 310                                  req->r_end_latency, rc);
 311
 312         kfree(osd_data->pages);
 313 }
 314
 315 /*
 316  * start an async read(ahead) operation.  return nr_pages we submitted
 317  * a read for on success, or negative error code.
 318  */
 319 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
 320                       struct list_head *page_list, int max)
 321 {
 322         struct ceph_osd_client *osdc =
 323                 &ceph_inode_to_client(inode)->client->osdc;
 324         struct ceph_inode_info *ci = ceph_inode(inode);
 325         struct page *page = lru_to_page(page_list);
 326         struct ceph_vino vino;
 327         struct ceph_osd_request *req;
 328         u64 off;
 329         u64 len;
 330         int i;
 331         struct page **pages;
 332         pgoff_t next_index;
 333         int nr_pages = 0;
 334         int got = 0;
 335         int ret = 0;
 336
 337         if (!rw_ctx) {
 338                 /* caller of readpages does not hold buffer and read caps
 339                  * (fadvise, madvise and readahead cases) */
 340                 int want = CEPH_CAP_FILE_CACHE;
 341                 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want,
 342                                         true, &got);
 343                 if (ret < 0) {
 344                         dout("start_read %p, error getting cap\n", inode);
 345                 } else if (!(got & want)) {
 346                         dout("start_read %p, no cache cap\n", inode);
 347                         ret = 0;
 348                 }
 349                 if (ret <= 0) {
 350                         if (got)
 351                                 ceph_put_cap_refs(ci, got);
 352                         while (!list_empty(page_list)) {
 353                                 page = lru_to_page(page_list);
 354                                 list_del(&page->lru);
 355                                 put_page(page);
 356                         }
 357                         return ret;
 358                 }
 359         }
 360
 361         off = (u64) page_offset(page);
 362
 363         /* count pages */
 364         next_index = page->index;
 365         list_for_each_entry_reverse(page, page_list, lru) {
 366                 if (page->index != next_index)
 367                         break;
 368                 nr_pages++;
 369                 next_index++;
 370                 if (max && nr_pages == max)
 371                         break;
 372         }
 373         len = nr_pages << PAGE_SHIFT;
 374         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 375              off, len);
 376         vino = ceph_vino(inode);
 377         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 378                                     0, 1, CEPH_OSD_OP_READ,
 379                                     CEPH_OSD_FLAG_READ, NULL,
 380                                     ci->i_truncate_seq, ci->i_truncate_size,
 381                                     false);
 382         if (IS_ERR(req)) {
 383                 ret = PTR_ERR(req);
 384                 goto out;
 385         }
 386
 387         /* build page vector */
 388         nr_pages = calc_pages_for(0, len);
 389         pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
 390         if (!pages) {
 391                 ret = -ENOMEM;
 392                 goto out_put;
 393         }
 394         for (i = 0; i < nr_pages; ++i) {
 395                 page = list_entry(page_list->prev, struct page, lru);
 396                 BUG_ON(PageLocked(page));
 397                 list_del(&page->lru);
 398
 399                 dout("start_read %p adding %p idx %lu\n", inode, page,
 400                      page->index);
 401                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 402                                           GFP_KERNEL)) {
 403                         ceph_fscache_uncache_page(inode, page);
 404                         put_page(page);
 405                         dout("start_read %p add_to_page_cache failed %p\n",
 406                              inode, page);
 407                         nr_pages = i;
 408                         if (nr_pages > 0) {
 409                                 len = nr_pages << PAGE_SHIFT;
 410                                 osd_req_op_extent_update(req, 0, len);
 411                                 break;
 412                         }
 413                         goto out_pages;
 414                 }
 415                 pages[i] = page;
 416         }
 417         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 418         req->r_callback = finish_read;
 419         req->r_inode = inode;
 420
 421         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 422         ret = ceph_osdc_start_request(osdc, req, false);
 423         if (ret < 0)
 424                 goto out_pages;
 425         ceph_osdc_put_request(req);
 426
 427         /* After adding locked pages to page cache, the inode holds cache cap.
 428          * So we can drop our cap refs. */
 429         if (got)
 430                 ceph_put_cap_refs(ci, got);
 431
 432         return nr_pages;
 433
 434 out_pages:
 435         for (i = 0; i < nr_pages; ++i) {
 436                 ceph_fscache_readpage_cancel(inode, pages[i]);
 437                 unlock_page(pages[i]);
 438         }
 439         ceph_put_page_vector(pages, nr_pages, false);
 440 out_put:
 441         ceph_osdc_put_request(req);
 442 out:
 443         if (got)
 444                 ceph_put_cap_refs(ci, got);
 445         return ret;
 446 }
 447
 448
 449 /*
 450  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 451  * the caller (VM) cleans them up.
 452  */
 453 static int ceph_readpages(struct file *file, struct address_space *mapping,
 454                           struct list_head *page_list, unsigned nr_pages)
 455 {
 456         struct inode *inode = file_inode(file);
 457         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 458         struct ceph_file_info *fi = file->private_data;
 459         struct ceph_rw_context *rw_ctx;
 460         int rc = 0;
 461         int max = 0;
 462
 463         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 464                 return -EINVAL;
 465
 466         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 467                                          &nr_pages);
 468
 469         if (rc == 0)
 470                 goto out;
 471
 472         rw_ctx = ceph_find_rw_context(fi);
 473         max = fsc->mount_options->rsize >> PAGE_SHIFT;
 474         dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
 475              inode, file, rw_ctx, nr_pages, max);
 476         while (!list_empty(page_list)) {
 477                 rc = start_read(inode, rw_ctx, page_list, max);
 478                 if (rc < 0)
 479                         goto out;
 480         }
 481 out:
 482         ceph_fscache_readpages_cancel(inode, page_list);
 483
 484         dout("readpages %p file %p ret %d\n", inode, file, rc);
 485         return rc;
 486 }
 487
 488 struct ceph_writeback_ctl
 489 {
 490         loff_t i_size;
 491         u64 truncate_size;
 492         u32 truncate_seq;
 493         bool size_stable;
 494         bool head_snapc;
 495 };
 496
 497 /*
 498  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 499  * only snap context we are allowed to write back.
