fs/ceph/xattr.c

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