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/swap.h>
   8 #include <linux/pagemap.h>
   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 #include <linux/netfs.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include "metric.h"
  21 #include "crypto.h"
  22 #include <linux/ceph/osd_client.h>
  23 #include <linux/ceph/striper.h>
  24
  25 /*
  26  * Ceph address space ops.
  27  *
  28  * There are a few funny things going on here.
  29  *
  30  * The page->private field is used to reference a struct
  31  * ceph_snap_context for _every_ dirty page.  This indicates which
  32  * snapshot the page was logically dirtied in, and thus which snap
  33  * context needs to be associated with the osd write during writeback.
  34  *
  35  * Similarly, struct ceph_inode_info maintains a set of counters to
  36  * count dirty pages on the inode.  In the absence of snapshots,
  37  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  38  *
  39  * When a snapshot is taken (that is, when the client receives
  40  * notification that a snapshot was taken), each inode with caps and
  41  * with dirty pages (dirty pages implies there is a cap) gets a new
  42  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  43  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  44  * moved to capsnap->dirty. (Unless a sync write is currently in
  45  * progress.  In that case, the capsnap is said to be "pending", new
  46  * writes cannot start, and the capsnap isn't "finalized" until the
  47  * write completes (or fails) and a final size/mtime for the inode for
  48  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  49  *
  50  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  51  * we look for the first capsnap in i_cap_snaps and write out pages in
  52  * that snap context _only_.  Then we move on to the next capsnap,
  53  * eventually reaching the "live" or "head" context (i.e., pages that
  54  * are not yet snapped) and are writing the most recently dirtied
  55  * pages.
  56  *
  57  * Invalidate and so forth must take care to ensure the dirty page
  58  * accounting is preserved.
  59  */
  60
  61 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  62 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  63         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  64          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  65
  66 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
  67                                         struct folio **foliop, void **_fsdata);
  68
  69 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  70 {
  71         if (PagePrivate(page))
  72                 return (void *)page->private;
  73         return NULL;
  74 }
  75
  76 /*
  77  * Dirty a page.  Optimistically adjust accounting, on the assumption
  78  * that we won't race with invalidate.  If we do, readjust.
  79  */
  80 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
  81 {
  82         struct inode *inode = mapping->host;
  83         struct ceph_client *cl = ceph_inode_to_client(inode);
  84         struct ceph_inode_info *ci;
  85         struct ceph_snap_context *snapc;
  86
  87         if (folio_test_dirty(folio)) {
  88                 doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
  89                       ceph_vinop(inode), folio, folio->index);
  90                 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
  91                 return false;
  92         }
  93
  94         ci = ceph_inode(inode);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
  99         if (__ceph_have_pending_cap_snap(ci)) {
 100                 struct ceph_cap_snap *capsnap =
 101                                 list_last_entry(&ci->i_cap_snaps,
 102                                                 struct ceph_cap_snap,
 103                                                 ci_item);
 104                 snapc = ceph_get_snap_context(capsnap->context);
 105                 capsnap->dirty_pages++;
 106         } else {
 107                 BUG_ON(!ci->i_head_snapc);
 108                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 109                 ++ci->i_wrbuffer_ref_head;
 110         }
 111         if (ci->i_wrbuffer_ref == 0)
 112                 ihold(inode);
 113         ++ci->i_wrbuffer_ref;
 114         doutc(cl, "%llx.%llx %p idx %lu head %d/%d -> %d/%d "
 115               "snapc %p seq %lld (%d snaps)\n",
 116               ceph_vinop(inode), folio, folio->index,
 117               ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 118               ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 119               snapc, snapc->seq, snapc->num_snaps);
 120         spin_unlock(&ci->i_ceph_lock);
 121
 122         /*
 123          * Reference snap context in folio->private.  Also set
 124          * PagePrivate so that we get invalidate_folio callback.
 125          */
 126         VM_WARN_ON_FOLIO(folio->private, folio);
 127         folio_attach_private(folio, snapc);
 128
 129         return ceph_fscache_dirty_folio(mapping, folio);
 130 }
 131
 132 /*
 133  * If we are truncating the full folio (i.e. offset == 0), adjust the
 134  * dirty folio counters appropriately.  Only called if there is private
 135  * data on the folio.
 136  */
 137 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
 138                                 size_t length)
 139 {
 140         struct inode *inode = folio->mapping->host;
 141         struct ceph_client *cl = ceph_inode_to_client(inode);
 142         struct ceph_inode_info *ci = ceph_inode(inode);
 143         struct ceph_snap_context *snapc;
 144
 145
 146         if (offset != 0 || length != folio_size(folio)) {
 147                 doutc(cl, "%llx.%llx idx %lu partial dirty page %zu~%zu\n",
 148                       ceph_vinop(inode), folio->index, offset, length);
 149                 return;
 150         }
 151
 152         WARN_ON(!folio_test_locked(folio));
 153         if (folio_test_private(folio)) {
 154                 doutc(cl, "%llx.%llx idx %lu full dirty page\n",
 155                       ceph_vinop(inode), folio->index);
 156
 157                 snapc = folio_detach_private(folio);
 158                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 159                 ceph_put_snap_context(snapc);
 160         }
 161
 162         netfs_invalidate_folio(folio, offset, length);
 163 }
 164
 165 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 166 {
 167         struct inode *inode = rreq->inode;
 168         struct ceph_inode_info *ci = ceph_inode(inode);
 169         struct ceph_file_layout *lo = &ci->i_layout;
 170         unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
 171         loff_t end = rreq->start + rreq->len, new_end;
 172         struct ceph_netfs_request_data *priv = rreq->netfs_priv;
 173         unsigned long max_len;
 174         u32 blockoff;
 175
 176         if (priv) {
 177                 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
 178                 if (priv->file_ra_disabled)
 179                         max_pages = 0;
 180                 else
 181                         max_pages = priv->file_ra_pages;
 182
 183         }
 184
 185         /* Readahead is disabled */
 186         if (!max_pages)
 187                 return;
 188
 189         max_len = max_pages << PAGE_SHIFT;
 190
 191         /*
 192          * Try to expand the length forward by rounding up it to the next
 193          * block, but do not exceed the file size, unless the original
 194          * request already exceeds it.
 195          */
 196         new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
 197         if (new_end > end && new_end <= rreq->start + max_len)
 198                 rreq->len = new_end - rreq->start;
 199
 200         /* Try to expand the start downward */
 201         div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
 202         if (rreq->len + blockoff <= max_len) {
 203                 rreq->start -= blockoff;
 204                 rreq->len += blockoff;
 205         }
 206 }
 207
 208 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
 209 {
 210         struct inode *inode = subreq->rreq->inode;
 211         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 212         struct ceph_inode_info *ci = ceph_inode(inode);
 213         u64 objno, objoff;
 214         u32 xlen;
 215
 216         /* Truncate the extent at the end of the current block */
 217         ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
 218                                       &objno, &objoff, &xlen);
 219         subreq->len = min(xlen, fsc->mount_options->rsize);
 220         return true;
 221 }
 222
 223 static void finish_netfs_read(struct ceph_osd_request *req)
 224 {
 225         struct inode *inode = req->r_inode;
 226         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 227         struct ceph_client *cl = fsc->client;
 228         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 229         struct netfs_io_subrequest *subreq = req->r_priv;
 230         struct ceph_osd_req_op *op = &req->r_ops[0];
 231         int err = req->r_result;
 232         bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
 233
 234         ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
 235                                  req->r_end_latency, osd_data->length, err);
 236
 237         doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
 238               subreq->len, i_size_read(req->r_inode));
 239
 240         /* no object means success but no data */
 241         if (err == -ENOENT)
 242                 err = 0;
 243         else if (err == -EBLOCKLISTED)
 244                 fsc->blocklisted = true;
 245
 246         if (err >= 0) {
 247                 if (sparse && err > 0)
 248                         err = ceph_sparse_ext_map_end(op);
 249                 if (err < subreq->len)
 250                         __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 251                 if (IS_ENCRYPTED(inode) && err > 0) {
 252                         err = ceph_fscrypt_decrypt_extents(inode,
 253                                         osd_data->pages, subreq->start,
 254                                         op->extent.sparse_ext,
 255                                         op->extent.sparse_ext_cnt);
 256                         if (err > subreq->len)
 257                                 err = subreq->len;
 258                 }
 259         }
 260
 261         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 262                 ceph_put_page_vector(osd_data->pages,
 263                                      calc_pages_for(osd_data->alignment,
 264                                         osd_data->length), false);
 265         }
 266         netfs_subreq_terminated(subreq, err, false);
 267         iput(req->r_inode);
 268         ceph_dec_osd_stopping_blocker(fsc->mdsc);
 269 }
 270
 271 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 272 {
 273         struct netfs_io_request *rreq = subreq->rreq;
 274         struct inode *inode = rreq->inode;
 275         struct ceph_mds_reply_info_parsed *rinfo;
 276         struct ceph_mds_reply_info_in *iinfo;
 277         struct ceph_mds_request *req;
 278         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 279         struct ceph_inode_info *ci = ceph_inode(inode);
 280         struct iov_iter iter;
 281         ssize_t err = 0;
 282         size_t len;
 283         int mode;
 284
 285         __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
 286         __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 287
 288         if (subreq->start >= inode->i_size)
 289                 goto out;
 290
 291         /* We need to fetch the inline data. */
 292         mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
 293         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
 294         if (IS_ERR(req)) {
 295                 err = PTR_ERR(req);
 296                 goto out;
 297         }
 298         req->r_ino1 = ci->i_vino;
 299         req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 300         req->r_num_caps = 2;
 301
 302         err = ceph_mdsc_do_request(mdsc, NULL, req);
 303         if (err < 0)
 304                 goto out;
 305
 306         rinfo = &req->r_reply_info;
 307         iinfo = &rinfo->targeti;
 308         if (iinfo->inline_version == CEPH_INLINE_NONE) {
 309                 /* The data got uninlined */
 310                 ceph_mdsc_put_request(req);
 311                 return false;
 312         }
 313
 314         len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
 315         iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
 316         err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
 317         if (err == 0)
 318                 err = -EFAULT;
 319
 320         ceph_mdsc_put_request(req);
 321 out:
 322         netfs_subreq_terminated(subreq, err, false);
 323         return true;
 324 }
 325
 326 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 327 {
 328         struct netfs_io_request *rreq = subreq->rreq;
 329         struct inode *inode = rreq->inode;
 330         struct ceph_inode_info *ci = ceph_inode(inode);
 331         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 332         struct ceph_client *cl = fsc->client;
 333         struct ceph_osd_request *req = NULL;
 334         struct ceph_vino vino = ceph_vino(inode);
 335         struct iov_iter iter;
 336         int err = 0;
 337         u64 len = subreq->len;
 338         bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
 339         u64 off = subreq->start;
 340         int extent_cnt;
 341
 342         if (ceph_inode_is_shutdown(inode)) {
 343                 err = -EIO;
 344                 goto out;
 345         }
 346
 347         if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
 348                 return;
 349
 350         ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
 351
 352         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
 353                         off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
 354                         CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
 355                         ci->i_truncate_size, false);
 356         if (IS_ERR(req)) {
 357                 err = PTR_ERR(req);
 358                 req = NULL;
 359                 goto out;
 360         }
 361
 362         if (sparse) {
 363                 extent_cnt = __ceph_sparse_read_ext_count(inode, len);
 364                 err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
 365                 if (err)
 366                         goto out;
 367         }
 368
 369         doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
 370               ceph_vinop(inode), subreq->start, subreq->len, len);
 371
 372         iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
 373
 374         /*
 375          * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
 376          * encrypted inodes. We'd need infrastructure that handles an iov_iter
 377          * instead of page arrays, and we don't have that as of yet. Once the
 378          * dust settles on the write helpers and encrypt/decrypt routines for
 379          * netfs, we should be able to rework this.
