fs/ceph/xattr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include "super.h"
   4 #include "mds_client.h"
   5
   6 #include <linux/ceph/decode.h>
   7
   8 #include <linux/xattr.h>
   9 #include <linux/slab.h>
  10
  11 static bool ceph_is_valid_xattr(const char *name)
  12 {
  13         return !strncmp(name, "ceph.", 5) ||
  14                !strncmp(name, XATTR_SECURITY_PREFIX,
  15                         XATTR_SECURITY_PREFIX_LEN) ||
  16                !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
  17                !strncmp(name, XATTR_USER_PREFIX, XATTR_USER_PREFIX_LEN);
  18 }
  19
  20 /*
  21  * These define virtual xattrs exposing the recursive directory
  22  * statistics and layout metadata.
  23  */
  24 struct ceph_vxattr_cb {
  25         bool readonly;
  26         char *name;
  27         size_t (*getxattr_cb)(struct ceph_inode_info *ci, char *val,
  28                               size_t size);
  29 };
  30
  31 /* directories */
  32
  33 static size_t ceph_vxattrcb_entries(struct ceph_inode_info *ci, char *val,
  34                                         size_t size)
  35 {
  36         return snprintf(val, size, "%lld", ci->i_files + ci->i_subdirs);
  37 }
  38
  39 static size_t ceph_vxattrcb_files(struct ceph_inode_info *ci, char *val,
  40                                       size_t size)
  41 {
  42         return snprintf(val, size, "%lld", ci->i_files);
  43 }
  44
  45 static size_t ceph_vxattrcb_subdirs(struct ceph_inode_info *ci, char *val,
  46                                         size_t size)
  47 {
  48         return snprintf(val, size, "%lld", ci->i_subdirs);
  49 }
  50
  51 static size_t ceph_vxattrcb_rentries(struct ceph_inode_info *ci, char *val,
  52                                          size_t size)
  53 {
  54         return snprintf(val, size, "%lld", ci->i_rfiles + ci->i_rsubdirs);
  55 }
  56
  57 static size_t ceph_vxattrcb_rfiles(struct ceph_inode_info *ci, char *val,
  58                                        size_t size)
  59 {
  60         return snprintf(val, size, "%lld", ci->i_rfiles);
  61 }
  62
  63 static size_t ceph_vxattrcb_rsubdirs(struct ceph_inode_info *ci, char *val,
  64                                          size_t size)
  65 {
  66         return snprintf(val, size, "%lld", ci->i_rsubdirs);
  67 }
  68
  69 static size_t ceph_vxattrcb_rbytes(struct ceph_inode_info *ci, char *val,
  70                                        size_t size)
  71 {
  72         return snprintf(val, size, "%lld", ci->i_rbytes);
  73 }
  74
  75 static size_t ceph_vxattrcb_rctime(struct ceph_inode_info *ci, char *val,
  76                                        size_t size)
  77 {
  78         return snprintf(val, size, "%ld.%ld", (long)ci->i_rctime.tv_sec,
  79                         (long)ci->i_rctime.tv_nsec);
  80 }
  81
  82 static struct ceph_vxattr_cb ceph_dir_vxattrs[] = {
  83         { true, "ceph.dir.entries", ceph_vxattrcb_entries},
  84         { true, "ceph.dir.files", ceph_vxattrcb_files},
  85         { true, "ceph.dir.subdirs", ceph_vxattrcb_subdirs},
  86         { true, "ceph.dir.rentries", ceph_vxattrcb_rentries},
  87         { true, "ceph.dir.rfiles", ceph_vxattrcb_rfiles},
  88         { true, "ceph.dir.rsubdirs", ceph_vxattrcb_rsubdirs},
  89         { true, "ceph.dir.rbytes", ceph_vxattrcb_rbytes},
  90         { true, "ceph.dir.rctime", ceph_vxattrcb_rctime},
  91         { true, NULL, NULL }
  92 };
  93
  94 /* files */
  95
  96 static size_t ceph_vxattrcb_layout(struct ceph_inode_info *ci, char *val,
  97                                    size_t size)
  98 {
  99         int ret;
 100
 101         ret = snprintf(val, size,
 102                 "chunk_bytes=%lld\nstripe_count=%lld\nobject_size=%lld\n",
 103                 (unsigned long long)ceph_file_layout_su(ci->i_layout),
 104                 (unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
 105                 (unsigned long long)ceph_file_layout_object_size(ci->i_layout));
 106         if (ceph_file_layout_pg_preferred(ci->i_layout))
 107                 ret += snprintf(val + ret, size, "preferred_osd=%lld\n",
 108                             (unsigned long long)ceph_file_layout_pg_preferred(
 109                                     ci->i_layout));
