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 #define XATTR_CEPH_PREFIX "ceph."
  12 #define XATTR_CEPH_PREFIX_LEN (sizeof (XATTR_CEPH_PREFIX) - 1)
  13
  14 static bool ceph_is_valid_xattr(const char *name)
  15 {
  16         return !strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN) ||
  17                !strncmp(name, XATTR_SECURITY_PREFIX,
  18                         XATTR_SECURITY_PREFIX_LEN) ||
  19                !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
  20                !strncmp(name, XATTR_USER_PREFIX, XATTR_USER_PREFIX_LEN);
  21 }
  22
  23 /*
  24  * These define virtual xattrs exposing the recursive directory
  25  * statistics and layout metadata.
  26  */
  27 struct ceph_vxattr {
  28         char *name;
  29         size_t name_size;       /* strlen(name) + 1 (for '\0') */
  30         size_t (*getxattr_cb)(struct ceph_inode_info *ci, char *val,
  31                               size_t size);
  32         bool readonly, hidden;
  33         bool (*exists_cb)(struct ceph_inode_info *ci);
  34 };
  35
  36 /* directories */
  37
  38 static size_t ceph_vxattrcb_dir_entries(struct ceph_inode_info *ci, char *val,
  39                                         size_t size)
  40 {
  41         return snprintf(val, size, "%lld", ci->i_files + ci->i_subdirs);
  42 }
  43
  44 static size_t ceph_vxattrcb_dir_files(struct ceph_inode_info *ci, char *val,
  45                                       size_t size)
  46 {
  47         return snprintf(val, size, "%lld", ci->i_files);
  48 }
  49
  50 static size_t ceph_vxattrcb_dir_subdirs(struct ceph_inode_info *ci, char *val,
  51                                         size_t size)
  52 {
  53         return snprintf(val, size, "%lld", ci->i_subdirs);
  54 }
  55
  56 static size_t ceph_vxattrcb_dir_rentries(struct ceph_inode_info *ci, char *val,
  57                                          size_t size)
  58 {
  59         return snprintf(val, size, "%lld", ci->i_rfiles + ci->i_rsubdirs);
  60 }
  61
  62 static size_t ceph_vxattrcb_dir_rfiles(struct ceph_inode_info *ci, char *val,
  63                                        size_t size)
  64 {
  65         return snprintf(val, size, "%lld", ci->i_rfiles);
  66 }
  67
  68 static size_t ceph_vxattrcb_dir_rsubdirs(struct ceph_inode_info *ci, char *val,
  69                                          size_t size)
  70 {
  71         return snprintf(val, size, "%lld", ci->i_rsubdirs);
  72 }
  73
  74 static size_t ceph_vxattrcb_dir_rbytes(struct ceph_inode_info *ci, char *val,
  75                                        size_t size)
  76 {
  77         return snprintf(val, size, "%lld", ci->i_rbytes);
  78 }
  79
  80 static size_t ceph_vxattrcb_dir_rctime(struct ceph_inode_info *ci, char *val,
  81                                        size_t size)
  82 {
  83         return snprintf(val, size, "%ld.09%ld", (long)ci->i_rctime.tv_sec,
  84                         (long)ci->i_rctime.tv_nsec);
  85 }
  86
  87 #define CEPH_XATTR_NAME(_type, _name)   XATTR_CEPH_PREFIX #_type "." #_name
  88
  89 #define XATTR_NAME_CEPH(_type, _name)                                   \
  90         {                                                               \
  91                 .name = CEPH_XATTR_NAME(_type, _name),                  \
  92                 .name_size = sizeof (CEPH_XATTR_NAME(_type, _name)), \
  93                 .getxattr_cb = ceph_vxattrcb_ ## _type ## _ ## _name, \
  94                 .readonly = true,                               \
  95                 .hidden = false,                                \
  96                 .exists_cb = NULL,                      \
  97         }
  98
  99 static struct ceph_vxattr ceph_dir_vxattrs[] = {
 100         XATTR_NAME_CEPH(dir, entries),
 101         XATTR_NAME_CEPH(dir, files),
 102         XATTR_NAME_CEPH(dir, subdirs),
 103         XATTR_NAME_CEPH(dir, rentries),
 104         XATTR_NAME_CEPH(dir, rfiles),
 105         XATTR_NAME_CEPH(dir, rsubdirs),
 106         XATTR_NAME_CEPH(dir, rbytes),
 107         XATTR_NAME_CEPH(dir, rctime),
 108         { 0 }   /* Required table terminator */
 109 };
 110 static size_t ceph_dir_vxattrs_name_size;       /* total size of all names */
 111
 112 /* files */
 113
 114 static size_t ceph_vxattrcb_file_layout(struct ceph_inode_info *ci, char *val,
 115                                    size_t size)