 380          */
 381         if (IS_ENCRYPTED(inode)) {
 382                 struct page **pages;
 383                 size_t page_off;
 384
 385                 err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
 386                 if (err < 0) {
 387                         doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
 388                               ceph_vinop(inode), err);
 389                         goto out;
 390                 }
 391
 392                 /* should always give us a page-aligned read */
 393                 WARN_ON_ONCE(page_off);
 394                 len = err;
 395                 err = 0;
 396
 397                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
 398                                                  false);
 399         } else {
 400                 osd_req_op_extent_osd_iter(req, 0, &iter);
 401         }
 402         if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
 403                 err = -EIO;
 404                 goto out;
 405         }
 406         req->r_callback = finish_netfs_read;
 407         req->r_priv = subreq;
 408         req->r_inode = inode;
 409         ihold(inode);
 410
 411         ceph_osdc_start_request(req->r_osdc, req);
 412 out:
 413         ceph_osdc_put_request(req);
 414         if (err)
 415                 netfs_subreq_terminated(subreq, err, false);
 416         doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
 417 }
 418
 419 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 420 {
 421         struct inode *inode = rreq->inode;
 422         struct ceph_client *cl = ceph_inode_to_client(inode);
 423         int got = 0, want = CEPH_CAP_FILE_CACHE;
 424         struct ceph_netfs_request_data *priv;
 425         int ret = 0;
 426
 427         if (rreq->origin != NETFS_READAHEAD)
 428                 return 0;
 429
 430         priv = kzalloc(sizeof(*priv), GFP_NOFS);
 431         if (!priv)
 432                 return -ENOMEM;
 433
 434         if (file) {
 435                 struct ceph_rw_context *rw_ctx;
 436                 struct ceph_file_info *fi = file->private_data;
 437
 438                 priv->file_ra_pages = file->f_ra.ra_pages;
 439                 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
 440
 441                 rw_ctx = ceph_find_rw_context(fi);
 442                 if (rw_ctx) {
 443                         rreq->netfs_priv = priv;
 444                         return 0;
 445                 }
 446         }
 447
 448         /*
 449          * readahead callers do not necessarily hold Fcb caps
 450          * (e.g. fadvise, madvise).
 451          */
 452         ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 453         if (ret < 0) {
 454                 doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
 455                 goto out;
 456         }
 457
 458         if (!(got & want)) {
 459                 doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
 460                 ret = -EACCES;
 461                 goto out;
 462         }
 463         if (ret == 0) {
 464                 ret = -EACCES;
 465                 goto out;
 466         }
 467
 468         priv->caps = got;
 469         rreq->netfs_priv = priv;
 470
 471 out:
 472         if (ret < 0)
 473                 kfree(priv);
 474
 475         return ret;
 476 }
 477
 478 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 479 {
 480         struct ceph_netfs_request_data *priv = rreq->netfs_priv;
 481
 482         if (!priv)
 483                 return;
 484
 485         if (priv->caps)
 486                 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
 487         kfree(priv);
 488         rreq->netfs_priv = NULL;
 489 }
 490
 491 const struct netfs_request_ops ceph_netfs_ops = {
 492         .init_request           = ceph_init_request,
 493         .free_request           = ceph_netfs_free_request,
 494         .issue_read             = ceph_netfs_issue_read,
 495         .expand_readahead       = ceph_netfs_expand_readahead,
 496         .clamp_length           = ceph_netfs_clamp_length,
 497         .check_write_begin      = ceph_netfs_check_write_begin,
 498 };
 499
 500 #ifdef CONFIG_CEPH_FSCACHE
 501 static void ceph_set_page_fscache(struct page *page)
 502 {
 503         set_page_fscache(page);
 504 }
 505
 506 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 507 {
 508         struct inode *inode = priv;
 509
 510         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 511                 ceph_fscache_invalidate(inode, false);
 512 }
 513
 514 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 515 {
 516         struct ceph_inode_info *ci = ceph_inode(inode);
 517         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 518
 519         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 520                                ceph_fscache_write_terminated, inode, caching);
 521 }
 522 #else
 523 static inline void ceph_set_page_fscache(struct page *page)
 524 {
 525 }
 526
 527 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 528 {
 529 }
 530 #endif /* CONFIG_CEPH_FSCACHE */
 531
 532 struct ceph_writeback_ctl
 533 {
 534         loff_t i_size;
 535         u64 truncate_size;
 536         u32 truncate_seq;
 537         bool size_stable;
 538         bool head_snapc;
 539 };
 540
 541 /*
 542  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 543  * only snap context we are allowed to write back.
 544  */
 545 static struct ceph_snap_context *
 546 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 547                    struct ceph_snap_context *page_snapc)
 548 {
 549         struct ceph_inode_info *ci = ceph_inode(inode);
 550         struct ceph_client *cl = ceph_inode_to_client(inode);
 551         struct ceph_snap_context *snapc = NULL;
 552         struct ceph_cap_snap *capsnap = NULL;
 553
 554         spin_lock(&ci->i_ceph_lock);
 555         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 556                 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
 557                       capsnap, capsnap->context, capsnap->dirty_pages);
 558                 if (!capsnap->dirty_pages)
 559                         continue;
 560
 561                 /* get i_size, truncate_{seq,size} for page_snapc? */
 562                 if (snapc && capsnap->context != page_snapc)
 563                         continue;
 564
 565                 if (ctl) {
 566                         if (capsnap->writing) {
 567                                 ctl->i_size = i_size_read(inode);
 568                                 ctl->size_stable = false;
 569                         } else {
 570                                 ctl->i_size = capsnap->size;
 571                                 ctl->size_stable = true;
 572                         }
 573                         ctl->truncate_size = capsnap->truncate_size;
 574                         ctl->truncate_seq = capsnap->truncate_seq;
 575                         ctl->head_snapc = false;
 576                 }
 577
 578                 if (snapc)
 579                         break;
 580
 581                 snapc = ceph_get_snap_context(capsnap->context);
 582                 if (!page_snapc ||
 583                     page_snapc == snapc ||
 584                     page_snapc->seq > snapc->seq)
 585                         break;
 586         }
 587         if (!snapc && ci->i_wrbuffer_ref_head) {
 588                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 589                 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
 590                       ci->i_wrbuffer_ref_head);
 591                 if (ctl) {
 592                         ctl->i_size = i_size_read(inode);
 593                         ctl->truncate_size = ci->i_truncate_size;
 594                         ctl->truncate_seq = ci->i_truncate_seq;
 595                         ctl->size_stable = false;
 596                         ctl->head_snapc = true;
 597                 }
 598         }
 599         spin_unlock(&ci->i_ceph_lock);
 600         return snapc;
 601 }
 602
 603 static u64 get_writepages_data_length(struct inode *inode,
 604                                       struct page *page, u64 start)
 605 {
 606         struct ceph_inode_info *ci = ceph_inode(inode);
 607         struct ceph_snap_context *snapc;
 608         struct ceph_cap_snap *capsnap = NULL;
 609         u64 end = i_size_read(inode);
 610         u64 ret;
 611
 612         snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
 613         if (snapc != ci->i_head_snapc) {
 614                 bool found = false;
 615                 spin_lock(&ci->i_ceph_lock);
 616                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 617                         if (capsnap->context == snapc) {
 618                                 if (!capsnap->writing)
 619                                         end = capsnap->size;
 620                                 found = true;
 621                                 break;
 622                         }
 623                 }
 624                 spin_unlock(&ci->i_ceph_lock);
 625                 WARN_ON(!found);
 626         }
 627         if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
 628                 end = ceph_fscrypt_page_offset(page) + thp_size(page);
 629         ret = end > start ? end - start : 0;
 630         if (ret && fscrypt_is_bounce_page(page))
 631                 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
 632         return ret;
 633 }
 634
 635 /*
 636  * Write a single page, but leave the page locked.
