drivers/staging/lustre/lustre/llite/xattr_cache.c

   1 /*
   2  * Copyright 2012 Xyratex Technology Limited
   3  *
   4  * Author: Andrew Perepechko <Andrew_Perepechko@xyratex.com>
   5  *
   6  */
   7
   8 #define DEBUG_SUBSYSTEM S_LLITE
   9
  10 #include <linux/fs.h>
  11 #include <linux/sched.h>
  12 #include <linux/mm.h>
  13 #include "../include/obd_support.h"
  14 #include "../include/lustre_lite.h"
  15 #include "../include/lustre_dlm.h"
  16 #include "../include/lustre_ver.h"
  17 #include "llite_internal.h"
  18
  19 /* If we ever have hundreds of extended attributes, we might want to consider
  20  * using a hash or a tree structure instead of list for faster lookups.
  21  */
  22 struct ll_xattr_entry {
  23         struct list_head        xe_list;    /* protected with
  24                                              * lli_xattrs_list_rwsem */
  25         char                    *xe_name;   /* xattr name, \0-terminated */
  26         char                    *xe_value;  /* xattr value */
  27         unsigned                xe_namelen; /* strlen(xe_name) + 1 */
  28         unsigned                xe_vallen;  /* xattr value length */
  29 };
  30
  31 static struct kmem_cache *xattr_kmem;
  32 static struct lu_kmem_descr xattr_caches[] = {
  33         {
  34                 .ckd_cache = &xattr_kmem,
  35                 .ckd_name  = "xattr_kmem",
  36                 .ckd_size  = sizeof(struct ll_xattr_entry)
  37         },
  38         {
  39                 .ckd_cache = NULL
  40         }
  41 };
  42
  43 int ll_xattr_init(void)
  44 {
  45         return lu_kmem_init(xattr_caches);
  46 }
  47
  48 void ll_xattr_fini(void)
  49 {
  50         lu_kmem_fini(xattr_caches);
  51 }
  52
  53 /**
  54  * Initializes xattr cache for an inode.
  55  *
  56  * This initializes the xattr list and marks cache presence.
  57  */
  58 static void ll_xattr_cache_init(struct ll_inode_info *lli)
  59 {
  60
  61
  62         LASSERT(lli != NULL);
  63
  64         INIT_LIST_HEAD(&lli->lli_xattrs);
  65         lli->lli_flags |= LLIF_XATTR_CACHE;
  66 }
  67
  68 /**
  69  *  This looks for a specific extended attribute.
  70  *
  71  *  Find in @cache and return @xattr_name attribute in @xattr,
  72  *  for the NULL @xattr_name return the first cached @xattr.
  73  *
  74  *  \retval 0        success
  75  *  \retval -ENODATA if not found
  76  */
  77 static int ll_xattr_cache_find(struct list_head *cache,
  78                                const char *xattr_name,
  79                                struct ll_xattr_entry **xattr)
  80 {
  81         struct ll_xattr_entry *entry;
  82
  83
  84
  85         list_for_each_entry(entry, cache, xe_list) {
  86                 /* xattr_name == NULL means look for any entry */
  87                 if (xattr_name == NULL ||
  88                     strcmp(xattr_name, entry->xe_name) == 0) {
  89                         *xattr = entry;
  90                         CDEBUG(D_CACHE, "find: [%s]=%.*s\n",
  91                                entry->xe_name, entry->xe_vallen,
  92                                entry->xe_value);
  93                         return 0;
  94                 }
  95         }
  96
  97         return -ENODATA;
  98 }
  99
 100 /**
 101  * This adds an xattr.
