blob: cd7ffad4041d81b605bdfbc97a4bc0072c19ca99 [file] [log] [blame]
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001#include <linux/ceph/ceph_debug.h>
Yan, Zheng25e6bae2014-09-16 19:15:28 +08002#include <linux/ceph/pagelist.h>
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07003
Sage Weil355da1e2009-10-06 11:31:08 -07004#include "super.h"
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07005#include "mds_client.h"
6
7#include <linux/ceph/decode.h>
Sage Weil355da1e2009-10-06 11:31:08 -07008
9#include <linux/xattr.h>
Linus Torvalds4db658e2014-01-28 18:06:18 -080010#include <linux/posix_acl_xattr.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090011#include <linux/slab.h>
Sage Weil355da1e2009-10-06 11:31:08 -070012
Alex Elder22891902012-01-23 15:49:28 -060013#define XATTR_CEPH_PREFIX "ceph."
14#define XATTR_CEPH_PREFIX_LEN (sizeof (XATTR_CEPH_PREFIX) - 1)
15
Yan, Zhengbcdfeb22014-02-11 13:04:19 +080016static int __remove_xattr(struct ceph_inode_info *ci,
17 struct ceph_inode_xattr *xattr);
18
Guangliang Zhao7221fe42013-11-11 15:18:03 +080019/*
20 * List of handlers for synthetic system.* attributes. Other
21 * attributes are handled directly.
22 */
23const struct xattr_handler *ceph_xattr_handlers[] = {
24#ifdef CONFIG_CEPH_FS_POSIX_ACL
Linus Torvalds4db658e2014-01-28 18:06:18 -080025 &posix_acl_access_xattr_handler,
26 &posix_acl_default_xattr_handler,
Guangliang Zhao7221fe42013-11-11 15:18:03 +080027#endif
28 NULL,
29};
30
Sage Weil355da1e2009-10-06 11:31:08 -070031static bool ceph_is_valid_xattr(const char *name)
32{
Alex Elder22891902012-01-23 15:49:28 -060033 return !strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN) ||
Sage Weil1a756272010-05-11 11:40:25 -070034 !strncmp(name, XATTR_SECURITY_PREFIX,
Sage Weil355da1e2009-10-06 11:31:08 -070035 XATTR_SECURITY_PREFIX_LEN) ||
Guangliang Zhao7221fe42013-11-11 15:18:03 +080036 !strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN) ||
Sage Weil355da1e2009-10-06 11:31:08 -070037 !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
38 !strncmp(name, XATTR_USER_PREFIX, XATTR_USER_PREFIX_LEN);
39}
40
41/*
42 * These define virtual xattrs exposing the recursive directory
43 * statistics and layout metadata.
44 */
Alex Elder881a5fa2012-01-23 15:49:28 -060045struct ceph_vxattr {
Sage Weil355da1e2009-10-06 11:31:08 -070046 char *name;
Alex Elder3ce6cd12012-01-23 15:49:28 -060047 size_t name_size; /* strlen(name) + 1 (for '\0') */
Sage Weil355da1e2009-10-06 11:31:08 -070048 size_t (*getxattr_cb)(struct ceph_inode_info *ci, char *val,
49 size_t size);
Sage Weil88601472013-01-31 11:53:27 -080050 bool readonly, hidden;
Sage Weilf36e4472013-01-20 21:59:29 -080051 bool (*exists_cb)(struct ceph_inode_info *ci);
Sage Weil355da1e2009-10-06 11:31:08 -070052};
53
Sage Weil32ab0bd2013-01-19 16:46:32 -080054/* layouts */
55
56static bool ceph_vxattrcb_layout_exists(struct ceph_inode_info *ci)
57{
58 size_t s;
59 char *p = (char *)&ci->i_layout;
60
61 for (s = 0; s < sizeof(ci->i_layout); s++, p++)
62 if (*p)
63 return true;
64 return false;
65}
66
67static size_t ceph_vxattrcb_layout(struct ceph_inode_info *ci, char *val,
Yan, Zheng1e5c6642014-03-24 13:00:54 +080068 size_t size)
Sage Weil32ab0bd2013-01-19 16:46:32 -080069{
70 int ret;
71 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
72 struct ceph_osd_client *osdc = &fsc->client->osdc;
73 s64 pool = ceph_file_layout_pg_pool(ci->i_layout);
74 const char *pool_name;
Yan, Zheng1e5c6642014-03-24 13:00:54 +080075 char buf[128];
Sage Weil32ab0bd2013-01-19 16:46:32 -080076
77 dout("ceph_vxattrcb_layout %p\n", &ci->vfs_inode);
78 down_read(&osdc->map_sem);
79 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, pool);
Yan, Zheng1e5c6642014-03-24 13:00:54 +080080 if (pool_name) {
81 size_t len = strlen(pool_name);
82 ret = snprintf(buf, sizeof(buf),
83 "stripe_unit=%lld stripe_count=%lld object_size=%lld pool=",
Sage Weil32ab0bd2013-01-19 16:46:32 -080084 (unsigned long long)ceph_file_layout_su(ci->i_layout),
85 (unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
Yan, Zheng1e5c6642014-03-24 13:00:54 +080086 (unsigned long long)ceph_file_layout_object_size(ci->i_layout));
87 if (!size) {
88 ret += len;
89 } else if (ret + len > size) {
90 ret = -ERANGE;
91 } else {
92 memcpy(val, buf, ret);
93 memcpy(val + ret, pool_name, len);
94 ret += len;
95 }
96 } else {
97 ret = snprintf(buf, sizeof(buf),
Sage Weil32ab0bd2013-01-19 16:46:32 -080098 "stripe_unit=%lld stripe_count=%lld object_size=%lld pool=%lld",
99 (unsigned long long)ceph_file_layout_su(ci->i_layout),
100 (unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
101 (unsigned long long)ceph_file_layout_object_size(ci->i_layout),
102 (unsigned long long)pool);
Yan, Zheng1e5c6642014-03-24 13:00:54 +0800103 if (size) {
104 if (ret <= size)
105 memcpy(val, buf, ret);
106 else
107 ret = -ERANGE;
108 }
109 }
Sage Weil32ab0bd2013-01-19 16:46:32 -0800110 up_read(&osdc->map_sem);
111 return ret;
112}
113
Sage Weil695b7112013-01-20 22:08:33 -0800114static size_t ceph_vxattrcb_layout_stripe_unit(struct ceph_inode_info *ci,
115 char *val, size_t size)
116{
117 return snprintf(val, size, "%lld",
118 (unsigned long long)ceph_file_layout_su(ci->i_layout));
119}
120
121static size_t ceph_vxattrcb_layout_stripe_count(struct ceph_inode_info *ci,
122 char *val, size_t size)
123{
124 return snprintf(val, size, "%lld",
125 (unsigned long long)ceph_file_layout_stripe_count(ci->i_layout));
126}
127
128static size_t ceph_vxattrcb_layout_object_size(struct ceph_inode_info *ci,
