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 = umin(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_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
 502 {
 503         struct inode *inode = priv;
 504
 505         if (IS_ERR_VALUE(error) && error != -ENOBUFS)
 506                 ceph_fscache_invalidate(inode, false);
 507 }
 508
 509 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 510 {
 511         struct ceph_inode_info *ci = ceph_inode(inode);
 512         struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
 513
 514         fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
 515                                ceph_fscache_write_terminated, inode, true, caching);
 516 }
 517 #else
 518 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
 519 {
 520 }
 521 #endif /* CONFIG_CEPH_FSCACHE */
 522
 523 struct ceph_writeback_ctl
 524 {
 525         loff_t i_size;
 526         u64 truncate_size;
 527         u32 truncate_seq;
 528         bool size_stable;
 529         bool head_snapc;
 530 };
 531
 532 /*
 533  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 534  * only snap context we are allowed to write back.
 535  */
 536 static struct ceph_snap_context *
 537 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
 538                    struct ceph_snap_context *page_snapc)
 539 {
 540         struct ceph_inode_info *ci = ceph_inode(inode);
 541         struct ceph_client *cl = ceph_inode_to_client(inode);
 542         struct ceph_snap_context *snapc = NULL;
 543         struct ceph_cap_snap *capsnap = NULL;
 544
 545         spin_lock(&ci->i_ceph_lock);
 546         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 547                 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
 548                       capsnap, capsnap->context, capsnap->dirty_pages);
 549                 if (!capsnap->dirty_pages)
 550                         continue;
 551
 552                 /* get i_size, truncate_{seq,size} for page_snapc? */
 553                 if (snapc && capsnap->context != page_snapc)
 554                         continue;
 555
 556                 if (ctl) {
 557                         if (capsnap->writing) {
 558                                 ctl->i_size = i_size_read(inode);
 559                                 ctl->size_stable = false;
 560                         } else {
 561                                 ctl->i_size = capsnap->size;
 562                                 ctl->size_stable = true;
 563                         }
 564                         ctl->truncate_size = capsnap->truncate_size;
 565                         ctl->truncate_seq = capsnap->truncate_seq;
 566                         ctl->head_snapc = false;
 567                 }
 568
 569                 if (snapc)
 570                         break;
 571
 572                 snapc = ceph_get_snap_context(capsnap->context);
 573                 if (!page_snapc ||
 574                     page_snapc == snapc ||
 575                     page_snapc->seq > snapc->seq)
 576                         break;
 577         }
 578         if (!snapc && ci->i_wrbuffer_ref_head) {
 579                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 580                 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
 581                       ci->i_wrbuffer_ref_head);
 582                 if (ctl) {
 583                         ctl->i_size = i_size_read(inode);
 584                         ctl->truncate_size = ci->i_truncate_size;
 585                         ctl->truncate_seq = ci->i_truncate_seq;
 586                         ctl->size_stable = false;
 587                         ctl->head_snapc = true;
 588                 }
 589         }
 590         spin_unlock(&ci->i_ceph_lock);
 591         return snapc;
 592 }
 593
 594 static u64 get_writepages_data_length(struct inode *inode,
 595                                       struct page *page, u64 start)
 596 {
 597         struct ceph_inode_info *ci = ceph_inode(inode);
 598         struct ceph_snap_context *snapc;
 599         struct ceph_cap_snap *capsnap = NULL;
 600         u64 end = i_size_read(inode);
 601         u64 ret;
 602
 603         snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
 604         if (snapc != ci->i_head_snapc) {
 605                 bool found = false;
 606                 spin_lock(&ci->i_ceph_lock);
 607                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 608                         if (capsnap->context == snapc) {
 609                                 if (!capsnap->writing)
 610                                         end = capsnap->size;
 611                                 found = true;
 612                                 break;
 613                         }
 614                 }
 615                 spin_unlock(&ci->i_ceph_lock);
 616                 WARN_ON(!found);
 617         }
 618         if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
 619                 end = ceph_fscrypt_page_offset(page) + thp_size(page);
 620         ret = end > start ? end - start : 0;
 621         if (ret && fscrypt_is_bounce_page(page))
 622                 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
 623         return ret;
 624 }
 625
 626 /*
 627  * Write a single page, but leave the page locked.
 628  *
 629  * If we get a write error, mark the mapping for error, but still adjust the
 630  * dirty page accounting (i.e., page is no longer dirty).
 631  */
 632 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 633 {
 634         struct folio *folio = page_folio(page);
 635         struct inode *inode = page->mapping->host;
 636         struct ceph_inode_info *ci = ceph_inode(inode);
 637         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 638         struct ceph_client *cl = fsc->client;
 639         struct ceph_snap_context *snapc, *oldest;
 640         loff_t page_off = page_offset(page);
 641         int err;
 642         loff_t len = thp_size(page);
 643         loff_t wlen;
 644         struct ceph_writeback_ctl ceph_wbc;
 645         struct ceph_osd_client *osdc = &fsc->client->osdc;
 646         struct ceph_osd_request *req;
 647         bool caching = ceph_is_cache_enabled(inode);
 648         struct page *bounce_page = NULL;
 649
 650         doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
 651               page->index);
 652
 653         if (ceph_inode_is_shutdown(inode))
 654                 return -EIO;
 655
 656         /* verify this is a writeable snap context */
 657         snapc = page_snap_context(page);
 658         if (!snapc) {
 659                 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
 660                       page);
 661                 return 0;
 662         }
 663         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
 664         if (snapc->seq > oldest->seq) {
 665                 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
 666                       ceph_vinop(inode), page, snapc);
 667                 /* we should only noop if called by kswapd */
 668                 WARN_ON(!(current->flags & PF_MEMALLOC));
 669                 ceph_put_snap_context(oldest);
 670                 redirty_page_for_writepage(wbc, page);
 671                 return 0;
 672         }
 673         ceph_put_snap_context(oldest);
 674
 675         /* is this a partial page at end of file? */
 676         if (page_off >= ceph_wbc.i_size) {
 677                 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
 678                       ceph_vinop(inode), folio->index, ceph_wbc.i_size);
 679                 folio_invalidate(folio, 0, folio_size(folio));
 680                 return 0;
 681         }
 682
 683         if (ceph_wbc.i_size < page_off + len)
 684                 len = ceph_wbc.i_size - page_off;
 685
 686         wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
 687         doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
 688               ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
 689               snapc->seq);
 690
 691         if (atomic_long_inc_return(&fsc->writeback_count) >
 692             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 693                 fsc->write_congested = true;
 694
 695         req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
 696                                     page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
 697                                     CEPH_OSD_FLAG_WRITE, snapc,
 698                                     ceph_wbc.truncate_seq,
 699                                     ceph_wbc.truncate_size, true);
 700         if (IS_ERR(req)) {
 701                 redirty_page_for_writepage(wbc, page);
 702                 return PTR_ERR(req);
 703         }
 704
 705         if (wlen < len)
 706                 len = wlen;
 707
 708         set_page_writeback(page);
 709         ceph_fscache_write_to_cache(inode, page_off, len, caching);
 710
 711         if (IS_ENCRYPTED(inode)) {
 712                 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
 713                                                     CEPH_FSCRYPT_BLOCK_SIZE, 0,
 714                                                     GFP_NOFS);
 715                 if (IS_ERR(bounce_page)) {
 716                         redirty_page_for_writepage(wbc, page);
 717                         end_page_writeback(page);
 718                         ceph_osdc_put_request(req);
 719                         return PTR_ERR(bounce_page);
 720                 }
 721         }
 722
 723         /* it may be a short write due to an object boundary */
 724         WARN_ON_ONCE(len > thp_size(page));
 725         osd_req_op_extent_osd_data_pages(req, 0,
 726                         bounce_page ? &bounce_page : &page, wlen, 0,
 727                         false, false);
 728         doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