 102  *
 103  * Add @xattr_name attr with @xattr_val value and @xattr_val_len length,
 104  *
 105  * \retval 0       success
 106  * \retval -ENOMEM if no memory could be allocated for the cached attr
 107  * \retval -EPROTO if duplicate xattr is being added
 108  */
 109 static int ll_xattr_cache_add(struct list_head *cache,
 110                               const char *xattr_name,
 111                               const char *xattr_val,
 112                               unsigned xattr_val_len)
 113 {
 114         struct ll_xattr_entry *xattr;
 115
 116
 117
 118         if (ll_xattr_cache_find(cache, xattr_name, &xattr) == 0) {
 119                 CDEBUG(D_CACHE, "duplicate xattr: [%s]\n", xattr_name);
 120                 return -EPROTO;
 121         }
 122
 123         OBD_SLAB_ALLOC_PTR_GFP(xattr, xattr_kmem, GFP_NOFS);
 124         if (xattr == NULL) {
 125                 CDEBUG(D_CACHE, "failed to allocate xattr\n");
 126                 return -ENOMEM;
 127         }
 128
 129         xattr->xe_namelen = strlen(xattr_name) + 1;
 130
 131         OBD_ALLOC(xattr->xe_name, xattr->xe_namelen);
 132         if (!xattr->xe_name) {
 133                 CDEBUG(D_CACHE, "failed to alloc xattr name %u\n",
 134                        xattr->xe_namelen);
 135                 goto err_name;
 136         }
 137         OBD_ALLOC(xattr->xe_value, xattr_val_len);
 138         if (!xattr->xe_value) {
 139                 CDEBUG(D_CACHE, "failed to alloc xattr value %d\n",
 140                        xattr_val_len);
 141                 goto err_value;
 142         }
 143
 144         memcpy(xattr->xe_name, xattr_name, xattr->xe_namelen);
 145         memcpy(xattr->xe_value, xattr_val, xattr_val_len);
 146         xattr->xe_vallen = xattr_val_len;
 147         list_add(&xattr->xe_list, cache);
 148
 149         CDEBUG(D_CACHE, "set: [%s]=%.*s\n", xattr_name,
 150                 xattr_val_len, xattr_val);
 151
 152         return 0;
 153 err_value:
 154         OBD_FREE(xattr->xe_name, xattr->xe_namelen);
 155 err_name:
 156         OBD_SLAB_FREE_PTR(xattr, xattr_kmem);
 157
 158         return -ENOMEM;
 159 }
 160
 161 /**
 162  * This removes an extended attribute from cache.
 163  *
 164  * Remove @xattr_name attribute from @cache.
 165  *
 166  * \retval 0        success
 167  * \retval -ENODATA if @xattr_name is not cached
 168  */
 169 static int ll_xattr_cache_del(struct list_head *cache,
 170                               const char *xattr_name)
 171 {
 172         struct ll_xattr_entry *xattr;
 173
 174
 175
 176         CDEBUG(D_CACHE, "del xattr: %s\n", xattr_name);
 177
 178         if (ll_xattr_cache_find(cache, xattr_name, &xattr) == 0) {
 179                 list_del(&xattr->xe_list);
 180                 OBD_FREE(xattr->xe_name, xattr->xe_namelen);
 181                 OBD_FREE(xattr->xe_value, xattr->xe_vallen);
 182                 OBD_SLAB_FREE_PTR(xattr, xattr_kmem);
 183
 184                 return 0;
 185         }
 186
 187         return -ENODATA;
 188 }
 189
 190 /**
 191  * This iterates cached extended attributes.
 192  *
 193  * Walk over cached attributes in @cache and
 194  * fill in @xld_buffer or only calculate buffer
 195  * size if @xld_buffer is NULL.
 196  *
 197  * \retval >= 0     buffer list size
 198  * \retval -ENODATA if the list cannot fit @xld_size buffer
 199  */
 200 static int ll_xattr_cache_list(struct list_head *cache,
 201                                char *xld_buffer,
 202                                int xld_size)
 203 {
 204         struct ll_xattr_entry *xattr, *tmp;
 205         int xld_tail = 0;
 206
 207
 208
 209         list_for_each_entry_safe(xattr, tmp, cache, xe_list) {
 210                 CDEBUG(D_CACHE, "list: buffer=%p[%d] name=%s\n",
 211                         xld_buffer, xld_tail, xattr->xe_name);
 212
 213                 if (xld_buffer) {
 214                         xld_size -= xattr->xe_namelen;
 215                         if (xld_size < 0)
 216                                 break;
 217                         memcpy(&xld_buffer[xld_tail],
 218                                xattr->xe_name, xattr->xe_namelen);
 219                 }
 220                 xld_tail += xattr->xe_namelen;
 221         }
 222
 223         if (xld_size < 0)
 224                 return -ERANGE;
 225
 226         return xld_tail;
 227 }
 228
 229 /**
 230  * Check if the xattr cache is initialized (filled).