129 char *val, size_t size)
130{
131 return snprintf(val, size, "%lld",
132 (unsigned long long)ceph_file_layout_object_size(ci->i_layout));
133}
134
135static size_t ceph_vxattrcb_layout_pool(struct ceph_inode_info *ci,
136 char *val, size_t size)
137{
138 int ret;
139 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
140 struct ceph_osd_client *osdc = &fsc->client->osdc;
141 s64 pool = ceph_file_layout_pg_pool(ci->i_layout);
142 const char *pool_name;
143
144 down_read(&osdc->map_sem);
145 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, pool);
146 if (pool_name)
147 ret = snprintf(val, size, "%s", pool_name);
148 else
149 ret = snprintf(val, size, "%lld", (unsigned long long)pool);
150 up_read(&osdc->map_sem);
151 return ret;
152}
153
Sage Weil355da1e2009-10-06 11:31:08 -0700154/* directories */
155
Alex Elderaa4066e2012-01-23 15:49:28 -0600156static size_t ceph_vxattrcb_dir_entries(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700157 size_t size)
158{
159 return snprintf(val, size, "%lld", ci->i_files + ci->i_subdirs);
160}
161
Alex Elderaa4066e2012-01-23 15:49:28 -0600162static size_t ceph_vxattrcb_dir_files(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700163 size_t size)
164{
165 return snprintf(val, size, "%lld", ci->i_files);
166}
167
Alex Elderaa4066e2012-01-23 15:49:28 -0600168static size_t ceph_vxattrcb_dir_subdirs(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700169 size_t size)
170{
171 return snprintf(val, size, "%lld", ci->i_subdirs);
172}
173
Alex Elderaa4066e2012-01-23 15:49:28 -0600174static size_t ceph_vxattrcb_dir_rentries(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700175 size_t size)
176{
177 return snprintf(val, size, "%lld", ci->i_rfiles + ci->i_rsubdirs);
178}
179
Alex Elderaa4066e2012-01-23 15:49:28 -0600180static size_t ceph_vxattrcb_dir_rfiles(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700181 size_t size)
182{
183 return snprintf(val, size, "%lld", ci->i_rfiles);
184}
185
Alex Elderaa4066e2012-01-23 15:49:28 -0600186static size_t ceph_vxattrcb_dir_rsubdirs(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700187 size_t size)
188{
189 return snprintf(val, size, "%lld", ci->i_rsubdirs);
190}
191
Alex Elderaa4066e2012-01-23 15:49:28 -0600192static size_t ceph_vxattrcb_dir_rbytes(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700193 size_t size)
194{
195 return snprintf(val, size, "%lld", ci->i_rbytes);
196}
197
Alex Elderaa4066e2012-01-23 15:49:28 -0600198static size_t ceph_vxattrcb_dir_rctime(struct ceph_inode_info *ci, char *val,
Sage Weil355da1e2009-10-06 11:31:08 -0700199 size_t size)
200{
Alex Elder3489b422012-03-08 16:50:09 -0600201 return snprintf(val, size, "%ld.09%ld", (long)ci->i_rctime.tv_sec,
Sage Weil355da1e2009-10-06 11:31:08 -0700202 (long)ci->i_rctime.tv_nsec);
203}
204
Sage Weil32ab0bd2013-01-19 16:46:32 -0800205
Alex Eldereb788082012-01-23 15:49:28 -0600206#define CEPH_XATTR_NAME(_type, _name) XATTR_CEPH_PREFIX #_type "." #_name
Sage Weil695b7112013-01-20 22:08:33 -0800207#define CEPH_XATTR_NAME2(_type, _name, _name2) \
208 XATTR_CEPH_PREFIX #_type "." #_name "." #_name2
Alex Eldereb788082012-01-23 15:49:28 -0600209
Sage Weil88601472013-01-31 11:53:27 -0800210#define XATTR_NAME_CEPH(_type, _name) \
211 { \
212 .name = CEPH_XATTR_NAME(_type, _name), \
213 .name_size = sizeof (CEPH_XATTR_NAME(_type, _name)), \
214 .getxattr_cb = ceph_vxattrcb_ ## _type ## _ ## _name, \
215 .readonly = true, \
216 .hidden = false, \
Sage Weilf36e4472013-01-20 21:59:29 -0800217 .exists_cb = NULL, \
Sage Weil88601472013-01-31 11:53:27 -0800218 }
Sage Weil695b7112013-01-20 22:08:33 -0800219#define XATTR_LAYOUT_FIELD(_type, _name, _field) \
220 { \
221 .name = CEPH_XATTR_NAME2(_type, _name, _field), \
222 .name_size = sizeof (CEPH_XATTR_NAME2(_type, _name, _field)), \
223 .getxattr_cb = ceph_vxattrcb_ ## _name ## _ ## _field, \
224 .readonly = false, \
225 .hidden = true, \
226 .exists_cb = ceph_vxattrcb_layout_exists, \
227 }
Alex Eldereb788082012-01-23 15:49:28 -0600228
Alex Elder881a5fa2012-01-23 15:49:28 -0600229static struct ceph_vxattr ceph_dir_vxattrs[] = {
Sage Weil1f08f2b2013-01-20 22:07:12 -0800230 {
231 .name = "ceph.dir.layout",
232 .name_size = sizeof("ceph.dir.layout"),
233 .getxattr_cb = ceph_vxattrcb_layout,
234 .readonly = false,
Yan, Zhengcc48c3e2014-03-24 19:15:05 +0800235 .hidden = true,
Sage Weil1f08f2b2013-01-20 22:07:12 -0800236 .exists_cb = ceph_vxattrcb_layout_exists,
237 },
Sage Weil695b7112013-01-20 22:08:33 -0800238 XATTR_LAYOUT_FIELD(dir, layout, stripe_unit),
239 XATTR_LAYOUT_FIELD(dir, layout, stripe_count),
240 XATTR_LAYOUT_FIELD(dir, layout, object_size),
241 XATTR_LAYOUT_FIELD(dir, layout, pool),
Alex Eldereb788082012-01-23 15:49:28 -0600242 XATTR_NAME_CEPH(dir, entries),
243 XATTR_NAME_CEPH(dir, files),
244 XATTR_NAME_CEPH(dir, subdirs),
245 XATTR_NAME_CEPH(dir, rentries),
246 XATTR_NAME_CEPH(dir, rfiles),
247 XATTR_NAME_CEPH(dir, rsubdirs),
248 XATTR_NAME_CEPH(dir, rbytes),
249 XATTR_NAME_CEPH(dir, rctime),
Alex Elder2c3dd4f2013-02-19 12:25:56 -0600250 { .name = NULL, 0 } /* Required table terminator */
Sage Weil355da1e2009-10-06 11:31:08 -0700251};
Alex Elder3ce6cd12012-01-23 15:49:28 -0600252static size_t ceph_dir_vxattrs_name_size; /* total size of all names */
Sage Weil355da1e2009-10-06 11:31:08 -0700253
254/* files */
255
Alex Elder881a5fa2012-01-23 15:49:28 -0600256static struct ceph_vxattr ceph_file_vxattrs[] = {
Sage Weil32ab0bd2013-01-19 16:46:32 -0800257 {
258 .name = "ceph.file.layout",