 729               ceph_vinop(inode), page_off, len, wlen,
 730               IS_ENCRYPTED(inode) ? "" : "not ");
 731
 732         req->r_mtime = inode_get_mtime(inode);
 733         ceph_osdc_start_request(osdc, req);
 734         err = ceph_osdc_wait_request(osdc, req);
 735
 736         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 737                                   req->r_end_latency, len, err);
 738         fscrypt_free_bounce_page(bounce_page);
 739         ceph_osdc_put_request(req);
 740         if (err == 0)
 741                 err = len;
 742
 743         if (err < 0) {
 744                 struct writeback_control tmp_wbc;
 745                 if (!wbc)
 746                         wbc = &tmp_wbc;
 747                 if (err == -ERESTARTSYS) {
 748                         /* killed by SIGKILL */
 749                         doutc(cl, "%llx.%llx interrupted page %p\n",
 750                               ceph_vinop(inode), page);
 751                         redirty_page_for_writepage(wbc, page);
 752                         end_page_writeback(page);
 753                         return err;
 754                 }
 755                 if (err == -EBLOCKLISTED)
 756                         fsc->blocklisted = true;
 757                 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
 758                       ceph_vinop(inode), err, page);
 759                 mapping_set_error(&inode->i_data, err);
 760                 wbc->pages_skipped++;
 761         } else {
 762                 doutc(cl, "%llx.%llx cleaned page %p\n",
 763                       ceph_vinop(inode), page);
 764                 err = 0;  /* vfs expects us to return 0 */
 765         }
 766         oldest = detach_page_private(page);
 767         WARN_ON_ONCE(oldest != snapc);
 768         end_page_writeback(page);
 769         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 770         ceph_put_snap_context(snapc);  /* page's reference */
 771
 772         if (atomic_long_dec_return(&fsc->writeback_count) <
 773             CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 774                 fsc->write_congested = false;
 775
 776         return err;
 777 }
 778
 779 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 780 {
 781         int err;
 782         struct inode *inode = page->mapping->host;
 783         BUG_ON(!inode);
 784         ihold(inode);
 785
 786         if (wbc->sync_mode == WB_SYNC_NONE &&
 787             ceph_inode_to_fs_client(inode)->write_congested) {
 788                 redirty_page_for_writepage(wbc, page);
 789                 return AOP_WRITEPAGE_ACTIVATE;
 790         }
 791
 792         err = writepage_nounlock(page, wbc);
 793         if (err == -ERESTARTSYS) {
 794                 /* direct memory reclaimer was killed by SIGKILL. return 0
 795                  * to prevent caller from setting mapping/page error */
 796                 err = 0;
 797         }
 798         unlock_page(page);
 799         iput(inode);
 800         return err;
 801 }
 802
 803 /*
 804  * async writeback completion handler.
 805  *
 806  * If we get an error, set the mapping error bit, but not the individual
 807  * page error bits.
 808  */
 809 static void writepages_finish(struct ceph_osd_request *req)
 810 {
 811         struct inode *inode = req->r_inode;
 812         struct ceph_inode_info *ci = ceph_inode(inode);
 813         struct ceph_client *cl = ceph_inode_to_client(inode);
 814         struct ceph_osd_data *osd_data;
 815         struct page *page;
 816         int num_pages, total_pages = 0;
 817         int i, j;
 818         int rc = req->r_result;
 819         struct ceph_snap_context *snapc = req->r_snapc;
 820         struct address_space *mapping = inode->i_mapping;
 821         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 822         unsigned int len = 0;
 823         bool remove_page;
 824
 825         doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
 826         if (rc < 0) {
 827                 mapping_set_error(mapping, rc);
 828                 ceph_set_error_write(ci);
 829                 if (rc == -EBLOCKLISTED)
 830                         fsc->blocklisted = true;
 831         } else {
 832                 ceph_clear_error_write(ci);
 833         }
 834
 835         /*
 836          * We lost the cache cap, need to truncate the page before
 837          * it is unlocked, otherwise we'd truncate it later in the
 838          * page truncation thread, possibly losing some data that
 839          * raced its way in
 840          */
 841         remove_page = !(ceph_caps_issued(ci) &
 842                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
 843
 844         /* clean all pages */
 845         for (i = 0; i < req->r_num_ops; i++) {
 846                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
 847                         pr_warn_client(cl,
 848                                 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
 849                                 ceph_vinop(inode), req->r_ops[i].op, req, i,
 850                                 req->r_tid);
 851                         break;
 852                 }
 853
 854                 osd_data = osd_req_op_extent_osd_data(req, i);
 855                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 856                 len += osd_data->length;
 857                 num_pages = calc_pages_for((u64)osd_data->alignment,
 858                                            (u64)osd_data->length);
 859                 total_pages += num_pages;
 860                 for (j = 0; j < num_pages; j++) {
 861                         page = osd_data->pages[j];
 862                         if (fscrypt_is_bounce_page(page)) {
 863                                 page = fscrypt_pagecache_page(page);
 864                                 fscrypt_free_bounce_page(osd_data->pages[j]);
 865                                 osd_data->pages[j] = page;
 866                         }
 867                         BUG_ON(!page);
 868                         WARN_ON(!PageUptodate(page));
 869
 870                         if (atomic_long_dec_return(&fsc->writeback_count) <
 871                              CONGESTION_OFF_THRESH(
 872                                         fsc->mount_options->congestion_kb))
 873                                 fsc->write_congested = false;
 874
 875                         ceph_put_snap_context(detach_page_private(page));
 876                         end_page_writeback(page);
 877                         doutc(cl, "unlocking %p\n", page);
 878
 879                         if (remove_page)
 880                                 generic_error_remove_folio(inode->i_mapping,
 881                                                           page_folio(page));
 882
 883                         unlock_page(page);
 884                 }
 885                 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
 886                       ceph_vinop(inode), osd_data->length,
 887                       rc >= 0 ? num_pages : 0);
 888
 889                 release_pages(osd_data->pages, num_pages);
 890         }
 891
 892         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
 893                                   req->r_end_latency, len, rc);
 894
 895         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
 896
 897         osd_data = osd_req_op_extent_osd_data(req, 0);
 898         if (osd_data->pages_from_pool)
 899                 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
 900         else
 901                 kfree(osd_data->pages);
 902         ceph_osdc_put_request(req);
 903         ceph_dec_osd_stopping_blocker(fsc->mdsc);
 904 }
 905
 906 /*
 907  * initiate async writeback
 908  */
 909 static int ceph_writepages_start(struct address_space *mapping,
 910                                  struct writeback_control *wbc)
 911 {
 912         struct inode *inode = mapping->host;
 913         struct ceph_inode_info *ci = ceph_inode(inode);
 914         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 915         struct ceph_client *cl = fsc->client;
 916         struct ceph_vino vino = ceph_vino(inode);
 917         pgoff_t index, start_index, end = -1;
 918         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 919         struct folio_batch fbatch;
 920         int rc = 0;
 921         unsigned int wsize = i_blocksize(inode);
 922         struct ceph_osd_request *req = NULL;
 923         struct ceph_writeback_ctl ceph_wbc;
 924         bool should_loop, range_whole = false;
 925         bool done = false;
 926         bool caching = ceph_is_cache_enabled(inode);
 927         xa_mark_t tag;
 928
 929         if (wbc->sync_mode == WB_SYNC_NONE &&
 930             fsc->write_congested)
 931                 return 0;
 932
 933         doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
 934               wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 935               (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 936
 937         if (ceph_inode_is_shutdown(inode)) {
 938                 if (ci->i_wrbuffer_ref > 0) {
 939                         pr_warn_ratelimited_client(cl,
 940                                 "%llx.%llx %lld forced umount\n",
 941                                 ceph_vinop(inode), ceph_ino(inode));
 942                 }
 943                 mapping_set_error(mapping, -EIO);
 944                 return -EIO; /* we're in a forced umount, don't write! */
 945         }
 946         if (fsc->mount_options->wsize < wsize)
 947                 wsize = fsc->mount_options->wsize;
 948
 949         folio_batch_init(&fbatch);
 950
 951         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
 952         index = start_index;
 953
 954         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
 955                 tag = PAGECACHE_TAG_TOWRITE;
 956         } else {
 957                 tag = PAGECACHE_TAG_DIRTY;