 231  *
 232  * \retval 0 @cache is not initialized
 233  * \retval 1 @cache is initialized
 234  */
 235 static int ll_xattr_cache_valid(struct ll_inode_info *lli)
 236 {
 237         return !!(lli->lli_flags & LLIF_XATTR_CACHE);
 238 }
 239
 240 /**
 241  * This finalizes the xattr cache.
 242  *
 243  * Free all xattr memory. @lli is the inode info pointer.
 244  *
 245  * \retval 0 no error occurred
 246  */
 247 static int ll_xattr_cache_destroy_locked(struct ll_inode_info *lli)
 248 {
 249
 250
 251         if (!ll_xattr_cache_valid(lli))
 252                 return 0;
 253
 254         while (ll_xattr_cache_del(&lli->lli_xattrs, NULL) == 0)
 255                 ; /* empty loop */
 256         lli->lli_flags &= ~LLIF_XATTR_CACHE;
 257
 258         return 0;
 259 }
 260
 261 int ll_xattr_cache_destroy(struct inode *inode)
 262 {
 263         struct ll_inode_info *lli = ll_i2info(inode);
 264         int rc;
 265
 266
 267
 268         down_write(&lli->lli_xattrs_list_rwsem);
 269         rc = ll_xattr_cache_destroy_locked(lli);
 270         up_write(&lli->lli_xattrs_list_rwsem);
 271
 272         return rc;
 273 }
 274
 275 /**
 276  * Match or enqueue a PR lock.
 277  *
 278  * Find or request an LDLM lock with xattr data.
 279  * Since LDLM does not provide API for atomic match_or_enqueue,
 280  * the function handles it with a separate enq lock.
 281  * If successful, the function exits with the list lock held.
 282  *
 283  * \retval 0       no error occurred
 284  * \retval -ENOMEM not enough memory
 285  */
 286 static int ll_xattr_find_get_lock(struct inode *inode,
 287                                   struct lookup_intent *oit,
 288                                   struct ptlrpc_request **req)
 289 {
 290         ldlm_mode_t mode;
 291         struct lustre_handle lockh = { 0 };
 292         struct md_op_data *op_data;
 293         struct ll_inode_info *lli = ll_i2info(inode);
 294         struct ldlm_enqueue_info einfo = { .ei_type = LDLM_IBITS,
 295                                            .ei_mode = it_to_lock_mode(oit),
 296                                            .ei_cb_bl = ll_md_blocking_ast,
 297                                            .ei_cb_cp = ldlm_completion_ast };
 298         struct ll_sb_info *sbi = ll_i2sbi(inode);
 299         struct obd_export *exp = sbi->ll_md_exp;
 300         int rc;
 301
 302
 303
 304         mutex_lock(&lli->lli_xattrs_enq_lock);
 305         /* Try matching first. */
 306         mode = ll_take_md_lock(inode, MDS_INODELOCK_XATTR, &lockh, 0, LCK_PR);
 307         if (mode != 0) {
 308                 /* fake oit in mdc_revalidate_lock() manner */
 309                 oit->d.lustre.it_lock_handle = lockh.cookie;
 310                 oit->d.lustre.it_lock_mode = mode;
 311                 goto out;
 312         }
 313
 314         /* Enqueue if the lock isn't cached locally. */
 315         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 316                                      LUSTRE_OPC_ANY, NULL);
 317         if (IS_ERR(op_data)) {
 318                 mutex_unlock(&lli->lli_xattrs_enq_lock);
 319                 return PTR_ERR(op_data);
 320         }
 321
 322         op_data->op_valid = OBD_MD_FLXATTR | OBD_MD_FLXATTRLS;
 323
 324         rc = md_enqueue(exp, &einfo, oit, op_data, &lockh, NULL, 0, NULL, 0);
 325         ll_finish_md_op_data(op_data);
 326
 327         if (rc < 0) {
 328                 CDEBUG(D_CACHE,
 329                        "md_intent_lock failed with %d for fid "DFID"\n",
 330                        rc, PFID(ll_inode2fid(inode)));
 331                 mutex_unlock(&lli->lli_xattrs_enq_lock);
 332                 return rc;
 333         }
 334
 335         *req = (struct ptlrpc_request *)oit->d.lustre.it_data;
 336 out:
 337         down_write(&lli->lli_xattrs_list_rwsem);
 338         mutex_unlock(&lli->lli_xattrs_enq_lock);
 339
 340         return 0;
 341 }
 342
 343 /**
 344  * Refill the xattr cache.