259 .name_size = sizeof("ceph.file.layout"),
260 .getxattr_cb = ceph_vxattrcb_layout,
261 .readonly = false,
Yan, Zhengcc48c3e2014-03-24 19:15:05 +0800262 .hidden = true,
Sage Weil32ab0bd2013-01-19 16:46:32 -0800263 .exists_cb = ceph_vxattrcb_layout_exists,
264 },
Sage Weil695b7112013-01-20 22:08:33 -0800265 XATTR_LAYOUT_FIELD(file, layout, stripe_unit),
266 XATTR_LAYOUT_FIELD(file, layout, stripe_count),
267 XATTR_LAYOUT_FIELD(file, layout, object_size),
268 XATTR_LAYOUT_FIELD(file, layout, pool),
Alex Elder2c3dd4f2013-02-19 12:25:56 -0600269 { .name = NULL, 0 } /* Required table terminator */
Sage Weil355da1e2009-10-06 11:31:08 -0700270};
Alex Elder3ce6cd12012-01-23 15:49:28 -0600271static size_t ceph_file_vxattrs_name_size; /* total size of all names */
Sage Weil355da1e2009-10-06 11:31:08 -0700272
Alex Elder881a5fa2012-01-23 15:49:28 -0600273static struct ceph_vxattr *ceph_inode_vxattrs(struct inode *inode)
Sage Weil355da1e2009-10-06 11:31:08 -0700274{
275 if (S_ISDIR(inode->i_mode))
276 return ceph_dir_vxattrs;
277 else if (S_ISREG(inode->i_mode))
278 return ceph_file_vxattrs;
279 return NULL;
280}
281
Alex Elder3ce6cd12012-01-23 15:49:28 -0600282static size_t ceph_vxattrs_name_size(struct ceph_vxattr *vxattrs)
283{
284 if (vxattrs == ceph_dir_vxattrs)
285 return ceph_dir_vxattrs_name_size;
286 if (vxattrs == ceph_file_vxattrs)
287 return ceph_file_vxattrs_name_size;
Yan, Zheng0abb43d2014-09-18 16:11:12 +0800288 BUG_ON(vxattrs);
Alex Elder3ce6cd12012-01-23 15:49:28 -0600289 return 0;
290}
291
292/*
293 * Compute the aggregate size (including terminating '\0') of all
294 * virtual extended attribute names in the given vxattr table.
295 */
296static size_t __init vxattrs_name_size(struct ceph_vxattr *vxattrs)
297{
298 struct ceph_vxattr *vxattr;
299 size_t size = 0;
300
301 for (vxattr = vxattrs; vxattr->name; vxattr++)
Sage Weil88601472013-01-31 11:53:27 -0800302 if (!vxattr->hidden)
303 size += vxattr->name_size;
Alex Elder3ce6cd12012-01-23 15:49:28 -0600304
305 return size;
306}
307
308/* Routines called at initialization and exit time */
309
310void __init ceph_xattr_init(void)
311{
312 ceph_dir_vxattrs_name_size = vxattrs_name_size(ceph_dir_vxattrs);
313 ceph_file_vxattrs_name_size = vxattrs_name_size(ceph_file_vxattrs);
314}
315
316void ceph_xattr_exit(void)
317{
318 ceph_dir_vxattrs_name_size = 0;
319 ceph_file_vxattrs_name_size = 0;
320}
321
Alex Elder881a5fa2012-01-23 15:49:28 -0600322static struct ceph_vxattr *ceph_match_vxattr(struct inode *inode,
Sage Weil355da1e2009-10-06 11:31:08 -0700323 const char *name)
324{
Alex Elder881a5fa2012-01-23 15:49:28 -0600325 struct ceph_vxattr *vxattr = ceph_inode_vxattrs(inode);
Alex Elder06476a62012-01-23 15:49:27 -0600326
327 if (vxattr) {
328 while (vxattr->name) {
329 if (!strcmp(vxattr->name, name))
330 return vxattr;
331 vxattr++;
332 }
333 }
334
Sage Weil355da1e2009-10-06 11:31:08 -0700335 return NULL;
336}
337
338static int __set_xattr(struct ceph_inode_info *ci,
339 const char *name, int name_len,
340 const char *val, int val_len,
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800341 int flags, int update_xattr,
Sage Weil355da1e2009-10-06 11:31:08 -0700342 struct ceph_inode_xattr **newxattr)
343{
344 struct rb_node **p;
345 struct rb_node *parent = NULL;
346 struct ceph_inode_xattr *xattr = NULL;
347 int c;
348 int new = 0;
349
350 p = &ci->i_xattrs.index.rb_node;
351 while (*p) {
352 parent = *p;
353 xattr = rb_entry(parent, struct ceph_inode_xattr, node);
354 c = strncmp(name, xattr->name, min(name_len, xattr->name_len));
355 if (c < 0)
356 p = &(*p)->rb_left;
357 else if (c > 0)
358 p = &(*p)->rb_right;
359 else {
360 if (name_len == xattr->name_len)
361 break;
362 else if (name_len < xattr->name_len)
363 p = &(*p)->rb_left;
364 else
365 p = &(*p)->rb_right;
366 }
367 xattr = NULL;
368 }
369
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800370 if (update_xattr) {
371 int err = 0;
372 if (xattr && (flags & XATTR_CREATE))
373 err = -EEXIST;
374 else if (!xattr && (flags & XATTR_REPLACE))
375 err = -ENODATA;
376 if (err) {
377 kfree(name);
378 kfree(val);
379 return err;
380 }
Yan, Zhengbcdfeb22014-02-11 13:04:19 +0800381 if (update_xattr < 0) {
382 if (xattr)
383 __remove_xattr(ci, xattr);
384 kfree(name);
385 return 0;
386 }
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800387 }
388
Sage Weil355da1e2009-10-06 11:31:08 -0700389 if (!xattr) {
390 new = 1;
391 xattr = *newxattr;
392 xattr->name = name;
393 xattr->name_len = name_len;
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800394 xattr->should_free_name = update_xattr;
Sage Weil355da1e2009-10-06 11:31:08 -0700395
396 ci->i_xattrs.count++;
397 dout("__set_xattr count=%d\n", ci->i_xattrs.count);
398 } else {
399 kfree(*newxattr);
400 *newxattr = NULL;
401 if (xattr->should_free_val)
402 kfree((void *)xattr->val);
403
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800404 if (update_xattr) {
Sage Weil355da1e2009-10-06 11:31:08 -0700405 kfree((void *)name);
406 name = xattr->name;
407 }
408 ci->i_xattrs.names_size -= xattr->name_len;
409 ci->i_xattrs.vals_size -= xattr->val_len;
410 }
Sage Weil355da1e2009-10-06 11:31:08 -0700411 ci->i_xattrs.names_size += name_len;
412 ci->i_xattrs.vals_size += val_len;
413 if (val)
414 xattr->val = val;
415 else
416 xattr->val = "";
417
418 xattr->val_len = val_len;
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800419 xattr->dirty = update_xattr;
420 xattr->should_free_val = (val && update_xattr);
Sage Weil355da1e2009-10-06 11:31:08 -0700421
