2 * Copyright 2012 Xyratex Technology Limited
4 * Copyright (c) 2013, 2015, Intel Corporation.
6 * Author: Andrew Perepechko <Andrew_Perepechko@xyratex.com>
10 #define DEBUG_SUBSYSTEM S_LLITE
13 #include <linux/sched.h>
15 #include "../include/obd_support.h"
16 #include "../include/lustre_lite.h"
17 #include "../include/lustre_dlm.h"
18 #include "../include/lustre_ver.h"
19 #include "llite_internal.h"
21 /* If we ever have hundreds of extended attributes, we might want to consider
22 * using a hash or a tree structure instead of list for faster lookups.
24 struct ll_xattr_entry {
25 struct list_head xe_list; /* protected with
26 * lli_xattrs_list_rwsem
28 char *xe_name; /* xattr name, \0-terminated */
29 char *xe_value; /* xattr value */
30 unsigned xe_namelen; /* strlen(xe_name) + 1 */
31 unsigned xe_vallen; /* xattr value length */
34 static struct kmem_cache *xattr_kmem;
35 static struct lu_kmem_descr xattr_caches[] = {
37 .ckd_cache = &xattr_kmem,
38 .ckd_name = "xattr_kmem",
39 .ckd_size = sizeof(struct ll_xattr_entry)
46 int ll_xattr_init(void)
48 return lu_kmem_init(xattr_caches);
51 void ll_xattr_fini(void)
53 lu_kmem_fini(xattr_caches);
57 * Initializes xattr cache for an inode.
59 * This initializes the xattr list and marks cache presence.
61 static void ll_xattr_cache_init(struct ll_inode_info *lli)
63 INIT_LIST_HEAD(&lli->lli_xattrs);
64 lli->lli_flags |= LLIF_XATTR_CACHE;
68 * This looks for a specific extended attribute.
70 * Find in @cache and return @xattr_name attribute in @xattr,
71 * for the NULL @xattr_name return the first cached @xattr.
74 * \retval -ENODATA if not found
76 static int ll_xattr_cache_find(struct list_head *cache,
77 const char *xattr_name,
78 struct ll_xattr_entry **xattr)
80 struct ll_xattr_entry *entry;
82 list_for_each_entry(entry, cache, xe_list) {
83 /* xattr_name == NULL means look for any entry */
84 if (!xattr_name || strcmp(xattr_name, entry->xe_name) == 0) {
86 CDEBUG(D_CACHE, "find: [%s]=%.*s\n",
87 entry->xe_name, entry->xe_vallen,
99 * Add @xattr_name attr with @xattr_val value and @xattr_val_len length,
102 * \retval -ENOMEM if no memory could be allocated for the cached attr
103 * \retval -EPROTO if duplicate xattr is being added
105 static int ll_xattr_cache_add(struct list_head *cache,
106 const char *xattr_name,
107 const char *xattr_val,
108 unsigned xattr_val_len)
110 struct ll_xattr_entry *xattr;
112 if (ll_xattr_cache_find(cache, xattr_name, &xattr) == 0) {
113 CDEBUG(D_CACHE, "duplicate xattr: [%s]\n", xattr_name);
117 xattr = kmem_cache_zalloc(xattr_kmem, GFP_NOFS);
119 CDEBUG(D_CACHE, "failed to allocate xattr\n");
123 xattr->xe_name = kstrdup(xattr_name, GFP_NOFS);
124 if (!xattr->xe_name) {
125 CDEBUG(D_CACHE, "failed to alloc xattr name %u\n",
129 xattr->xe_value = kmemdup(xattr_val, xattr_val_len, GFP_NOFS);
130 if (!xattr->xe_value)
133 xattr->xe_vallen = xattr_val_len;
134 list_add(&xattr->xe_list, cache);
136 CDEBUG(D_CACHE, "set: [%s]=%.*s\n", xattr_name, xattr_val_len,
141 kfree(xattr->xe_name);
143 kmem_cache_free(xattr_kmem, xattr);
149 * This removes an extended attribute from cache.
151 * Remove @xattr_name attribute from @cache.
154 * \retval -ENODATA if @xattr_name is not cached
156 static int ll_xattr_cache_del(struct list_head *cache,
157 const char *xattr_name)
159 struct ll_xattr_entry *xattr;
161 CDEBUG(D_CACHE, "del xattr: %s\n", xattr_name);
163 if (ll_xattr_cache_find(cache, xattr_name, &xattr) == 0) {
164 list_del(&xattr->xe_list);
165 kfree(xattr->xe_name);
166 kfree(xattr->xe_value);
167 kmem_cache_free(xattr_kmem, xattr);
176 * This iterates cached extended attributes.
178 * Walk over cached attributes in @cache and
179 * fill in @xld_buffer or only calculate buffer
180 * size if @xld_buffer is NULL.