 500  */
 501 static struct ceph_snap_context *
 502 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 503                    struct ceph_snap_context *page_snapc)
 504 {
 505         struct ceph_inode_info *ci = ceph_inode(inode);
 506         struct ceph_snap_context *snapc = NULL;
 507         struct ceph_cap_snap *capsnap = NULL;
 508
 509         spin_lock(&ci->i_ceph_lock);
 510         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 511                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 512                      capsnap->context, capsnap->dirty_pages);
 513                 if (!capsnap->dirty_pages)
 514                         continue;
 515
 516                 /* get i_size, truncate_{seq,size} for page_snapc? */
 517                 if (snapc && capsnap->context != page_snapc)
 518                         continue;
 519
 520                 if (ctl) {
 521                         if (capsnap->writing) {
 522                                 ctl->i_size = i_size_read(inode);
 523                                 ctl->size_stable = false;
 524                         } else {
 525                                 ctl->i_size = capsnap->size;
 526                                 ctl->size_stable = true;
 527                         }
 528                         ctl->truncate_size = capsnap->truncate_size;
 529                         ctl->truncate_seq = capsnap->truncate_seq;
 530                         ctl->head_snapc = false;
 531                 }
 532
 533                 if (snapc)
 534                         break;
 535
 536                 snapc = ceph_get_snap_context(capsnap->context);
 537                 if (!page_snapc ||
 538                     page_snapc == snapc ||
 539                     page_snapc->seq > snapc->seq)
 540                         break;
 541         }
 542         if (!snapc && ci->i_wrbuffer_ref_head) {
 543                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 544                 dout(" head snapc %p has %d dirty pages\n",
 545                      snapc, ci->i_wrbuffer_ref_head);
 546                 if (ctl) {
 547                         ctl->i_size = i_size_read(inode);
 548                         ctl->truncate_size = ci->i_truncate_size;
 549                         ctl->truncate_seq = ci->i_truncate_seq;
 550                         ctl->size_stable = false;
 551                         ctl->head_snapc = true;
 552                 }
 553         }
 554         spin_unlock(&ci->i_ceph_lock);
 555         return snapc;
 556 }
 557
 558 static u64 get_writepages_data_length(struct inode *inode,
 559                                       struct page *page, u64 start)
 560 {
 561         struct ceph_inode_info *ci = ceph_inode(inode);
 562         struct ceph_snap_context *snapc = page_snap_context(page);
 563         struct ceph_cap_snap *capsnap = NULL;
 564         u64 end = i_size_read(inode);
 565
 566         if (snapc != ci->i_head_snapc) {
 567                 bool found = false;
 568                 spin_lock(&ci->i_ceph_lock);
 569                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 570                         if (capsnap->context == snapc) {
 571                                 if (!capsnap->writing)
 572                                         end = capsnap->size;
 573                                 found = true;
 574                                 break;
 575                         }
 576                 }
 577                 spin_unlock(&ci->i_ceph_lock);
 578                 WARN_ON(!found);
 579         }
 580         if (end > page_offset(page) + PAGE_SIZE)
 581                 end = page_offset(page) + PAGE_SIZE;
 582         return end > start ? end - start : 0;
 583 }
 584
 585 /*
 586  * Write a single page, but leave the page locked.
 587  *
 588  * If we get a write error, mark the mapping for error, but still adjust the
 589  * dirty page accounting (i.e., page is no longer dirty).
 590  */
 591 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 592 {
 593         struct inode *inode = page->mapping->host;
 594         struct ceph_inode_info *ci = ceph_inode(inode);
 595         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 596         struct ceph_snap_context *snapc, *oldest;
 597         loff_t page_off = page_offset(page);
 598         int err;
 599         loff_t len = PAGE_SIZE;
 600         struct ceph_writeback_ctl ceph_wbc;
 601         struct ceph_osd_client *osdc = &fsc->client->osdc;
 602         struct ceph_osd_request *req;
 603
 604         dout("writepage %p idx %lu\n", page, page->index);
 605
 606         /* verify this is a writeable snap context */
 607         snapc = page_snap_context(page);
 608         if (!snapc) {
 609                 dout("writepage %p page %p not dirty?\n", inode, page);
 610                 return 0;
 611         }
 612         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 613         if (snapc->seq > oldest->seq) {
 614                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 615                      inode, page, snapc);
 616                 /* we should only noop if called by kswapd */
 617                 WARN_ON(!(current->flags & PF_MEMALLOC));
 618                 ceph_put_snap_context(oldest);
 619                 redirty_page_for_writepage(wbc, page);
 620                 return 0;
 621         }
 622         ceph_put_snap_context(oldest);
 623
 624         /* is this a partial page at end of file? */
 625         if (page_off >= ceph_wbc.i_size) {
 626                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
 627                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 628                 return 0;
 629         }
 630
 631         if (ceph_wbc.i_size < page_off + len)
 632                 len = ceph_wbc.i_size - page_off;
 633
 634         dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 635              inode, page, page->index, page_off, len, snapc, snapc->seq);
 636
 637         if (atomic_long_inc_return(&fsc->writeback_count) >
 638             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 639                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 640
 641         set_page_writeback(page);
 642         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), page_off, &len, 0, 1,
 643                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, snapc,
 644                                     ceph_wbc.truncate_seq, ceph_wbc.truncate_size,
 645                                     true);
 646         if (IS_ERR(req)) {
 647                 redirty_page_for_writepage(wbc, page);
 648                 end_page_writeback(page);
 649                 return PTR_ERR(req);
 650         }
 651
 652         /* it may be a short write due to an object boundary */
 653         WARN_ON_ONCE(len > PAGE_SIZE);
 654         osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false);
 655         dout("writepage %llu~%llu (%llu bytes)\n", page_off, len, len);
 656
 657         req->r_mtime = inode->i_mtime;
 658         err = ceph_osdc_start_request(osdc, req, true);
 659         if (!err)
 660                 err = ceph_osdc_wait_request(osdc, req);
 661
 662         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
 663                                   req->r_end_latency, err);
 664
 665         ceph_osdc_put_request(req);
 666         if (err == 0)
 667                 err = len;
 668
 669         if (err < 0) {
 670                 struct writeback_control tmp_wbc;
 671                 if (!wbc)
 672                         wbc = &tmp_wbc;
 673                 if (err == -ERESTARTSYS) {
 674                         /* killed by SIGKILL */
 675                         dout("writepage interrupted page %p\n", page);
 676                         redirty_page_for_writepage(wbc, page);
 677                         end_page_writeback(page);
 678                         return err;
 679                 }
 680                 if (err == -EBLOCKLISTED)
 681                         fsc->blocklisted = true;
 682                 dout("writepage setting page/mapping error %d %p\n",
 683                      err, page);
 684                 mapping_set_error(&inode->i_data, err);
 685                 wbc->pages_skipped++;
 686         } else {
 687                 dout("writepage cleaned page %p\n", page);
 688                 err = 0;  /* vfs expects us to return 0 */
 689         }
 690         page->private = 0;
 691         ClearPagePrivate(page);
 692         end_page_writeback(page);
 693         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 694         ceph_put_snap_context(snapc);  /* page's reference */
 695
 696         if (atomic_long_dec_return(&fsc->writeback_count) <
 697             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 698                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 699
 700         return err;
 701 }
 702
 703 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 704 {
 705         int err;
 706         struct inode *inode = page->mapping->host;
 707         BUG_ON(!inode);
 708         ihold(inode);
 709         err = writepage_nounlock(page, wbc);
 710         if (err == -ERESTARTSYS) {
 711                 /* direct memory reclaimer was killed by SIGKILL. return 0
 712                  * to prevent caller from setting mapping/page error */
 713                 err = 0;
 714         }
 715         unlock_page(page);
 716         iput(inode);
 717         return err;
 718 }
 719
 720 /*
 721  * async writeback completion handler.
 722  *
 723  * If we get an error, set the mapping error bit, but not the individual
 724  * page error bits.