422 if (new) {
423 rb_link_node(&xattr->node, parent, p);
424 rb_insert_color(&xattr->node, &ci->i_xattrs.index);
425 dout("__set_xattr_val p=%p\n", p);
426 }
427
428 dout("__set_xattr_val added %llx.%llx xattr %p %s=%.*s\n",
429 ceph_vinop(&ci->vfs_inode), xattr, name, val_len, val);
430
431 return 0;
432}
433
434static struct ceph_inode_xattr *__get_xattr(struct ceph_inode_info *ci,
435 const char *name)
436{
437 struct rb_node **p;
438 struct rb_node *parent = NULL;
439 struct ceph_inode_xattr *xattr = NULL;
Sage Weil17db1432011-01-13 15:27:29 -0800440 int name_len = strlen(name);
Sage Weil355da1e2009-10-06 11:31:08 -0700441 int c;
442
443 p = &ci->i_xattrs.index.rb_node;
444 while (*p) {
445 parent = *p;
446 xattr = rb_entry(parent, struct ceph_inode_xattr, node);
447 c = strncmp(name, xattr->name, xattr->name_len);
Sage Weil17db1432011-01-13 15:27:29 -0800448 if (c == 0 && name_len > xattr->name_len)
449 c = 1;
Sage Weil355da1e2009-10-06 11:31:08 -0700450 if (c < 0)
451 p = &(*p)->rb_left;
452 else if (c > 0)
453 p = &(*p)->rb_right;
454 else {
455 dout("__get_xattr %s: found %.*s\n", name,
456 xattr->val_len, xattr->val);
457 return xattr;
458 }
459 }
460
461 dout("__get_xattr %s: not found\n", name);
462
463 return NULL;
464}
465
466static void __free_xattr(struct ceph_inode_xattr *xattr)
467{
468 BUG_ON(!xattr);
469
470 if (xattr->should_free_name)
471 kfree((void *)xattr->name);
472 if (xattr->should_free_val)
473 kfree((void *)xattr->val);
474
475 kfree(xattr);
476}
477
478static int __remove_xattr(struct ceph_inode_info *ci,
479 struct ceph_inode_xattr *xattr)
480{
481 if (!xattr)
Yan, Zheng524186a2014-02-11 13:23:09 +0800482 return -ENODATA;
Sage Weil355da1e2009-10-06 11:31:08 -0700483
484 rb_erase(&xattr->node, &ci->i_xattrs.index);
485
486 if (xattr->should_free_name)
487 kfree((void *)xattr->name);
488 if (xattr->should_free_val)
489 kfree((void *)xattr->val);
490
491 ci->i_xattrs.names_size -= xattr->name_len;
492 ci->i_xattrs.vals_size -= xattr->val_len;
493 ci->i_xattrs.count--;
494 kfree(xattr);
495
496 return 0;
497}
498
499static int __remove_xattr_by_name(struct ceph_inode_info *ci,
500 const char *name)
501{
502 struct rb_node **p;
503 struct ceph_inode_xattr *xattr;
504 int err;
505
506 p = &ci->i_xattrs.index.rb_node;
507 xattr = __get_xattr(ci, name);
508 err = __remove_xattr(ci, xattr);
509 return err;
510}
511
512static char *__copy_xattr_names(struct ceph_inode_info *ci,
513 char *dest)
514{
515 struct rb_node *p;
516 struct ceph_inode_xattr *xattr = NULL;
517
518 p = rb_first(&ci->i_xattrs.index);
519 dout("__copy_xattr_names count=%d\n", ci->i_xattrs.count);
520
521 while (p) {
522 xattr = rb_entry(p, struct ceph_inode_xattr, node);
523 memcpy(dest, xattr->name, xattr->name_len);
524 dest[xattr->name_len] = '\0';
525
526 dout("dest=%s %p (%s) (%d/%d)\n", dest, xattr, xattr->name,
527 xattr->name_len, ci->i_xattrs.names_size);
528
529 dest += xattr->name_len + 1;
530 p = rb_next(p);
531 }
532
533 return dest;
534}
535
536void __ceph_destroy_xattrs(struct ceph_inode_info *ci)
537{
538 struct rb_node *p, *tmp;
539 struct ceph_inode_xattr *xattr = NULL;
540
541 p = rb_first(&ci->i_xattrs.index);
542
543 dout("__ceph_destroy_xattrs p=%p\n", p);
544
545 while (p) {
546 xattr = rb_entry(p, struct ceph_inode_xattr, node);
547 tmp = p;
548 p = rb_next(tmp);
549 dout("__ceph_destroy_xattrs next p=%p (%.*s)\n", p,
550 xattr->name_len, xattr->name);
551 rb_erase(tmp, &ci->i_xattrs.index);
552
553 __free_xattr(xattr);
554 }
555
556 ci->i_xattrs.names_size = 0;
557 ci->i_xattrs.vals_size = 0;
558 ci->i_xattrs.index_version = 0;
559 ci->i_xattrs.count = 0;
560 ci->i_xattrs.index = RB_ROOT;
561}
562
563static int __build_xattrs(struct inode *inode)
Sage Weilbe655592011-11-30 09:47:09 -0800564 __releases(ci->i_ceph_lock)
565 __acquires(ci->i_ceph_lock)
Sage Weil355da1e2009-10-06 11:31:08 -0700566{
567 u32 namelen;
568 u32 numattr = 0;
569 void *p, *end;
570 u32 len;
571 const char *name, *val;
572 struct ceph_inode_info *ci = ceph_inode(inode);
573 int xattr_version;
574 struct ceph_inode_xattr **xattrs = NULL;
Sage Weil63ff78b2009-11-01 17:51:15 -0800575 int err = 0;
Sage Weil355da1e2009-10-06 11:31:08 -0700576 int i;
577
578 dout("__build_xattrs() len=%d\n",
579 ci->i_xattrs.blob ? (int)ci->i_xattrs.blob->vec.iov_len : 0);
580
581 if (ci->i_xattrs.index_version >= ci->i_xattrs.version)
582 return 0; /* already built */
583
584 __ceph_destroy_xattrs(ci);
585
586start:
587 /* updated internal xattr rb tree */
588 if (ci->i_xattrs.blob && ci->i_xattrs.blob->vec.iov_len > 4) {
589 p = ci->i_xattrs.blob->vec.iov_base;
590 end = p + ci->i_xattrs.blob->vec.iov_len;
591 ceph_decode_32_safe(&p, end, numattr, bad);
592 xattr_version = ci->i_xattrs.version;
Sage Weilbe655592011-11-30 09:47:09 -0800593 spin_unlock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700594
Ilya Dryomov7e8a2952014-07-25 11:47:21 +0400595 xattrs = kcalloc(numattr, sizeof(struct ceph_inode_xattr *),
Sage Weil355da1e2009-10-06 11:31:08 -0700596 GFP_NOFS);
597 err = -ENOMEM;
598 if (!xattrs)
599 goto bad_lock;
Ilya Dryomov1a295bd2014-07-25 12:44:58 +0400600
Sage Weil355da1e2009-10-06 11:31:08 -0700601 for (i = 0; i < numattr; i++) {
602 xattrs[i] = kmalloc(sizeof(struct ceph_inode_xattr),
603 GFP_NOFS);
604 if (!xattrs[i])
605 goto bad_lock;
606 }
607
Sage Weilbe655592011-11-30 09:47:09 -0800608 spin_lock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700609 if (ci->i_xattrs.version != xattr_version) {
610 /* lost a race, retry */