 958         }
 959 retry:
 960         /* find oldest snap context with dirty data */
 961         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
 962         if (!snapc) {
 963                 /* hmm, why does writepages get called when there
 964                    is no dirty data? */
 965                 doutc(cl, " no snap context with dirty data?\n");
 966                 goto out;
 967         }
 968         doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
 969               snapc->seq, snapc->num_snaps);
 970
 971         should_loop = false;
 972         if (ceph_wbc.head_snapc && snapc != last_snapc) {
 973                 /* where to start/end? */
 974                 if (wbc->range_cyclic) {
 975                         index = start_index;
 976                         end = -1;
 977                         if (index > 0)
 978                                 should_loop = true;
 979                         doutc(cl, " cyclic, start at %lu\n", index);
 980                 } else {
 981                         index = wbc->range_start >> PAGE_SHIFT;
 982                         end = wbc->range_end >> PAGE_SHIFT;
 983                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 984                                 range_whole = true;
 985                         doutc(cl, " not cyclic, %lu to %lu\n", index, end);
 986                 }
 987         } else if (!ceph_wbc.head_snapc) {
 988                 /* Do not respect wbc->range_{start,end}. Dirty pages
 989                  * in that range can be associated with newer snapc.
 990                  * They are not writeable until we write all dirty pages
 991                  * associated with 'snapc' get written */
 992                 if (index > 0)
 993                         should_loop = true;
 994                 doutc(cl, " non-head snapc, range whole\n");
 995         }
 996
 997         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 998                 tag_pages_for_writeback(mapping, index, end);
 999
1000         ceph_put_snap_context(last_snapc);
1001         last_snapc = snapc;
1002
1003         while (!done && index <= end) {
1004                 int num_ops = 0, op_idx;
1005                 unsigned i, nr_folios, max_pages, locked_pages = 0;
1006                 struct page **pages = NULL, **data_pages;
1007                 struct page *page;
1008                 pgoff_t strip_unit_end = 0;
1009                 u64 offset = 0, len = 0;
1010                 bool from_pool = false;
1011
1012                 max_pages = wsize >> PAGE_SHIFT;
1013
1014 get_more_pages:
1015                 nr_folios = filemap_get_folios_tag(mapping, &index,
1016                                                    end, tag, &fbatch);
1017                 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1018                 if (!nr_folios && !locked_pages)
1019                         break;
1020                 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1021                         page = &fbatch.folios[i]->page;
1022                         doutc(cl, "? %p idx %lu\n", page, page->index);
1023                         if (locked_pages == 0)
1024                                 lock_page(page);  /* first page */
1025                         else if (!trylock_page(page))
1026                                 break;
1027
1028                         /* only dirty pages, or our accounting breaks */
1029                         if (unlikely(!PageDirty(page)) ||
1030                             unlikely(page->mapping != mapping)) {
1031                                 doutc(cl, "!dirty or !mapping %p\n", page);
1032                                 unlock_page(page);
1033                                 continue;
1034                         }
1035                         /* only if matching snap context */
1036                         pgsnapc = page_snap_context(page);
1037                         if (pgsnapc != snapc) {
1038                                 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1039                                       pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1040                                 if (!should_loop &&
1041                                     !ceph_wbc.head_snapc &&
1042                                     wbc->sync_mode != WB_SYNC_NONE)
1043                                         should_loop = true;
1044                                 unlock_page(page);
1045                                 continue;
1046                         }
1047                         if (page_offset(page) >= ceph_wbc.i_size) {
1048                                 struct folio *folio = page_folio(page);
1049
1050                                 doutc(cl, "folio at %lu beyond eof %llu\n",
1051                                       folio->index, ceph_wbc.i_size);
1052                                 if ((ceph_wbc.size_stable ||
1053                                     folio_pos(folio) >= i_size_read(inode)) &&
1054                                     folio_clear_dirty_for_io(folio))
1055                                         folio_invalidate(folio, 0,
1056                                                         folio_size(folio));
1057                                 folio_unlock(folio);
1058                                 continue;
1059                         }
1060                         if (strip_unit_end && (page->index > strip_unit_end)) {
1061                                 doutc(cl, "end of strip unit %p\n", page);
1062                                 unlock_page(page);
1063                                 break;
1064                         }
1065                         if (PageWriteback(page)) {
1066                                 if (wbc->sync_mode == WB_SYNC_NONE) {
1067                                         doutc(cl, "%p under writeback\n", page);
1068                                         unlock_page(page);
1069                                         continue;
1070                                 }
1071                                 doutc(cl, "waiting on writeback %p\n", page);
1072                                 wait_on_page_writeback(page);
1073                         }
1074
1075                         if (!clear_page_dirty_for_io(page)) {
1076                                 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1077                                 unlock_page(page);
1078                                 continue;
1079                         }
1080
1081                         /*
1082                          * We have something to write.  If this is
1083                          * the first locked page this time through,
1084                          * calculate max possinle write size and
1085                          * allocate a page array
1086                          */
1087                         if (locked_pages == 0) {
1088                                 u64 objnum;
1089                                 u64 objoff;
1090                                 u32 xlen;
1091
1092                                 /* prepare async write request */
1093                                 offset = (u64)page_offset(page);
1094                                 ceph_calc_file_object_mapping(&ci->i_layout,
1095                                                               offset, wsize,
1096                                                               &objnum, &objoff,
1097                                                               &xlen);
1098                                 len = xlen;
1099
1100                                 num_ops = 1;
1101                                 strip_unit_end = page->index +
1102                                         ((len - 1) >> PAGE_SHIFT);
1103
1104                                 BUG_ON(pages);
1105                                 max_pages = calc_pages_for(0, (u64)len);
1106                                 pages = kmalloc_array(max_pages,
1107                                                       sizeof(*pages),
1108                                                       GFP_NOFS);
1109                                 if (!pages) {
1110                                         from_pool = true;
1111                                         pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1112                                         BUG_ON(!pages);
1113                                 }
1114
1115                                 len = 0;
1116                         } else if (page->index !=
1117                                    (offset + len) >> PAGE_SHIFT) {
1118                                 if (num_ops >= (from_pool ?  CEPH_OSD_SLAB_OPS :
1119                                                              CEPH_OSD_MAX_OPS)) {
1120                                         redirty_page_for_writepage(wbc, page);
1121                                         unlock_page(page);
1122                                         break;
1123                                 }
1124
1125                                 num_ops++;
1126                                 offset = (u64)page_offset(page);
1127                                 len = 0;
1128                         }
1129
1130                         /* note position of first page in fbatch */
1131                         doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1132                               ceph_vinop(inode), page, page->index);
1133
1134                         if (atomic_long_inc_return(&fsc->writeback_count) >
1135                             CONGESTION_ON_THRESH(
1136                                     fsc->mount_options->congestion_kb))
1137                                 fsc->write_congested = true;
1138
1139                         if (IS_ENCRYPTED(inode)) {
1140                                 pages[locked_pages] =
1141                                         fscrypt_encrypt_pagecache_blocks(page,