 725  */
 726 static void writepages_finish(struct ceph_osd_request *req)
 727 {
 728         struct inode *inode = req->r_inode;
 729         struct ceph_inode_info *ci = ceph_inode(inode);
 730         struct ceph_osd_data *osd_data;
 731         struct page *page;
 732         int num_pages, total_pages = 0;
 733         int i, j;
 734         int rc = req->r_result;
 735         struct ceph_snap_context *snapc = req->r_snapc;
 736         struct address_space *mapping = inode->i_mapping;
 737         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 738         bool remove_page;
 739
 740         dout("writepages_finish %p rc %d\n", inode, rc);
 741         if (rc < 0) {
 742                 mapping_set_error(mapping, rc);
 743                 ceph_set_error_write(ci);
 744                 if (rc == -EBLOCKLISTED)
 745                         fsc->blocklisted = true;
 746         } else {
 747                 ceph_clear_error_write(ci);
 748         }
 749
 750         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
 751                                   req->r_end_latency, rc);
 752
 753         /*
 754          * We lost the cache cap, need to truncate the page before
 755          * it is unlocked, otherwise we'd truncate it later in the
 756          * page truncation thread, possibly losing some data that
 757          * raced its way in
 758          */
 759         remove_page = !(ceph_caps_issued(ci) &
 760                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 761
 762         /* clean all pages */
 763         for (i = 0; i < req->r_num_ops; i++) {
 764                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
 765                         break;
 766
 767                 osd_data = osd_req_op_extent_osd_data(req, i);
 768                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 769                 num_pages = calc_pages_for((u64)osd_data->alignment,
 770                                            (u64)osd_data->length);
 771                 total_pages += num_pages;
 772                 for (j = 0; j < num_pages; j++) {
 773                         page = osd_data->pages[j];
 774                         BUG_ON(!page);
 775                         WARN_ON(!PageUptodate(page));
 776
 777                         if (atomic_long_dec_return(&fsc->writeback_count) <
 778                              CONGESTION_OFF_THRESH(
 779                                         fsc->mount_options->congestion_kb))
 780                                 clear_bdi_congested(inode_to_bdi(inode),
 781                                                     BLK_RW_ASYNC);
 782
 783                         ceph_put_snap_context(page_snap_context(page));
 784                         page->private = 0;
 785                         ClearPagePrivate(page);
 786                         dout("unlocking %p\n", page);
 787                         end_page_writeback(page);
 788
 789                         if (remove_page)
 790                                 generic_error_remove_page(inode->i_mapping,
 791                                                           page);
 792
 793                         unlock_page(page);
 794                 }
 795                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
 796                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
 797
 798                 release_pages(osd_data->pages, num_pages);
 799         }
 800
 801         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 802
 803         osd_data = osd_req_op_extent_osd_data(req, 0);
 804         if (osd_data->pages_from_pool)
 805                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 806         else
 807                 kfree(osd_data->pages);
 808         ceph_osdc_put_request(req);
 809 }
 810
 811 /*
 812  * initiate async writeback
 813  */
 814 static int ceph_writepages_start(struct address_space *mapping,
 815                                  struct writeback_control *wbc)
 816 {
 817         struct inode *inode = mapping->host;
 818         struct ceph_inode_info *ci = ceph_inode(inode);
 819         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 820         struct ceph_vino vino = ceph_vino(inode);
 821         pgoff_t index, start_index, end = -1;
 822         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 823         struct pagevec pvec;
 824         int rc = 0;
 825         unsigned int wsize = i_blocksize(inode);
 826         struct ceph_osd_request *req = NULL;
 827         struct ceph_writeback_ctl ceph_wbc;
 828         bool should_loop, range_whole = false;
 829         bool done = false;
 830
 831         dout("writepages_start %p (mode=%s)\n", inode,
 832              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 833              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 834
 835         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
 836                 if (ci->i_wrbuffer_ref > 0) {
 837                         pr_warn_ratelimited(
 838                                 "writepage_start %p %lld forced umount\n",
 839                                 inode, ceph_ino(inode));
 840                 }
 841                 mapping_set_error(mapping, -EIO);
 842                 return -EIO; /* we're in a forced umount, don't write! */
 843         }
 844         if (fsc->mount_options->wsize < wsize)
 845                 wsize = fsc->mount_options->wsize;
 846
 847         pagevec_init(&pvec);
 848
 849         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 850         index = start_index;
 851
 852 retry:
 853         /* find oldest snap context with dirty data */
 854         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 855         if (!snapc) {
 856                 /* hmm, why does writepages get called when there
 857                    is no dirty data? */
 858                 dout(" no snap context with dirty data?\n");
 859                 goto out;
 860         }
 861         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 862              snapc, snapc->seq, snapc->num_snaps);
 863
 864         should_loop = false;
 865         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 866                 /* where to start/end? */
 867                 if (wbc->range_cyclic) {
 868                         index = start_index;
 869                         end = -1;
 870                         if (index > 0)
 871                                 should_loop = true;
 872                         dout(" cyclic, start at %lu\n", index);
 873                 } else {
 874                         index = wbc->range_start >> PAGE_SHIFT;
 875                         end = wbc->range_end >> PAGE_SHIFT;
 876                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 877                                 range_whole = true;
 878                         dout(" not cyclic, %lu to %lu\n", index, end);
 879                 }
 880         } else if (!ceph_wbc.head_snapc) {
 881                 /* Do not respect wbc->range_{start,end}. Dirty pages
 882                  * in that range can be associated with newer snapc.
 883                  * They are not writeable until we write all dirty pages
 884                  * associated with 'snapc' get written */
 885                 if (index > 0)
 886                         should_loop = true;
 887                 dout(" non-head snapc, range whole\n");
 888         }
 889
 890         ceph_put_snap_context(last_snapc);
 891         last_snapc = snapc;
 892
 893         while (!done && index <= end) {
 894                 int num_ops = 0, op_idx;
 895                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
 896                 struct page **pages = NULL, **data_pages;
 897                 struct page *page;
 898                 pgoff_t strip_unit_end = 0;
 899                 u64 offset = 0, len = 0;
 900                 bool from_pool = false;
 901
 902                 max_pages = wsize >> PAGE_SHIFT;
 903
 904 get_more_pages:
 905                 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
 906                                                 end, PAGECACHE_TAG_DIRTY);
 907                 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
 908                 if (!pvec_pages && !locked_pages)
 909                         break;
 910                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 911                         page = pvec.pages[i];
 912                         dout("? %p idx %lu\n", page, page->index);
 913                         if (locked_pages == 0)
 914                                 lock_page(page);  /* first page */
 915                         else if (!trylock_page(page))
 916                                 break;
 917
 918                         /* only dirty pages, or our accounting breaks */
 919                         if (unlikely(!PageDirty(page)) ||
 920                             unlikely(page->mapping != mapping)) {
 921                                 dout("!dirty or !mapping %p\n", page);
 922                                 unlock_page(page);
 923                                 continue;
 924                         }
 925                         /* only if matching snap context */
 926                         pgsnapc = page_snap_context(page);
 927                         if (pgsnapc != snapc) {
 928                                 dout("page snapc %p %lld != oldest %p %lld\n",
 929                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 930                                 if (!should_loop &&
 931                                     !ceph_wbc.head_snapc &&
 932                                     wbc->sync_mode != WB_SYNC_NONE)
 933                                         should_loop = true;
 934                                 unlock_page(page);
 935                                 continue;
 936                         }
 937                         if (page_offset(page) >= ceph_wbc.i_size) {
 938                                 dout("%p page eof %llu\n",
 939                                      page, ceph_wbc.i_size);
 940                                 if ((ceph_wbc.size_stable ||
 941                                     page_offset(page) >= i_size_read(inode)) &&
 942                                     clear_page_dirty_for_io(page))
 943                                         mapping->a_ops->invalidatepage(page,
 944                                                                 0, PAGE_SIZE);
 945                                 unlock_page(page);
 946                                 continue;
 947                         }
 948                         if (strip_unit_end && (page->index > strip_unit_end)) {
 949                                 dout("end of strip unit %p\n", page);
 950                                 unlock_page(page);
 951                                 break;
 952                         }
 953                         if (PageWriteback(page)) {
 954                                 if (wbc->sync_mode == WB_SYNC_NONE) {
 955                                         dout("%p under writeback\n", page);
 956                                         unlock_page(page);