 116 {
 117         int ret;
 118
 119         ret = snprintf(val, size,
 120                 "chunk_bytes=%lld\nstripe_count=%lld\nobject_size=%lld\n",
 121                 (unsigned long long)ceph_file_layout_su(ci->i_layout),
 122                 (unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
 123                 (unsigned long long)ceph_file_layout_object_size(ci->i_layout));
 124         return ret;
 125 }
 126
 127 static struct ceph_vxattr ceph_file_vxattrs[] = {
 128         XATTR_NAME_CEPH(file, layout),
 129         { 0 }   /* Required table terminator */
 130 };
 131 static size_t ceph_file_vxattrs_name_size;      /* total size of all names */
 132
 133 static struct ceph_vxattr *ceph_inode_vxattrs(struct inode *inode)
 134 {
 135         if (S_ISDIR(inode->i_mode))
 136                 return ceph_dir_vxattrs;
 137         else if (S_ISREG(inode->i_mode))
 138                 return ceph_file_vxattrs;
 139         return NULL;
 140 }
 141
 142 static size_t ceph_vxattrs_name_size(struct ceph_vxattr *vxattrs)
 143 {
 144         if (vxattrs == ceph_dir_vxattrs)
 145                 return ceph_dir_vxattrs_name_size;
 146         if (vxattrs == ceph_file_vxattrs)
 147                 return ceph_file_vxattrs_name_size;
 148         BUG();
 149
 150         return 0;
 151 }
 152
 153 /*
 154  * Compute the aggregate size (including terminating '\0') of all
 155  * virtual extended attribute names in the given vxattr table.
 156  */
 157 static size_t __init vxattrs_name_size(struct ceph_vxattr *vxattrs)
 158 {
 159         struct ceph_vxattr *vxattr;
 160         size_t size = 0;
 161
 162         for (vxattr = vxattrs; vxattr->name; vxattr++)
 163                 if (!vxattr->hidden)
 164                         size += vxattr->name_size;
 165
 166         return size;
 167 }
 168
 169 /* Routines called at initialization and exit time */
 170
 171 void __init ceph_xattr_init(void)
 172 {
 173         ceph_dir_vxattrs_name_size = vxattrs_name_size(ceph_dir_vxattrs);
 174         ceph_file_vxattrs_name_size = vxattrs_name_size(ceph_file_vxattrs);
 175 }
 176
 177 void ceph_xattr_exit(void)
 178 {
 179         ceph_dir_vxattrs_name_size = 0;
 180         ceph_file_vxattrs_name_size = 0;
 181 }
 182
 183 static struct ceph_vxattr *ceph_match_vxattr(struct inode *inode,
 184                                                 const char *name)
 185 {
 186         struct ceph_vxattr *vxattr = ceph_inode_vxattrs(inode);
 187
 188         if (vxattr) {
 189                 while (vxattr->name) {
 190                         if (!strcmp(vxattr->name, name))
 191                                 return vxattr;
 192                         vxattr++;
 193                 }
 194         }
 195
 196         return NULL;
 197 }
 198
 199 static int __set_xattr(struct ceph_inode_info *ci,
 200                            const char *name, int name_len,
 201                            const char *val, int val_len,
 202                            int dirty,
 203                            int should_free_name, int should_free_val,
 204                            struct ceph_inode_xattr **newxattr)
 205 {
 206         struct rb_node **p;
 207         struct rb_node *parent = NULL;
 208         struct ceph_inode_xattr *xattr = NULL;
 209         int c;
 210         int new = 0;
 211
 212         p = &ci->i_xattrs.index.rb_node;
 213         while (*p) {
 214                 parent = *p;
 215                 xattr = rb_entry(parent, struct ceph_inode_xattr, node);
 216                 c = strncmp(name, xattr->name, min(name_len, xattr->name_len));
 217                 if (c < 0)
 218                         p = &(*p)->rb_left;
 219                 else if (c > 0)
 220                         p = &(*p)->rb_right;
 221                 else {
 222                         if (name_len == xattr->name_len)
 223                                 break;
 224                         else if (name_len < xattr->name_len)
 225                                 p = &(*p)->rb_left;
 226                         else
 227                                 p = &(*p)->rb_right;
 228                 }
 229                 xattr = NULL;
 230         }
 231
 232         if (!xattr) {
 233                 new = 1;
 234                 xattr = *newxattr;
 235                 xattr->name = name;