 637  *
 638  * If we get a write error, mark the mapping for error, but still adjust the
 639  * dirty page accounting (i.e., page is no longer dirty).
 640  */
 641 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 642 {
 643         struct folio *folio = page_folio(page);
 644         struct inode *inode = page->mapping->host;
 645         struct ceph_inode_info *ci = ceph_inode(inode);
 646         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 647         struct ceph_client *cl = fsc->client;
 648         struct ceph_snap_context *snapc, *oldest;
 649         loff_t page_off = page_offset(page);
 650         int err;
 651         loff_t len = thp_size(page);
 652         loff_t wlen;
 653         struct ceph_writeback_ctl ceph_wbc;
 654         struct ceph_osd_client *osdc = &fsc->client->osdc;
 655         struct ceph_osd_request *req;
 656         bool caching = ceph_is_cache_enabled(inode);
 657         struct page *bounce_page = NULL;
 658
 659         doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
 660               page->index);
 661
 662         if (ceph_inode_is_shutdown(inode))
 663                 return -EIO;
 664
 665         /* verify this is a writeable snap context */
 666         snapc = page_snap_context(page);
 667         if (!snapc) {
 668                 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
 669                       page);
 670                 return 0;
 671         }
 672         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 673         if (snapc->seq > oldest->seq) {
 674                 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
 675                       ceph_vinop(inode), page, snapc);
 676                 /* we should only noop if called by kswapd */
 677                 WARN_ON(!(current->flags & PF_MEMALLOC));
 678                 ceph_put_snap_context(oldest);
 679                 redirty_page_for_writepage(wbc, page);
 680                 return 0;
 681         }
 682         ceph_put_snap_context(oldest);
 683
 684         /* is this a partial page at end of file? */
 685         if (page_off >= ceph_wbc.i_size) {
 686                 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
 687                       ceph_vinop(inode), folio->index, ceph_wbc.i_size);
 688                 folio_invalidate(folio, 0, folio_size(folio));
 689                 return 0;
 690         }
 691
 692         if (ceph_wbc.i_size < page_off + len)
 693                 len = ceph_wbc.i_size - page_off;
 694
 695         wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
 696         doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 697               ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
 698               snapc->seq);
 699
 700         if (atomic_long_inc_return(&fsc->writeback_count) >
 701             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 702                 fsc->write_congested = true;
 703
 704         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
 705                                     page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
 706                                     CEPH_OSD_FLAG_WRITE, snapc,
 707                                     ceph_wbc.truncate_seq,
 708                                     ceph_wbc.truncate_size, true);
 709         if (IS_ERR(req)) {
 710                 redirty_page_for_writepage(wbc, page);
 711                 return PTR_ERR(req);
 712         }
 713
 714         if (wlen < len)
 715                 len = wlen;
 716
 717         set_page_writeback(page);
 718         if (caching)
 719                 ceph_set_page_fscache(page);
 720         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 721
 722         if (IS_ENCRYPTED(inode)) {
 723                 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
 724                                                     CEPH_FSCRYPT_BLOCK_SIZE, 0,
 725                                                     GFP_NOFS);
 726                 if (IS_ERR(bounce_page)) {
 727                         redirty_page_for_writepage(wbc, page);
 728                         end_page_writeback(page);
 729                         ceph_osdc_put_request(req);
 730                         return PTR_ERR(bounce_page);
 731                 }
 732         }
 733
 734         /* it may be a short write due to an object boundary */
 735         WARN_ON_ONCE(len > thp_size(page));
 736         osd_req_op_extent_osd_data_pages(req, 0,
 737                         bounce_page ? &bounce_page : &page, wlen, 0,
 738                         false, false);
 739         doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
 740               ceph_vinop(inode), page_off, len, wlen,
 741               IS_ENCRYPTED(inode) ? "" : "not ");
 742
 743         req->r_mtime = inode_get_mtime(inode);
 744         ceph_osdc_start_request(osdc, req);
 745         err = ceph_osdc_wait_request(osdc, req);
 746
 747         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 748                                   req->r_end_latency, len, err);
 749         fscrypt_free_bounce_page(bounce_page);
 750         ceph_osdc_put_request(req);
 751         if (err == 0)
 752                 err = len;
 753
 754         if (err < 0) {
 755                 struct writeback_control tmp_wbc;
 756                 if (!wbc)
 757                         wbc = &tmp_wbc;
 758                 if (err == -ERESTARTSYS) {
 759                         /* killed by SIGKILL */
 760                         doutc(cl, "%llx.%llx interrupted page %p\n",
 761                               ceph_vinop(inode), page);
 762                         redirty_page_for_writepage(wbc, page);
 763                         end_page_writeback(page);
 764                         return err;
 765                 }
 766                 if (err == -EBLOCKLISTED)
 767                         fsc->blocklisted = true;
 768                 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
 769                       ceph_vinop(inode), err, page);
 770                 mapping_set_error(&inode->i_data, err);
 771                 wbc->pages_skipped++;
 772         } else {
 773                 doutc(cl, "%llx.%llx cleaned page %p\n",
 774                       ceph_vinop(inode), page);
 775                 err = 0;  /* vfs expects us to return 0 */
 776         }
 777         oldest = detach_page_private(page);
 778         WARN_ON_ONCE(oldest != snapc);
 779         end_page_writeback(page);
 780         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 781         ceph_put_snap_context(snapc);  /* page's reference */
 782
 783         if (atomic_long_dec_return(&fsc->writeback_count) <
 784             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 785                 fsc->write_congested = false;
 786
 787         return err;
 788 }
 789
 790 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 791 {
 792         int err;
 793         struct inode *inode = page->mapping->host;
 794         BUG_ON(!inode);
 795         ihold(inode);
 796
 797         if (wbc->sync_mode == WB_SYNC_NONE &&
 798             ceph_inode_to_fs_client(inode)->write_congested) {
 799                 redirty_page_for_writepage(wbc, page);
 800                 return AOP_WRITEPAGE_ACTIVATE;
 801         }
 802
 803         wait_on_page_fscache(page);
 804
 805         err = writepage_nounlock(page, wbc);
 806         if (err == -ERESTARTSYS) {
 807                 /* direct memory reclaimer was killed by SIGKILL. return 0
 808                  * to prevent caller from setting mapping/page error */
 809                 err = 0;
 810         }
 811         unlock_page(page);
 812         iput(inode);
 813         return err;
 814 }
 815
 816 /*
 817  * async writeback completion handler.
 818  *
 819  * If we get an error, set the mapping error bit, but not the individual
 820  * page error bits.
