xref: /openbmc/linux/fs/ceph/xattr.c (revision 7221fe4c2ed72804b28633c8e0217d65abb0023f)
13d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h>
23d14c5d2SYehuda Sadeh 
3355da1ebSSage Weil #include "super.h"
43d14c5d2SYehuda Sadeh #include "mds_client.h"
53d14c5d2SYehuda Sadeh 
63d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h>
7355da1ebSSage Weil 
8355da1ebSSage Weil #include <linux/xattr.h>
95a0e3ad6STejun Heo #include <linux/slab.h>
10355da1ebSSage Weil 
1122891907SAlex Elder #define XATTR_CEPH_PREFIX "ceph."
1222891907SAlex Elder #define XATTR_CEPH_PREFIX_LEN (sizeof (XATTR_CEPH_PREFIX) - 1)
1322891907SAlex Elder 
14*7221fe4cSGuangliang Zhao /*
15*7221fe4cSGuangliang Zhao  * List of handlers for synthetic system.* attributes. Other
16*7221fe4cSGuangliang Zhao  * attributes are handled directly.
17*7221fe4cSGuangliang Zhao  */
18*7221fe4cSGuangliang Zhao const struct xattr_handler *ceph_xattr_handlers[] = {
19*7221fe4cSGuangliang Zhao #ifdef CONFIG_CEPH_FS_POSIX_ACL
20*7221fe4cSGuangliang Zhao 	&ceph_xattr_acl_access_handler,
21*7221fe4cSGuangliang Zhao 	&ceph_xattr_acl_default_handler,
22*7221fe4cSGuangliang Zhao #endif
23*7221fe4cSGuangliang Zhao 	NULL,
24*7221fe4cSGuangliang Zhao };
25*7221fe4cSGuangliang Zhao 
26355da1ebSSage Weil static bool ceph_is_valid_xattr(const char *name)
27355da1ebSSage Weil {
2822891907SAlex Elder 	return !strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN) ||
291a756278SSage Weil 	       !strncmp(name, XATTR_SECURITY_PREFIX,
30355da1ebSSage Weil 			XATTR_SECURITY_PREFIX_LEN) ||
31*7221fe4cSGuangliang Zhao 	       !strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN) ||
32355da1ebSSage Weil 	       !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
33355da1ebSSage Weil 	       !strncmp(name, XATTR_USER_PREFIX, XATTR_USER_PREFIX_LEN);
34355da1ebSSage Weil }
35355da1ebSSage Weil 
36355da1ebSSage Weil /*
37355da1ebSSage Weil  * These define virtual xattrs exposing the recursive directory
38355da1ebSSage Weil  * statistics and layout metadata.
39355da1ebSSage Weil  */
40881a5fa2SAlex Elder struct ceph_vxattr {
41355da1ebSSage Weil 	char *name;
423ce6cd12SAlex Elder 	size_t name_size;	/* strlen(name) + 1 (for '\0') */
43355da1ebSSage Weil 	size_t (*getxattr_cb)(struct ceph_inode_info *ci, char *val,
44355da1ebSSage Weil 			      size_t size);
458860147aSSage Weil 	bool readonly, hidden;
46f36e4472SSage Weil 	bool (*exists_cb)(struct ceph_inode_info *ci);
47355da1ebSSage Weil };
48355da1ebSSage Weil 
4932ab0bd7SSage Weil /* layouts */
5032ab0bd7SSage Weil 
5132ab0bd7SSage Weil static bool ceph_vxattrcb_layout_exists(struct ceph_inode_info *ci)
5232ab0bd7SSage Weil {
5332ab0bd7SSage Weil 	size_t s;
5432ab0bd7SSage Weil 	char *p = (char *)&ci->i_layout;
5532ab0bd7SSage Weil 
5632ab0bd7SSage Weil 	for (s = 0; s < sizeof(ci->i_layout); s++, p++)
5732ab0bd7SSage Weil 		if (*p)
5832ab0bd7SSage Weil 			return true;
5932ab0bd7SSage Weil 	return false;
6032ab0bd7SSage Weil }
6132ab0bd7SSage Weil 
6232ab0bd7SSage Weil static size_t ceph_vxattrcb_layout(struct ceph_inode_info *ci, char *val,
6332ab0bd7SSage Weil 					size_t size)
6432ab0bd7SSage Weil {
6532ab0bd7SSage Weil 	int ret;
6632ab0bd7SSage Weil 	struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
6732ab0bd7SSage Weil 	struct ceph_osd_client *osdc = &fsc->client->osdc;
6832ab0bd7SSage Weil 	s64 pool = ceph_file_layout_pg_pool(ci->i_layout);
6932ab0bd7SSage Weil 	const char *pool_name;
7032ab0bd7SSage Weil 
7132ab0bd7SSage Weil 	dout("ceph_vxattrcb_layout %p\n", &ci->vfs_inode);
7232ab0bd7SSage Weil 	down_read(&osdc->map_sem);
7332ab0bd7SSage Weil 	pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, pool);
7432ab0bd7SSage Weil 	if (pool_name)
7532ab0bd7SSage Weil 		ret = snprintf(val, size,
7632ab0bd7SSage Weil 		"stripe_unit=%lld stripe_count=%lld object_size=%lld pool=%s",
7732ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_su(ci->i_layout),
7832ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
7932ab0bd7SSage Weil 	        (unsigned long long)ceph_file_layout_object_size(ci->i_layout),