 345  *
 346  * Fetch and cache the whole of xattrs for @inode, acquiring
 347  * a read or a write xattr lock depending on operation in @oit.
 348  * Intent is dropped on exit unless the operation is setxattr.
 349  *
 350  * \retval 0       no error occurred
 351  * \retval -EPROTO network protocol error
 352  * \retval -ENOMEM not enough memory for the cache
 353  */
 354 static int ll_xattr_cache_refill(struct inode *inode, struct lookup_intent *oit)
 355 {
 356         struct ll_sb_info *sbi = ll_i2sbi(inode);
 357         struct ptlrpc_request *req = NULL;
 358         const char *xdata, *xval, *xtail, *xvtail;
 359         struct ll_inode_info *lli = ll_i2info(inode);
 360         struct mdt_body *body;
 361         __u32 *xsizes;
 362         int rc = 0, i;
 363
 364
 365
 366         rc = ll_xattr_find_get_lock(inode, oit, &req);
 367         if (rc)
 368                 goto out_no_unlock;
 369
 370         /* Do we have the data at this point? */
 371         if (ll_xattr_cache_valid(lli)) {
 372                 ll_stats_ops_tally(sbi, LPROC_LL_GETXATTR_HITS, 1);
 373                 rc = 0;
 374                 goto out_maybe_drop;
 375         }
 376
 377         /* Matched but no cache? Cancelled on error by a parallel refill. */
 378         if (unlikely(req == NULL)) {
 379                 CDEBUG(D_CACHE, "cancelled by a parallel getxattr\n");
 380                 rc = -EIO;
 381                 goto out_maybe_drop;
 382         }
 383
 384         if (oit->d.lustre.it_status < 0) {
 385                 CDEBUG(D_CACHE, "getxattr intent returned %d for fid "DFID"\n",
 386                        oit->d.lustre.it_status, PFID(ll_inode2fid(inode)));
 387                 rc = oit->d.lustre.it_status;
 388                 /* xattr data is so large that we don't want to cache it */
 389                 if (rc == -ERANGE)
 390                         rc = -EAGAIN;
 391                 goto out_destroy;
 392         }
 393
 394         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 395         if (body == NULL) {
 396                 CERROR("no MDT BODY in the refill xattr reply\n");
 397                 rc = -EPROTO;
 398                 goto out_destroy;
 399         }
 400         /* do not need swab xattr data */
 401         xdata = req_capsule_server_sized_get(&req->rq_pill, &RMF_EADATA,
 402                                                 body->eadatasize);
 403         xval = req_capsule_server_sized_get(&req->rq_pill, &RMF_EAVALS,
 404                                                 body->aclsize);
 405         xsizes = req_capsule_server_sized_get(&req->rq_pill, &RMF_EAVALS_LENS,
 406                                               body->max_mdsize * sizeof(__u32));
 407         if (xdata == NULL || xval == NULL || xsizes == NULL) {
 408                 CERROR("wrong setxattr reply\n");
 409                 rc = -EPROTO;
 410                 goto out_destroy;
 411         }
 412
 413         xtail = xdata + body->eadatasize;
 414         xvtail = xval + body->aclsize;
 415
 416         CDEBUG(D_CACHE, "caching: xdata=%p xtail=%p\n", xdata, xtail);
 417
 418         ll_xattr_cache_init(lli);
 419
 420         for (i = 0; i < body->max_mdsize; i++) {
 421                 CDEBUG(D_CACHE, "caching [%s]=%.*s\n", xdata, *xsizes, xval);
 422                 /* Perform consistency checks: attr names and vals in pill */
 423                 if (memchr(xdata, 0, xtail - xdata) == NULL) {
 424                         CERROR("xattr protocol violation (names are broken)\n");
 425                         rc = -EPROTO;
 426                 } else if (xval + *xsizes > xvtail) {
 427                         CERROR("xattr protocol violation (vals are broken)\n");
 428                         rc = -EPROTO;
 429                 } else if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_XATTR_ENOMEM)) {
 430                         rc = -ENOMEM;
 431                 } else if (!strcmp(xdata, XATTR_NAME_ACL_ACCESS)) {
 432                         /* Filter out ACL ACCESS since it's cached separately */
 433                         CDEBUG(D_CACHE, "not caching %s\n",