 821  */
 822 static void writepages_finish(struct ceph_osd_request *req)
 823 {
 824         struct inode *inode = req->r_inode;
 825         struct ceph_inode_info *ci = ceph_inode(inode);
 826         struct ceph_client *cl = ceph_inode_to_client(inode);
 827         struct ceph_osd_data *osd_data;
 828         struct page *page;
 829         int num_pages, total_pages = 0;
 830         int i, j;
 831         int rc = req->r_result;
 832         struct ceph_snap_context *snapc = req->r_snapc;
 833         struct address_space *mapping = inode->i_mapping;
 834         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 835         unsigned int len = 0;
 836         bool remove_page;
 837
 838         doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
 839         if (rc < 0) {
 840                 mapping_set_error(mapping, rc);
 841                 ceph_set_error_write(ci);
 842                 if (rc == -EBLOCKLISTED)
 843                         fsc->blocklisted = true;
 844         } else {
 845                 ceph_clear_error_write(ci);
 846         }
 847
 848         /*
 849          * We lost the cache cap, need to truncate the page before
 850          * it is unlocked, otherwise we'd truncate it later in the
 851          * page truncation thread, possibly losing some data that
 852          * raced its way in
 853          */
 854         remove_page = !(ceph_caps_issued(ci) &
 855                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 856
 857         /* clean all pages */
 858         for (i = 0; i < req->r_num_ops; i++) {
 859                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 860                         pr_warn_client(cl,
 861                                 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
 862                                 ceph_vinop(inode), req->r_ops[i].op, req, i,
 863                                 req->r_tid);
 864                         break;
 865                 }
 866
 867                 osd_data = osd_req_op_extent_osd_data(req, i);
 868                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 869                 len += osd_data->length;
 870                 num_pages = calc_pages_for((u64)osd_data->alignment,
 871                                            (u64)osd_data->length);
 872                 total_pages += num_pages;
 873                 for (j = 0; j < num_pages; j++) {
 874                         page = osd_data->pages[j];
 875                         if (fscrypt_is_bounce_page(page)) {
 876                                 page = fscrypt_pagecache_page(page);
 877                                 fscrypt_free_bounce_page(osd_data->pages[j]);
 878                                 osd_data->pages[j] = page;
 879                         }
 880                         BUG_ON(!page);
 881                         WARN_ON(!PageUptodate(page));
 882
 883                         if (atomic_long_dec_return(&fsc->writeback_count) <
 884                              CONGESTION_OFF_THRESH(
 885                                         fsc->mount_options->congestion_kb))
 886                                 fsc->write_congested = false;
 887
 888                         ceph_put_snap_context(detach_page_private(page));
 889                         end_page_writeback(page);
 890                         doutc(cl, "unlocking %p\n", page);
 891
 892                         if (remove_page)
 893                                 generic_error_remove_folio(inode->i_mapping,
 894                                                           page_folio(page));
 895
 896                         unlock_page(page);
 897                 }
 898                 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
 899                       ceph_vinop(inode), osd_data->length,
 900                       rc >= 0 ? num_pages : 0);
 901
 902                 release_pages(osd_data->pages, num_pages);
 903         }
 904
 905         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 906                                   req->r_end_latency, len, rc);
 907
 908         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 909
 910         osd_data = osd_req_op_extent_osd_data(req, 0);
 911         if (osd_data->pages_from_pool)
 912                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 913         else
 914                 kfree(osd_data->pages);
 915         ceph_osdc_put_request(req);
 916         ceph_dec_osd_stopping_blocker(fsc->mdsc);
 917 }
 918
 919 /*
 920  * initiate async writeback
 921  */
 922 static int ceph_writepages_start(struct address_space *mapping,
 923                                  struct writeback_control *wbc)
 924 {
 925         struct inode *inode = mapping->host;
 926         struct ceph_inode_info *ci = ceph_inode(inode);
 927         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 928         struct ceph_client *cl = fsc->client;
 929         struct ceph_vino vino = ceph_vino(inode);
 930         pgoff_t index, start_index, end = -1;
 931         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 932         struct folio_batch fbatch;
 933         int rc = 0;
 934         unsigned int wsize = i_blocksize(inode);
 935         struct ceph_osd_request *req = NULL;
 936         struct ceph_writeback_ctl ceph_wbc;
 937         bool should_loop, range_whole = false;
 938         bool done = false;
 939         bool caching = ceph_is_cache_enabled(inode);
 940         xa_mark_t tag;
 941
 942         if (wbc->sync_mode == WB_SYNC_NONE &&
 943             fsc->write_congested)
 944                 return 0;
 945
 946         doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
 947               wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 948               (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 949
 950         if (ceph_inode_is_shutdown(inode)) {
 951                 if (ci->i_wrbuffer_ref > 0) {
 952                         pr_warn_ratelimited_client(cl,
 953                                 "%llx.%llx %lld forced umount\n",
 954                                 ceph_vinop(inode), ceph_ino(inode));
 955                 }
 956                 mapping_set_error(mapping, -EIO);
 957                 return -EIO; /* we're in a forced umount, don't write! */
 958         }
 959         if (fsc->mount_options->wsize < wsize)
 960                 wsize = fsc->mount_options->wsize;
 961
 962         folio_batch_init(&fbatch);
 963
 964         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 965         index = start_index;
 966
 967         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
 968                 tag = PAGECACHE_TAG_TOWRITE;
 969         } else {
 970                 tag = PAGECACHE_TAG_DIRTY;
 971         }
 972 retry:
 973         /* find oldest snap context with dirty data */
 974         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 975         if (!snapc) {
 976                 /* hmm, why does writepages get called when there
 977                    is no dirty data? */
 978                 doutc(cl, " no snap context with dirty data?\n");
 979                 goto out;
 980         }
 981         doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
 982               snapc->seq, snapc->num_snaps);
 983
 984         should_loop = false;
 985         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 986                 /* where to start/end? */
 987                 if (wbc->range_cyclic) {
 988                         index = start_index;
 989                         end = -1;
 990                         if (index > 0)
 991                                 should_loop = true;
 992                         doutc(cl, " cyclic, start at %lu\n", index);
 993                 } else {
 994                         index = wbc->range_start >> PAGE_SHIFT;
 995                         end = wbc->range_end >> PAGE_SHIFT;
 996                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 997                                 range_whole = true;
 998                         doutc(cl, " not cyclic, %lu to %lu\n", index, end);
 999                 }
1000         } else if (!ceph_wbc.head_snapc) {
1001                 /* Do not respect wbc->range_{start,end}. Dirty pages
1002                  * in that range can be associated with newer snapc.
1003                  * They are not writeable until we write all dirty pages
1004                  * associated with 'snapc' get written */
1005                 if (index > 0)
1006                         should_loop = true;
1007                 doutc(cl, " non-head snapc, range whole\n");
1008         }
1009
1010         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1011                 tag_pages_for_writeback(mapping, index, end);
1012
1013         ceph_put_snap_context(last_snapc);
1014         last_snapc = snapc;
1015
1016         while (!done && index <= end) {
1017                 int num_ops = 0, op_idx;
1018                 unsigned i, nr_folios, max_pages, locked_pages = 0;
1019                 struct page **pages = NULL, **data_pages;
1020                 struct page *page;
1021                 pgoff_t strip_unit_end = 0;
1022                 u64 offset = 0, len = 0;
1023                 bool from_pool = false;
1024
1025                 max_pages = wsize >> PAGE_SHIFT;
1026
1027 get_more_pages:
1028                 nr_folios = filemap_get_folios_tag(mapping, &index,
1029                                                    end, tag, &fbatch);
1030                 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1031                 if (!nr_folios && !locked_pages)
1032                         break;
1033                 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1034                         page = &fbatch.folios[i]->page;
1035                         doutc(cl, "? %p idx %lu\n", page, page->index);
1036                         if (locked_pages == 0)
1037                                 lock_page(page);  /* first page */
1038                         else if (!trylock_page(page))
1039                                 break;
1040
1041                         /* only dirty pages, or our accounting breaks */
1042                         if (unlikely(!PageDirty(page)) ||
1043                             unlikely(page->mapping != mapping)) {
1044                                 doutc(cl, "!dirty or !mapping %p\n", page);
1045                                 unlock_page(page);
1046                                 continue;
1047                         }
1048                         /* only if matching snap context */
1049                         pgsnapc = page_snap_context(page);
1050                         if (pgsnapc != snapc) {
1051                                 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1052                                       pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1053                                 if (!should_loop &&
1054                                     !ceph_wbc.head_snapc &&
1055                                     wbc->sync_mode != WB_SYNC_NONE)
1056                                         should_loop = true;
1057                                 unlock_page(page);
1058                                 continue;
1059                         }
1060                         if (page_offset(page) >= ceph_wbc.i_size) {
1061                                 struct folio *folio = page_folio(page);
1062
1063                                 doutc(cl, "folio at %lu beyond eof %llu\n",
1064                                       folio->index, ceph_wbc.i_size);
1065                                 if ((ceph_wbc.size_stable ||
1066                                     folio_pos(folio) >= i_size_read(inode)) &&
1067                                     folio_clear_dirty_for_io(folio))
1068                                         folio_invalidate(folio, 0,
1069                                                         folio_size(folio));
1070                                 folio_unlock(folio);
1071                                 continue;
1072                         }
1073                         if (strip_unit_end && (page->index > strip_unit_end)) {
1074                                 doutc(cl, "end of strip unit %p\n", page);
1075                                 unlock_page(page);
1076                                 break;
1077                         }
1078                         if (PageWriteback(page) || PageFsCache(page)) {
1079                                 if (wbc->sync_mode == WB_SYNC_NONE) {
1080                                         doutc(cl, "%p under writeback\n", page);
1081                                         unlock_page(page);
1082                                         continue;
1083                                 }
1084                                 doutc(cl, "waiting on writeback %p\n", page);
1085                                 wait_on_page_writeback(page);
1086                                 wait_on_page_fscache(page);
1087                         }
1088
1089                         if (!clear_page_dirty_for_io(page)) {
1090                                 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1091                                 unlock_page(page);
1092                                 continue;
1093                         }
1094
1095                         /*
1096                          * We have something to write.  If this is
1097                          * the first locked page this time through,
1098                          * calculate max possinle write size and
1099                          * allocate a page array
1100                          */
1101                         if (locked_pages == 0) {
1102                                 u64 objnum;
1103                                 u64 objoff;
1104                                 u32 xlen;
1105
1106                                 /* prepare async write request */
1107                                 offset = (u64)page_offset(page);
1108                                 ceph_calc_file_object_mapping(&ci->i_layout,
1109                                                               offset, wsize,
1110                                                               &objnum, &objoff,
1111                                                               &xlen);
1112                                 len = xlen;
1113
1114                                 num_ops = 1;
1115                                 strip_unit_end = page->index +
1116                                         ((len - 1) >> PAGE_SHIFT);
1117
1118                                 BUG_ON(pages);