 236                 xattr->name_len = name_len;
 237                 xattr->should_free_name = should_free_name;
 238
 239                 ci->i_xattrs.count++;
 240                 dout("__set_xattr count=%d\n", ci->i_xattrs.count);
 241         } else {
 242                 kfree(*newxattr);
 243                 *newxattr = NULL;
 244                 if (xattr->should_free_val)
 245                         kfree((void *)xattr->val);
 246
 247                 if (should_free_name) {
 248                         kfree((void *)name);
 249                         name = xattr->name;
 250                 }
 251                 ci->i_xattrs.names_size -= xattr->name_len;
 252                 ci->i_xattrs.vals_size -= xattr->val_len;
 253         }
 254         ci->i_xattrs.names_size += name_len;
 255         ci->i_xattrs.vals_size += val_len;
 256         if (val)
 257                 xattr->val = val;
 258         else
 259                 xattr->val = "";
 260
 261         xattr->val_len = val_len;
 262         xattr->dirty = dirty;
 263         xattr->should_free_val = (val && should_free_val);
 264
 265         if (new) {
 266                 rb_link_node(&xattr->node, parent, p);
 267                 rb_insert_color(&xattr->node, &ci->i_xattrs.index);
 268                 dout("__set_xattr_val p=%p\n", p);
 269         }
 270
 271         dout("__set_xattr_val added %llx.%llx xattr %p %s=%.*s\n",
 272              ceph_vinop(&ci->vfs_inode), xattr, name, val_len, val);
 273
 274         return 0;
 275 }
 276
 277 static struct ceph_inode_xattr *__get_xattr(struct ceph_inode_info *ci,
 278                            const char *name)
 279 {
 280         struct rb_node **p;
 281         struct rb_node *parent = NULL;
 282         struct ceph_inode_xattr *xattr = NULL;
 283         int name_len = strlen(name);
 284         int c;
 285
 286         p = &ci->i_xattrs.index.rb_node;
 287         while (*p) {
 288                 parent = *p;
 289                 xattr = rb_entry(parent, struct ceph_inode_xattr, node);
 290                 c = strncmp(name, xattr->name, xattr->name_len);
 291                 if (c == 0 && name_len > xattr->name_len)
 292                         c = 1;
 293                 if (c < 0)
 294                         p = &(*p)->rb_left;
 295                 else if (c > 0)
 296                         p = &(*p)->rb_right;
 297                 else {
 298                         dout("__get_xattr %s: found %.*s\n", name,
 299                              xattr->val_len, xattr->val);
 300                         return xattr;
 301                 }
 302         }
 303
 304         dout("__get_xattr %s: not found\n", name);
 305
 306         return NULL;
 307 }
 308
 309 static void __free_xattr(struct ceph_inode_xattr *xattr)
 310 {
 311         BUG_ON(!xattr);
 312
 313         if (xattr->should_free_name)
 314                 kfree((void *)xattr->name);
 315         if (xattr->should_free_val)
 316                 kfree((void *)xattr->val);
 317
 318         kfree(xattr);
 319 }
 320
 321 static int __remove_xattr(struct ceph_inode_info *ci,
 322                           struct ceph_inode_xattr *xattr)
 323 {
 324         if (!xattr)
 325                 return -EOPNOTSUPP;
 326
 327         rb_erase(&xattr->node, &ci->i_xattrs.index);
 328
 329         if (xattr->should_free_name)
 330                 kfree((void *)xattr->name);
 331         if (xattr->should_free_val)
 332                 kfree((void *)xattr->val);
 333
 334         ci->i_xattrs.names_size -= xattr->name_len;
 335         ci->i_xattrs.vals_size -= xattr->val_len;
 336         ci->i_xattrs.count--;
 337         kfree(xattr);
 338
 339         return 0;
 340 }
 341
 342 static int __remove_xattr_by_name(struct ceph_inode_info *ci,
 343                            const char *name)
 344 {
 345         struct rb_node **p;
 346         struct ceph_inode_xattr *xattr;
 347         int err;
 348
 349         p = &ci->i_xattrs.index.rb_node;
 350         xattr = __get_xattr(ci, name);
 351         err = __remove_xattr(ci, xattr);
 352         return err;
 353 }
 354
 355 static char *__copy_xattr_names(struct ceph_inode_info *ci,
 356                                 char *dest)
 357 {
 358         struct rb_node *p;
 359         struct ceph_inode_xattr *xattr = NULL;
 360
 361         p = rb_first(&ci->i_xattrs.index);