1142                                                 PAGE_SIZE, 0,
1143                                                 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1144                                 if (IS_ERR(pages[locked_pages])) {
1145                                         if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1146                                                 pr_err_client(cl,
1147                                                         "inode->i_blkbits=%hhu\n",
1148                                                         inode->i_blkbits);
1149                                         /* better not fail on first page! */
1150                                         BUG_ON(locked_pages == 0);
1151                                         pages[locked_pages] = NULL;
1152                                         redirty_page_for_writepage(wbc, page);
1153                                         unlock_page(page);
1154                                         break;
1155                                 }
1156                                 ++locked_pages;
1157                         } else {
1158                                 pages[locked_pages++] = page;
1159                         }
1160
1161                         fbatch.folios[i] = NULL;
1162                         len += thp_size(page);
1163                 }
1164
1165                 /* did we get anything? */
1166                 if (!locked_pages)
1167                         goto release_folios;
1168                 if (i) {
1169                         unsigned j, n = 0;
1170                         /* shift unused page to beginning of fbatch */
1171                         for (j = 0; j < nr_folios; j++) {
1172                                 if (!fbatch.folios[j])
1173                                         continue;
1174                                 if (n < j)
1175                                         fbatch.folios[n] = fbatch.folios[j];
1176                                 n++;
1177                         }
1178                         fbatch.nr = n;
1179
1180                         if (nr_folios && i == nr_folios &&
1181                             locked_pages < max_pages) {
1182                                 doutc(cl, "reached end fbatch, trying for more\n");
1183                                 folio_batch_release(&fbatch);
1184                                 goto get_more_pages;
1185                         }
1186                 }
1187
1188 new_request:
1189                 offset = ceph_fscrypt_page_offset(pages[0]);
1190                 len = wsize;
1191
1192                 req = ceph_osdc_new_request(&fsc->client->osdc,
1193                                         &ci->i_layout, vino,
1194                                         offset, &len, 0, num_ops,
1195                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1196                                         snapc, ceph_wbc.truncate_seq,
1197                                         ceph_wbc.truncate_size, false);
1198                 if (IS_ERR(req)) {
1199                         req = ceph_osdc_new_request(&fsc->client->osdc,
1200                                                 &ci->i_layout, vino,
1201                                                 offset, &len, 0,
1202                                                 min(num_ops,
1203                                                     CEPH_OSD_SLAB_OPS),
1204                                                 CEPH_OSD_OP_WRITE,
1205                                                 CEPH_OSD_FLAG_WRITE,
1206                                                 snapc, ceph_wbc.truncate_seq,
1207                                                 ceph_wbc.truncate_size, true);
1208                         BUG_ON(IS_ERR(req));
1209                 }
1210                 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1211                              thp_size(pages[locked_pages - 1]) - offset);
1212
1213                 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1214                         rc = -EIO;
1215                         goto release_folios;
1216                 }
1217                 req->r_callback = writepages_finish;
1218                 req->r_inode = inode;
1219
1220                 /* Format the osd request message and submit the write */
1221                 len = 0;
1222                 data_pages = pages;
1223                 op_idx = 0;
1224                 for (i = 0; i < locked_pages; i++) {
1225                         struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1226
1227                         u64 cur_offset = page_offset(page);
1228                         /*
1229                          * Discontinuity in page range? Ceph can handle that by just passing
1230                          * multiple extents in the write op.
1231                          */
1232                         if (offset + len != cur_offset) {
1233                                 /* If it's full, stop here */
1234                                 if (op_idx + 1 == req->r_num_ops)
1235                                         break;
1236
1237                                 /* Kick off an fscache write with what we have so far. */
1238                                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1239
1240                                 /* Start a new extent */
1241                                 osd_req_op_extent_dup_last(req, op_idx,
1242                                                            cur_offset - offset);
1243                                 doutc(cl, "got pages at %llu~%llu\n", offset,
1244                                       len);
1245                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1246                                                         data_pages, len, 0,
1247                                                         from_pool, false);
1248                                 osd_req_op_extent_update(req, op_idx, len);
1249
1250                                 len = 0;
1251                                 offset = cur_offset;
1252                                 data_pages = pages + i;
1253                                 op_idx++;
1254                         }
1255
1256                         set_page_writeback(page);
1257                         len += thp_size(page);
1258                 }
1259                 ceph_fscache_write_to_cache(inode, offset, len, caching);
1260
1261                 if (ceph_wbc.size_stable) {
1262                         len = min(len, ceph_wbc.i_size - offset);
1263                 } else if (i == locked_pages) {
1264                         /* writepages_finish() clears writeback pages
1265                          * according to the data length, so make sure
1266                          * data length covers all locked pages */
1267                         u64 min_len = len + 1 - thp_size(page);
1268                         len = get_writepages_data_length(inode, pages[i - 1],
1269                                                          offset);
1270                         len = max(len, min_len);
1271                 }
1272                 if (IS_ENCRYPTED(inode))
1273                         len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1274
1275                 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1276
1277                 if (IS_ENCRYPTED(inode) &&
1278                     ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1279                         pr_warn_client(cl,
1280                                 "bad encrypted write offset=%lld len=%llu\n",
1281                                 offset, len);
1282
1283                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1284                                                  0, from_pool, false);
1285                 osd_req_op_extent_update(req, op_idx, len);
1286
1287                 BUG_ON(op_idx + 1 != req->r_num_ops);
1288
1289                 from_pool = false;
1290                 if (i < locked_pages) {
1291                         BUG_ON(num_ops <= req->r_num_ops);
1292                         num_ops -= req->r_num_ops;
1293                         locked_pages -= i;
1294
1295                         /* allocate new pages array for next request */
1296                         data_pages = pages;
1297                         pages = kmalloc_array(locked_pages, sizeof(*pages),
1298                                               GFP_NOFS);
1299                         if (!pages) {
1300                                 from_pool = true;
1301                                 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1302                                 BUG_ON(!pages);
1303                         }
1304                         memcpy(pages, data_pages + i,
1305                                locked_pages * sizeof(*pages));
1306                         memset(data_pages + i, 0,
1307                                locked_pages * sizeof(*pages));
1308                 } else {
1309                         BUG_ON(num_ops != req->r_num_ops);
1310                         index = pages[i - 1]->index + 1;
1311                         /* request message now owns the pages array */
1312                         pages = NULL;
1313                 }
1314
1315                 req->r_mtime = inode_get_mtime(inode);
1316                 ceph_osdc_start_request(&fsc->client->osdc, req);
1317                 req = NULL;
1318
1319                 wbc->nr_to_write -= i;
1320                 if (pages)
1321                         goto new_request;
1322
1323                 /*
1324                  * We stop writing back only if we are not doing
1325                  * integrity sync. In case of integrity sync we have to
1326                  * keep going until we have written all the pages
1327                  * we tagged for writeback prior to entering this loop.