 110         return ret;
 111 }
 112
 113 static struct ceph_vxattr_cb ceph_file_vxattrs[] = {
 114         { true, "ceph.file.layout", ceph_vxattrcb_layout},
 115         /* The following extended attribute name is deprecated */
 116         { true, "ceph.layout", ceph_vxattrcb_layout},
 117         { true, NULL, NULL }
 118 };
 119
 120 static struct ceph_vxattr_cb *ceph_inode_vxattrs(struct inode *inode)
 121 {
 122         if (S_ISDIR(inode->i_mode))
 123                 return ceph_dir_vxattrs;
 124         else if (S_ISREG(inode->i_mode))
 125                 return ceph_file_vxattrs;
 126         return NULL;
 127 }
 128
 129 static struct ceph_vxattr_cb *ceph_match_vxattr(struct ceph_vxattr_cb *vxattr,
 130                                                 const char *name)
 131 {
 132         do {
 133                 if (strcmp(vxattr->name, name) == 0)
 134                         return vxattr;
 135                 vxattr++;
 136         } while (vxattr->name);
 137         return NULL;
 138 }
 139
 140 static int __set_xattr(struct ceph_inode_info *ci,
 141                            const char *name, int name_len,
 142                            const char *val, int val_len,
 143                            int dirty,
 144                            int should_free_name, int should_free_val,
 145                            struct ceph_inode_xattr **newxattr)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct ceph_inode_xattr *xattr = NULL;
 150         int c;
 151         int new = 0;
 152
 153         p = &ci->i_xattrs.index.rb_node;
 154         while (*p) {
 155                 parent = *p;
 156                 xattr = rb_entry(parent, struct ceph_inode_xattr, node);
 157                 c = strncmp(name, xattr->name, min(name_len, xattr->name_len));
 158                 if (c < 0)
 159                         p = &(*p)->rb_left;
 160                 else if (c > 0)
 161                         p = &(*p)->rb_right;
 162                 else {
 163                         if (name_len == xattr->name_len)
 164                                 break;
 165                         else if (name_len < xattr->name_len)
 166                                 p = &(*p)->rb_left;
 167                         else
 168                                 p = &(*p)->rb_right;
 169                 }
 170                 xattr = NULL;
 171         }
 172
 173         if (!xattr) {
 174                 new = 1;
 175                 xattr = *newxattr;
 176                 xattr->name = name;
 177                 xattr->name_len = name_len;
 178                 xattr->should_free_name = should_free_name;
 179
 180                 ci->i_xattrs.count++;
 181                 dout("__set_xattr count=%d\n", ci->i_xattrs.count);
 182         } else {
 183                 kfree(*newxattr);
 184                 *newxattr = NULL;
 185                 if (xattr->should_free_val)
 186                         kfree((void *)xattr->val);
 187
 188                 if (should_free_name) {
 189                         kfree((void *)name);
 190                         name = xattr->name;
 191                 }
 192                 ci->i_xattrs.names_size -= xattr->name_len;
 193                 ci->i_xattrs.vals_size -= xattr->val_len;
 194         }
 195         ci->i_xattrs.names_size += name_len;
 196         ci->i_xattrs.vals_size += val_len;
 197         if (val)
 198                 xattr->val = val;
 199         else
 200                 xattr->val = "";
 201
 202         xattr->val_len = val_len;
 203         xattr->dirty = dirty;
 204         xattr->should_free_val = (val && should_free_val);
 205
 206         if (new) {
 207                 rb_link_node(&xattr->node, parent, p);
 208                 rb_insert_color(&xattr->node, &ci->i_xattrs.index);
 209                 dout("__set_xattr_val p=%p\n", p);
 210         }
 211
 212         dout("__set_xattr_val added %llx.%llx xattr %p %s=%.*s\n",
 213              ceph_vinop(&ci->vfs_inode), xattr, name, val_len, val);
 214
 215         return 0;
 216 }
 217
 218 static struct ceph_inode_xattr *__get_xattr(struct ceph_inode_info *ci,
 219                            const char *name)
 220 {
 221         struct rb_node **p;