611 for (i = 0; i < numattr; i++)
612 kfree(xattrs[i]);
613 kfree(xattrs);
Alan Cox21ec6ff2012-07-20 08:18:36 -0500614 xattrs = NULL;
Sage Weil355da1e2009-10-06 11:31:08 -0700615 goto start;
616 }
617 err = -EIO;
618 while (numattr--) {
619 ceph_decode_32_safe(&p, end, len, bad);
620 namelen = len;
621 name = p;
622 p += len;
623 ceph_decode_32_safe(&p, end, len, bad);
624 val = p;
625 p += len;
626
627 err = __set_xattr(ci, name, namelen, val, len,
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800628 0, 0, &xattrs[numattr]);
Sage Weil355da1e2009-10-06 11:31:08 -0700629
630 if (err < 0)
631 goto bad;
632 }
633 kfree(xattrs);
634 }
635 ci->i_xattrs.index_version = ci->i_xattrs.version;
636 ci->i_xattrs.dirty = false;
637
638 return err;
639bad_lock:
Sage Weilbe655592011-11-30 09:47:09 -0800640 spin_lock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700641bad:
642 if (xattrs) {
643 for (i = 0; i < numattr; i++)
644 kfree(xattrs[i]);
645 kfree(xattrs);
646 }
647 ci->i_xattrs.names_size = 0;
648 return err;
649}
650
651static int __get_required_blob_size(struct ceph_inode_info *ci, int name_size,
652 int val_size)
653{
654 /*
655 * 4 bytes for the length, and additional 4 bytes per each xattr name,
656 * 4 bytes per each value
657 */
658 int size = 4 + ci->i_xattrs.count*(4 + 4) +
659 ci->i_xattrs.names_size +
660 ci->i_xattrs.vals_size;
661 dout("__get_required_blob_size c=%d names.size=%d vals.size=%d\n",
662 ci->i_xattrs.count, ci->i_xattrs.names_size,
663 ci->i_xattrs.vals_size);
664
665 if (name_size)
666 size += 4 + 4 + name_size + val_size;
667
668 return size;
669}
670
671/*
672 * If there are dirty xattrs, reencode xattrs into the prealloc_blob
673 * and swap into place.
674 */
675void __ceph_build_xattrs_blob(struct ceph_inode_info *ci)
676{
677 struct rb_node *p;
678 struct ceph_inode_xattr *xattr = NULL;
679 void *dest;
680
681 dout("__build_xattrs_blob %p\n", &ci->vfs_inode);
682 if (ci->i_xattrs.dirty) {
683 int need = __get_required_blob_size(ci, 0, 0);
684
685 BUG_ON(need > ci->i_xattrs.prealloc_blob->alloc_len);
686
687 p = rb_first(&ci->i_xattrs.index);
688 dest = ci->i_xattrs.prealloc_blob->vec.iov_base;
689
690 ceph_encode_32(&dest, ci->i_xattrs.count);
691 while (p) {
692 xattr = rb_entry(p, struct ceph_inode_xattr, node);
693
694 ceph_encode_32(&dest, xattr->name_len);
695 memcpy(dest, xattr->name, xattr->name_len);
696 dest += xattr->name_len;
697 ceph_encode_32(&dest, xattr->val_len);
698 memcpy(dest, xattr->val, xattr->val_len);
699 dest += xattr->val_len;
700
701 p = rb_next(p);
702 }
703
704 /* adjust buffer len; it may be larger than we need */
705 ci->i_xattrs.prealloc_blob->vec.iov_len =
706 dest - ci->i_xattrs.prealloc_blob->vec.iov_base;
707
Sage Weilb6c1d5b2009-12-07 12:17:17 -0800708 if (ci->i_xattrs.blob)
709 ceph_buffer_put(ci->i_xattrs.blob);
Sage Weil355da1e2009-10-06 11:31:08 -0700710 ci->i_xattrs.blob = ci->i_xattrs.prealloc_blob;
711 ci->i_xattrs.prealloc_blob = NULL;
712 ci->i_xattrs.dirty = false;
Sage Weil4a625be2010-08-22 15:03:56 -0700713 ci->i_xattrs.version++;
Sage Weil355da1e2009-10-06 11:31:08 -0700714 }
715}
716
Guangliang Zhao7221fe42013-11-11 15:18:03 +0800717ssize_t __ceph_getxattr(struct inode *inode, const char *name, void *value,
Sage Weil355da1e2009-10-06 11:31:08 -0700718 size_t size)
719{
Sage Weil355da1e2009-10-06 11:31:08 -0700720 struct ceph_inode_info *ci = ceph_inode(inode);
Sage Weil355da1e2009-10-06 11:31:08 -0700721 int err;
722 struct ceph_inode_xattr *xattr;
Alex Elder881a5fa2012-01-23 15:49:28 -0600723 struct ceph_vxattr *vxattr = NULL;
Sage Weil355da1e2009-10-06 11:31:08 -0700724
725 if (!ceph_is_valid_xattr(name))
726 return -ENODATA;
727
Sage Weil0bee82f2013-01-20 22:00:58 -0800728 /* let's see if a virtual xattr was requested */
729 vxattr = ceph_match_vxattr(inode, name);
730 if (vxattr && !(vxattr->exists_cb && !vxattr->exists_cb(ci))) {
731 err = vxattr->getxattr_cb(ci, value, size);
majianpenga1dc1932013-06-19 14:58:10 +0800732 return err;
Sage Weil0bee82f2013-01-20 22:00:58 -0800733 }
734
majianpenga1dc1932013-06-19 14:58:10 +0800735 spin_lock(&ci->i_ceph_lock);
736 dout("getxattr %p ver=%lld index_ver=%lld\n", inode,
737 ci->i_xattrs.version, ci->i_xattrs.index_version);
738
Yan, Zheng508b32d2014-09-16 21:46:17 +0800739 if (ci->i_xattrs.version == 0 ||
740 !__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1)) {
Sage Weilbe655592011-11-30 09:47:09 -0800741 spin_unlock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700742 /* get xattrs from mds (if we don't already have them) */
Yan, Zheng508b32d2014-09-16 21:46:17 +0800743 err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR, true);
Sage Weil355da1e2009-10-06 11:31:08 -0700744 if (err)
745 return err;
Yan, Zheng508b32d2014-09-16 21:46:17 +0800746 spin_lock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700747 }
748
Sage Weil355da1e2009-10-06 11:31:08 -0700749 err = __build_xattrs(inode);
750 if (err < 0)
751 goto out;
752
Sage Weil355da1e2009-10-06 11:31:08 -0700753 err = -ENODATA; /* == ENOATTR */
754 xattr = __get_xattr(ci, name);
Sage Weil0bee82f2013-01-20 22:00:58 -0800755 if (!xattr)
Sage Weil355da1e2009-10-06 11:31:08 -0700756 goto out;
Sage Weil355da1e2009-10-06 11:31:08 -0700757
758 err = -ERANGE;
759 if (size && size < xattr->val_len)
760 goto out;
761
762 err = xattr->val_len;
763 if (size == 0)
764 goto out;
765
766 memcpy(value, xattr->val, xattr->val_len);
767
768out:
Sage Weilbe655592011-11-30 09:47:09 -0800769 spin_unlock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700770 return err;