 362         dout("__copy_xattr_names count=%d\n", ci->i_xattrs.count);
 363
 364         while (p) {
 365                 xattr = rb_entry(p, struct ceph_inode_xattr, node);
 366                 memcpy(dest, xattr->name, xattr->name_len);
 367                 dest[xattr->name_len] = '\0';
 368
 369                 dout("dest=%s %p (%s) (%d/%d)\n", dest, xattr, xattr->name,
 370                      xattr->name_len, ci->i_xattrs.names_size);
 371
 372                 dest += xattr->name_len + 1;
 373                 p = rb_next(p);
 374         }
 375
 376         return dest;
 377 }
 378
 379 void __ceph_destroy_xattrs(struct ceph_inode_info *ci)
 380 {
 381         struct rb_node *p, *tmp;
 382         struct ceph_inode_xattr *xattr = NULL;
 383
 384         p = rb_first(&ci->i_xattrs.index);
 385
 386         dout("__ceph_destroy_xattrs p=%p\n", p);
 387
 388         while (p) {
 389                 xattr = rb_entry(p, struct ceph_inode_xattr, node);
 390                 tmp = p;
 391                 p = rb_next(tmp);
 392                 dout("__ceph_destroy_xattrs next p=%p (%.*s)\n", p,
 393                      xattr->name_len, xattr->name);
 394                 rb_erase(tmp, &ci->i_xattrs.index);
 395
 396                 __free_xattr(xattr);
 397         }
 398
 399         ci->i_xattrs.names_size = 0;
 400         ci->i_xattrs.vals_size = 0;
 401         ci->i_xattrs.index_version = 0;
 402         ci->i_xattrs.count = 0;
 403         ci->i_xattrs.index = RB_ROOT;
 404 }
 405
 406 static int __build_xattrs(struct inode *inode)
 407         __releases(ci->i_ceph_lock)
 408         __acquires(ci->i_ceph_lock)
 409 {
 410         u32 namelen;
 411         u32 numattr = 0;
 412         void *p, *end;
 413         u32 len;
 414         const char *name, *val;
 415         struct ceph_inode_info *ci = ceph_inode(inode);
 416         int xattr_version;
 417         struct ceph_inode_xattr **xattrs = NULL;
 418         int err = 0;
 419         int i;
 420
 421         dout("__build_xattrs() len=%d\n",
 422              ci->i_xattrs.blob ? (int)ci->i_xattrs.blob->vec.iov_len : 0);
 423
 424         if (ci->i_xattrs.index_version >= ci->i_xattrs.version)
 425                 return 0; /* already built */
 426
 427         __ceph_destroy_xattrs(ci);
 428
 429 start:
 430         /* updated internal xattr rb tree */
 431         if (ci->i_xattrs.blob && ci->i_xattrs.blob->vec.iov_len > 4) {
 432                 p = ci->i_xattrs.blob->vec.iov_base;
 433                 end = p + ci->i_xattrs.blob->vec.iov_len;
 434                 ceph_decode_32_safe(&p, end, numattr, bad);
 435                 xattr_version = ci->i_xattrs.version;
 436                 spin_unlock(&ci->i_ceph_lock);
 437
 438                 xattrs = kcalloc(numattr, sizeof(struct ceph_xattr *),
 439                                  GFP_NOFS);
 440                 err = -ENOMEM;
 441                 if (!xattrs)
 442                         goto bad_lock;
 443                 memset(xattrs, 0, numattr*sizeof(struct ceph_xattr *));
 444                 for (i = 0; i < numattr; i++) {
 445                         xattrs[i] = kmalloc(sizeof(struct ceph_inode_xattr),
 446                                             GFP_NOFS);
 447                         if (!xattrs[i])
 448                                 goto bad_lock;
 449                 }
 450
 451                 spin_lock(&ci->i_ceph_lock);
 452                 if (ci->i_xattrs.version != xattr_version) {
 453                         /* lost a race, retry */
 454                         for (i = 0; i < numattr; i++)
 455                                 kfree(xattrs[i]);
 456                         kfree(xattrs);
 457                         xattrs = NULL;
 458                         goto start;
 459                 }
 460                 err = -EIO;
 461                 while (numattr--) {
 462                         ceph_decode_32_safe(&p, end, len, bad);
 463                         namelen = len;
 464                         name = p;
 465                         p += len;
 466                         ceph_decode_32_safe(&p, end, len, bad);
 467                         val = p;
 468                         p += len;
 469
 470                         err = __set_xattr(ci, name, namelen, val, len,
 471                                           0, 0, 0, &xattrs[numattr]);
 472
