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