771}
772
Guangliang Zhao7221fe42013-11-11 15:18:03 +0800773ssize_t ceph_getxattr(struct dentry *dentry, const char *name, void *value,
774 size_t size)
775{
776 if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
777 return generic_getxattr(dentry, name, value, size);
778
David Howells2b0143b2015-03-17 22:25:59 +0000779 return __ceph_getxattr(d_inode(dentry), name, value, size);
Guangliang Zhao7221fe42013-11-11 15:18:03 +0800780}
781
Sage Weil355da1e2009-10-06 11:31:08 -0700782ssize_t ceph_listxattr(struct dentry *dentry, char *names, size_t size)
783{
David Howells2b0143b2015-03-17 22:25:59 +0000784 struct inode *inode = d_inode(dentry);
Sage Weil355da1e2009-10-06 11:31:08 -0700785 struct ceph_inode_info *ci = ceph_inode(inode);
Alex Elder881a5fa2012-01-23 15:49:28 -0600786 struct ceph_vxattr *vxattrs = ceph_inode_vxattrs(inode);
Sage Weil355da1e2009-10-06 11:31:08 -0700787 u32 vir_namelen = 0;
788 u32 namelen;
789 int err;
790 u32 len;
791 int i;
792
Sage Weilbe655592011-11-30 09:47:09 -0800793 spin_lock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700794 dout("listxattr %p ver=%lld index_ver=%lld\n", inode,
795 ci->i_xattrs.version, ci->i_xattrs.index_version);
796
Yan, Zheng508b32d2014-09-16 21:46:17 +0800797 if (ci->i_xattrs.version == 0 ||
798 !__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1)) {
Sage Weilbe655592011-11-30 09:47:09 -0800799 spin_unlock(&ci->i_ceph_lock);
Yan, Zheng508b32d2014-09-16 21:46:17 +0800800 err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR, true);
Sage Weil355da1e2009-10-06 11:31:08 -0700801 if (err)
802 return err;
Yan, Zheng508b32d2014-09-16 21:46:17 +0800803 spin_lock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700804 }
805
Sage Weil355da1e2009-10-06 11:31:08 -0700806 err = __build_xattrs(inode);
807 if (err < 0)
808 goto out;
Alex Elder3ce6cd12012-01-23 15:49:28 -0600809 /*
810 * Start with virtual dir xattr names (if any) (including
811 * terminating '\0' characters for each).
812 */
813 vir_namelen = ceph_vxattrs_name_size(vxattrs);
814
Sage Weil355da1e2009-10-06 11:31:08 -0700815 /* adding 1 byte per each variable due to the null termination */
Sage Weilb65917d2013-01-20 22:02:39 -0800816 namelen = ci->i_xattrs.names_size + ci->i_xattrs.count;
Sage Weil355da1e2009-10-06 11:31:08 -0700817 err = -ERANGE;
Sage Weilb65917d2013-01-20 22:02:39 -0800818 if (size && vir_namelen + namelen > size)
Sage Weil355da1e2009-10-06 11:31:08 -0700819 goto out;
820
Sage Weilb65917d2013-01-20 22:02:39 -0800821 err = namelen + vir_namelen;
Sage Weil355da1e2009-10-06 11:31:08 -0700822 if (size == 0)
823 goto out;
824
825 names = __copy_xattr_names(ci, names);
826
827 /* virtual xattr names, too */
Sage Weilb65917d2013-01-20 22:02:39 -0800828 err = namelen;
829 if (vxattrs) {
Sage Weil355da1e2009-10-06 11:31:08 -0700830 for (i = 0; vxattrs[i].name; i++) {
Sage Weilb65917d2013-01-20 22:02:39 -0800831 if (!vxattrs[i].hidden &&
832 !(vxattrs[i].exists_cb &&
833 !vxattrs[i].exists_cb(ci))) {
834 len = sprintf(names, "%s", vxattrs[i].name);
835 names += len + 1;
836 err += len + 1;
837 }
Sage Weil355da1e2009-10-06 11:31:08 -0700838 }
Sage Weilb65917d2013-01-20 22:02:39 -0800839 }
Sage Weil355da1e2009-10-06 11:31:08 -0700840
841out:
Sage Weilbe655592011-11-30 09:47:09 -0800842 spin_unlock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700843 return err;
844}
845
846static int ceph_sync_setxattr(struct dentry *dentry, const char *name,
847 const char *value, size_t size, int flags)
848{
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700849 struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
David Howells2b0143b2015-03-17 22:25:59 +0000850 struct inode *inode = d_inode(dentry);
Sage Weil355da1e2009-10-06 11:31:08 -0700851 struct ceph_inode_info *ci = ceph_inode(inode);
Sage Weil355da1e2009-10-06 11:31:08 -0700852 struct ceph_mds_request *req;
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700853 struct ceph_mds_client *mdsc = fsc->mdsc;
Yan, Zheng25e6bae2014-09-16 19:15:28 +0800854 struct ceph_pagelist *pagelist = NULL;
Sage Weil355da1e2009-10-06 11:31:08 -0700855 int err;
Sage Weil355da1e2009-10-06 11:31:08 -0700856
Yan, Zheng0aeff372014-12-17 21:26:47 +0800857 if (size > 0) {
Yan, Zheng25e6bae2014-09-16 19:15:28 +0800858 /* copy value into pagelist */
859 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
860 if (!pagelist)
Sage Weil355da1e2009-10-06 11:31:08 -0700861 return -ENOMEM;
Yan, Zheng25e6bae2014-09-16 19:15:28 +0800862
863 ceph_pagelist_init(pagelist);
864 err = ceph_pagelist_append(pagelist, value, size);
865 if (err)
866 goto out;
Yan, Zheng0aeff372014-12-17 21:26:47 +0800867 } else if (!value) {
Yan, Zheng25e6bae2014-09-16 19:15:28 +0800868 flags |= CEPH_XATTR_REMOVE;
Sage Weil355da1e2009-10-06 11:31:08 -0700869 }
870
871 dout("setxattr value=%.*s\n", (int)size, value);
872
873 /* do request */
874 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETXATTR,
875 USE_AUTH_MDS);
Julia Lawall60d87732009-11-21 12:53:08 +0100876 if (IS_ERR(req)) {
877 err = PTR_ERR(req);
878 goto out;
879 }
Sanidhya Kashyapa149bb92015-03-21 12:54:58 -0400880
Sage Weil355da1e2009-10-06 11:31:08 -0700881 req->r_args.setxattr.flags = cpu_to_le32(flags);
882 req->r_path2 = kstrdup(name, GFP_NOFS);
Sanidhya Kashyapa149bb92015-03-21 12:54:58 -0400883 if (!req->r_path2) {
884 ceph_mdsc_put_request(req);
885 err = -ENOMEM;
886 goto out;
887 }
Sage Weil355da1e2009-10-06 11:31:08 -0700888
Yan, Zheng25e6bae2014-09-16 19:15:28 +0800889 req->r_pagelist = pagelist;