1119                                 max_pages = calc_pages_for(0, (u64)len);
1120                                 pages = kmalloc_array(max_pages,
1121                                                       sizeof(*pages),
1122                                                       GFP_NOFS);
1123                                 if (!pages) {
1124                                         from_pool = true;
1125                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1126                                         BUG_ON(!pages);
1127                                 }
1128
1129                                 len = 0;
1130                         } else if (page->index !=
1131                                    (offset + len) >> PAGE_SHIFT) {
1132                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1133                                                              CEPH_OSD_MAX_OPS)) {
1134                                         redirty_page_for_writepage(wbc, page);
1135                                         unlock_page(page);
1136                                         break;
1137                                 }
1138
1139                                 num_ops++;
1140                                 offset = (u64)page_offset(page);
1141                                 len = 0;
1142                         }
1143
1144                         /* note position of first page in fbatch */
1145                         doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1146                               ceph_vinop(inode), page, page->index);
1147
1148                         if (atomic_long_inc_return(&fsc->writeback_count) >
1149                             CONGESTION_ON_THRESH(
1150                                     fsc->mount_options->congestion_kb))
1151                                 fsc->write_congested = true;
1152
1153                         if (IS_ENCRYPTED(inode)) {
1154                                 pages[locked_pages] =
1155                                         fscrypt_encrypt_pagecache_blocks(page,
1156                                                 PAGE_SIZE, 0,
1157                                                 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1158                                 if (IS_ERR(pages[locked_pages])) {
1159                                         if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1160                                                 pr_err_client(cl,
1161                                                         "inode->i_blkbits=%hhu\n",
1162                                                         inode->i_blkbits);
1163                                         /* better not fail on first page! */
1164                                         BUG_ON(locked_pages == 0);
1165                                         pages[locked_pages] = NULL;
1166                                         redirty_page_for_writepage(wbc, page);
1167                                         unlock_page(page);
1168                                         break;
1169                                 }
1170                                 ++locked_pages;
1171                         } else {
1172                                 pages[locked_pages++] = page;
1173                         }
1174
1175                         fbatch.folios[i] = NULL;
1176                         len += thp_size(page);
1177                 }
1178
1179                 /* did we get anything? */
1180                 if (!locked_pages)
1181                         goto release_folios;
1182                 if (i) {
1183                         unsigned j, n = 0;
1184                         /* shift unused page to beginning of fbatch */
1185                         for (j = 0; j < nr_folios; j++) {
1186                                 if (!fbatch.folios[j])
1187                                         continue;
1188                                 if (n < j)
1189                                         fbatch.folios[n] = fbatch.folios[j];
1190                                 n++;
1191                         }
1192                         fbatch.nr = n;
1193
1194                         if (nr_folios && i == nr_folios &&
1195                             locked_pages < max_pages) {
1196                                 doutc(cl, "reached end fbatch, trying for more\n");
1197                                 folio_batch_release(&fbatch);
1198                                 goto get_more_pages;
1199                         }
1200                 }
1201
1202 new_request:
1203                 offset = ceph_fscrypt_page_offset(pages[0]);
1204                 len = wsize;
1205
1206                 req = ceph_osdc_new_request(&fsc->client->osdc,
1207                                         &ci->i_layout, vino,
1208                                         offset, &len, 0, num_ops,
1209                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1210                                         snapc, ceph_wbc.truncate_seq,
1211                                         ceph_wbc.truncate_size, false);
1212                 if (IS_ERR(req)) {
1213                         req = ceph_osdc_new_request(&fsc->client->osdc,
1214                                                 &ci->i_layout, vino,
1215                                                 offset, &len, 0,
1216                                                 min(num_ops,
1217                                                     CEPH_OSD_SLAB_OPS),
1218                                                 CEPH_OSD_OP_WRITE,
1219                                                 CEPH_OSD_FLAG_WRITE,
1220                                                 snapc, ceph_wbc.truncate_seq,
1221                                                 ceph_wbc.truncate_size, true);
1222                         BUG_ON(IS_ERR(req));
1223                 }
1224                 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1225                              thp_size(pages[locked_pages - 1]) - offset);
1226
1227                 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1228                         rc = -EIO;
1229                         goto release_folios;
1230                 }
1231                 req->r_callback = writepages_finish;
1232                 req->r_inode = inode;
1233
1234                 /* Format the osd request message and submit the write */
1235                 len = 0;
1236                 data_pages = pages;
1237                 op_idx = 0;
1238                 for (i = 0; i < locked_pages; i++) {
1239                         struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1240
1241                         u64 cur_offset = page_offset(page);
1242                         /*
1243                          * Discontinuity in page range? Ceph can handle that by just passing
1244                          * multiple extents in the write op.
1245                          */
1246                         if (offset + len != cur_offset) {
1247                                 /* If it's full, stop here */
1248                                 if (op_idx + 1 == req->r_num_ops)
1249                                         break;
1250
1251                                 /* Kick off an fscache write with what we have so far. */
1252                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1253
1254                                 /* Start a new extent */
1255                                 osd_req_op_extent_dup_last(req, op_idx,
1256                                                            cur_offset - offset);
1257                                 doutc(cl, "got pages at %llu~%llu\n", offset,
1258                                       len);
1259                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1260                                                         data_pages, len, 0,
1261                                                         from_pool, false);
1262                                 osd_req_op_extent_update(req, op_idx, len);
1263
1264                                 len = 0;
1265                                 offset = cur_offset;
1266                                 data_pages = pages + i;
1267                                 op_idx++;
1268                         }
1269
1270                         set_page_writeback(page);
1271                         if (caching)
1272                                 ceph_set_page_fscache(page);
1273                         len += thp_size(page);
1274                 }
1275                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1276
1277                 if (ceph_wbc.size_stable) {
1278                         len = min(len, ceph_wbc.i_size - offset);
1279                 } else if (i == locked_pages) {
1280                         /* writepages_finish() clears writeback pages
1281                          * according to the data length, so make sure
1282                          * data length covers all locked pages */
1283                         u64 min_len = len + 1 - thp_size(page);
1284                         len = get_writepages_data_length(inode, pages[i - 1],
1285                                                          offset);
1286                         len = max(len, min_len);
1287                 }
1288                 if (IS_ENCRYPTED(inode))
1289                         len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1290
1291                 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1292
1293                 if (IS_ENCRYPTED(inode) &&
1294                     ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1295                         pr_warn_client(cl,
1296                                 "bad encrypted write offset=%lld len=%llu\n",
1297                                 offset, len);
1298
1299                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1300                                                  0, from_pool, false);
1301                 osd_req_op_extent_update(req, op_idx, len);
1302
1303                 BUG_ON(op_idx + 1 != req->r_num_ops);
1304
1305                 from_pool = false;
1306                 if (i < locked_pages) {
1307                         BUG_ON(num_ops <= req->r_num_ops);
1308                         num_ops -= req->r_num_ops;
1309                         locked_pages -= i;
1310
1311                         /* allocate new pages array for next request */
1312                         data_pages = pages;
1313                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1314                                               GFP_NOFS);
1315                         if (!pages) {
1316                                 from_pool = true;
1317                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1318                                 BUG_ON(!pages);
1319                         }
1320                         memcpy(pages, data_pages + i,
1321                                locked_pages * sizeof(*pages));
1322                         memset(data_pages + i, 0,
1323                                locked_pages * sizeof(*pages));
1324                 } else {
1325                         BUG_ON(num_ops != req->r_num_ops);
1326                         index = pages[i - 1]->index + 1;
1327                         /* request message now owns the pages array */
1328                         pages = NULL;
1329                 }
1330
1331                 req->r_mtime = inode_get_mtime(inode);
1332                 ceph_osdc_start_request(&fsc->client->osdc, req);
1333                 req = NULL;
1334
1335                 wbc->nr_to_write -= i;
1336                 if (pages)
1337                         goto new_request;
1338
1339                 /*
1340                  * We stop writing back only if we are not doing
1341                  * integrity sync. In case of integrity sync we have to
1342                  * keep going until we have written all the pages
1343                  * we tagged for writeback prior to entering this loop.
1344                  */
1345                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1346                         done = true;
1347
1348 release_folios:
1349                 doutc(cl, "folio_batch release on %d folios (%p)\n",
1350                       (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1351                 folio_batch_release(&fbatch);
1352         }
1353
1354         if (should_loop && !done) {
1355                 /* more to do; loop back to beginning of file */
1356                 doutc(cl, "looping back to beginning of file\n");
1357                 end = start_index - 1; /* OK even when start_index == 0 */
1358
1359                 /* to write dirty pages associated with next snapc,
1360                  * we need to wait until current writes complete */
1361                 if (wbc->sync_mode != WB_SYNC_NONE &&
1362                     start_index == 0 && /* all dirty pages were checked */
1363                     !ceph_wbc.head_snapc) {
1364                         struct page *page;
1365                         unsigned i, nr;
1366                         index = 0;
1367                         while ((index <= end) &&
1368                                (nr = filemap_get_folios_tag(mapping, &index,
1369                                                 (pgoff_t)-1,
1370                                                 PAGECACHE_TAG_WRITEBACK,
1371                                                 &fbatch))) {
1372                                 for (i = 0; i < nr; i++) {
1373                                         page = &fbatch.folios[i]->page;
1374                                         if (page_snap_context(page) != snapc)
1375                                                 continue;
1376                                         wait_on_page_writeback(page);
1377                                 }
1378                                 folio_batch_release(&fbatch);
1379                                 cond_resched();
1380                         }
1381                 }
1382
1383                 start_index = 0;
1384                 index = 0;
1385                 goto retry;
1386         }
1387
1388         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1389                 mapping->writeback_index = index;
1390
1391 out:
1392         ceph_osdc_put_request(req);
1393         ceph_put_snap_context(last_snapc);
1394         doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1395               rc);
1396         return rc;
1397 }
1398
1399
1400
1401 /*
1402  * See if a given @snapc is either writeable, or already written.