 957                                         continue;
 958                                 }
 959                                 dout("waiting on writeback %p\n", page);
 960                                 wait_on_page_writeback(page);
 961                         }
 962
 963                         if (!clear_page_dirty_for_io(page)) {
 964                                 dout("%p !clear_page_dirty_for_io\n", page);
 965                                 unlock_page(page);
 966                                 continue;
 967                         }
 968
 969                         /*
 970                          * We have something to write.  If this is
 971                          * the first locked page this time through,
 972                          * calculate max possinle write size and
 973                          * allocate a page array
 974                          */
 975                         if (locked_pages == 0) {
 976                                 u64 objnum;
 977                                 u64 objoff;
 978                                 u32 xlen;
 979
 980                                 /* prepare async write request */
 981                                 offset = (u64)page_offset(page);
 982                                 ceph_calc_file_object_mapping(&ci->i_layout,
 983                                                               offset, wsize,
 984                                                               &objnum, &objoff,
 985                                                               &xlen);
 986                                 len = xlen;
 987
 988                                 num_ops = 1;
 989                                 strip_unit_end = page->index +
 990                                         ((len - 1) >> PAGE_SHIFT);
 991
 992                                 BUG_ON(pages);
 993                                 max_pages = calc_pages_for(0, (u64)len);
 994                                 pages = kmalloc_array(max_pages,
 995                                                       sizeof(*pages),
 996                                                       GFP_NOFS);
 997                                 if (!pages) {
 998                                         from_pool = true;
 999                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1000                                         BUG_ON(!pages);
1001                                 }
1002
1003                                 len = 0;
1004                         } else if (page->index !=
1005                                    (offset + len) >> PAGE_SHIFT) {
1006                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1007                                                              CEPH_OSD_MAX_OPS)) {
1008                                         redirty_page_for_writepage(wbc, page);
1009                                         unlock_page(page);
1010                                         break;
1011                                 }
1012
1013                                 num_ops++;
1014                                 offset = (u64)page_offset(page);
1015                                 len = 0;
1016                         }
1017
1018                         /* note position of first page in pvec */
1019                         dout("%p will write page %p idx %lu\n",
1020                              inode, page, page->index);
1021
1022                         if (atomic_long_inc_return(&fsc->writeback_count) >
1023                             CONGESTION_ON_THRESH(
1024                                     fsc->mount_options->congestion_kb)) {
1025                                 set_bdi_congested(inode_to_bdi(inode),
1026                                                   BLK_RW_ASYNC);
1027                         }
1028
1029
1030                         pages[locked_pages++] = page;
1031                         pvec.pages[i] = NULL;
1032
1033                         len += PAGE_SIZE;
1034                 }
1035
1036                 /* did we get anything? */
1037                 if (!locked_pages)
1038                         goto release_pvec_pages;
1039                 if (i) {
1040                         unsigned j, n = 0;
1041                         /* shift unused page to beginning of pvec */
1042                         for (j = 0; j < pvec_pages; j++) {
1043                                 if (!pvec.pages[j])
1044                                         continue;
1045                                 if (n < j)
1046                                         pvec.pages[n] = pvec.pages[j];
1047                                 n++;
1048                         }
1049                         pvec.nr = n;
1050
1051                         if (pvec_pages && i == pvec_pages &&
1052                             locked_pages < max_pages) {
1053                                 dout("reached end pvec, trying for more\n");
1054                                 pagevec_release(&pvec);
1055                                 goto get_more_pages;
1056                         }
1057                 }
1058
1059 new_request:
1060                 offset = page_offset(pages[0]);
1061                 len = wsize;
1062
1063                 req = ceph_osdc_new_request(&fsc->client->osdc,
1064                                         &ci->i_layout, vino,
1065                                         offset, &len, 0, num_ops,
1066                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1067                                         snapc, ceph_wbc.truncate_seq,
1068                                         ceph_wbc.truncate_size, false);
1069                 if (IS_ERR(req)) {
1070                         req = ceph_osdc_new_request(&fsc->client->osdc,
1071                                                 &ci->i_layout, vino,
1072                                                 offset, &len, 0,
1073                                                 min(num_ops,
1074                                                     CEPH_OSD_SLAB_OPS),
1075                                                 CEPH_OSD_OP_WRITE,
1076                                                 CEPH_OSD_FLAG_WRITE,
1077                                                 snapc, ceph_wbc.truncate_seq,
1078                                                 ceph_wbc.truncate_size, true);
1079                         BUG_ON(IS_ERR(req));
1080                 }
1081                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1082                              PAGE_SIZE - offset);
1083
1084                 req->r_callback = writepages_finish;
1085                 req->r_inode = inode;
1086
1087                 /* Format the osd request message and submit the write */
1088                 len = 0;
1089                 data_pages = pages;
1090                 op_idx = 0;
1091                 for (i = 0; i < locked_pages; i++) {
1092                         u64 cur_offset = page_offset(pages[i]);
1093                         if (offset + len != cur_offset) {
1094                                 if (op_idx + 1 == req->r_num_ops)
1095                                         break;
1096                                 osd_req_op_extent_dup_last(req, op_idx,
1097                                                            cur_offset - offset);
1098                                 dout("writepages got pages at %llu~%llu\n",
1099                                      offset, len);
1100                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1101                                                         data_pages, len, 0,
1102                                                         from_pool, false);
1103                                 osd_req_op_extent_update(req, op_idx, len);
1104
1105                                 len = 0;
1106                                 offset = cur_offset;
1107                                 data_pages = pages + i;
1108                                 op_idx++;
1109                         }
1110
1111                         set_page_writeback(pages[i]);
1112                         len += PAGE_SIZE;
1113                 }
1114
1115                 if (ceph_wbc.size_stable) {
1116                         len = min(len, ceph_wbc.i_size - offset);
1117                 } else if (i == locked_pages) {
1118                         /* writepages_finish() clears writeback pages
1119                          * according to the data length, so make sure
1120                          * data length covers all locked pages */
1121                         u64 min_len = len + 1 - PAGE_SIZE;
1122                         len = get_writepages_data_length(inode, pages[i - 1],
1123                                                          offset);
1124                         len = max(len, min_len);
1125                 }
1126                 dout("writepages got pages at %llu~%llu\n", offset, len);
1127
1128                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1129                                                  0, from_pool, false);
1130                 osd_req_op_extent_update(req, op_idx, len);
1131
1132                 BUG_ON(op_idx + 1 != req->r_num_ops);
1133
1134                 from_pool = false;
1135                 if (i < locked_pages) {
1136                         BUG_ON(num_ops <= req->r_num_ops);
1137                         num_ops -= req->r_num_ops;
1138                         locked_pages -= i;
1139
1140                         /* allocate new pages array for next request */
1141                         data_pages = pages;
1142                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1143                                               GFP_NOFS);
1144                         if (!pages) {
1145                                 from_pool = true;
1146                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1147                                 BUG_ON(!pages);
1148                         }
1149                         memcpy(pages, data_pages + i,
1150                                locked_pages * sizeof(*pages));
1151                         memset(data_pages + i, 0,
1152                                locked_pages * sizeof(*pages));
1153                 } else {
1154                         BUG_ON(num_ops != req->r_num_ops);
1155                         index = pages[i - 1]->index + 1;
1156                         /* request message now owns the pages array */
1157                         pages = NULL;
1158                 }
1159
1160                 req->r_mtime = inode->i_mtime;
1161                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1162                 BUG_ON(rc);
1163                 req = NULL;
1164
1165                 wbc->nr_to_write -= i;
1166                 if (pages)
1167                         goto new_request;
1168
1169                 /*
1170                  * We stop writing back only if we are not doing
1171                  * integrity sync. In case of integrity sync we have to
1172                  * keep going until we have written all the pages
1173                  * we tagged for writeback prior to entering this loop.