890 pagelist = NULL;
Sage Weil355da1e2009-10-06 11:31:08 -0700891
Sanidhya Kashyapa149bb92015-03-21 12:54:58 -0400892 req->r_inode = inode;
893 ihold(inode);
894 req->r_num_caps = 1;
895 req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
896
Sage Weil355da1e2009-10-06 11:31:08 -0700897 dout("xattr.ver (before): %lld\n", ci->i_xattrs.version);
Sage Weil752c8bd2013-02-05 13:52:29 -0800898 err = ceph_mdsc_do_request(mdsc, NULL, req);
Sage Weil355da1e2009-10-06 11:31:08 -0700899 ceph_mdsc_put_request(req);
900 dout("xattr.ver (after): %lld\n", ci->i_xattrs.version);
901
902out:
Yan, Zheng25e6bae2014-09-16 19:15:28 +0800903 if (pagelist)
904 ceph_pagelist_release(pagelist);
Sage Weil355da1e2009-10-06 11:31:08 -0700905 return err;
906}
907
Guangliang Zhao7221fe42013-11-11 15:18:03 +0800908int __ceph_setxattr(struct dentry *dentry, const char *name,
909 const void *value, size_t size, int flags)
Sage Weil355da1e2009-10-06 11:31:08 -0700910{
David Howells2b0143b2015-03-17 22:25:59 +0000911 struct inode *inode = d_inode(dentry);
Alex Elder881a5fa2012-01-23 15:49:28 -0600912 struct ceph_vxattr *vxattr;
Sage Weil355da1e2009-10-06 11:31:08 -0700913 struct ceph_inode_info *ci = ceph_inode(inode);
Alex Elder18fa8b32012-01-23 15:49:28 -0600914 int issued;
Sage Weil355da1e2009-10-06 11:31:08 -0700915 int err;
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800916 int dirty = 0;
Sage Weil355da1e2009-10-06 11:31:08 -0700917 int name_len = strlen(name);
918 int val_len = size;
919 char *newname = NULL;
920 char *newval = NULL;
921 struct ceph_inode_xattr *xattr = NULL;
Sage Weil355da1e2009-10-06 11:31:08 -0700922 int required_blob_size;
923
Sage Weil355da1e2009-10-06 11:31:08 -0700924 if (!ceph_is_valid_xattr(name))
925 return -EOPNOTSUPP;
926
Alex Elder06476a62012-01-23 15:49:27 -0600927 vxattr = ceph_match_vxattr(inode, name);
928 if (vxattr && vxattr->readonly)
929 return -EOPNOTSUPP;
Sage Weil355da1e2009-10-06 11:31:08 -0700930
Sage Weil3adf6542013-01-31 11:53:41 -0800931 /* pass any unhandled ceph.* xattrs through to the MDS */
932 if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
933 goto do_sync_unlocked;
934
Sage Weil355da1e2009-10-06 11:31:08 -0700935 /* preallocate memory for xattr name, value, index node */
936 err = -ENOMEM;
Julia Lawall61413c22010-10-17 21:55:21 +0200937 newname = kmemdup(name, name_len + 1, GFP_NOFS);
Sage Weil355da1e2009-10-06 11:31:08 -0700938 if (!newname)
939 goto out;
Sage Weil355da1e2009-10-06 11:31:08 -0700940
941 if (val_len) {
Alex Elderb829c192012-01-23 15:49:27 -0600942 newval = kmemdup(value, val_len, GFP_NOFS);
Sage Weil355da1e2009-10-06 11:31:08 -0700943 if (!newval)
944 goto out;
Sage Weil355da1e2009-10-06 11:31:08 -0700945 }
946
947 xattr = kmalloc(sizeof(struct ceph_inode_xattr), GFP_NOFS);
948 if (!xattr)
949 goto out;
950
Sage Weilbe655592011-11-30 09:47:09 -0800951 spin_lock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700952retry:
953 issued = __ceph_caps_issued(ci, NULL);
Alex Elder18fa8b32012-01-23 15:49:28 -0600954 dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
Yan, Zheng508b32d2014-09-16 21:46:17 +0800955 if (ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL))
Sage Weil355da1e2009-10-06 11:31:08 -0700956 goto do_sync;
957 __build_xattrs(inode);
958
959 required_blob_size = __get_required_blob_size(ci, name_len, val_len);
960
961 if (!ci->i_xattrs.prealloc_blob ||
962 required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
Alex Elder18fa8b32012-01-23 15:49:28 -0600963 struct ceph_buffer *blob;
Sage Weil355da1e2009-10-06 11:31:08 -0700964
Sage Weilbe655592011-11-30 09:47:09 -0800965 spin_unlock(&ci->i_ceph_lock);
Sage Weil355da1e2009-10-06 11:31:08 -0700966 dout(" preaallocating new blob size=%d\n", required_blob_size);
Sage Weilb6c1d5b2009-12-07 12:17:17 -0800967 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
Sage Weil355da1e2009-10-06 11:31:08 -0700968 if (!blob)
969 goto out;
Sage Weilbe655592011-11-30 09:47:09 -0800970 spin_lock(&ci->i_ceph_lock);
Sage Weilb6c1d5b2009-12-07 12:17:17 -0800971 if (ci->i_xattrs.prealloc_blob)
972 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
Sage Weil355da1e2009-10-06 11:31:08 -0700973 ci->i_xattrs.prealloc_blob = blob;
974 goto retry;
975 }
976
Yan, Zhengbcdfeb22014-02-11 13:04:19 +0800977 err = __set_xattr(ci, newname, name_len, newval, val_len,
978 flags, value ? 1 : -1, &xattr);
Alex Elder18fa8b32012-01-23 15:49:28 -0600979
Yan, Zhengfbc0b972014-02-11 13:01:19 +0800980 if (!err) {
981 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
982 ci->i_xattrs.dirty = true;
983 inode->i_ctime = CURRENT_TIME;
984 }
Alex Elder18fa8b32012-01-23 15:49:28 -0600985
Sage Weilbe655592011-11-30 09:47:09 -0800986 spin_unlock(&ci->i_ceph_lock);
Sage Weilfca65b42011-05-04 11:33:47 -0700987 if (dirty)
988 __mark_inode_dirty(inode, dirty);
Sage Weil355da1e2009-10-06 11:31:08 -0700989 return err;
990
991do_sync:
Sage Weilbe655592011-11-30 09:47:09 -0800992 spin_unlock(&ci->i_ceph_lock);
Sage Weil3adf6542013-01-31 11:53:41 -0800993do_sync_unlocked:
Sage Weil355da1e2009-10-06 11:31:08 -0700994 err = ceph_sync_setxattr(dentry, name, value, size, flags);
995out:
996 kfree(newname);
997 kfree(newval);
998 kfree(xattr);
999 return err;
1000}
1001
Guangliang Zhao7221fe42013-11-11 15:18:03 +08001002int ceph_setxattr(struct dentry *dentry, const char *name,
1003 const void *value, size_t size, int flags)
1004{
David Howells2b0143b2015-03-17 22:25:59 +00001005 if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