182 * \retval >= 0 buffer list size
183 * \retval -ENODATA if the list cannot fit @xld_size buffer
185 static int ll_xattr_cache_list(struct list_head *cache,
189 struct ll_xattr_entry *xattr, *tmp;
192 list_for_each_entry_safe(xattr, tmp, cache, xe_list) {
193 CDEBUG(D_CACHE, "list: buffer=%p[%d] name=%s\n",
194 xld_buffer, xld_tail, xattr->xe_name);
197 xld_size -= xattr->xe_namelen;
200 memcpy(&xld_buffer[xld_tail],
201 xattr->xe_name, xattr->xe_namelen);
203 xld_tail += xattr->xe_namelen;
213 * Check if the xattr cache is initialized (filled).
215 * \retval 0 @cache is not initialized
216 * \retval 1 @cache is initialized
218 static int ll_xattr_cache_valid(struct ll_inode_info *lli)
220 return !!(lli->lli_flags & LLIF_XATTR_CACHE);
224 * This finalizes the xattr cache.
226 * Free all xattr memory. @lli is the inode info pointer.
228 * \retval 0 no error occurred
230 static int ll_xattr_cache_destroy_locked(struct ll_inode_info *lli)
232 if (!ll_xattr_cache_valid(lli))
235 while (ll_xattr_cache_del(&lli->lli_xattrs, NULL) == 0)
237 lli->lli_flags &= ~LLIF_XATTR_CACHE;
242 int ll_xattr_cache_destroy(struct inode *inode)
244 struct ll_inode_info *lli = ll_i2info(inode);
247 down_write(&lli->lli_xattrs_list_rwsem);
248 rc = ll_xattr_cache_destroy_locked(lli);
249 up_write(&lli->lli_xattrs_list_rwsem);
255 * Match or enqueue a PR lock.
257 * Find or request an LDLM lock with xattr data.
258 * Since LDLM does not provide API for atomic match_or_enqueue,
259 * the function handles it with a separate enq lock.
260 * If successful, the function exits with the list lock held.
262 * \retval 0 no error occurred
263 * \retval -ENOMEM not enough memory
265 static int ll_xattr_find_get_lock(struct inode *inode,
266 struct lookup_intent *oit,
267 struct ptlrpc_request **req)
270 struct lustre_handle lockh = { 0 };
271 struct md_op_data *op_data;
272 struct ll_inode_info *lli = ll_i2info(inode);
273 struct ldlm_enqueue_info einfo = { .ei_type = LDLM_IBITS,
274 .ei_mode = it_to_lock_mode(oit),
275 .ei_cb_bl = ll_md_blocking_ast,
276 .ei_cb_cp = ldlm_completion_ast };
277 struct ll_sb_info *sbi = ll_i2sbi(inode);
278 struct obd_export *exp = sbi->ll_md_exp;
281 mutex_lock(&lli->lli_xattrs_enq_lock);
282 /* inode may have been shrunk and recreated, so data is gone, match lock
283 * only when data exists.
285 if (ll_xattr_cache_valid(lli)) {
286 /* Try matching first. */
287 mode = ll_take_md_lock(inode, MDS_INODELOCK_XATTR, &lockh, 0,
290 /* fake oit in mdc_revalidate_lock() manner */
291 oit->it_lock_handle = lockh.cookie;
292 oit->it_lock_mode = mode;
297 /* Enqueue if the lock isn't cached locally. */
298 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
299 LUSTRE_OPC_ANY, NULL);
300 if (IS_ERR(op_data)) {
301 mutex_unlock(&lli->lli_xattrs_enq_lock);
302 return PTR_ERR(op_data);
305 op_data->op_valid = OBD_MD_FLXATTR | OBD_MD_FLXATTRLS;
307 rc = md_enqueue(exp, &einfo, oit, op_data, &lockh, NULL, 0, NULL, 0);
308 ll_finish_md_op_data(op_data);
312 "md_intent_lock failed with %d for fid "DFID"\n",
313 rc, PFID(ll_inode2fid(inode)));
314 mutex_unlock(&lli->lli_xattrs_enq_lock);
318 *req = oit->it_request;
320 down_write(&lli->lli_xattrs_list_rwsem);
321 mutex_unlock(&lli->lli_xattrs_enq_lock);
327 * Refill the xattr cache.
329 * Fetch and cache the whole of xattrs for @inode, acquiring
330 * a read or a write xattr lock depending on operation in @oit.
331 * Intent is dropped on exit unless the operation is setxattr.
