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