1174                  */
1175                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1176                         done = true;
1177
1178 release_pvec_pages:
1179                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1180                      pvec.nr ? pvec.pages[0] : NULL);
1181                 pagevec_release(&pvec);
1182         }
1183
1184         if (should_loop && !done) {
1185                 /* more to do; loop back to beginning of file */
1186                 dout("writepages looping back to beginning of file\n");
1187                 end = start_index - 1; /* OK even when start_index == 0 */
1188
1189                 /* to write dirty pages associated with next snapc,
1190                  * we need to wait until current writes complete */
1191                 if (wbc->sync_mode != WB_SYNC_NONE &&
1192                     start_index == 0 && /* all dirty pages were checked */
1193                     !ceph_wbc.head_snapc) {
1194                         struct page *page;
1195                         unsigned i, nr;
1196                         index = 0;
1197                         while ((index <= end) &&
1198                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1199                                                 PAGECACHE_TAG_WRITEBACK))) {
1200                                 for (i = 0; i < nr; i++) {
1201                                         page = pvec.pages[i];
1202                                         if (page_snap_context(page) != snapc)
1203                                                 continue;
1204                                         wait_on_page_writeback(page);
1205                                 }
1206                                 pagevec_release(&pvec);
1207                                 cond_resched();
1208                         }
1209                 }
1210
1211                 start_index = 0;
1212                 index = 0;
1213                 goto retry;
1214         }
1215
1216         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1217                 mapping->writeback_index = index;
1218
1219 out:
1220         ceph_osdc_put_request(req);
1221         ceph_put_snap_context(last_snapc);
1222         dout("writepages dend - startone, rc = %d\n", rc);
1223         return rc;
1224 }
1225
1226
1227
1228 /*
1229  * See if a given @snapc is either writeable, or already written.
1230  */
1231 static int context_is_writeable_or_written(struct inode *inode,
1232                                            struct ceph_snap_context *snapc)
1233 {
1234         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1235         int ret = !oldest || snapc->seq <= oldest->seq;
1236
1237         ceph_put_snap_context(oldest);
1238         return ret;
1239 }
1240
1241 /**
1242  * ceph_find_incompatible - find an incompatible context and return it
1243  * @page: page being dirtied
1244  *
1245  * We are only allowed to write into/dirty a page if the page is
1246  * clean, or already dirty within the same snap context. Returns a
1247  * conflicting context if there is one, NULL if there isn't, or a
1248  * negative error code on other errors.
1249  *
1250  * Must be called with page lock held.
1251  */
1252 static struct ceph_snap_context *
1253 ceph_find_incompatible(struct page *page)
1254 {
1255         struct inode *inode = page->mapping->host;
1256         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1257         struct ceph_inode_info *ci = ceph_inode(inode);
1258
1259         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1260                 dout(" page %p forced umount\n", page);
1261                 return ERR_PTR(-EIO);
1262         }
1263
1264         for (;;) {
1265                 struct ceph_snap_context *snapc, *oldest;
1266
1267                 wait_on_page_writeback(page);
1268
1269                 snapc = page_snap_context(page);
1270                 if (!snapc || snapc == ci->i_head_snapc)
1271                         break;
1272
1273                 /*
1274                  * this page is already dirty in another (older) snap
1275                  * context!  is it writeable now?
1276                  */
1277                 oldest = get_oldest_context(inode, NULL, NULL);
1278                 if (snapc->seq > oldest->seq) {
1279                         /* not writeable -- return it for the caller to deal with */
1280                         ceph_put_snap_context(oldest);
1281                         dout(" page %p snapc %p not current or oldest\n", page, snapc);
1282                         return ceph_get_snap_context(snapc);
1283                 }
1284                 ceph_put_snap_context(oldest);
1285
1286                 /* yay, writeable, do it now (without dropping page lock) */
1287                 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1288                 if (clear_page_dirty_for_io(page)) {
1289                         int r = writepage_nounlock(page, NULL);
1290                         if (r < 0)
1291                                 return ERR_PTR(r);
1292                 }
1293         }
1294         return NULL;
1295 }
1296
1297 /**
1298  * prep_noread_page - prep a page for writing without reading first
1299  * @page: page being prepared
1300  * @pos: starting position for the write
1301  * @len: length of write
1302  *
1303  * In some cases, write_begin doesn't need to read at all:
1304  * - full page write
1305  * - file is currently zero-length
1306  * - write that lies in a page that is completely beyond EOF
1307  * - write that covers the the page from start to EOF or beyond it
1308  *
1309  * If any of these criteria are met, then zero out the unwritten parts
1310  * of the page and return true. Otherwise, return false.
1311  */
1312 static bool skip_page_read(struct page *page, loff_t pos, size_t len)
1313 {
1314         struct inode *inode = page->mapping->host;
1315         loff_t i_size = i_size_read(inode);
1316         size_t offset = offset_in_page(pos);
1317
1318         /* Full page write */
1319         if (offset == 0 && len >= PAGE_SIZE)
1320                 return true;
1321
1322         /* pos beyond last page in the file */
1323         if (pos - offset >= i_size)
1324                 goto zero_out;
1325
1326         /* write that covers the whole page from start to EOF or beyond it */
1327         if (offset == 0 && (pos + len) >= i_size)
1328                 goto zero_out;
1329
1330         return false;
1331 zero_out:
1332         zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1333         return true;
1334 }
1335
1336 /*
1337  * We are only allowed to write into/dirty the page if the page is
1338  * clean, or already dirty within the same snap context.
1339  */
1340 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1341                             loff_t pos, unsigned len, unsigned flags,
1342                             struct page **pagep, void **fsdata)
1343 {
1344         struct inode *inode = file_inode(file);
1345         struct ceph_inode_info *ci = ceph_inode(inode);
1346         struct ceph_snap_context *snapc;
1347         struct page *page = NULL;
1348         pgoff_t index = pos >> PAGE_SHIFT;
1349         int r = 0;
1350
1351         dout("write_begin file %p inode %p page %p %d~%d\n", file, inode, page, (int)pos, (int)len);
1352
1353         for (;;) {
1354                 page = grab_cache_page_write_begin(mapping, index, 0);
1355                 if (!page) {
1356                         r = -ENOMEM;
1357                         break;
1358                 }
1359
1360                 snapc = ceph_find_incompatible(page);
1361                 if (snapc) {
1362                         if (IS_ERR(snapc)) {
1363                                 r = PTR_ERR(snapc);
1364                                 break;
1365                         }
1366                         unlock_page(page);
1367                         put_page(page);
1368                         page = NULL;
1369                         ceph_queue_writeback(inode);
1370                         r = wait_event_killable(ci->i_cap_wq,
1371                                                 context_is_writeable_or_written(inode, snapc));
1372                         ceph_put_snap_context(snapc);
1373                         if (r != 0)
1374                                 break;
1375                         continue;
1376                 }
1377
1378                 if (PageUptodate(page)) {
1379                         dout(" page %p already uptodate\n", page);
1380                         break;
1381                 }
1382
1383                 /* No need to read in some cases */
1384                 if (skip_page_read(page, pos, len))
1385                         break;
1386
1387                 /*
1388                  * We need to read it. If we get back -EINPROGRESS, then the page was
1389                  * handed off to fscache and it will be unlocked when the read completes.
1390                  * Refind the page in that case so we can reacquire the page lock. Otherwise
1391                  * we got a hard error or the read was completed synchronously.