 222         struct rb_node *parent = NULL;
 223         struct ceph_inode_xattr *xattr = NULL;
 224         int name_len = strlen(name);
 225         int c;
 226
 227         p = &ci->i_xattrs.index.rb_node;
 228         while (*p) {
 229                 parent = *p;
 230                 xattr = rb_entry(parent, struct ceph_inode_xattr, node);
 231                 c = strncmp(name, xattr->name, xattr->name_len);
 232                 if (c == 0 && name_len > xattr->name_len)
 233                         c = 1;
 234                 if (c < 0)
 235                         p = &(*p)->rb_left;
 236                 else if (c > 0)
 237                         p = &(*p)->rb_right;
 238                 else {
 239                         dout("__get_xattr %s: found %.*s\n", name,
 240                              xattr->val_len, xattr->val);
 241                         return xattr;
 242                 }
 243         }
 244
 245         dout("__get_xattr %s: not found\n", name);
 246
 247         return NULL;
 248 }
 249
 250 static void __free_xattr(struct ceph_inode_xattr *xattr)
 251 {
 252         BUG_ON(!xattr);
 253
 254         if (xattr->should_free_name)
 255                 kfree((void *)xattr->name);
 256         if (xattr->should_free_val)
 257                 kfree((void *)xattr->val);
 258
 259         kfree(xattr);
 260 }
 261
 262 static int __remove_xattr(struct ceph_inode_info *ci,
 263                           struct ceph_inode_xattr *xattr)
 264 {
 265         if (!xattr)
 266                 return -EOPNOTSUPP;
 267
 268         rb_erase(&xattr->node, &ci->i_xattrs.index);
 269
 270         if (xattr->should_free_name)
 271                 kfree((void *)xattr->name);
 272         if (xattr->should_free_val)
 273                 kfree((void *)xattr->val);
 274
 275         ci->i_xattrs.names_size -= xattr->name_len;
 276         ci->i_xattrs.vals_size -= xattr->val_len;
 277         ci->i_xattrs.count--;
 278         kfree(xattr);
 279
 280         return 0;
 281 }
 282
 283 static int __remove_xattr_by_name(struct ceph_inode_info *ci,
 284                            const char *name)
 285 {
 286         struct rb_node **p;
 287         struct ceph_inode_xattr *xattr;
 288         int err;
 289
 290         p = &ci->i_xattrs.index.rb_node;
 291         xattr = __get_xattr(ci, name);
 292         err = __remove_xattr(ci, xattr);
 293         return err;
 294 }
 295
 296 static char *__copy_xattr_names(struct ceph_inode_info *ci,
 297                                 char *dest)
 298 {
 299         struct rb_node *p;
 300         struct ceph_inode_xattr *xattr = NULL;
 301
 302         p = rb_first(&ci->i_xattrs.index);
 303         dout("__copy_xattr_names count=%d\n", ci->i_xattrs.count);
 304
 305         while (p) {
 306                 xattr = rb_entry(p, struct ceph_inode_xattr, node);
 307                 memcpy(dest, xattr->name, xattr->name_len);
 308                 dest[xattr->name_len] = '\0';
 309
 310                 dout("dest=%s %p (%s) (%d/%d)\n", dest, xattr, xattr->name,
 311                      xattr->name_len, ci->i_xattrs.names_size);
 312
 313                 dest += xattr->name_len + 1;
 314                 p = rb_next(p);
 315         }
 316
 317         return dest;
 318 }
 319
 320 void __ceph_destroy_xattrs(struct ceph_inode_info *ci)
 321 {
 322         struct rb_node *p, *tmp;
 323         struct ceph_inode_xattr *xattr = NULL;
 324
 325         p = rb_first(&ci->i_xattrs.index);
 326
 327         dout("__ceph_destroy_xattrs p=%p\n", p);
 328
 329         while (p) {
 330                 xattr = rb_entry(p, struct ceph_inode_xattr, node);
 331                 tmp = p;
 332                 p = rb_next(tmp);
 333                 dout("__ceph_destroy_xattrs next p=%p (%.*s)\n", p,
 334                      xattr->name_len, xattr->name);
 335                 rb_erase(tmp, &ci->i_xattrs.index);
 336
 337                 __free_xattr(xattr);
 338         }
 339
 340         ci->i_xattrs.names_size = 0;
 341         ci->i_xattrs.vals_size = 0;
 342         ci->i_xattrs.index_version = 0;
 343         ci->i_xattrs.count = 0;
 344         ci->i_xattrs.index = RB_ROOT;
 345 }
 346
 347 static int __build_xattrs(struct inode *inode)
 348         __releases(ci->i_ceph_lock)