 473                         if (err < 0)
 474                                 goto bad;
 475                 }
 476                 kfree(xattrs);
 477         }
 478         ci->i_xattrs.index_version = ci->i_xattrs.version;
 479         ci->i_xattrs.dirty = false;
 480
 481         return err;
 482 bad_lock:
 483         spin_lock(&ci->i_ceph_lock);
 484 bad:
 485         if (xattrs) {
 486                 for (i = 0; i < numattr; i++)
 487                         kfree(xattrs[i]);
 488                 kfree(xattrs);
 489         }
 490         ci->i_xattrs.names_size = 0;
 491         return err;
 492 }
 493
 494 static int __get_required_blob_size(struct ceph_inode_info *ci, int name_size,
 495                                     int val_size)
 496 {
 497         /*
 498          * 4 bytes for the length, and additional 4 bytes per each xattr name,
 499          * 4 bytes per each value
 500          */
 501         int size = 4 + ci->i_xattrs.count*(4 + 4) +
 502                              ci->i_xattrs.names_size +
 503                              ci->i_xattrs.vals_size;
 504         dout("__get_required_blob_size c=%d names.size=%d vals.size=%d\n",
 505              ci->i_xattrs.count, ci->i_xattrs.names_size,
 506              ci->i_xattrs.vals_size);
 507
 508         if (name_size)
 509                 size += 4 + 4 + name_size + val_size;
 510
 511         return size;
 512 }
 513
 514 /*
 515  * If there are dirty xattrs, reencode xattrs into the prealloc_blob
 516  * and swap into place.
 517  */
 518 void __ceph_build_xattrs_blob(struct ceph_inode_info *ci)
 519 {
 520         struct rb_node *p;
 521         struct ceph_inode_xattr *xattr = NULL;
 522         void *dest;
 523
 524         dout("__build_xattrs_blob %p\n", &ci->vfs_inode);
 525         if (ci->i_xattrs.dirty) {
 526                 int need = __get_required_blob_size(ci, 0, 0);
 527
 528                 BUG_ON(need > ci->i_xattrs.prealloc_blob->alloc_len);
 529
 530                 p = rb_first(&ci->i_xattrs.index);
 531                 dest = ci->i_xattrs.prealloc_blob->vec.iov_base;
 532
 533                 ceph_encode_32(&dest, ci->i_xattrs.count);
 534                 while (p) {
 535                         xattr = rb_entry(p, struct ceph_inode_xattr, node);
 536
 537                         ceph_encode_32(&dest, xattr->name_len);
 538                         memcpy(dest, xattr->name, xattr->name_len);
 539                         dest += xattr->name_len;
 540                         ceph_encode_32(&dest, xattr->val_len);
 541                         memcpy(dest, xattr->val, xattr->val_len);
 542                         dest += xattr->val_len;
 543
 544                         p = rb_next(p);
 545                 }
 546
 547                 /* adjust buffer len; it may be larger than we need */
 548                 ci->i_xattrs.prealloc_blob->vec.iov_len =
 549                         dest - ci->i_xattrs.prealloc_blob->vec.iov_base;
 550
 551                 if (ci->i_xattrs.blob)
 552                         ceph_buffer_put(ci->i_xattrs.blob);
 553                 ci->i_xattrs.blob = ci->i_xattrs.prealloc_blob;
 554                 ci->i_xattrs.prealloc_blob = NULL;
 555                 ci->i_xattrs.dirty = false;
 556                 ci->i_xattrs.version++;
 557         }
 558 }
 559
 560 ssize_t ceph_getxattr(struct dentry *dentry, const char *name, void *value,
 561                       size_t size)
 562 {
 563         struct inode *inode = dentry->d_inode;
 564         struct ceph_inode_info *ci = ceph_inode(inode);
 565         int err;
 566         struct ceph_inode_xattr *xattr;
 567         struct ceph_vxattr *vxattr = NULL;
 568
 569         if (!ceph_is_valid_xattr(name))
 570                 return -ENODATA;
 571
 572         /* let's see if a virtual xattr was requested */
 573         vxattr = ceph_match_vxattr(inode, name);
 574
 575         spin_lock(&ci->i_ceph_lock);
 576         dout("getxattr %p ver=%lld index_ver=%lld\n", inode,
 577              ci->i_xattrs.version, ci->i_xattrs.index_version);
 578
 579         if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
 580             (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
 581                 goto get_xattr;
 582         } else {
 583                 spin_unlock(&ci->i_ceph_lock);
 584                 /* get xattrs from mds (if we don't already have them) */