8032ab0bd7SSage Weil 		pool_name);
8132ab0bd7SSage Weil 	else
8232ab0bd7SSage Weil 		ret = snprintf(val, size,
8332ab0bd7SSage Weil 		"stripe_unit=%lld stripe_count=%lld object_size=%lld pool=%lld",
8432ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_su(ci->i_layout),
8532ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
8632ab0bd7SSage Weil 	        (unsigned long long)ceph_file_layout_object_size(ci->i_layout),
8732ab0bd7SSage Weil 		(unsigned long long)pool);
8832ab0bd7SSage Weil 
8932ab0bd7SSage Weil 	up_read(&osdc->map_sem);
9032ab0bd7SSage Weil 	return ret;
9132ab0bd7SSage Weil }
9232ab0bd7SSage Weil 
93695b7119SSage Weil static size_t ceph_vxattrcb_layout_stripe_unit(struct ceph_inode_info *ci,
94695b7119SSage Weil 					       char *val, size_t size)
95695b7119SSage Weil {
96695b7119SSage Weil 	return snprintf(val, size, "%lld",
97695b7119SSage Weil 			(unsigned long long)ceph_file_layout_su(ci->i_layout));
98695b7119SSage Weil }
99695b7119SSage Weil 
100695b7119SSage Weil static size_t ceph_vxattrcb_layout_stripe_count(struct ceph_inode_info *ci,
101695b7119SSage Weil 						char *val, size_t size)
102695b7119SSage Weil {
103695b7119SSage Weil 	return snprintf(val, size, "%lld",
104695b7119SSage Weil 	       (unsigned long long)ceph_file_layout_stripe_count(ci->i_layout));
105695b7119SSage Weil }
106695b7119SSage Weil 
107695b7119SSage Weil static size_t ceph_vxattrcb_layout_object_size(struct ceph_inode_info *ci,
108695b7119SSage Weil 					       char *val, size_t size)
109695b7119SSage Weil {
110695b7119SSage Weil 	return snprintf(val, size, "%lld",
111695b7119SSage Weil 	       (unsigned long long)ceph_file_layout_object_size(ci->i_layout));
112695b7119SSage Weil }
113695b7119SSage Weil 
114695b7119SSage Weil static size_t ceph_vxattrcb_layout_pool(struct ceph_inode_info *ci,
115695b7119SSage Weil 					char *val, size_t size)
116695b7119SSage Weil {
117695b7119SSage Weil 	int ret;
118695b7119SSage Weil 	struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
119695b7119SSage Weil 	struct ceph_osd_client *osdc = &fsc->client->osdc;
120695b7119SSage Weil 	s64 pool = ceph_file_layout_pg_pool(ci->i_layout);
121695b7119SSage Weil 	const char *pool_name;
122695b7119SSage Weil 
123695b7119SSage Weil 	down_read(&osdc->map_sem);
124695b7119SSage Weil 	pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, pool);
125695b7119SSage Weil 	if (pool_name)
126695b7119SSage Weil 		ret = snprintf(val, size, "%s", pool_name);
127695b7119SSage Weil 	else
128695b7119SSage Weil 		ret = snprintf(val, size, "%lld", (unsigned long long)pool);
129695b7119SSage Weil 	up_read(&osdc->map_sem);
130695b7119SSage Weil 	return ret;
131695b7119SSage Weil }
132695b7119SSage Weil 
133355da1ebSSage Weil /* directories */
134355da1ebSSage Weil 
135aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_entries(struct ceph_inode_info *ci, char *val,
136355da1ebSSage Weil 					size_t size)
137355da1ebSSage Weil {
138355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_files + ci->i_subdirs);
139355da1ebSSage Weil }
140355da1ebSSage Weil 
141aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_files(struct ceph_inode_info *ci, char *val,
142355da1ebSSage Weil 				      size_t size)
143355da1ebSSage Weil {
144355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_files);
145355da1ebSSage Weil }
146355da1ebSSage Weil 
147aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_subdirs(struct ceph_inode_info *ci, char *val,
148355da1ebSSage Weil 					size_t size)
149355da1ebSSage Weil {
150355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_subdirs);
151355da1ebSSage Weil }
152355da1ebSSage Weil 
153aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rentries(struct ceph_inode_info *ci, char *val,
154355da1ebSSage Weil 					 size_t size)
155355da1ebSSage Weil {
156355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rfiles + ci->i_rsubdirs);
157355da1ebSSage Weil }
158355da1ebSSage Weil 
159aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rfiles(struct ceph_inode_info *ci, char *val,