 349         __acquires(ci->i_ceph_lock)
 350 {
 351         u32 namelen;
 352         u32 numattr = 0;
 353         void *p, *end;
 354         u32 len;
 355         const char *name, *val;
 356         struct ceph_inode_info *ci = ceph_inode(inode);
 357         int xattr_version;
 358         struct ceph_inode_xattr **xattrs = NULL;
 359         int err = 0;
 360         int i;
 361
 362         dout("__build_xattrs() len=%d\n",
 363              ci->i_xattrs.blob ? (int)ci->i_xattrs.blob->vec.iov_len : 0);
 364
 365         if (ci->i_xattrs.index_version >= ci->i_xattrs.version)
 366                 return 0; /* already built */
 367
 368         __ceph_destroy_xattrs(ci);
 369
 370 start:
 371         /* updated internal xattr rb tree */
 372         if (ci->i_xattrs.blob && ci->i_xattrs.blob->vec.iov_len > 4) {
 373                 p = ci->i_xattrs.blob->vec.iov_base;
 374                 end = p + ci->i_xattrs.blob->vec.iov_len;
 375                 ceph_decode_32_safe(&p, end, numattr, bad);
 376                 xattr_version = ci->i_xattrs.version;
 377                 spin_unlock(&ci->i_ceph_lock);
 378
 379                 xattrs = kcalloc(numattr, sizeof(struct ceph_xattr *),
 380                                  GFP_NOFS);
 381                 err = -ENOMEM;
 382                 if (!xattrs)
 383                         goto bad_lock;
 384                 memset(xattrs, 0, numattr*sizeof(struct ceph_xattr *));
 385                 for (i = 0; i < numattr; i++) {
 386                         xattrs[i] = kmalloc(sizeof(struct ceph_inode_xattr),
 387                                             GFP_NOFS);
 388                         if (!xattrs[i])
 389                                 goto bad_lock;
 390                 }
 391
 392                 spin_lock(&ci->i_ceph_lock);
 393                 if (ci->i_xattrs.version != xattr_version) {
 394                         /* lost a race, retry */
 395                         for (i = 0; i < numattr; i++)
 396                                 kfree(xattrs[i]);
 397                         kfree(xattrs);
 398                         goto start;
 399                 }
 400                 err = -EIO;
 401                 while (numattr--) {
 402                         ceph_decode_32_safe(&p, end, len, bad);
 403                         namelen = len;
 404                         name = p;
 405                         p += len;
 406                         ceph_decode_32_safe(&p, end, len, bad);
 407                         val = p;
 408                         p += len;
 409
 410                         err = __set_xattr(ci, name, namelen, val, len,
 411                                           0, 0, 0, &xattrs[numattr]);
 412
 413                         if (err < 0)
 414                                 goto bad;
 415                 }
 416                 kfree(xattrs);
 417         }
 418         ci->i_xattrs.index_version = ci->i_xattrs.version;
 419         ci->i_xattrs.dirty = false;
 420
 421         return err;
 422 bad_lock:
 423         spin_lock(&ci->i_ceph_lock);
 424 bad:
 425         if (xattrs) {
 426                 for (i = 0; i < numattr; i++)
 427                         kfree(xattrs[i]);
 428                 kfree(xattrs);
 429         }
 430         ci->i_xattrs.names_size = 0;
 431         return err;
 432 }
 433
 434 static int __get_required_blob_size(struct ceph_inode_info *ci, int name_size,
 435                                     int val_size)
 436 {
 437         /*
 438          * 4 bytes for the length, and additional 4 bytes per each xattr name,
 439          * 4 bytes per each value
 440          */
 441         int size = 4 + ci->i_xattrs.count*(4 + 4) +
 442                              ci->i_xattrs.names_size +
 443                              ci->i_xattrs.vals_size;
 444         dout("__get_required_blob_size c=%d names.size=%d vals.size=%d\n",
 445              ci->i_xattrs.count, ci->i_xattrs.names_size,
 446              ci->i_xattrs.vals_size);
 447
 448         if (name_size)
 449                 size += 4 + 4 + name_size + val_size;
 450
 451         return size;
 452 }
 453
 454 /*
 455  * If there are dirty xattrs, reencode xattrs into the prealloc_blob
 456  * and swap into place.