1403  */
1404 static int context_is_writeable_or_written(struct inode *inode,
1405                                            struct ceph_snap_context *snapc)
1406 {
1407         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1408         int ret = !oldest || snapc->seq <= oldest->seq;
1409
1410         ceph_put_snap_context(oldest);
1411         return ret;
1412 }
1413
1414 /**
1415  * ceph_find_incompatible - find an incompatible context and return it
1416  * @page: page being dirtied
1417  *
1418  * We are only allowed to write into/dirty a page if the page is
1419  * clean, or already dirty within the same snap context. Returns a
1420  * conflicting context if there is one, NULL if there isn't, or a
1421  * negative error code on other errors.
1422  *
1423  * Must be called with page lock held.
1424  */
1425 static struct ceph_snap_context *
1426 ceph_find_incompatible(struct page *page)
1427 {
1428         struct inode *inode = page->mapping->host;
1429         struct ceph_client *cl = ceph_inode_to_client(inode);
1430         struct ceph_inode_info *ci = ceph_inode(inode);
1431
1432         if (ceph_inode_is_shutdown(inode)) {
1433                 doutc(cl, " %llx.%llx page %p is shutdown\n",
1434                       ceph_vinop(inode), page);
1435                 return ERR_PTR(-ESTALE);
1436         }
1437
1438         for (;;) {
1439                 struct ceph_snap_context *snapc, *oldest;
1440
1441                 wait_on_page_writeback(page);
1442
1443                 snapc = page_snap_context(page);
1444                 if (!snapc || snapc == ci->i_head_snapc)
1445                         break;
1446
1447                 /*
1448                  * this page is already dirty in another (older) snap
1449                  * context!  is it writeable now?
1450                  */
1451                 oldest = get_oldest_context(inode, NULL, NULL);
1452                 if (snapc->seq > oldest->seq) {
1453                         /* not writeable -- return it for the caller to deal with */
1454                         ceph_put_snap_context(oldest);
1455                         doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1456                               ceph_vinop(inode), page, snapc);
1457                         return ceph_get_snap_context(snapc);
1458                 }
1459                 ceph_put_snap_context(oldest);
1460
1461                 /* yay, writeable, do it now (without dropping page lock) */
1462                 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1463                       ceph_vinop(inode), page, snapc);
1464                 if (clear_page_dirty_for_io(page)) {
1465                         int r = writepage_nounlock(page, NULL);
1466                         if (r < 0)
1467                                 return ERR_PTR(r);
1468                 }
1469         }
1470         return NULL;
1471 }
1472
1473 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1474                                         struct folio **foliop, void **_fsdata)
1475 {
1476         struct inode *inode = file_inode(file);
1477         struct ceph_inode_info *ci = ceph_inode(inode);
1478         struct ceph_snap_context *snapc;
1479
1480         snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1481         if (snapc) {
1482                 int r;
1483
1484                 folio_unlock(*foliop);
1485                 folio_put(*foliop);
1486                 *foliop = NULL;
1487                 if (IS_ERR(snapc))
1488                         return PTR_ERR(snapc);
1489
1490                 ceph_queue_writeback(inode);
1491                 r = wait_event_killable(ci->i_cap_wq,
1492                                         context_is_writeable_or_written(inode, snapc));
1493                 ceph_put_snap_context(snapc);
1494                 return r == 0 ? -EAGAIN : r;
1495         }
1496         return 0;
1497 }
1498
1499 /*
1500  * We are only allowed to write into/dirty the page if the page is
1501  * clean, or already dirty within the same snap context.
1502  */
1503 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1504                             loff_t pos, unsigned len,
1505                             struct page **pagep, void **fsdata)
1506 {
1507         struct inode *inode = file_inode(file);
1508         struct ceph_inode_info *ci = ceph_inode(inode);
1509         struct folio *folio = NULL;
1510         int r;
1511
1512         r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1513         if (r < 0)
1514                 return r;
1515
1516         folio_wait_fscache(folio);
1517         WARN_ON_ONCE(!folio_test_locked(folio));
1518         *pagep = &folio->page;
1519         return 0;
1520 }
1521
1522 /*
1523  * we don't do anything in here that simple_write_end doesn't do
1524  * except adjust dirty page accounting
1525  */
1526 static int ceph_write_end(struct file *file, struct address_space *mapping,
1527                           loff_t pos, unsigned len, unsigned copied,
1528                           struct page *subpage, void *fsdata)
1529 {
1530         struct folio *folio = page_folio(subpage);
1531         struct inode *inode = file_inode(file);
1532         struct ceph_client *cl = ceph_inode_to_client(inode);
1533         bool check_cap = false;
1534
1535         doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1536               file, folio, (int)pos, (int)copied, (int)len);
1537
1538         if (!folio_test_uptodate(folio)) {
1539                 /* just return that nothing was copied on a short copy */
1540                 if (copied < len) {
1541                         copied = 0;
1542                         goto out;
1543                 }
1544                 folio_mark_uptodate(folio);
1545         }
1546
1547         /* did file size increase? */
1548         if (pos+copied > i_size_read(inode))
1549                 check_cap = ceph_inode_set_size(inode, pos+copied);
1550
1551         folio_mark_dirty(folio);
1552
1553 out:
1554         folio_unlock(folio);
1555         folio_put(folio);
1556
1557         if (check_cap)
1558                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1559
1560         return copied;
1561 }
1562
1563 const struct address_space_operations ceph_aops = {
1564         .read_folio = netfs_read_folio,
1565         .readahead = netfs_readahead,
1566         .writepage = ceph_writepage,
1567         .writepages = ceph_writepages_start,
1568         .write_begin = ceph_write_begin,
1569         .write_end = ceph_write_end,
1570         .dirty_folio = ceph_dirty_folio,
1571         .invalidate_folio = ceph_invalidate_folio,
1572         .release_folio = netfs_release_folio,
1573         .direct_IO = noop_direct_IO,
1574 };
1575
1576 static void ceph_block_sigs(sigset_t *oldset)
1577 {
1578         sigset_t mask;
1579         siginitsetinv(&mask, sigmask(SIGKILL));
1580         sigprocmask(SIG_BLOCK, &mask, oldset);
1581 }
1582
1583 static void ceph_restore_sigs(sigset_t *oldset)
1584 {
1585         sigprocmask(SIG_SETMASK, oldset, NULL);
1586 }
1587
1588 /*
1589  * vm ops
1590  */
1591 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1592 {
1593         struct vm_area_struct *vma = vmf->vma;
1594         struct inode *inode = file_inode(vma->vm_file);
1595         struct ceph_inode_info *ci = ceph_inode(inode);
1596         struct ceph_client *cl = ceph_inode_to_client(inode);
1597         struct ceph_file_info *fi = vma->vm_file->private_data;
1598         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1599         int want, got, err;
1600         sigset_t oldset;
1601         vm_fault_t ret = VM_FAULT_SIGBUS;
1602
1603         if (ceph_inode_is_shutdown(inode))
1604                 return ret;
1605
1606         ceph_block_sigs(&oldset);
1607
1608         doutc(cl, "%llx.%llx %llu trying to get caps\n",
1609               ceph_vinop(inode), off);
1610         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1611                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1612         else
1613                 want = CEPH_CAP_FILE_CACHE;
1614
1615         got = 0;
1616         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1617         if (err < 0)
1618                 goto out_restore;
1619
1620         doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1621               off, ceph_cap_string(got));
1622
1623         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1624             !ceph_has_inline_data(ci)) {
1625                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1626                 ceph_add_rw_context(fi, &rw_ctx);
1627                 ret = filemap_fault(vmf);
1628                 ceph_del_rw_context(fi, &rw_ctx);
1629                 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1630                       ceph_vinop(inode), off, ceph_cap_string(got), ret);
1631         } else
1632                 err = -EAGAIN;
1633
1634         ceph_put_cap_refs(ci, got);
1635
1636         if (err != -EAGAIN)
1637                 goto out_restore;
1638
1639         /* read inline data */
1640         if (off >= PAGE_SIZE) {
1641                 /* does not support inline data > PAGE_SIZE */
1642                 ret = VM_FAULT_SIGBUS;
1643         } else {
1644                 struct address_space *mapping = inode->i_mapping;
1645                 struct page *page;
1646
1647                 filemap_invalidate_lock_shared(mapping);
1648                 page = find_or_create_page(mapping, 0,
1649                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1650                 if (!page) {
1651                         ret = VM_FAULT_OOM;
1652                         goto out_inline;
1653                 }
1654                 err = __ceph_do_getattr(inode, page,
1655                                          CEPH_STAT_CAP_INLINE_DATA, true);
1656                 if (err < 0 || off >= i_size_read(inode)) {
1657                         unlock_page(page);
1658                         put_page(page);
1659                         ret = vmf_error(err);
1660                         goto out_inline;
1661                 }
1662                 if (err < PAGE_SIZE)
1663                         zero_user_segment(page, err, PAGE_SIZE);
1664                 else
1665                         flush_dcache_page(page);
1666                 SetPageUptodate(page);
1667                 vmf->page = page;
1668                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1669 out_inline:
1670                 filemap_invalidate_unlock_shared(mapping);
1671                 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1672                       ceph_vinop(inode), off, ret);
1673         }
1674 out_restore:
1675         ceph_restore_sigs(&oldset);
1676         if (err < 0)
1677                 ret = vmf_error(err);
1678