 585                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
 586                 if (err)
 587                         return err;
 588         }
 589
 590         spin_lock(&ci->i_ceph_lock);
 591
 592         if (vxattr && vxattr->readonly) {
 593                 err = vxattr->getxattr_cb(ci, value, size);
 594                 goto out;
 595         }
 596
 597         err = __build_xattrs(inode);
 598         if (err < 0)
 599                 goto out;
 600
 601 get_xattr:
 602         err = -ENODATA;  /* == ENOATTR */
 603         xattr = __get_xattr(ci, name);
 604         if (!xattr) {
 605                 if (vxattr)
 606                         err = vxattr->getxattr_cb(ci, value, size);
 607                 goto out;
 608         }
 609
 610         err = -ERANGE;
 611         if (size && size < xattr->val_len)
 612                 goto out;
 613
 614         err = xattr->val_len;
 615         if (size == 0)
 616                 goto out;
 617
 618         memcpy(value, xattr->val, xattr->val_len);
 619
 620 out:
 621         spin_unlock(&ci->i_ceph_lock);
 622         return err;
 623 }
 624
 625 ssize_t ceph_listxattr(struct dentry *dentry, char *names, size_t size)
 626 {
 627         struct inode *inode = dentry->d_inode;
 628         struct ceph_inode_info *ci = ceph_inode(inode);
 629         struct ceph_vxattr *vxattrs = ceph_inode_vxattrs(inode);
 630         u32 vir_namelen = 0;
 631         u32 namelen;
 632         int err;
 633         u32 len;
 634         int i;
 635
 636         spin_lock(&ci->i_ceph_lock);
 637         dout("listxattr %p ver=%lld index_ver=%lld\n", inode,
 638              ci->i_xattrs.version, ci->i_xattrs.index_version);
 639
 640         if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
 641             (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
 642                 goto list_xattr;
 643         } else {
 644                 spin_unlock(&ci->i_ceph_lock);
 645                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
 646                 if (err)
 647                         return err;
 648         }
 649
 650         spin_lock(&ci->i_ceph_lock);
 651
 652         err = __build_xattrs(inode);
 653         if (err < 0)
 654                 goto out;
 655
 656 list_xattr:
 657         /*
 658          * Start with virtual dir xattr names (if any) (including
 659          * terminating '\0' characters for each).
 660          */
 661         vir_namelen = ceph_vxattrs_name_size(vxattrs);
 662
 663         /* adding 1 byte per each variable due to the null termination */
 664         namelen = vir_namelen + ci->i_xattrs.names_size + ci->i_xattrs.count;
 665         err = -ERANGE;
 666         if (size && namelen > size)
 667                 goto out;
 668
 669         err = namelen;
 670         if (size == 0)
 671                 goto out;
 672
 673         names = __copy_xattr_names(ci, names);
 674
 675         /* virtual xattr names, too */
 676         if (vxattrs)
 677                 for (i = 0; vxattrs[i].name; i++) {
 678                         len = sprintf(names, "%s", vxattrs[i].name);
 679                         names += len + 1;
 680                 }
 681
 682 out:
 683         spin_unlock(&ci->i_ceph_lock);
 684         return err;
 685 }
 686
 687 static int ceph_sync_setxattr(struct dentry *dentry, const char *name,
 688                               const char *value, size_t size, int flags)
 689 {
 690         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 691         struct inode *inode = dentry->d_inode;
 692         struct ceph_inode_info *ci = ceph_inode(inode);
 693         struct inode *parent_inode;
 694         struct ceph_mds_request *req;
 695         struct ceph_mds_client *mdsc = fsc->mdsc;
 696         int err;
 697         int i, nr_pages;
 698         struct page **pages = NULL;
 699         void *kaddr;
 700
 701         /* copy value into some pages */
 702         nr_pages = calc_pages_for(0, size);
 703         if (nr_pages) {
 704                 pages = kmalloc(sizeof(pages[0])*nr_pages, GFP_NOFS);
 705                 if (!pages)
 706                         return -ENOMEM;
 707                 err = -ENOMEM;
 708                 for (i = 0; i < nr_pages; i++) {
 709                         pages[i] = __page_cache_alloc(GFP_NOFS);