 457  */
 458 void __ceph_build_xattrs_blob(struct ceph_inode_info *ci)
 459 {
 460         struct rb_node *p;
 461         struct ceph_inode_xattr *xattr = NULL;
 462         void *dest;
 463
 464         dout("__build_xattrs_blob %p\n", &ci->vfs_inode);
 465         if (ci->i_xattrs.dirty) {
 466                 int need = __get_required_blob_size(ci, 0, 0);
 467
 468                 BUG_ON(need > ci->i_xattrs.prealloc_blob->alloc_len);
 469
 470                 p = rb_first(&ci->i_xattrs.index);
 471                 dest = ci->i_xattrs.prealloc_blob->vec.iov_base;
 472
 473                 ceph_encode_32(&dest, ci->i_xattrs.count);
 474                 while (p) {
 475                         xattr = rb_entry(p, struct ceph_inode_xattr, node);
 476
 477                         ceph_encode_32(&dest, xattr->name_len);
 478                         memcpy(dest, xattr->name, xattr->name_len);
 479                         dest += xattr->name_len;
 480                         ceph_encode_32(&dest, xattr->val_len);
 481                         memcpy(dest, xattr->val, xattr->val_len);
 482                         dest += xattr->val_len;
 483
 484                         p = rb_next(p);
 485                 }
 486
 487                 /* adjust buffer len; it may be larger than we need */
 488                 ci->i_xattrs.prealloc_blob->vec.iov_len =
 489                         dest - ci->i_xattrs.prealloc_blob->vec.iov_base;
 490
 491                 if (ci->i_xattrs.blob)
 492                         ceph_buffer_put(ci->i_xattrs.blob);
 493                 ci->i_xattrs.blob = ci->i_xattrs.prealloc_blob;
 494                 ci->i_xattrs.prealloc_blob = NULL;
 495                 ci->i_xattrs.dirty = false;
 496                 ci->i_xattrs.version++;
 497         }
 498 }
 499
 500 ssize_t ceph_getxattr(struct dentry *dentry, const char *name, void *value,
 501                       size_t size)
 502 {
 503         struct inode *inode = dentry->d_inode;
 504         struct ceph_inode_info *ci = ceph_inode(inode);
 505         struct ceph_vxattr_cb *vxattrs = ceph_inode_vxattrs(inode);
 506         int err;
 507         struct ceph_inode_xattr *xattr;
 508         struct ceph_vxattr_cb *vxattr = NULL;
 509
 510         if (!ceph_is_valid_xattr(name))
 511                 return -ENODATA;
 512
 513         /* let's see if a virtual xattr was requested */
 514         if (vxattrs)
 515                 vxattr = ceph_match_vxattr(vxattrs, name);
 516
 517         spin_lock(&ci->i_ceph_lock);
 518         dout("getxattr %p ver=%lld index_ver=%lld\n", inode,
 519              ci->i_xattrs.version, ci->i_xattrs.index_version);
 520
 521         if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
 522             (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
 523                 goto get_xattr;
 524         } else {
 525                 spin_unlock(&ci->i_ceph_lock);
 526                 /* get xattrs from mds (if we don't already have them) */
 527                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
 528                 if (err)
 529                         return err;
 530         }
 531
 532         spin_lock(&ci->i_ceph_lock);
 533
 534         if (vxattr && vxattr->readonly) {
 535                 err = vxattr->getxattr_cb(ci, value, size);
 536                 goto out;
 537         }
 538
 539         err = __build_xattrs(inode);
 540         if (err < 0)
 541                 goto out;
 542
 543 get_xattr:
 544         err = -ENODATA;  /* == ENOATTR */
 545         xattr = __get_xattr(ci, name);
 546         if (!xattr) {
 547                 if (vxattr)
 548                         err = vxattr->getxattr_cb(ci, value, size);
 549                 goto out;
 550         }
 551
 552         err = -ERANGE;
 553         if (size && size < xattr->val_len)
 554                 goto out;
 555
 556         err = xattr->val_len;
 557         if (size == 0)
 558                 goto out;
 559
 560         memcpy(value, xattr->val, xattr->val_len);
 561
 562 out:
 563         spin_unlock(&ci->i_ceph_lock);
 564         return err;
 565 }
 566
 567 ssize_t ceph_listxattr(struct dentry *dentry, char *names, size_t size)
 568 {
 569         struct inode *inode = dentry->d_inode;
 570         struct ceph_inode_info *ci = ceph_inode(inode);
 571         struct ceph_vxattr_cb *vxattrs = ceph_inode_vxattrs(inode);
 572         u32 vir_namelen = 0;
 573         u32 namelen;
 574         int err;
 575         u32 len;
 576         int i;
 577
 578         spin_lock(&ci->i_ceph_lock);
 579         dout("listxattr %p ver=%lld index_ver=%lld\n", inode,
 580              ci->i_xattrs.version, ci->i_xattrs.index_version);
 581
 582         if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
 583             (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
 584                 goto list_xattr;
 585         } else {
 586                 spin_unlock(&ci->i_ceph_lock);
 587                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
 588                 if (err)
 589                         return err;
 590         }
 591
 592         spin_lock(&ci->i_ceph_lock);
 593
 594         err = __build_xattrs(inode);
 595         if (err < 0)
 596                 goto out;
 597
 598 list_xattr:
 599         vir_namelen = 0;
 600         /* include virtual dir xattrs */
 601         if (vxattrs)
 602                 for (i = 0; vxattrs[i].name; i++)
 603                         vir_namelen += strlen(vxattrs[i].name) + 1;
 604         /* adding 1 byte per each variable due to the null termination */
 605         namelen = vir_namelen + ci->i_xattrs.names_size + ci->i_xattrs.count;
 606         err = -ERANGE;
 607         if (size && namelen > size)
 608                 goto out;
 609
 610         err = namelen;
 611         if (size == 0)
 612                 goto out;
 613
 614         names = __copy_xattr_names(ci, names);
 615
 616         /* virtual xattr names, too */
 617         if (vxattrs)
 618                 for (i = 0; vxattrs[i].name; i++) {
 619                         len = sprintf(names, "%s", vxattrs[i].name);
 620                         names += len + 1;
 621                 }
 622
 623 out:
 624         spin_unlock(&ci->i_ceph_lock);
 625         return err;
 626 }
 627
 628 static int ceph_sync_setxattr(struct dentry *dentry, const char *name,
 629                               const char *value, size_t size, int flags)
 630 {
 631         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 632         struct inode *inode = dentry->d_inode;
 633         struct ceph_inode_info *ci = ceph_inode(inode);
 634         struct inode *parent_inode;
 635         struct ceph_mds_request *req;
 636         struct ceph_mds_client *mdsc = fsc->mdsc;
 637         int err;
 638         int i, nr_pages;
 639         struct page **pages = NULL;
 640         void *kaddr;
 641
 642         /* copy value into some pages */
 643         nr_pages = calc_pages_for(0, size);
 644         if (nr_pages) {
 645                 pages = kmalloc(sizeof(pages[0])*nr_pages, GFP_NOFS);
 646                 if (!pages)
 647                         return -ENOMEM;
 648                 err = -ENOMEM;
 649                 for (i = 0; i < nr_pages; i++) {
 650                         pages[i] = __page_cache_alloc(GFP_NOFS);
 651                         if (!pages[i]) {
 652                                 nr_pages = i;
 653                                 goto out;
 654                         }
 655                         kaddr = kmap(pages[i]);
 656                         memcpy(kaddr, value + i*PAGE_CACHE_SIZE,
 657                                min(PAGE_CACHE_SIZE, size-i*PAGE_CACHE_SIZE));
 658                 }
 659         }
 660
 661         dout("setxattr value=%.*s\n", (int)size, value);
 662
 663         /* do request */
 664         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETXATTR,
 665                                        USE_AUTH_MDS);
 666         if (IS_ERR(req)) {
 667                 err = PTR_ERR(req);
 668                 goto out;