333 * \retval 0 no error occurred
334 * \retval -EPROTO network protocol error
335 * \retval -ENOMEM not enough memory for the cache
337 static int ll_xattr_cache_refill(struct inode *inode, struct lookup_intent *oit)
339 struct ll_sb_info *sbi = ll_i2sbi(inode);
340 struct ptlrpc_request *req = NULL;
341 const char *xdata, *xval, *xtail, *xvtail;
342 struct ll_inode_info *lli = ll_i2info(inode);
343 struct mdt_body *body;
347 rc = ll_xattr_find_get_lock(inode, oit, &req);
351 /* Do we have the data at this point? */
352 if (ll_xattr_cache_valid(lli)) {
353 ll_stats_ops_tally(sbi, LPROC_LL_GETXATTR_HITS, 1);
358 /* Matched but no cache? Cancelled on error by a parallel refill. */
359 if (unlikely(!req)) {
360 CDEBUG(D_CACHE, "cancelled by a parallel getxattr\n");
365 if (oit->it_status < 0) {
366 CDEBUG(D_CACHE, "getxattr intent returned %d for fid "DFID"\n",
367 oit->it_status, PFID(ll_inode2fid(inode)));
369 /* xattr data is so large that we don't want to cache it */
375 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
377 CERROR("no MDT BODY in the refill xattr reply\n");
381 /* do not need swab xattr data */
382 xdata = req_capsule_server_sized_get(&req->rq_pill, &RMF_EADATA,
384 xval = req_capsule_server_sized_get(&req->rq_pill, &RMF_EAVALS,
386 xsizes = req_capsule_server_sized_get(&req->rq_pill, &RMF_EAVALS_LENS,
387 body->max_mdsize * sizeof(__u32));
388 if (!xdata || !xval || !xsizes) {
389 CERROR("wrong setxattr reply\n");
394 xtail = xdata + body->eadatasize;
395 xvtail = xval + body->aclsize;
397 CDEBUG(D_CACHE, "caching: xdata=%p xtail=%p\n", xdata, xtail);
399 ll_xattr_cache_init(lli);
401 for (i = 0; i < body->max_mdsize; i++) {
402 CDEBUG(D_CACHE, "caching [%s]=%.*s\n", xdata, *xsizes, xval);
403 /* Perform consistency checks: attr names and vals in pill */
404 if (!memchr(xdata, 0, xtail - xdata)) {
405 CERROR("xattr protocol violation (names are broken)\n");
407 } else if (xval + *xsizes > xvtail) {
408 CERROR("xattr protocol violation (vals are broken)\n");
410 } else if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_XATTR_ENOMEM)) {
412 } else if (!strcmp(xdata, XATTR_NAME_ACL_ACCESS)) {
413 /* Filter out ACL ACCESS since it's cached separately */
414 CDEBUG(D_CACHE, "not caching %s\n",
415 XATTR_NAME_ACL_ACCESS);
418 rc = ll_xattr_cache_add(&lli->lli_xattrs, xdata, xval,
422 ll_xattr_cache_destroy_locked(lli);
425 xdata += strlen(xdata) + 1;
430 if (xdata != xtail || xval != xvtail)
431 CERROR("a hole in xattr data\n");
433 ll_set_lock_data(sbi->ll_md_exp, inode, oit, NULL);
438 ll_intent_drop_lock(oit);
441 up_write(&lli->lli_xattrs_list_rwsem);
443 ptlrpc_req_finished(req);
448 up_write(&lli->lli_xattrs_list_rwsem);
450 ldlm_lock_decref_and_cancel((struct lustre_handle *)
451 &oit->it_lock_handle,
458 * Get an xattr value or list xattrs using the write-through cache.
460 * Get the xattr value (@valid has OBD_MD_FLXATTR set) of @name or
461 * list xattr names (@valid has OBD_MD_FLXATTRLS set) for @inode.
462 * The resulting value/list is stored in @buffer if the former
463 * is not larger than @size.
465 * \retval 0 no error occurred
466 * \retval -EPROTO network protocol error
467 * \retval -ENOMEM not enough memory for the cache
468 * \retval -ERANGE the buffer is not large enough
469 * \retval -ENODATA no such attr or the list is empty
471 int ll_xattr_cache_get(struct inode *inode, const char *name, char *buffer,
472 size_t size, __u64 valid)
474 struct lookup_intent oit = { .it_op = IT_GETXATTR };
475 struct ll_inode_info *lli = ll_i2info(inode);
478 LASSERT(!!(valid & OBD_MD_FLXATTR) ^ !!(valid & OBD_MD_FLXATTRLS));
480 down_read(&lli->lli_xattrs_list_rwsem);
481 if (!ll_xattr_cache_valid(lli)) {
482 up_read(&lli->lli_xattrs_list_rwsem);
483 rc = ll_xattr_cache_refill(inode, &oit);
486 downgrade_write(&lli->lli_xattrs_list_rwsem);
488 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_GETXATTR_HITS, 1);
491 if (valid & OBD_MD_FLXATTR) {
492 struct ll_xattr_entry *xattr;
494 rc = ll_xattr_cache_find(&lli->lli_xattrs, name, &xattr);
496 rc = xattr->xe_vallen;
497 /* zero size means we are only requested size in rc */
499 if (size >= xattr->xe_vallen)
500 memcpy(buffer, xattr->xe_value,
506 } else if (valid & OBD_MD_FLXATTRLS) {
507 rc = ll_xattr_cache_list(&lli->lli_xattrs,
508 size ? buffer : NULL, size);
513 up_read(&lli->lli_xattrs_list_rwsem);