Guangliang Zhao7221fe42013-11-11 15:18:03 +08001006 return -EROFS;
1007
1008 if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
1009 return generic_setxattr(dentry, name, value, size, flags);
1010
Yan, Zheng0aeff372014-12-17 21:26:47 +08001011 if (size == 0)
1012 value = ""; /* empty EA, do not remove */
1013
Guangliang Zhao7221fe42013-11-11 15:18:03 +08001014 return __ceph_setxattr(dentry, name, value, size, flags);
1015}
1016
Sage Weil355da1e2009-10-06 11:31:08 -07001017static int ceph_send_removexattr(struct dentry *dentry, const char *name)
1018{
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001019 struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1020 struct ceph_mds_client *mdsc = fsc->mdsc;
David Howells2b0143b2015-03-17 22:25:59 +00001021 struct inode *inode = d_inode(dentry);
Sage Weil355da1e2009-10-06 11:31:08 -07001022 struct ceph_mds_request *req;
1023 int err;
1024
1025 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RMXATTR,
1026 USE_AUTH_MDS);
1027 if (IS_ERR(req))
1028 return PTR_ERR(req);
Sanidhya Kashyapa149bb92015-03-21 12:54:58 -04001029 req->r_path2 = kstrdup(name, GFP_NOFS);
1030 if (!req->r_path2)
1031 return -ENOMEM;
1032
Sage Weil70b666c2011-05-27 09:24:26 -07001033 req->r_inode = inode;
1034 ihold(inode);
Sage Weil355da1e2009-10-06 11:31:08 -07001035 req->r_num_caps = 1;
Sanidhya Kashyapa149bb92015-03-21 12:54:58 -04001036 req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
Sage Weil752c8bd2013-02-05 13:52:29 -08001037 err = ceph_mdsc_do_request(mdsc, NULL, req);
Sage Weil355da1e2009-10-06 11:31:08 -07001038 ceph_mdsc_put_request(req);
1039 return err;
1040}
1041
Guangliang Zhao7221fe42013-11-11 15:18:03 +08001042int __ceph_removexattr(struct dentry *dentry, const char *name)
Sage Weil355da1e2009-10-06 11:31:08 -07001043{
David Howells2b0143b2015-03-17 22:25:59 +00001044 struct inode *inode = d_inode(dentry);
Alex Elder881a5fa2012-01-23 15:49:28 -06001045 struct ceph_vxattr *vxattr;
Sage Weil355da1e2009-10-06 11:31:08 -07001046 struct ceph_inode_info *ci = ceph_inode(inode);
Sage Weil355da1e2009-10-06 11:31:08 -07001047 int issued;
1048 int err;
Alex Elder83eb26a2012-01-11 17:41:01 -08001049 int required_blob_size;
Sage Weilfca65b42011-05-04 11:33:47 -07001050 int dirty;
Sage Weil355da1e2009-10-06 11:31:08 -07001051
Sage Weil355da1e2009-10-06 11:31:08 -07001052 if (!ceph_is_valid_xattr(name))
1053 return -EOPNOTSUPP;
1054
Alex Elder06476a62012-01-23 15:49:27 -06001055 vxattr = ceph_match_vxattr(inode, name);
1056 if (vxattr && vxattr->readonly)
1057 return -EOPNOTSUPP;
Sage Weil355da1e2009-10-06 11:31:08 -07001058
Sage Weild421acb2013-01-20 21:55:30 -08001059 /* pass any unhandled ceph.* xattrs through to the MDS */
1060 if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
1061 goto do_sync_unlocked;
1062
Alex Elder83eb26a2012-01-11 17:41:01 -08001063 err = -ENOMEM;
Sage Weilbe655592011-11-30 09:47:09 -08001064 spin_lock(&ci->i_ceph_lock);
Alex Elder83eb26a2012-01-11 17:41:01 -08001065retry:
Sage Weil355da1e2009-10-06 11:31:08 -07001066 issued = __ceph_caps_issued(ci, NULL);
1067 dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
1068
Yan, Zheng508b32d2014-09-16 21:46:17 +08001069 if (ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL))
Sage Weil355da1e2009-10-06 11:31:08 -07001070 goto do_sync;
Alex Elder18fa8b32012-01-23 15:49:28 -06001071 __build_xattrs(inode);
Sage Weil355da1e2009-10-06 11:31:08 -07001072
Alex Elder83eb26a2012-01-11 17:41:01 -08001073 required_blob_size = __get_required_blob_size(ci, 0, 0);
1074
1075 if (!ci->i_xattrs.prealloc_blob ||
1076 required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
1077 struct ceph_buffer *blob;
1078
1079 spin_unlock(&ci->i_ceph_lock);
1080 dout(" preaallocating new blob size=%d\n", required_blob_size);
1081 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
1082 if (!blob)
1083 goto out;
1084 spin_lock(&ci->i_ceph_lock);
1085 if (ci->i_xattrs.prealloc_blob)
1086 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
1087 ci->i_xattrs.prealloc_blob = blob;
1088 goto retry;
1089 }
1090
Sage Weil355da1e2009-10-06 11:31:08 -07001091 err = __remove_xattr_by_name(ceph_inode(inode), name);
Alex Elder18fa8b32012-01-23 15:49:28 -06001092
Sage Weilfca65b42011-05-04 11:33:47 -07001093 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
Sage Weil355da1e2009-10-06 11:31:08 -07001094 ci->i_xattrs.dirty = true;
1095 inode->i_ctime = CURRENT_TIME;
Sage Weilbe655592011-11-30 09:47:09 -08001096 spin_unlock(&ci->i_ceph_lock);
Sage Weilfca65b42011-05-04 11:33:47 -07001097 if (dirty)
1098 __mark_inode_dirty(inode, dirty);
Sage Weil355da1e2009-10-06 11:31:08 -07001099 return err;
1100do_sync:
Sage Weilbe655592011-11-30 09:47:09 -08001101 spin_unlock(&ci->i_ceph_lock);
Sage Weild421acb2013-01-20 21:55:30 -08001102do_sync_unlocked:
Sage Weil355da1e2009-10-06 11:31:08 -07001103 err = ceph_send_removexattr(dentry, name);
Alex Elder83eb26a2012-01-11 17:41:01 -08001104out:
Sage Weil355da1e2009-10-06 11:31:08 -07001105 return err;
1106}
1107
Guangliang Zhao7221fe42013-11-11 15:18:03 +08001108int ceph_removexattr(struct dentry *dentry, const char *name)
1109{
David Howells2b0143b2015-03-17 22:25:59 +00001110 if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
Guangliang Zhao7221fe42013-11-11 15:18:03 +08001111 return -EROFS;
1112
1113 if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
1114 return generic_removexattr(dentry, name);
1115
1116 return __ceph_removexattr(dentry, name);
1117}