 669         }
 670         req->r_inode = inode;
 671         ihold(inode);
 672         req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
 673         req->r_num_caps = 1;
 674         req->r_args.setxattr.flags = cpu_to_le32(flags);
 675         req->r_path2 = kstrdup(name, GFP_NOFS);
 676
 677         req->r_pages = pages;
 678         req->r_num_pages = nr_pages;
 679         req->r_data_len = size;
 680
 681         dout("xattr.ver (before): %lld\n", ci->i_xattrs.version);
 682         parent_inode = ceph_get_dentry_parent_inode(dentry);
 683         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 684         iput(parent_inode);
 685         ceph_mdsc_put_request(req);
 686         dout("xattr.ver (after): %lld\n", ci->i_xattrs.version);
 687
 688 out:
 689         if (pages) {
 690                 for (i = 0; i < nr_pages; i++)
 691                         __free_page(pages[i]);
 692                 kfree(pages);
 693         }
 694         return err;
 695 }
 696
 697 int ceph_setxattr(struct dentry *dentry, const char *name,
 698                   const void *value, size_t size, int flags)
 699 {
 700         struct inode *inode = dentry->d_inode;
 701         struct ceph_inode_info *ci = ceph_inode(inode);
 702         struct ceph_vxattr_cb *vxattrs = ceph_inode_vxattrs(inode);
 703         int err;
 704         int name_len = strlen(name);
 705         int val_len = size;
 706         char *newname = NULL;
 707         char *newval = NULL;
 708         struct ceph_inode_xattr *xattr = NULL;
 709         int issued;
 710         int required_blob_size;
 711         int dirty;
 712
 713         if (ceph_snap(inode) != CEPH_NOSNAP)
 714                 return -EROFS;
 715
 716         if (!ceph_is_valid_xattr(name))
 717                 return -EOPNOTSUPP;
 718
 719         if (vxattrs) {
 720                 struct ceph_vxattr_cb *vxattr =
 721                         ceph_match_vxattr(vxattrs, name);
 722                 if (vxattr && vxattr->readonly)
 723                         return -EOPNOTSUPP;
 724         }
 725
 726         /* preallocate memory for xattr name, value, index node */
 727         err = -ENOMEM;
 728         newname = kmemdup(name, name_len + 1, GFP_NOFS);
 729         if (!newname)
 730                 goto out;
 731
 732         if (val_len) {
 733                 newval = kmalloc(val_len + 1, GFP_NOFS);
 734                 if (!newval)
 735                         goto out;
 736                 memcpy(newval, value, val_len);
 737                 newval[val_len] = '\0';
 738         }
 739
 740         xattr = kmalloc(sizeof(struct ceph_inode_xattr), GFP_NOFS);
 741         if (!xattr)
 742                 goto out;
 743
 744         spin_lock(&ci->i_ceph_lock);
 745 retry:
 746         issued = __ceph_caps_issued(ci, NULL);
 747         if (!(issued & CEPH_CAP_XATTR_EXCL))
 748                 goto do_sync;
 749         __build_xattrs(inode);
 750
 751         required_blob_size = __get_required_blob_size(ci, name_len, val_len);
 752
 753         if (!ci->i_xattrs.prealloc_blob ||
 754             required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
 755                 struct ceph_buffer *blob = NULL;
 756
 757                 spin_unlock(&ci->i_ceph_lock);
 758                 dout(" preaallocating new blob size=%d\n", required_blob_size);
 759                 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
 760                 if (!blob)
 761                         goto out;
 762                 spin_lock(&ci->i_ceph_lock);
 763                 if (ci->i_xattrs.prealloc_blob)
 764                         ceph_buffer_put(ci->i_xattrs.prealloc_blob);
 765                 ci->i_xattrs.prealloc_blob = blob;
 766                 goto retry;
 767         }
 768
 769         dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
 770         err = __set_xattr(ci, newname, name_len, newval,
 771                           val_len, 1, 1, 1, &xattr);
 772         dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