1328                  */
1329                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1330                         done = true;
1331
1332 release_folios:
1333                 doutc(cl, "folio_batch release on %d folios (%p)\n",
1334                       (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1335                 folio_batch_release(&fbatch);
1336         }
1337
1338         if (should_loop && !done) {
1339                 /* more to do; loop back to beginning of file */
1340                 doutc(cl, "looping back to beginning of file\n");
1341                 end = start_index - 1; /* OK even when start_index == 0 */
1342
1343                 /* to write dirty pages associated with next snapc,
1344                  * we need to wait until current writes complete */
1345                 if (wbc->sync_mode != WB_SYNC_NONE &&
1346                     start_index == 0 && /* all dirty pages were checked */
1347                     !ceph_wbc.head_snapc) {
1348                         struct page *page;
1349                         unsigned i, nr;
1350                         index = 0;
1351                         while ((index <= end) &&
1352                                (nr = filemap_get_folios_tag(mapping, &index,
1353                                                 (pgoff_t)-1,
1354                                                 PAGECACHE_TAG_WRITEBACK,
1355                                                 &fbatch))) {
1356                                 for (i = 0; i < nr; i++) {
1357                                         page = &fbatch.folios[i]->page;
1358                                         if (page_snap_context(page) != snapc)
1359                                                 continue;
1360                                         wait_on_page_writeback(page);
1361                                 }
1362                                 folio_batch_release(&fbatch);
1363                                 cond_resched();
1364                         }
1365                 }
1366
1367                 start_index = 0;
1368                 index = 0;
1369                 goto retry;
1370         }
1371
1372         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1373                 mapping->writeback_index = index;
1374
1375 out:
1376         ceph_osdc_put_request(req);
1377         ceph_put_snap_context(last_snapc);
1378         doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1379               rc);
1380         return rc;
1381 }
1382
1383
1384
1385 /*
1386  * See if a given @snapc is either writeable, or already written.
1387  */
1388 static int context_is_writeable_or_written(struct inode *inode,
1389                                            struct ceph_snap_context *snapc)
1390 {
1391         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1392         int ret = !oldest || snapc->seq <= oldest->seq;
1393
1394         ceph_put_snap_context(oldest);
1395         return ret;
1396 }
1397
1398 /**
1399  * ceph_find_incompatible - find an incompatible context and return it
1400  * @page: page being dirtied
1401  *
1402  * We are only allowed to write into/dirty a page if the page is
1403  * clean, or already dirty within the same snap context. Returns a
1404  * conflicting context if there is one, NULL if there isn't, or a
1405  * negative error code on other errors.
1406  *
1407  * Must be called with page lock held.
1408  */
1409 static struct ceph_snap_context *
1410 ceph_find_incompatible(struct page *page)
1411 {
1412         struct inode *inode = page->mapping->host;
1413         struct ceph_client *cl = ceph_inode_to_client(inode);
1414         struct ceph_inode_info *ci = ceph_inode(inode);
1415
1416         if (ceph_inode_is_shutdown(inode)) {
1417                 doutc(cl, " %llx.%llx page %p is shutdown\n",
1418                       ceph_vinop(inode), page);
1419                 return ERR_PTR(-ESTALE);
1420         }
1421
1422         for (;;) {
1423                 struct ceph_snap_context *snapc, *oldest;
1424
1425                 wait_on_page_writeback(page);
1426
1427                 snapc = page_snap_context(page);
1428                 if (!snapc || snapc == ci->i_head_snapc)
1429                         break;
1430
1431                 /*
1432                  * this page is already dirty in another (older) snap
1433                  * context!  is it writeable now?
1434                  */
1435                 oldest = get_oldest_context(inode, NULL, NULL);
1436                 if (snapc->seq > oldest->seq) {
1437                         /* not writeable -- return it for the caller to deal with */
1438                         ceph_put_snap_context(oldest);
1439                         doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1440                               ceph_vinop(inode), page, snapc);
1441                         return ceph_get_snap_context(snapc);
1442                 }
1443                 ceph_put_snap_context(oldest);
1444
1445                 /* yay, writeable, do it now (without dropping page lock) */
1446                 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1447                       ceph_vinop(inode), page, snapc);
1448                 if (clear_page_dirty_for_io(page)) {
1449                         int r = writepage_nounlock(page, NULL);
1450                         if (r < 0)
1451                                 return ERR_PTR(r);
1452                 }
1453         }
1454         return NULL;
1455 }
1456
1457 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1458                                         struct folio **foliop, void **_fsdata)
1459 {
1460         struct inode *inode = file_inode(file);
1461         struct ceph_inode_info *ci = ceph_inode(inode);
1462         struct ceph_snap_context *snapc;
1463
1464         snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1465         if (snapc) {
1466                 int r;
1467
1468                 folio_unlock(*foliop);
1469                 folio_put(*foliop);
1470                 *foliop = NULL;
1471                 if (IS_ERR(snapc))
1472                         return PTR_ERR(snapc);
1473
1474                 ceph_queue_writeback(inode);
1475                 r = wait_event_killable(ci->i_cap_wq,
1476                                         context_is_writeable_or_written(inode, snapc));
1477                 ceph_put_snap_context(snapc);
1478                 return r == 0 ? -EAGAIN : r;
1479         }
1480         return 0;
1481 }
1482
1483 /*
1484  * We are only allowed to write into/dirty the page if the page is
1485  * clean, or already dirty within the same snap context.