1679         return ret;
1680 }
1681
1682 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1683 {
1684         struct vm_area_struct *vma = vmf->vma;
1685         struct inode *inode = file_inode(vma->vm_file);
1686         struct ceph_client *cl = ceph_inode_to_client(inode);
1687         struct ceph_inode_info *ci = ceph_inode(inode);
1688         struct ceph_file_info *fi = vma->vm_file->private_data;
1689         struct ceph_cap_flush *prealloc_cf;
1690         struct page *page = vmf->page;
1691         loff_t off = page_offset(page);
1692         loff_t size = i_size_read(inode);
1693         size_t len;
1694         int want, got, err;
1695         sigset_t oldset;
1696         vm_fault_t ret = VM_FAULT_SIGBUS;
1697
1698         if (ceph_inode_is_shutdown(inode))
1699                 return ret;
1700
1701         prealloc_cf = ceph_alloc_cap_flush();
1702         if (!prealloc_cf)
1703                 return VM_FAULT_OOM;
1704
1705         sb_start_pagefault(inode->i_sb);
1706         ceph_block_sigs(&oldset);
1707
1708         if (off + thp_size(page) <= size)
1709                 len = thp_size(page);
1710         else
1711                 len = offset_in_thp(page, size);
1712
1713         doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1714               ceph_vinop(inode), off, len, size);
1715         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1716                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1717         else
1718                 want = CEPH_CAP_FILE_BUFFER;
1719
1720         got = 0;
1721         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1722         if (err < 0)
1723                 goto out_free;
1724
1725         doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1726               off, len, ceph_cap_string(got));
1727
1728         /* Update time before taking page lock */
1729         file_update_time(vma->vm_file);
1730         inode_inc_iversion_raw(inode);
1731
1732         do {
1733                 struct ceph_snap_context *snapc;
1734
1735                 lock_page(page);
1736
1737                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1738                         unlock_page(page);
1739                         ret = VM_FAULT_NOPAGE;
1740                         break;
1741                 }
1742
1743                 snapc = ceph_find_incompatible(page);
1744                 if (!snapc) {
1745                         /* success.  we'll keep the page locked. */
1746                         set_page_dirty(page);
1747                         ret = VM_FAULT_LOCKED;
1748                         break;
1749                 }
1750
1751                 unlock_page(page);
1752
1753                 if (IS_ERR(snapc)) {
1754                         ret = VM_FAULT_SIGBUS;
1755                         break;
1756                 }
1757
1758                 ceph_queue_writeback(inode);
1759                 err = wait_event_killable(ci->i_cap_wq,
1760                                 context_is_writeable_or_written(inode, snapc));
1761                 ceph_put_snap_context(snapc);
1762         } while (err == 0);
1763
1764         if (ret == VM_FAULT_LOCKED) {
1765                 int dirty;
1766                 spin_lock(&ci->i_ceph_lock);
1767                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1768                                                &prealloc_cf);
1769                 spin_unlock(&ci->i_ceph_lock);
1770                 if (dirty)
1771                         __mark_inode_dirty(inode, dirty);
1772         }
1773
1774         doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1775               ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1776         ceph_put_cap_refs_async(ci, got);
1777 out_free:
1778         ceph_restore_sigs(&oldset);
1779         sb_end_pagefault(inode->i_sb);
1780         ceph_free_cap_flush(prealloc_cf);
1781         if (err < 0)
1782                 ret = vmf_error(err);
1783         return ret;
1784 }
1785
1786 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1787                            char *data, size_t len)
1788 {
1789         struct ceph_client *cl = ceph_inode_to_client(inode);
1790         struct address_space *mapping = inode->i_mapping;
1791         struct page *page;
1792
1793         if (locked_page) {
1794                 page = locked_page;
1795         } else {
1796                 if (i_size_read(inode) == 0)
1797                         return;
1798                 page = find_or_create_page(mapping, 0,
1799                                            mapping_gfp_constraint(mapping,
1800                                            ~__GFP_FS));
1801                 if (!page)
1802                         return;
1803                 if (PageUptodate(page)) {
1804                         unlock_page(page);
1805                         put_page(page);
1806                         return;
1807                 }
1808         }
1809
1810         doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1811               ceph_vinop(inode), len, locked_page);
1812
1813         if (len > 0) {
1814                 void *kaddr = kmap_atomic(page);
1815                 memcpy(kaddr, data, len);
1816                 kunmap_atomic(kaddr);
1817         }
1818
1819         if (page != locked_page) {
1820                 if (len < PAGE_SIZE)
1821                         zero_user_segment(page, len, PAGE_SIZE);
1822                 else
1823                         flush_dcache_page(page);
1824
1825                 SetPageUptodate(page);
1826                 unlock_page(page);
1827                 put_page(page);
1828         }
1829 }
1830
1831 int ceph_uninline_data(struct file *file)
1832 {
1833         struct inode *inode = file_inode(file);
1834         struct ceph_inode_info *ci = ceph_inode(inode);
1835         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1836         struct ceph_client *cl = fsc->client;
1837         struct ceph_osd_request *req = NULL;
1838         struct ceph_cap_flush *prealloc_cf = NULL;
1839         struct folio *folio = NULL;
1840         u64 inline_version = CEPH_INLINE_NONE;
1841         struct page *pages[1];
1842         int err = 0;
1843         u64 len;
1844
1845         spin_lock(&ci->i_ceph_lock);
1846         inline_version = ci->i_inline_version;
1847         spin_unlock(&ci->i_ceph_lock);
1848
1849         doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1850               inline_version);
1851
1852         if (ceph_inode_is_shutdown(inode)) {
1853                 err = -EIO;
1854                 goto out;
1855         }
1856
1857         if (inline_version == CEPH_INLINE_NONE)
1858                 return 0;
1859
1860         prealloc_cf = ceph_alloc_cap_flush();
1861         if (!prealloc_cf)
1862                 return -ENOMEM;
1863
1864         if (inline_version == 1) /* initial version, no data */
1865                 goto out_uninline;
1866
1867         folio = read_mapping_folio(inode->i_mapping, 0, file);
1868         if (IS_ERR(folio)) {
1869                 err = PTR_ERR(folio);
1870                 goto out;
1871         }
1872
1873         folio_lock(folio);
1874
1875         len = i_size_read(inode);
1876         if (len > folio_size(folio))
1877                 len = folio_size(folio);
1878
1879         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1880                                     ceph_vino(inode), 0, &len, 0, 1,
1881                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1882                                     NULL, 0, 0, false);
1883         if (IS_ERR(req)) {
1884                 err = PTR_ERR(req);
1885                 goto out_unlock;
1886         }
1887
1888         req->r_mtime = inode_get_mtime(inode);
1889         ceph_osdc_start_request(&fsc->client->osdc, req);
1890         err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1891         ceph_osdc_put_request(req);
1892         if (err < 0)
1893                 goto out_unlock;
1894
1895         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1896                                     ceph_vino(inode), 0, &len, 1, 3,
1897                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1898                                     NULL, ci->i_truncate_seq,
1899                                     ci->i_truncate_size, false);
1900         if (IS_ERR(req)) {
1901                 err = PTR_ERR(req);
1902                 goto out_unlock;
1903         }
1904
1905         pages[0] = folio_page(folio, 0);
1906         osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1907
1908         {
1909                 __le64 xattr_buf = cpu_to_le64(inline_version);
1910                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1911                                             "inline_version", &xattr_buf,
1912                                             sizeof(xattr_buf),
1913                                             CEPH_OSD_CMPXATTR_OP_GT,
1914                                             CEPH_OSD_CMPXATTR_MODE_U64);
1915                 if (err)
1916                         goto out_put_req;
1917         }
1918
1919         {
1920                 char xattr_buf[32];
1921                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1922                                          "%llu", inline_version);
1923                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1924                                             "inline_version",
1925                                             xattr_buf, xattr_len, 0, 0);
1926                 if (err)
1927                         goto out_put_req;
1928         }
1929
1930         req->r_mtime = inode_get_mtime(inode);
1931         ceph_osdc_start_request(&fsc->client->osdc, req);
1932         err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1933
1934         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1935                                   req->r_end_latency, len, err);
1936
1937 out_uninline:
1938         if (!err) {
1939                 int dirty;
1940
1941                 /* Set to CAP_INLINE_NONE and dirty the caps */
1942                 down_read(&fsc->mdsc->snap_rwsem);
1943                 spin_lock(&ci->i_ceph_lock);
1944                 ci->i_inline_version = CEPH_INLINE_NONE;
1945                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1946                 spin_unlock(&ci->i_ceph_lock);
1947                 up_read(&fsc->mdsc->snap_rwsem);
1948                 if (dirty)
1949                         __mark_inode_dirty(inode, dirty);
1950         }
1951 out_put_req:
1952         ceph_osdc_put_request(req);
1953         if (err == -ECANCELED)
1954                 err = 0;
1955 out_unlock:
1956         if (folio) {
1957                 folio_unlock(folio);
1958                 folio_put(folio);