 710                         if (!pages[i]) {
 711                                 nr_pages = i;
 712                                 goto out;
 713                         }
 714                         kaddr = kmap(pages[i]);
 715                         memcpy(kaddr, value + i*PAGE_CACHE_SIZE,
 716                                min(PAGE_CACHE_SIZE, size-i*PAGE_CACHE_SIZE));
 717                 }
 718         }
 719
 720         dout("setxattr value=%.*s\n", (int)size, value);
 721
 722         /* do request */
 723         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETXATTR,
 724                                        USE_AUTH_MDS);
 725         if (IS_ERR(req)) {
 726                 err = PTR_ERR(req);
 727                 goto out;
 728         }
 729         req->r_inode = inode;
 730         ihold(inode);
 731         req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
 732         req->r_num_caps = 1;
 733         req->r_args.setxattr.flags = cpu_to_le32(flags);
 734         req->r_path2 = kstrdup(name, GFP_NOFS);
 735
 736         req->r_pages = pages;
 737         req->r_num_pages = nr_pages;
 738         req->r_data_len = size;
 739
 740         dout("xattr.ver (before): %lld\n", ci->i_xattrs.version);
 741         parent_inode = ceph_get_dentry_parent_inode(dentry);
 742         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 743         iput(parent_inode);
 744         ceph_mdsc_put_request(req);
 745         dout("xattr.ver (after): %lld\n", ci->i_xattrs.version);
 746
 747 out:
 748         if (pages) {
 749                 for (i = 0; i < nr_pages; i++)
 750                         __free_page(pages[i]);
 751                 kfree(pages);
 752         }
 753         return err;
 754 }
 755
 756 int ceph_setxattr(struct dentry *dentry, const char *name,
 757                   const void *value, size_t size, int flags)
 758 {
 759         struct inode *inode = dentry->d_inode;
 760         struct ceph_vxattr *vxattr;
 761         struct ceph_inode_info *ci = ceph_inode(inode);
 762         int issued;
 763         int err;
 764         int dirty;
 765         int name_len = strlen(name);
 766         int val_len = size;
 767         char *newname = NULL;
 768         char *newval = NULL;
 769         struct ceph_inode_xattr *xattr = NULL;
 770         int required_blob_size;
 771
 772         if (ceph_snap(inode) != CEPH_NOSNAP)
 773                 return -EROFS;
 774
 775         if (!ceph_is_valid_xattr(name))
 776                 return -EOPNOTSUPP;
 777
 778         vxattr = ceph_match_vxattr(inode, name);
 779         if (vxattr && vxattr->readonly)
 780                 return -EOPNOTSUPP;
 781
 782         /* pass any unhandled ceph.* xattrs through to the MDS */
 783         if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
 784                 goto do_sync_unlocked;
 785
 786         /* preallocate memory for xattr name, value, index node */
 787         err = -ENOMEM;
 788         newname = kmemdup(name, name_len + 1, GFP_NOFS);
 789         if (!newname)
 790                 goto out;
 791
 792         if (val_len) {
 793                 newval = kmemdup(value, val_len, GFP_NOFS);
 794                 if (!newval)
 795                         goto out;
 796         }
 797
 798         xattr = kmalloc(sizeof(struct ceph_inode_xattr), GFP_NOFS);
 799         if (!xattr)
 800                 goto out;
 801
 802         spin_lock(&ci->i_ceph_lock);
 803 retry:
 804         issued = __ceph_caps_issued(ci, NULL);
 805         dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
 806         if (!(issued & CEPH_CAP_XATTR_EXCL))
 807                 goto do_sync;
 808         __build_xattrs(inode);
 809
 810         required_blob_size = __get_required_blob_size(ci, name_len, val_len);
 811
 812         if (!ci->i_xattrs.prealloc_blob ||
 813             required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
 814                 struct ceph_buffer *blob;
 815
 816                 spin_unlock(&ci->i_ceph_lock);
 817                 dout(" preaallocating new blob size=%d\n", required_blob_size);
 818                 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
 819                 if (!blob)
 820                         goto out;
 821                 spin_lock(&ci->i_ceph_lock);
 822                 if (ci->i_xattrs.prealloc_blob)