1486  */
1487 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1488                             loff_t pos, unsigned len,
1489                             struct page **pagep, void **fsdata)
1490 {
1491         struct inode *inode = file_inode(file);
1492         struct ceph_inode_info *ci = ceph_inode(inode);
1493         struct folio *folio = NULL;
1494         int r;
1495
1496         r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1497         if (r < 0)
1498                 return r;
1499
1500         folio_wait_private_2(folio); /* [DEPRECATED] */
1501         WARN_ON_ONCE(!folio_test_locked(folio));
1502         *pagep = &folio->page;
1503         return 0;
1504 }
1505
1506 /*
1507  * we don't do anything in here that simple_write_end doesn't do
1508  * except adjust dirty page accounting
1509  */
1510 static int ceph_write_end(struct file *file, struct address_space *mapping,
1511                           loff_t pos, unsigned len, unsigned copied,
1512                           struct page *subpage, void *fsdata)
1513 {
1514         struct folio *folio = page_folio(subpage);
1515         struct inode *inode = file_inode(file);
1516         struct ceph_client *cl = ceph_inode_to_client(inode);
1517         bool check_cap = false;
1518
1519         doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1520               file, folio, (int)pos, (int)copied, (int)len);
1521
1522         if (!folio_test_uptodate(folio)) {
1523                 /* just return that nothing was copied on a short copy */
1524                 if (copied < len) {
1525                         copied = 0;
1526                         goto out;
1527                 }
1528                 folio_mark_uptodate(folio);
1529         }
1530
1531         /* did file size increase? */
1532         if (pos+copied > i_size_read(inode))
1533                 check_cap = ceph_inode_set_size(inode, pos+copied);
1534
1535         folio_mark_dirty(folio);
1536
1537 out:
1538         folio_unlock(folio);
1539         folio_put(folio);
1540
1541         if (check_cap)
1542                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1543
1544         return copied;
1545 }
1546
1547 const struct address_space_operations ceph_aops = {
1548         .read_folio = netfs_read_folio,
1549         .readahead = netfs_readahead,
1550         .writepage = ceph_writepage,
1551         .writepages = ceph_writepages_start,
1552         .write_begin = ceph_write_begin,
1553         .write_end = ceph_write_end,
1554         .dirty_folio = ceph_dirty_folio,
1555         .invalidate_folio = ceph_invalidate_folio,
1556         .release_folio = netfs_release_folio,
1557         .direct_IO = noop_direct_IO,
1558 };
1559
1560 static void ceph_block_sigs(sigset_t *oldset)
1561 {
1562         sigset_t mask;
1563         siginitsetinv(&mask, sigmask(SIGKILL));
1564         sigprocmask(SIG_BLOCK, &mask, oldset);
1565 }
1566
1567 static void ceph_restore_sigs(sigset_t *oldset)
1568 {
1569         sigprocmask(SIG_SETMASK, oldset, NULL);
1570 }
1571
1572 /*
1573  * vm ops
1574  */
1575 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1576 {
1577         struct vm_area_struct *vma = vmf->vma;
1578         struct inode *inode = file_inode(vma->vm_file);
1579         struct ceph_inode_info *ci = ceph_inode(inode);
1580         struct ceph_client *cl = ceph_inode_to_client(inode);
1581         struct ceph_file_info *fi = vma->vm_file->private_data;
1582         loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1583         int want, got, err;
1584         sigset_t oldset;
1585         vm_fault_t ret = VM_FAULT_SIGBUS;
1586
1587         if (ceph_inode_is_shutdown(inode))
1588                 return ret;
1589
1590         ceph_block_sigs(&oldset);
1591
1592         doutc(cl, "%llx.%llx %llu trying to get caps\n",
1593               ceph_vinop(inode), off);
1594         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1595                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1596         else
1597                 want = CEPH_CAP_FILE_CACHE;
1598
1599         got = 0;
1600         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1601         if (err < 0)
1602                 goto out_restore;
1603
1604         doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1605               off, ceph_cap_string(got));
1606
1607         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1608             !ceph_has_inline_data(ci)) {
1609                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1610                 ceph_add_rw_context(fi, &rw_ctx);
1611                 ret = filemap_fault(vmf);
1612                 ceph_del_rw_context(fi, &rw_ctx);
1613                 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1614                       ceph_vinop(inode), off, ceph_cap_string(got), ret);
1615         } else
1616                 err = -EAGAIN;
1617
1618         ceph_put_cap_refs(ci, got);
1619
1620         if (err != -EAGAIN)
1621                 goto out_restore;
1622
1623         /* read inline data */
1624         if (off >= PAGE_SIZE) {
1625                 /* does not support inline data > PAGE_SIZE */
1626                 ret = VM_FAULT_SIGBUS;
1627         } else {
1628                 struct address_space *mapping = inode->i_mapping;
1629                 struct page *page;
1630
1631                 filemap_invalidate_lock_shared(mapping);
1632                 page = find_or_create_page(mapping, 0,
1633                                 mapping_gfp_constraint(mapping, ~__GFP_FS));
1634                 if (!page) {
1635                         ret = VM_FAULT_OOM;
1636                         goto out_inline;
1637                 }
1638                 err = __ceph_do_getattr(inode, page,
1639                                          CEPH_STAT_CAP_INLINE_DATA, true);
1640                 if (err < 0 || off >= i_size_read(inode)) {
1641                         unlock_page(page);
1642                         put_page(page);
1643                         ret = vmf_error(err);
1644                         goto out_inline;
1645                 }
1646                 if (err < PAGE_SIZE)
1647                         zero_user_segment(page, err, PAGE_SIZE);
1648                 else
1649                         flush_dcache_page(page);
1650                 SetPageUptodate(page);
1651                 vmf->page = page;
1652                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1653 out_inline:
1654                 filemap_invalidate_unlock_shared(mapping);
1655                 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1656                       ceph_vinop(inode), off, ret);
1657         }
1658 out_restore:
1659         ceph_restore_sigs(&oldset);
1660         if (err < 0)
1661                 ret = vmf_error(err);
1662
1663         return ret;
1664 }
1665
1666 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1667 {
1668         struct vm_area_struct *vma = vmf->vma;
1669         struct inode *inode = file_inode(vma->vm_file);
1670         struct ceph_client *cl = ceph_inode_to_client(inode);
1671         struct ceph_inode_info *ci = ceph_inode(inode);
1672         struct ceph_file_info *fi = vma->vm_file->private_data;
1673         struct ceph_cap_flush *prealloc_cf;
1674         struct page *page = vmf->page;
1675         loff_t off = page_offset(page);
1676         loff_t size = i_size_read(inode);
1677         size_t len;
1678         int want, got, err;
1679         sigset_t oldset;
1680         vm_fault_t ret = VM_FAULT_SIGBUS;
1681
1682         if (ceph_inode_is_shutdown(inode))
1683                 return ret;
1684
1685         prealloc_cf = ceph_alloc_cap_flush();
1686         if (!prealloc_cf)
1687                 return VM_FAULT_OOM;
1688
1689         sb_start_pagefault(inode->i_sb);
1690         ceph_block_sigs(&oldset);
1691
1692         if (off + thp_size(page) <= size)
1693                 len = thp_size(page);
1694         else
1695                 len = offset_in_thp(page, size);
1696
1697         doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1698               ceph_vinop(inode), off, len, size);
1699         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1700                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1701         else
1702                 want = CEPH_CAP_FILE_BUFFER;
1703
1704         got = 0;
1705         err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1706         if (err < 0)
1707                 goto out_free;
1708
1709         doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1710               off, len, ceph_cap_string(got));
1711
1712         /* Update time before taking page lock */
1713         file_update_time(vma->vm_file);
1714         inode_inc_iversion_raw(inode);
1715
1716         do {
1717                 struct ceph_snap_context *snapc;
1718
1719                 lock_page(page);
1720
1721                 if (page_mkwrite_check_truncate(page, inode) < 0) {
1722                         unlock_page(page);
1723                         ret = VM_FAULT_NOPAGE;
1724                         break;
1725                 }
1726
1727                 snapc = ceph_find_incompatible(page);
1728                 if (!snapc) {
1729                         /* success.  we'll keep the page locked. */
1730                         set_page_dirty(page);
1731                         ret = VM_FAULT_LOCKED;
1732                         break;
1733                 }
1734
1735                 unlock_page(page);