1392                  */
1393                 r = ceph_do_readpage(file, page);
1394                 if (r != -EINPROGRESS)
1395                         break;
1396         }
1397
1398         if (r < 0) {
1399                 if (page) {
1400                         unlock_page(page);
1401                         put_page(page);
1402                 }
1403         } else {
1404                 *pagep = page;
1405         }
1406         return r;
1407 }
1408
1409 /*
1410  * we don't do anything in here that simple_write_end doesn't do
1411  * except adjust dirty page accounting
1412  */
1413 static int ceph_write_end(struct file *file, struct address_space *mapping,
1414                           loff_t pos, unsigned len, unsigned copied,
1415                           struct page *page, void *fsdata)
1416 {
1417         struct inode *inode = file_inode(file);
1418         bool check_cap = false;
1419
1420         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1421              inode, page, (int)pos, (int)copied, (int)len);
1422
1423         /* zero the stale part of the page if we did a short copy */
1424         if (!PageUptodate(page)) {
1425                 if (copied < len) {
1426                         copied = 0;
1427                         goto out;
1428                 }
1429                 SetPageUptodate(page);
1430         }
1431
1432         /* did file size increase? */
1433         if (pos+copied > i_size_read(inode))
1434                 check_cap = ceph_inode_set_size(inode, pos+copied);
1435
1436         set_page_dirty(page);
1437
1438 out:
1439         unlock_page(page);
1440         put_page(page);
1441
1442         if (check_cap)
1443                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1444
1445         return copied;
1446 }
1447
1448 /*
1449  * we set .direct_IO to indicate direct io is supported, but since we
1450  * intercept O_DIRECT reads and writes early, this function should
1451  * never get called.
1452  */
1453 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1454 {
1455         WARN_ON(1);
1456         return -EINVAL;
1457 }
1458
1459 const struct address_space_operations ceph_aops = {
1460         .readpage = ceph_readpage,
1461         .readpages = ceph_readpages,
1462         .writepage = ceph_writepage,
1463         .writepages = ceph_writepages_start,
1464         .write_begin = ceph_write_begin,
1465         .write_end = ceph_write_end,
1466         .set_page_dirty = ceph_set_page_dirty,
1467         .invalidatepage = ceph_invalidatepage,
1468         .releasepage = ceph_releasepage,
1469         .direct_IO = ceph_direct_io,
1470 };
1471
1472 static void ceph_block_sigs(sigset_t *oldset)
1473 {
1474         sigset_t mask;
1475         siginitsetinv(&mask, sigmask(SIGKILL));
1476         sigprocmask(SIG_BLOCK, &mask, oldset);
1477 }
1478
1479 static void ceph_restore_sigs(sigset_t *oldset)
1480 {
1481         sigprocmask(SIG_SETMASK, oldset, NULL);
1482 }
1483
1484 /*
1485  * vm ops
1486  */
1487 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1488 {
1489         struct vm_area_struct *vma = vmf->vma;
1490         struct inode *inode = file_inode(vma->vm_file);
1491         struct ceph_inode_info *ci = ceph_inode(inode);
1492         struct ceph_file_info *fi = vma->vm_file->private_data;
1493         struct page *pinned_page = NULL;
1494         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1495         int want, got, err;
1496         sigset_t oldset;
1497         vm_fault_t ret = VM_FAULT_SIGBUS;
1498
1499         ceph_block_sigs(&oldset);
1500
1501         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1502              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1503         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1504                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1505         else
1506                 want = CEPH_CAP_FILE_CACHE;
1507
1508         got = 0;
1509         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1,
1510                             &got, &pinned_page);
1511         if (err < 0)
1512                 goto out_restore;
1513
1514         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1515              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1516
1517         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1518             ci->i_inline_version == CEPH_INLINE_NONE) {
1519                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1520                 ceph_add_rw_context(fi, &rw_ctx);
1521                 ret = filemap_fault(vmf);
1522                 ceph_del_rw_context(fi, &rw_ctx);
1523                 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n",
1524                         inode, off, (size_t)PAGE_SIZE,
1525                                 ceph_cap_string(got), ret);
1526         } else
1527                 err = -EAGAIN;
1528
1529         if (pinned_page)
1530                 put_page(pinned_page);
1531         ceph_put_cap_refs(ci, got);
1532
1533         if (err != -EAGAIN)
1534                 goto out_restore;
1535
1536         /* read inline data */
1537         if (off >= PAGE_SIZE) {
1538                 /* does not support inline data > PAGE_SIZE */
1539                 ret = VM_FAULT_SIGBUS;
1540         } else {
1541                 struct address_space *mapping = inode->i_mapping;
1542                 struct page *page = find_or_create_page(mapping, 0,
1543                                                 mapping_gfp_constraint(mapping,
1544                                                 ~__GFP_FS));
1545                 if (!page) {
1546                         ret = VM_FAULT_OOM;
1547                         goto out_inline;
1548                 }
1549                 err = __ceph_do_getattr(inode, page,
1550                                          CEPH_STAT_CAP_INLINE_DATA, true);
1551                 if (err < 0 || off >= i_size_read(inode)) {
1552                         unlock_page(page);
1553                         put_page(page);
1554                         ret = vmf_error(err);
1555                         goto out_inline;
1556                 }
1557                 if (err < PAGE_SIZE)
1558                         zero_user_segment(page, err, PAGE_SIZE);
1559                 else
1560                         flush_dcache_page(page);
1561                 SetPageUptodate(page);
1562                 vmf->page = page;
1563                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1564 out_inline:
1565                 dout("filemap_fault %p %llu~%zd read inline data ret %x\n",
1566                      inode, off, (size_t)PAGE_SIZE, ret);
1567         }
1568 out_restore:
1569         ceph_restore_sigs(&oldset);
1570         if (err < 0)
1571                 ret = vmf_error(err);
1572
1573         return ret;
1574 }
1575
1576 /*
1577  * Reuse write_begin here for simplicity.