1959         }
1960 out:
1961         ceph_free_cap_flush(prealloc_cf);
1962         doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1963               ceph_vinop(inode), inline_version, err);
1964         return err;
1965 }
1966
1967 static const struct vm_operations_struct ceph_vmops = {
1968         .fault          = ceph_filemap_fault,
1969         .page_mkwrite   = ceph_page_mkwrite,
1970 };
1971
1972 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1973 {
1974         struct address_space *mapping = file->f_mapping;
1975
1976         if (!mapping->a_ops->read_folio)
1977                 return -ENOEXEC;
1978         vma->vm_ops = &ceph_vmops;
1979         return 0;
1980 }
1981
1982 enum {
1983         POOL_READ       = 1,
1984         POOL_WRITE      = 2,
1985 };
1986
1987 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1988                                 s64 pool, struct ceph_string *pool_ns)
1989 {
1990         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
1991         struct ceph_mds_client *mdsc = fsc->mdsc;
1992         struct ceph_client *cl = fsc->client;
1993         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1994         struct rb_node **p, *parent;
1995         struct ceph_pool_perm *perm;
1996         struct page **pages;
1997         size_t pool_ns_len;
1998         int err = 0, err2 = 0, have = 0;
1999
2000         down_read(&mdsc->pool_perm_rwsem);
2001         p = &mdsc->pool_perm_tree.rb_node;
2002         while (*p) {
2003                 perm = rb_entry(*p, struct ceph_pool_perm, node);
2004                 if (pool < perm->pool)
2005                         p = &(*p)->rb_left;
2006                 else if (pool > perm->pool)
2007                         p = &(*p)->rb_right;
2008                 else {
2009                         int ret = ceph_compare_string(pool_ns,
2010                                                 perm->pool_ns,
2011                                                 perm->pool_ns_len);
2012                         if (ret < 0)
2013                                 p = &(*p)->rb_left;
2014                         else if (ret > 0)
2015                                 p = &(*p)->rb_right;
2016                         else {
2017                                 have = perm->perm;
2018                                 break;
2019                         }
2020                 }
2021         }
2022         up_read(&mdsc->pool_perm_rwsem);
2023         if (*p)
2024                 goto out;
2025
2026         if (pool_ns)
2027                 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2028                       (int)pool_ns->len, pool_ns->str);
2029         else
2030                 doutc(cl, "pool %lld no perm cached\n", pool);
2031
2032         down_write(&mdsc->pool_perm_rwsem);
2033         p = &mdsc->pool_perm_tree.rb_node;
2034         parent = NULL;
2035         while (*p) {
2036                 parent = *p;
2037                 perm = rb_entry(parent, struct ceph_pool_perm, node);
2038                 if (pool < perm->pool)
2039                         p = &(*p)->rb_left;
2040                 else if (pool > perm->pool)
2041                         p = &(*p)->rb_right;
2042                 else {
2043                         int ret = ceph_compare_string(pool_ns,
2044                                                 perm->pool_ns,
2045                                                 perm->pool_ns_len);
2046                         if (ret < 0)
2047                                 p = &(*p)->rb_left;
2048                         else if (ret > 0)
2049                                 p = &(*p)->rb_right;
2050                         else {
2051                                 have = perm->perm;
2052                                 break;
2053                         }
2054                 }
2055         }
2056         if (*p) {
2057                 up_write(&mdsc->pool_perm_rwsem);
2058                 goto out;
2059         }
2060
2061         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2062                                          1, false, GFP_NOFS);
2063         if (!rd_req) {
2064                 err = -ENOMEM;
2065                 goto out_unlock;
2066         }
2067
2068         rd_req->r_flags = CEPH_OSD_FLAG_READ;
2069         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2070         rd_req->r_base_oloc.pool = pool;
2071         if (pool_ns)
2072                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2073         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2074
2075         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2076         if (err)
2077                 goto out_unlock;
2078
2079         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2080                                          1, false, GFP_NOFS);
2081         if (!wr_req) {
2082                 err = -ENOMEM;
2083                 goto out_unlock;
2084         }
2085
2086         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2087         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2088         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2089         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2090
2091         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2092         if (err)
2093                 goto out_unlock;
2094
2095         /* one page should be large enough for STAT data */
2096         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2097         if (IS_ERR(pages)) {
2098                 err = PTR_ERR(pages);
2099                 goto out_unlock;
2100         }
2101
2102         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2103                                      0, false, true);
2104         ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2105
2106         wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2107         ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2108
2109         err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2110         err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2111
2112         if (err >= 0 || err == -ENOENT)
2113                 have |= POOL_READ;
2114         else if (err != -EPERM) {
2115                 if (err == -EBLOCKLISTED)
2116                         fsc->blocklisted = true;
2117                 goto out_unlock;
2118         }
2119
2120         if (err2 == 0 || err2 == -EEXIST)
2121                 have |= POOL_WRITE;
2122         else if (err2 != -EPERM) {
2123                 if (err2 == -EBLOCKLISTED)
2124                         fsc->blocklisted = true;
2125                 err = err2;
2126                 goto out_unlock;
2127         }
2128
2129         pool_ns_len = pool_ns ? pool_ns->len : 0;
2130         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2131         if (!perm) {
2132                 err = -ENOMEM;
2133                 goto out_unlock;
2134         }
2135
2136         perm->pool = pool;
2137         perm->perm = have;
2138         perm->pool_ns_len = pool_ns_len;
2139         if (pool_ns_len > 0)
2140                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2141         perm->pool_ns[pool_ns_len] = 0;
2142
2143         rb_link_node(&perm->node, parent, p);
2144         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2145         err = 0;
2146 out_unlock:
2147         up_write(&mdsc->pool_perm_rwsem);
2148
2149         ceph_osdc_put_request(rd_req);
2150         ceph_osdc_put_request(wr_req);
2151 out:
2152         if (!err)
2153                 err = have;
2154         if (pool_ns)
2155                 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2156                       (int)pool_ns->len, pool_ns->str, err);
2157         else
2158                 doutc(cl, "pool %lld result = %d\n", pool, err);
2159         return err;
2160 }
2161
2162 int ceph_pool_perm_check(struct inode *inode, int need)
2163 {
2164         struct ceph_client *cl = ceph_inode_to_client(inode);
2165         struct ceph_inode_info *ci = ceph_inode(inode);
2166         struct ceph_string *pool_ns;
2167         s64 pool;
2168         int ret, flags;
2169
2170         /* Only need to do this for regular files */
2171         if (!S_ISREG(inode->i_mode))
2172                 return 0;
2173
2174         if (ci->i_vino.snap != CEPH_NOSNAP) {
2175                 /*
2176                  * Pool permission check needs to write to the first object.
2177                  * But for snapshot, head of the first object may have alread
2178                  * been deleted. Skip check to avoid creating orphan object.
2179                  */
2180                 return 0;
2181         }
2182
2183         if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2184                                 NOPOOLPERM))
2185                 return 0;
2186
2187         spin_lock(&ci->i_ceph_lock);
2188         flags = ci->i_ceph_flags;
2189         pool = ci->i_layout.pool_id;
2190         spin_unlock(&ci->i_ceph_lock);
2191 check:
2192         if (flags & CEPH_I_POOL_PERM) {
2193                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2194                         doutc(cl, "pool %lld no read perm\n", pool);
2195                         return -EPERM;
2196                 }
2197                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2198                         doutc(cl, "pool %lld no write perm\n", pool);
2199                         return -EPERM;
2200                 }
2201                 return 0;
2202         }
2203
2204         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2205         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2206         ceph_put_string(pool_ns);
2207         if (ret < 0)
2208                 return ret;
2209
2210         flags = CEPH_I_POOL_PERM;
2211         if (ret & POOL_READ)
2212                 flags |= CEPH_I_POOL_RD;
2213         if (ret & POOL_WRITE)
2214                 flags |= CEPH_I_POOL_WR;
2215
2216         spin_lock(&ci->i_ceph_lock);
2217         if (pool == ci->i_layout.pool_id &&
2218             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2219                 ci->i_ceph_flags |= flags;
2220         } else {
2221                 pool = ci->i_layout.pool_id;
2222                 flags = ci->i_ceph_flags;
2223         }
2224         spin_unlock(&ci->i_ceph_lock);
2225         goto check;
2226 }
2227
2228 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2229 {
2230         struct ceph_pool_perm *perm;
2231         struct rb_node *n;
2232
2233         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2234                 n = rb_first(&mdsc->pool_perm_tree);
2235                 perm = rb_entry(n, struct ceph_pool_perm, node);
2236                 rb_erase(n, &mdsc->pool_perm_tree);
2237                 kfree(perm);
2238         }
2239 }