 773         ci->i_xattrs.dirty = true;
 774         inode->i_ctime = CURRENT_TIME;
 775         spin_unlock(&ci->i_ceph_lock);
 776         if (dirty)
 777                 __mark_inode_dirty(inode, dirty);
 778         return err;
 779
 780 do_sync:
 781         spin_unlock(&ci->i_ceph_lock);
 782         err = ceph_sync_setxattr(dentry, name, value, size, flags);
 783 out:
 784         kfree(newname);
 785         kfree(newval);
 786         kfree(xattr);
 787         return err;
 788 }
 789
 790 static int ceph_send_removexattr(struct dentry *dentry, const char *name)
 791 {
 792         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 793         struct ceph_mds_client *mdsc = fsc->mdsc;
 794         struct inode *inode = dentry->d_inode;
 795         struct inode *parent_inode;
 796         struct ceph_mds_request *req;
 797         int err;
 798
 799         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RMXATTR,
 800                                        USE_AUTH_MDS);
 801         if (IS_ERR(req))
 802                 return PTR_ERR(req);
 803         req->r_inode = inode;
 804         ihold(inode);
 805         req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
 806         req->r_num_caps = 1;
 807         req->r_path2 = kstrdup(name, GFP_NOFS);
 808
 809         parent_inode = ceph_get_dentry_parent_inode(dentry);
 810         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 811         iput(parent_inode);
 812         ceph_mdsc_put_request(req);
 813         return err;
 814 }
 815
 816 int ceph_removexattr(struct dentry *dentry, const char *name)
 817 {
 818         struct inode *inode = dentry->d_inode;
 819         struct ceph_inode_info *ci = ceph_inode(inode);
 820         struct ceph_vxattr_cb *vxattrs = ceph_inode_vxattrs(inode);
 821         int issued;
 822         int err;
 823         int required_blob_size;
 824         int dirty;
 825
 826         if (ceph_snap(inode) != CEPH_NOSNAP)
 827                 return -EROFS;
 828
 829         if (!ceph_is_valid_xattr(name))
 830                 return -EOPNOTSUPP;
 831
 832         if (vxattrs) {
 833                 struct ceph_vxattr_cb *vxattr =
 834                         ceph_match_vxattr(vxattrs, name);
 835                 if (vxattr && vxattr->readonly)
 836                         return -EOPNOTSUPP;
 837         }
 838
 839         err = -ENOMEM;
 840         spin_lock(&ci->i_ceph_lock);
 841         __build_xattrs(inode);
 842 retry:
 843         issued = __ceph_caps_issued(ci, NULL);
 844         dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
 845
 846         if (!(issued & CEPH_CAP_XATTR_EXCL))
 847                 goto do_sync;
 848
 849         required_blob_size = __get_required_blob_size(ci, 0, 0);
 850
 851         if (!ci->i_xattrs.prealloc_blob ||
 852             required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
 853                 struct ceph_buffer *blob;
 854
 855                 spin_unlock(&ci->i_ceph_lock);
 856                 dout(" preaallocating new blob size=%d\n", required_blob_size);
 857                 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
 858                 if (!blob)
 859                         goto out;
 860                 spin_lock(&ci->i_ceph_lock);
 861                 if (ci->i_xattrs.prealloc_blob)
 862                         ceph_buffer_put(ci->i_xattrs.prealloc_blob);
 863                 ci->i_xattrs.prealloc_blob = blob;
 864                 goto retry;
 865         }
 866
 867         err = __remove_xattr_by_name(ceph_inode(inode), name);
 868         dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
 869         ci->i_xattrs.dirty = true;
 870         inode->i_ctime = CURRENT_TIME;
 871
 872         spin_unlock(&ci->i_ceph_lock);
 873         if (dirty)
 874                 __mark_inode_dirty(inode, dirty);
 875         return err;
 876 do_sync:
 877         spin_unlock(&ci->i_ceph_lock);
 878         err = ceph_send_removexattr(dentry, name);
 879 out:
 880         return err;
 881 }
 882