 823                         ceph_buffer_put(ci->i_xattrs.prealloc_blob);
 824                 ci->i_xattrs.prealloc_blob = blob;
 825                 goto retry;
 826         }
 827
 828         err = __set_xattr(ci, newname, name_len, newval,
 829                           val_len, 1, 1, 1, &xattr);
 830
 831         dirty = __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(&ci->i_ceph_lock);
 836         if (dirty)
 837                 __mark_inode_dirty(inode, dirty);
 838         return err;
 839
 840 do_sync:
 841         spin_unlock(&ci->i_ceph_lock);
 842 do_sync_unlocked:
 843         err = ceph_sync_setxattr(dentry, name, value, size, flags);
 844 out:
 845         kfree(newname);
 846         kfree(newval);
 847         kfree(xattr);
 848         return err;
 849 }
 850
 851 static int ceph_send_removexattr(struct dentry *dentry, const char *name)
 852 {
 853         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 854         struct ceph_mds_client *mdsc = fsc->mdsc;
 855         struct inode *inode = dentry->d_inode;
 856         struct inode *parent_inode;
 857         struct ceph_mds_request *req;
 858         int err;
 859
 860         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RMXATTR,
 861                                        USE_AUTH_MDS);
 862         if (IS_ERR(req))
 863                 return PTR_ERR(req);
 864         req->r_inode = inode;
 865         ihold(inode);
 866         req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
 867         req->r_num_caps = 1;
 868         req->r_path2 = kstrdup(name, GFP_NOFS);
 869
 870         parent_inode = ceph_get_dentry_parent_inode(dentry);
 871         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 872         iput(parent_inode);
 873         ceph_mdsc_put_request(req);
 874         return err;
 875 }
 876
 877 int ceph_removexattr(struct dentry *dentry, const char *name)
 878 {
 879         struct inode *inode = dentry->d_inode;
 880         struct ceph_vxattr *vxattr;
 881         struct ceph_inode_info *ci = ceph_inode(inode);
 882         int issued;
 883         int err;
 884         int required_blob_size;
 885         int dirty;
 886
 887         if (ceph_snap(inode) != CEPH_NOSNAP)
 888                 return -EROFS;
 889
 890         if (!ceph_is_valid_xattr(name))
 891                 return -EOPNOTSUPP;
 892
 893         vxattr = ceph_match_vxattr(inode, name);
 894         if (vxattr && vxattr->readonly)
 895                 return -EOPNOTSUPP;
 896
 897         /* pass any unhandled ceph.* xattrs through to the MDS */
 898         if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
 899                 goto do_sync_unlocked;
 900
 901         err = -ENOMEM;
 902         spin_lock(&ci->i_ceph_lock);
 903 retry:
 904         issued = __ceph_caps_issued(ci, NULL);
 905         dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
 906
 907         if (!(issued & CEPH_CAP_XATTR_EXCL))
 908                 goto do_sync;
 909         __build_xattrs(inode);
 910
 911         required_blob_size = __get_required_blob_size(ci, 0, 0);
 912
 913         if (!ci->i_xattrs.prealloc_blob ||
 914             required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
 915                 struct ceph_buffer *blob;
 916
 917                 spin_unlock(&ci->i_ceph_lock);
 918                 dout(" preaallocating new blob size=%d\n", required_blob_size);
 919                 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
 920                 if (!blob)
 921                         goto out;
 922                 spin_lock(&ci->i_ceph_lock);
 923                 if (ci->i_xattrs.prealloc_blob)
 924                         ceph_buffer_put(ci->i_xattrs.prealloc_blob);
 925                 ci->i_xattrs.prealloc_blob = blob;
 926                 goto retry;
 927         }
 928
 929         err = __remove_xattr_by_name(ceph_inode(inode), name);
 930
 931         dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
 932         ci->i_xattrs.dirty = true;
 933         inode->i_ctime = CURRENT_TIME;
 934         spin_unlock(&ci->i_ceph_lock);
 935         if (dirty)
 936                 __mark_inode_dirty(inode, dirty);
 937         return err;
 938 do_sync:
 939         spin_unlock(&ci->i_ceph_lock);
 940 do_sync_unlocked:
 941         err = ceph_send_removexattr(dentry, name);
 942 out:
 943         return err;
 944 }
 945