1578  */
1579 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1580 {
1581         struct vm_area_struct *vma = vmf->vma;
1582         struct inode *inode = file_inode(vma->vm_file);
1583         struct ceph_inode_info *ci = ceph_inode(inode);
1584         struct ceph_file_info *fi = vma->vm_file->private_data;
1585         struct ceph_cap_flush *prealloc_cf;
1586         struct page *page = vmf->page;
1587         loff_t off = page_offset(page);
1588         loff_t size = i_size_read(inode);
1589         size_t len;
1590         int want, got, err;
1591         sigset_t oldset;
1592         vm_fault_t ret = VM_FAULT_SIGBUS;
1593
1594         prealloc_cf = ceph_alloc_cap_flush();
1595         if (!prealloc_cf)
1596                 return VM_FAULT_OOM;
1597
1598         sb_start_pagefault(inode->i_sb);
1599         ceph_block_sigs(&oldset);
1600
1601         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1602                 struct page *locked_page = NULL;
1603                 if (off == 0) {
1604                         lock_page(page);
1605                         locked_page = page;
1606                 }
1607                 err = ceph_uninline_data(vma->vm_file, locked_page);
1608                 if (locked_page)
1609                         unlock_page(locked_page);
1610                 if (err < 0)
1611                         goto out_free;
1612         }
1613
1614         if (off + PAGE_SIZE <= size)
1615                 len = PAGE_SIZE;
1616         else
1617                 len = size & ~PAGE_MASK;
1618
1619         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1620              inode, ceph_vinop(inode), off, len, size);
1621         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1622                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1623         else
1624                 want = CEPH_CAP_FILE_BUFFER;
1625
1626         got = 0;
1627         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len,
1628                             &got, NULL);
1629         if (err < 0)
1630                 goto out_free;
1631
1632         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1633              inode, off, len, ceph_cap_string(got));
1634
1635         /* Update time before taking page lock */
1636         file_update_time(vma->vm_file);
1637         inode_inc_iversion_raw(inode);
1638
1639         do {
1640                 struct ceph_snap_context *snapc;
1641
1642                 lock_page(page);
1643
1644                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1645                         unlock_page(page);
1646                         ret = VM_FAULT_NOPAGE;
1647                         break;
1648                 }
1649
1650                 snapc = ceph_find_incompatible(page);
1651                 if (!snapc) {
1652                         /* success.  we'll keep the page locked. */
1653                         set_page_dirty(page);
1654                         ret = VM_FAULT_LOCKED;
1655                         break;
1656                 }
1657
1658                 unlock_page(page);
1659
1660                 if (IS_ERR(snapc)) {
1661                         ret = VM_FAULT_SIGBUS;
1662                         break;
1663                 }
1664
1665                 ceph_queue_writeback(inode);
1666                 err = wait_event_killable(ci->i_cap_wq,
1667                                 context_is_writeable_or_written(inode, snapc));
1668                 ceph_put_snap_context(snapc);
1669         } while (err == 0);
1670
1671         if (ret == VM_FAULT_LOCKED ||
1672             ci->i_inline_version != CEPH_INLINE_NONE) {
1673                 int dirty;
1674                 spin_lock(&ci->i_ceph_lock);
1675                 ci->i_inline_version = CEPH_INLINE_NONE;
1676                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1677                                                &prealloc_cf);
1678                 spin_unlock(&ci->i_ceph_lock);
1679                 if (dirty)
1680                         __mark_inode_dirty(inode, dirty);
1681         }
1682
1683         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1684              inode, off, len, ceph_cap_string(got), ret);
1685         ceph_put_cap_refs(ci, got);
1686 out_free:
1687         ceph_restore_sigs(&oldset);
1688         sb_end_pagefault(inode->i_sb);
1689         ceph_free_cap_flush(prealloc_cf);
1690         if (err < 0)
1691                 ret = vmf_error(err);
1692         return ret;
1693 }
1694
1695 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1696                            char *data, size_t len)
1697 {
1698         struct address_space *mapping = inode->i_mapping;
1699         struct page *page;
1700
1701         if (locked_page) {
1702                 page = locked_page;
1703         } else {
1704                 if (i_size_read(inode) == 0)
1705                         return;
1706                 page = find_or_create_page(mapping, 0,
1707                                            mapping_gfp_constraint(mapping,
1708                                            ~__GFP_FS));
1709                 if (!page)
1710                         return;
1711                 if (PageUptodate(page)) {
1712                         unlock_page(page);
1713                         put_page(page);
1714                         return;
1715                 }
1716         }
1717
1718         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1719              inode, ceph_vinop(inode), len, locked_page);
1720
1721         if (len > 0) {
1722                 void *kaddr = kmap_atomic(page);
1723                 memcpy(kaddr, data, len);
1724                 kunmap_atomic(kaddr);
1725         }
1726
1727         if (page != locked_page) {
1728                 if (len < PAGE_SIZE)
1729                         zero_user_segment(page, len, PAGE_SIZE);
1730                 else
1731                         flush_dcache_page(page);
1732
1733                 SetPageUptodate(page);
1734                 unlock_page(page);
1735                 put_page(page);
1736         }
1737 }
1738
1739 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1740 {
1741         struct inode *inode = file_inode(filp);
1742         struct ceph_inode_info *ci = ceph_inode(inode);
1743         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1744         struct ceph_osd_request *req;
1745         struct page *page = NULL;
1746         u64 len, inline_version;
1747         int err = 0;
1748         bool from_pagecache = false;
1749
1750         spin_lock(&ci->i_ceph_lock);
1751         inline_version = ci->i_inline_version;
1752         spin_unlock(&ci->i_ceph_lock);
1753
1754         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1755              inode, ceph_vinop(inode), inline_version);
1756
1757         if (inline_version == 1 || /* initial version, no data */
1758             inline_version == CEPH_INLINE_NONE)
1759                 goto out;
1760
1761         if (locked_page) {
1762                 page = locked_page;
1763                 WARN_ON(!PageUptodate(page));
1764         } else if (ceph_caps_issued(ci) &
1765                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1766                 page = find_get_page(inode->i_mapping, 0);
1767                 if (page) {
1768                         if (PageUptodate(page)) {
1769                                 from_pagecache = true;
1770                                 lock_page(page);
1771                         } else {
1772                                 put_page(page);
1773                                 page = NULL;
1774                         }
1775                 }
1776         }
1777
1778         if (page) {
1779                 len = i_size_read(inode);
1780                 if (len > PAGE_SIZE)
1781                         len = PAGE_SIZE;
1782         } else {
1783                 page = __page_cache_alloc(GFP_NOFS);
1784                 if (!page) {
1785                         err = -ENOMEM;
1786                         goto out;
1787                 }
1788                 err = __ceph_do_getattr(inode, page,
1789                                         CEPH_STAT_CAP_INLINE_DATA, true);
1790                 if (err < 0) {
1791                         /* no inline data */
1792                         if (err == -ENODATA)
1793                                 err = 0;
1794                         goto out;
1795                 }
1796                 len = err;
1797         }
1798
1799         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1800                                     ceph_vino(inode), 0, &len, 0, 1,
1801                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1802                                     NULL, 0, 0, false);
1803         if (IS_ERR(req)) {
1804                 err = PTR_ERR(req);
1805                 goto out;
1806         }
1807
1808         req->r_mtime = inode->i_mtime;
1809         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1810         if (!err)
1811                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1812         ceph_osdc_put_request(req);
1813         if (err < 0)
1814                 goto out;
1815
1816         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1817                                     ceph_vino(inode), 0, &len, 1, 3,
1818                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1819                                     NULL, ci->i_truncate_seq,
1820                                     ci->i_truncate_size, false);
1821         if (IS_ERR(req)) {
1822                 err = PTR_ERR(req);
1823                 goto out;
1824         }
1825
1826         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1827
1828         {
1829                 __le64 xattr_buf = cpu_to_le64(inline_version);
1830                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1831                                             "inline_version", &xattr_buf,
1832                                             sizeof(xattr_buf),
1833                                             CEPH_OSD_CMPXATTR_OP_GT,
1834                                             CEPH_OSD_CMPXATTR_MODE_U64);
1835                 if (err)
1836                         goto out_put;
1837         }
1838
1839         {
1840                 char xattr_buf[32];
1841                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1842                                          "%llu", inline_version);
1843                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1844                                             "inline_version",
1845                                             xattr_buf, xattr_len, 0, 0);
1846                 if (err)
1847                         goto out_put;
1848         }
1849
1850         req->r_mtime = inode->i_mtime;
1851         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1852         if (!err)
1853                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1854
1855         ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
1856                                   req->r_end_latency, err);