 434                                XATTR_NAME_ACL_ACCESS);
 435                         rc = 0;
 436                 } else {
 437                         rc = ll_xattr_cache_add(&lli->lli_xattrs, xdata, xval,
 438                                                 *xsizes);
 439                 }
 440                 if (rc < 0) {
 441                         ll_xattr_cache_destroy_locked(lli);
 442                         goto out_destroy;
 443                 }
 444                 xdata += strlen(xdata) + 1;
 445                 xval  += *xsizes;
 446                 xsizes++;
 447         }
 448
 449         if (xdata != xtail || xval != xvtail)
 450                 CERROR("a hole in xattr data\n");
 451
 452         ll_set_lock_data(sbi->ll_md_exp, inode, oit, NULL);
 453
 454         goto out_maybe_drop;
 455 out_maybe_drop:
 456
 457                 ll_intent_drop_lock(oit);
 458
 459         if (rc != 0)
 460                 up_write(&lli->lli_xattrs_list_rwsem);
 461 out_no_unlock:
 462         ptlrpc_req_finished(req);
 463
 464         return rc;
 465
 466 out_destroy:
 467         up_write(&lli->lli_xattrs_list_rwsem);
 468
 469         ldlm_lock_decref_and_cancel((struct lustre_handle *)
 470                                         &oit->d.lustre.it_lock_handle,
 471                                         oit->d.lustre.it_lock_mode);
 472
 473         goto out_no_unlock;
 474 }
 475
 476 /**
 477  * Get an xattr value or list xattrs using the write-through cache.
 478  *
 479  * Get the xattr value (@valid has OBD_MD_FLXATTR set) of @name or
 480  * list xattr names (@valid has OBD_MD_FLXATTRLS set) for @inode.
 481  * The resulting value/list is stored in @buffer if the former
 482  * is not larger than @size.
 483  *
 484  * \retval 0        no error occurred
 485  * \retval -EPROTO  network protocol error
 486  * \retval -ENOMEM  not enough memory for the cache
 487  * \retval -ERANGE  the buffer is not large enough
 488  * \retval -ENODATA no such attr or the list is empty
 489  */
 490 int ll_xattr_cache_get(struct inode *inode,
 491                         const char *name,
 492                         char *buffer,
 493                         size_t size,
 494                         __u64 valid)
 495 {
 496         struct lookup_intent oit = { .it_op = IT_GETXATTR };
 497         struct ll_inode_info *lli = ll_i2info(inode);
 498         int rc = 0;
 499
 500
 501
 502         LASSERT(!!(valid & OBD_MD_FLXATTR) ^ !!(valid & OBD_MD_FLXATTRLS));
 503
 504         down_read(&lli->lli_xattrs_list_rwsem);
 505         if (!ll_xattr_cache_valid(lli)) {
 506                 up_read(&lli->lli_xattrs_list_rwsem);
 507                 rc = ll_xattr_cache_refill(inode, &oit);
 508                 if (rc)
 509                         return rc;
 510                 downgrade_write(&lli->lli_xattrs_list_rwsem);
 511         } else {
 512                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_GETXATTR_HITS, 1);
 513         }
 514
 515         if (valid & OBD_MD_FLXATTR) {
 516                 struct ll_xattr_entry *xattr;
 517
 518                 rc = ll_xattr_cache_find(&lli->lli_xattrs, name, &xattr);
 519                 if (rc == 0) {
 520                         rc = xattr->xe_vallen;
 521                         /* zero size means we are only requested size in rc */
 522                         if (size != 0) {
 523                                 if (size >= xattr->xe_vallen)
 524                                         memcpy(buffer, xattr->xe_value,
 525                                                 xattr->xe_vallen);
 526                                 else
 527                                         rc = -ERANGE;
 528                         }
 529                 }
 530         } else if (valid & OBD_MD_FLXATTRLS) {
 531                 rc = ll_xattr_cache_list(&lli->lli_xattrs,
 532                                          size ? buffer : NULL, size);
 533         }
 534
 535         goto out;
 536 out:
 537         up_read(&lli->lli_xattrs_list_rwsem);
 538
 539         return rc;
 540 }