1736
1737                 if (IS_ERR(snapc)) {
1738                         ret = VM_FAULT_SIGBUS;
1739                         break;
1740                 }
1741
1742                 ceph_queue_writeback(inode);
1743                 err = wait_event_killable(ci->i_cap_wq,
1744                                 context_is_writeable_or_written(inode, snapc));
1745                 ceph_put_snap_context(snapc);
1746         } while (err == 0);
1747
1748         if (ret == VM_FAULT_LOCKED) {
1749                 int dirty;
1750                 spin_lock(&ci->i_ceph_lock);
1751                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1752                                                &prealloc_cf);
1753                 spin_unlock(&ci->i_ceph_lock);
1754                 if (dirty)
1755                         __mark_inode_dirty(inode, dirty);
1756         }
1757
1758         doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1759               ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1760         ceph_put_cap_refs_async(ci, got);
1761 out_free:
1762         ceph_restore_sigs(&oldset);
1763         sb_end_pagefault(inode->i_sb);
1764         ceph_free_cap_flush(prealloc_cf);
1765         if (err < 0)
1766                 ret = vmf_error(err);
1767         return ret;
1768 }
1769
1770 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1771                            char *data, size_t len)
1772 {
1773         struct ceph_client *cl = ceph_inode_to_client(inode);
1774         struct address_space *mapping = inode->i_mapping;
1775         struct page *page;
1776
1777         if (locked_page) {
1778                 page = locked_page;
1779         } else {
1780                 if (i_size_read(inode) == 0)
1781                         return;
1782                 page = find_or_create_page(mapping, 0,
1783                                            mapping_gfp_constraint(mapping,
1784                                            ~__GFP_FS));
1785                 if (!page)
1786                         return;
1787                 if (PageUptodate(page)) {
1788                         unlock_page(page);
1789                         put_page(page);
1790                         return;
1791                 }
1792         }
1793
1794         doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1795               ceph_vinop(inode), len, locked_page);
1796
1797         if (len > 0) {
1798                 void *kaddr = kmap_atomic(page);
1799                 memcpy(kaddr, data, len);
1800                 kunmap_atomic(kaddr);
1801         }
1802
1803         if (page != locked_page) {
1804                 if (len < PAGE_SIZE)
1805                         zero_user_segment(page, len, PAGE_SIZE);
1806                 else
1807                         flush_dcache_page(page);
1808
1809                 SetPageUptodate(page);
1810                 unlock_page(page);
1811                 put_page(page);
1812         }
1813 }
1814
1815 int ceph_uninline_data(struct file *file)
1816 {
1817         struct inode *inode = file_inode(file);
1818         struct ceph_inode_info *ci = ceph_inode(inode);
1819         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1820         struct ceph_client *cl = fsc->client;
1821         struct ceph_osd_request *req = NULL;
1822         struct ceph_cap_flush *prealloc_cf = NULL;
1823         struct folio *folio = NULL;
1824         u64 inline_version = CEPH_INLINE_NONE;
1825         struct page *pages[1];
1826         int err = 0;
1827         u64 len;
1828
1829         spin_lock(&ci->i_ceph_lock);
1830         inline_version = ci->i_inline_version;
1831         spin_unlock(&ci->i_ceph_lock);
1832
1833         doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1834               inline_version);
1835
1836         if (ceph_inode_is_shutdown(inode)) {
1837                 err = -EIO;
1838                 goto out;
1839         }
1840
1841         if (inline_version == CEPH_INLINE_NONE)
1842                 return 0;
1843
1844         prealloc_cf = ceph_alloc_cap_flush();
1845         if (!prealloc_cf)
1846                 return -ENOMEM;
1847
1848         if (inline_version == 1) /* initial version, no data */
1849                 goto out_uninline;
1850
1851         folio = read_mapping_folio(inode->i_mapping, 0, file);
1852         if (IS_ERR(folio)) {
1853                 err = PTR_ERR(folio);
1854                 goto out;
1855         }
1856
1857         folio_lock(folio);
1858
1859         len = i_size_read(inode);
1860         if (len > folio_size(folio))
1861                 len = folio_size(folio);
1862
1863         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1864                                     ceph_vino(inode), 0, &len, 0, 1,
1865                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1866                                     NULL, 0, 0, false);
1867         if (IS_ERR(req)) {
1868                 err = PTR_ERR(req);
1869                 goto out_unlock;
1870         }
1871
1872         req->r_mtime = inode_get_mtime(inode);
1873         ceph_osdc_start_request(&fsc->client->osdc, req);
1874         err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1875         ceph_osdc_put_request(req);
1876         if (err < 0)
1877                 goto out_unlock;
1878
1879         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1880                                     ceph_vino(inode), 0, &len, 1, 3,
1881                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1882                                     NULL, ci->i_truncate_seq,
1883                                     ci->i_truncate_size, false);
1884         if (IS_ERR(req)) {
1885                 err = PTR_ERR(req);
1886                 goto out_unlock;
1887         }
1888
1889         pages[0] = folio_page(folio, 0);
1890         osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1891
1892         {
1893                 __le64 xattr_buf = cpu_to_le64(inline_version);
1894                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1895                                             "inline_version", &xattr_buf,
1896                                             sizeof(xattr_buf),
1897                                             CEPH_OSD_CMPXATTR_OP_GT,
1898                                             CEPH_OSD_CMPXATTR_MODE_U64);
1899                 if (err)
1900                         goto out_put_req;
1901         }
1902
1903         {
1904                 char xattr_buf[32];
1905                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1906                                          "%llu", inline_version);
1907                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1908                                             "inline_version",
1909                                             xattr_buf, xattr_len, 0, 0);
1910                 if (err)
1911                         goto out_put_req;
1912         }
1913
1914         req->r_mtime = inode_get_mtime(inode);
1915         ceph_osdc_start_request(&fsc->client->osdc, req);
1916         err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1917
1918         ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1919                                   req->r_end_latency, len, err);
1920
1921 out_uninline:
1922         if (!err) {
1923                 int dirty;
1924
1925                 /* Set to CAP_INLINE_NONE and dirty the caps */
1926                 down_read(&fsc->mdsc->snap_rwsem);
1927                 spin_lock(&ci->i_ceph_lock);
1928                 ci->i_inline_version = CEPH_INLINE_NONE;
1929                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1930                 spin_unlock(&ci->i_ceph_lock);
1931                 up_read(&fsc->mdsc->snap_rwsem);
1932                 if (dirty)
1933                         __mark_inode_dirty(inode, dirty);
1934         }
1935 out_put_req:
1936         ceph_osdc_put_request(req);
1937         if (err == -ECANCELED)
1938                 err = 0;
1939 out_unlock:
1940         if (folio) {
1941                 folio_unlock(folio);
1942                 folio_put(folio);
1943         }
1944 out:
1945         ceph_free_cap_flush(prealloc_cf);
1946         doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1947               ceph_vinop(inode), inline_version, err);
1948         return err;
1949 }
1950
1951 static const struct vm_operations_struct ceph_vmops = {
1952         .fault          = ceph_filemap_fault,
1953         .page_mkwrite   = ceph_page_mkwrite,
1954 };
1955
1956 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1957 {
1958         struct address_space *mapping = file->f_mapping;
1959
1960         if (!mapping->a_ops->read_folio)
1961                 return -ENOEXEC;
1962         vma->vm_ops = &ceph_vmops;
1963         return 0;
1964 }
1965
1966 enum {
1967         POOL_READ       = 1,
1968         POOL_WRITE      = 2,
1969 };
1970
1971 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1972                                 s64 pool, struct ceph_string *pool_ns)
1973 {
1974         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
1975         struct ceph_mds_client *mdsc = fsc->mdsc;
1976         struct ceph_client *cl = fsc->client;
1977         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1978         struct rb_node **p, *parent;
1979         struct ceph_pool_perm *perm;
1980         struct page **pages;
1981         size_t pool_ns_len;