1857
1858 out_put:
1859         ceph_osdc_put_request(req);
1860         if (err == -ECANCELED)
1861                 err = 0;
1862 out:
1863         if (page && page != locked_page) {
1864                 if (from_pagecache) {
1865                         unlock_page(page);
1866                         put_page(page);
1867                 } else
1868                         __free_pages(page, 0);
1869         }
1870
1871         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1872              inode, ceph_vinop(inode), inline_version, err);
1873         return err;
1874 }
1875
1876 static const struct vm_operations_struct ceph_vmops = {
1877         .fault          = ceph_filemap_fault,
1878         .page_mkwrite   = ceph_page_mkwrite,
1879 };
1880
1881 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1882 {
1883         struct address_space *mapping = file->f_mapping;
1884
1885         if (!mapping->a_ops->readpage)
1886                 return -ENOEXEC;
1887         file_accessed(file);
1888         vma->vm_ops = &ceph_vmops;
1889         return 0;
1890 }
1891
1892 enum {
1893         POOL_READ       = 1,
1894         POOL_WRITE      = 2,
1895 };
1896
1897 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1898                                 s64 pool, struct ceph_string *pool_ns)
1899 {
1900         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1901         struct ceph_mds_client *mdsc = fsc->mdsc;
1902         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1903         struct rb_node **p, *parent;
1904         struct ceph_pool_perm *perm;
1905         struct page **pages;
1906         size_t pool_ns_len;
1907         int err = 0, err2 = 0, have = 0;
1908
1909         down_read(&mdsc->pool_perm_rwsem);
1910         p = &mdsc->pool_perm_tree.rb_node;
1911         while (*p) {
1912                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1913                 if (pool < perm->pool)
1914                         p = &(*p)->rb_left;
1915                 else if (pool > perm->pool)
1916                         p = &(*p)->rb_right;
1917                 else {
1918                         int ret = ceph_compare_string(pool_ns,
1919                                                 perm->pool_ns,
1920                                                 perm->pool_ns_len);
1921                         if (ret < 0)
1922                                 p = &(*p)->rb_left;
1923                         else if (ret > 0)
1924                                 p = &(*p)->rb_right;
1925                         else {
1926                                 have = perm->perm;
1927                                 break;
1928                         }
1929                 }
1930         }
1931         up_read(&mdsc->pool_perm_rwsem);
1932         if (*p)
1933                 goto out;
1934
1935         if (pool_ns)
1936                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1937                      pool, (int)pool_ns->len, pool_ns->str);
1938         else
1939                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1940
1941         down_write(&mdsc->pool_perm_rwsem);
1942         p = &mdsc->pool_perm_tree.rb_node;
1943         parent = NULL;
1944         while (*p) {
1945                 parent = *p;
1946                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1947                 if (pool < perm->pool)
1948                         p = &(*p)->rb_left;
1949                 else if (pool > perm->pool)
1950                         p = &(*p)->rb_right;
1951                 else {
1952                         int ret = ceph_compare_string(pool_ns,
1953                                                 perm->pool_ns,
1954                                                 perm->pool_ns_len);
1955                         if (ret < 0)
1956                                 p = &(*p)->rb_left;
1957                         else if (ret > 0)
1958                                 p = &(*p)->rb_right;
1959                         else {
1960                                 have = perm->perm;
1961                                 break;
1962                         }
1963                 }
1964         }
1965         if (*p) {
1966                 up_write(&mdsc->pool_perm_rwsem);
1967                 goto out;
1968         }
1969
1970         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1971                                          1, false, GFP_NOFS);
1972         if (!rd_req) {
1973                 err = -ENOMEM;
1974                 goto out_unlock;
1975         }
1976
1977         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1978         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1979         rd_req->r_base_oloc.pool = pool;
1980         if (pool_ns)
1981                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1982         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1983
1984         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1985         if (err)
1986                 goto out_unlock;
1987
1988         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1989                                          1, false, GFP_NOFS);
1990         if (!wr_req) {
1991                 err = -ENOMEM;
1992                 goto out_unlock;
1993         }
1994
1995         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1996         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1997         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1998         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1999
2000         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2001         if (err)
2002                 goto out_unlock;
2003
2004         /* one page should be large enough for STAT data */
2005         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2006         if (IS_ERR(pages)) {
2007                 err = PTR_ERR(pages);
2008                 goto out_unlock;
2009         }
2010
2011         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2012                                      0, false, true);
2013         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
2014
2015         wr_req->r_mtime = ci->vfs_inode.i_mtime;
2016         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
2017
2018         if (!err)
2019                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2020         if (!err2)
2021                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2022
2023         if (err >= 0 || err == -ENOENT)
2024                 have |= POOL_READ;
2025         else if (err != -EPERM) {
2026                 if (err == -EBLOCKLISTED)
2027                         fsc->blocklisted = true;
2028                 goto out_unlock;
2029         }
2030
2031         if (err2 == 0 || err2 == -EEXIST)
2032                 have |= POOL_WRITE;
2033         else if (err2 != -EPERM) {
2034                 if (err2 == -EBLOCKLISTED)
2035                         fsc->blocklisted = true;
2036                 err = err2;
2037                 goto out_unlock;
2038         }
2039
2040         pool_ns_len = pool_ns ? pool_ns->len : 0;
2041         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2042         if (!perm) {
2043                 err = -ENOMEM;
2044                 goto out_unlock;
2045         }
2046
2047         perm->pool = pool;
2048         perm->perm = have;
2049         perm->pool_ns_len = pool_ns_len;
2050         if (pool_ns_len > 0)
2051                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2052         perm->pool_ns[pool_ns_len] = 0;
2053
2054         rb_link_node(&perm->node, parent, p);
2055         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2056         err = 0;
2057 out_unlock:
2058         up_write(&mdsc->pool_perm_rwsem);
2059
2060         ceph_osdc_put_request(rd_req);
2061         ceph_osdc_put_request(wr_req);
2062 out:
2063         if (!err)
2064                 err = have;
2065         if (pool_ns)
2066                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2067                      pool, (int)pool_ns->len, pool_ns->str, err);
2068         else
2069                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2070         return err;
2071 }
2072
2073 int ceph_pool_perm_check(struct inode *inode, int need)
2074 {
2075         struct ceph_inode_info *ci = ceph_inode(inode);
2076         struct ceph_string *pool_ns;
2077         s64 pool;
2078         int ret, flags;
2079
2080         if (ci->i_vino.snap != CEPH_NOSNAP) {
2081                 /*
2082                  * Pool permission check needs to write to the first object.
2083                  * But for snapshot, head of the first object may have alread
2084                  * been deleted. Skip check to avoid creating orphan object.
2085                  */
2086                 return 0;
2087         }
2088
2089         if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2090                                 NOPOOLPERM))
2091                 return 0;
2092
2093         spin_lock(&ci->i_ceph_lock);
2094         flags = ci->i_ceph_flags;
2095         pool = ci->i_layout.pool_id;
2096         spin_unlock(&ci->i_ceph_lock);
2097 check:
2098         if (flags & CEPH_I_POOL_PERM) {
2099                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2100                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2101                              pool);
2102                         return -EPERM;
2103                 }
2104                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2105                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2106                              pool);
2107                         return -EPERM;
2108                 }
2109                 return 0;
2110         }
2111
2112         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2113         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2114         ceph_put_string(pool_ns);
2115         if (ret < 0)
2116                 return ret;
2117
2118         flags = CEPH_I_POOL_PERM;
2119         if (ret & POOL_READ)
2120                 flags |= CEPH_I_POOL_RD;
2121         if (ret & POOL_WRITE)
2122                 flags |= CEPH_I_POOL_WR;
2123
2124         spin_lock(&ci->i_ceph_lock);
2125         if (pool == ci->i_layout.pool_id &&
2126             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2127                 ci->i_ceph_flags |= flags;
2128         } else {
2129                 pool = ci->i_layout.pool_id;
2130                 flags = ci->i_ceph_flags;
2131         }
2132         spin_unlock(&ci->i_ceph_lock);
2133         goto check;
2134 }
2135
2136 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2137 {
2138         struct ceph_pool_perm *perm;
2139         struct rb_node *n;
2140
2141         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2142                 n = rb_first(&mdsc->pool_perm_tree);
2143                 perm = rb_entry(n, struct ceph_pool_perm, node);
2144                 rb_erase(n, &mdsc->pool_perm_tree);
2145                 kfree(perm);
2146         }
2147 }