1982         int err = 0, err2 = 0, have = 0;
1983
1984         down_read(&mdsc->pool_perm_rwsem);
1985         p = &mdsc->pool_perm_tree.rb_node;
1986         while (*p) {
1987                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1988                 if (pool < perm->pool)
1989                         p = &(*p)->rb_left;
1990                 else if (pool > perm->pool)
1991                         p = &(*p)->rb_right;
1992                 else {
1993                         int ret = ceph_compare_string(pool_ns,
1994                                                 perm->pool_ns,
1995                                                 perm->pool_ns_len);
1996                         if (ret < 0)
1997                                 p = &(*p)->rb_left;
1998                         else if (ret > 0)
1999                                 p = &(*p)->rb_right;
2000                         else {
2001                                 have = perm->perm;
2002                                 break;
2003                         }
2004                 }
2005         }
2006         up_read(&mdsc->pool_perm_rwsem);
2007         if (*p)
2008                 goto out;
2009
2010         if (pool_ns)
2011                 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2012                       (int)pool_ns->len, pool_ns->str);
2013         else
2014                 doutc(cl, "pool %lld no perm cached\n", pool);
2015
2016         down_write(&mdsc->pool_perm_rwsem);
2017         p = &mdsc->pool_perm_tree.rb_node;
2018         parent = NULL;
2019         while (*p) {
2020                 parent = *p;
2021                 perm = rb_entry(parent, struct ceph_pool_perm, node);
2022                 if (pool < perm->pool)
2023                         p = &(*p)->rb_left;
2024                 else if (pool > perm->pool)
2025                         p = &(*p)->rb_right;
2026                 else {
2027                         int ret = ceph_compare_string(pool_ns,
2028                                                 perm->pool_ns,
2029                                                 perm->pool_ns_len);
2030                         if (ret < 0)
2031                                 p = &(*p)->rb_left;
2032                         else if (ret > 0)
2033                                 p = &(*p)->rb_right;
2034                         else {
2035                                 have = perm->perm;
2036                                 break;
2037                         }
2038                 }
2039         }
2040         if (*p) {
2041                 up_write(&mdsc->pool_perm_rwsem);
2042                 goto out;
2043         }
2044
2045         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2046                                          1, false, GFP_NOFS);
2047         if (!rd_req) {
2048                 err = -ENOMEM;
2049                 goto out_unlock;
2050         }
2051
2052         rd_req->r_flags = CEPH_OSD_FLAG_READ;
2053         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2054         rd_req->r_base_oloc.pool = pool;
2055         if (pool_ns)
2056                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2057         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2058
2059         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2060         if (err)
2061                 goto out_unlock;
2062
2063         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2064                                          1, false, GFP_NOFS);
2065         if (!wr_req) {
2066                 err = -ENOMEM;
2067                 goto out_unlock;
2068         }
2069
2070         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2071         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2072         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2073         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2074
2075         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2076         if (err)
2077                 goto out_unlock;
2078
2079         /* one page should be large enough for STAT data */
2080         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2081         if (IS_ERR(pages)) {
2082                 err = PTR_ERR(pages);
2083                 goto out_unlock;
2084         }
2085
2086         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2087                                      0, false, true);
2088         ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2089
2090         wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2091         ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2092
2093         err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2094         err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2095
2096         if (err >= 0 || err == -ENOENT)
2097                 have |= POOL_READ;
2098         else if (err != -EPERM) {
2099                 if (err == -EBLOCKLISTED)
2100                         fsc->blocklisted = true;
2101                 goto out_unlock;
2102         }
2103
2104         if (err2 == 0 || err2 == -EEXIST)
2105                 have |= POOL_WRITE;
2106         else if (err2 != -EPERM) {
2107                 if (err2 == -EBLOCKLISTED)
2108                         fsc->blocklisted = true;
2109                 err = err2;
2110                 goto out_unlock;
2111         }
2112
2113         pool_ns_len = pool_ns ? pool_ns->len : 0;
2114         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2115         if (!perm) {
2116                 err = -ENOMEM;
2117                 goto out_unlock;
2118         }
2119
2120         perm->pool = pool;
2121         perm->perm = have;
2122         perm->pool_ns_len = pool_ns_len;
2123         if (pool_ns_len > 0)
2124                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2125         perm->pool_ns[pool_ns_len] = 0;
2126
2127         rb_link_node(&perm->node, parent, p);
2128         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2129         err = 0;
2130 out_unlock:
2131         up_write(&mdsc->pool_perm_rwsem);
2132
2133         ceph_osdc_put_request(rd_req);
2134         ceph_osdc_put_request(wr_req);
2135 out:
2136         if (!err)
2137                 err = have;
2138         if (pool_ns)
2139                 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2140                       (int)pool_ns->len, pool_ns->str, err);
2141         else
2142                 doutc(cl, "pool %lld result = %d\n", pool, err);
2143         return err;
2144 }
2145
2146 int ceph_pool_perm_check(struct inode *inode, int need)
2147 {
2148         struct ceph_client *cl = ceph_inode_to_client(inode);
2149         struct ceph_inode_info *ci = ceph_inode(inode);
2150         struct ceph_string *pool_ns;
2151         s64 pool;
2152         int ret, flags;
2153
2154         /* Only need to do this for regular files */
2155         if (!S_ISREG(inode->i_mode))
2156                 return 0;
2157
2158         if (ci->i_vino.snap != CEPH_NOSNAP) {
2159                 /*
2160                  * Pool permission check needs to write to the first object.
2161                  * But for snapshot, head of the first object may have alread
2162                  * been deleted. Skip check to avoid creating orphan object.
2163                  */
2164                 return 0;
2165         }
2166
2167         if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2168                                 NOPOOLPERM))
2169                 return 0;
2170
2171         spin_lock(&ci->i_ceph_lock);
2172         flags = ci->i_ceph_flags;
2173         pool = ci->i_layout.pool_id;
2174         spin_unlock(&ci->i_ceph_lock);
2175 check:
2176         if (flags & CEPH_I_POOL_PERM) {
2177                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2178                         doutc(cl, "pool %lld no read perm\n", pool);
2179                         return -EPERM;
2180                 }
2181                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2182                         doutc(cl, "pool %lld no write perm\n", pool);
2183                         return -EPERM;
2184                 }
2185                 return 0;
2186         }
2187
2188         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2189         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2190         ceph_put_string(pool_ns);
2191         if (ret < 0)
2192                 return ret;
2193
2194         flags = CEPH_I_POOL_PERM;
2195         if (ret & POOL_READ)
2196                 flags |= CEPH_I_POOL_RD;
2197         if (ret & POOL_WRITE)
2198                 flags |= CEPH_I_POOL_WR;
2199
2200         spin_lock(&ci->i_ceph_lock);
2201         if (pool == ci->i_layout.pool_id &&
2202             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2203                 ci->i_ceph_flags |= flags;
2204         } else {
2205                 pool = ci->i_layout.pool_id;
2206                 flags = ci->i_ceph_flags;
2207         }
2208         spin_unlock(&ci->i_ceph_lock);
2209         goto check;
2210 }
2211
2212 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2213 {
2214         struct ceph_pool_perm *perm;
2215         struct rb_node *n;
2216
2217         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2218                 n = rb_first(&mdsc->pool_perm_tree);
2219                 perm = rb_entry(n, struct ceph_pool_perm, node);
2220                 rb_erase(n, &mdsc->pool_perm_tree);
2221                 kfree(perm);
2222         }
2223 }