160355da1ebSSage Weil 				       size_t size)
161355da1ebSSage Weil {
162355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rfiles);
163355da1ebSSage Weil }
164355da1ebSSage Weil 
165aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rsubdirs(struct ceph_inode_info *ci, char *val,
166355da1ebSSage Weil 					 size_t size)
167355da1ebSSage Weil {
168355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rsubdirs);
169355da1ebSSage Weil }
170355da1ebSSage Weil 
171aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rbytes(struct ceph_inode_info *ci, char *val,
172355da1ebSSage Weil 				       size_t size)
173355da1ebSSage Weil {
174355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rbytes);
175355da1ebSSage Weil }
176355da1ebSSage Weil 
177aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rctime(struct ceph_inode_info *ci, char *val,
178355da1ebSSage Weil 				       size_t size)
179355da1ebSSage Weil {
1803489b42aSAlex Elder 	return snprintf(val, size, "%ld.09%ld", (long)ci->i_rctime.tv_sec,
181355da1ebSSage Weil 			(long)ci->i_rctime.tv_nsec);
182355da1ebSSage Weil }
183355da1ebSSage Weil 
18432ab0bd7SSage Weil 
185eb788084SAlex Elder #define CEPH_XATTR_NAME(_type, _name)	XATTR_CEPH_PREFIX #_type "." #_name
186695b7119SSage Weil #define CEPH_XATTR_NAME2(_type, _name, _name2)	\
187695b7119SSage Weil 	XATTR_CEPH_PREFIX #_type "." #_name "." #_name2
188eb788084SAlex Elder 
189eb788084SAlex Elder #define XATTR_NAME_CEPH(_type, _name)					\
190eb788084SAlex Elder 	{								\
191eb788084SAlex Elder 		.name = CEPH_XATTR_NAME(_type, _name),			\
1923ce6cd12SAlex Elder 		.name_size = sizeof (CEPH_XATTR_NAME(_type, _name)), \
193aa4066edSAlex Elder 		.getxattr_cb = ceph_vxattrcb_ ## _type ## _ ## _name, \
194eb788084SAlex Elder 		.readonly = true,				\
1958860147aSSage Weil 		.hidden = false,				\
196f36e4472SSage Weil 		.exists_cb = NULL,			\
197eb788084SAlex Elder 	}
198695b7119SSage Weil #define XATTR_LAYOUT_FIELD(_type, _name, _field)			\
199695b7119SSage Weil 	{								\
200695b7119SSage Weil 		.name = CEPH_XATTR_NAME2(_type, _name, _field),	\
201695b7119SSage Weil 		.name_size = sizeof (CEPH_XATTR_NAME2(_type, _name, _field)), \
202695b7119SSage Weil 		.getxattr_cb = ceph_vxattrcb_ ## _name ## _ ## _field, \
203695b7119SSage Weil 		.readonly = false,				\
204695b7119SSage Weil 		.hidden = true,			\
205695b7119SSage Weil 		.exists_cb = ceph_vxattrcb_layout_exists,	\
206695b7119SSage Weil 	}
207eb788084SAlex Elder 
208881a5fa2SAlex Elder static struct ceph_vxattr ceph_dir_vxattrs[] = {
2091f08f2b0SSage Weil 	{
2101f08f2b0SSage Weil 		.name = "ceph.dir.layout",
2111f08f2b0SSage Weil 		.name_size = sizeof("ceph.dir.layout"),
2121f08f2b0SSage Weil 		.getxattr_cb = ceph_vxattrcb_layout,
2131f08f2b0SSage Weil 		.readonly = false,
2141f08f2b0SSage Weil 		.hidden = false,
2151f08f2b0SSage Weil 		.exists_cb = ceph_vxattrcb_layout_exists,
2161f08f2b0SSage Weil 	},
217695b7119SSage Weil 	XATTR_LAYOUT_FIELD(dir, layout, stripe_unit),
218695b7119SSage Weil 	XATTR_LAYOUT_FIELD(dir, layout, stripe_count),
219695b7119SSage Weil 	XATTR_LAYOUT_FIELD(dir, layout, object_size),
220695b7119SSage Weil 	XATTR_LAYOUT_FIELD(dir, layout, pool),
221eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, entries),
222eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, files),
223eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, subdirs),
224eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rentries),
225eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rfiles),
226eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rsubdirs),
227eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rbytes),
228eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rctime),
2292c3dd4ffSAlex Elder 	{ .name = NULL, 0 }	/* Required table terminator */
230355da1ebSSage Weil };
2313ce6cd12SAlex Elder static size_t ceph_dir_vxattrs_name_size;	/* total size of all names */
232355da1ebSSage Weil 
233355da1ebSSage Weil /* files */
234355da1ebSSage Weil 
235881a5fa2SAlex Elder static struct ceph_vxattr ceph_file_vxattrs[] = {
23632ab0bd7SSage Weil 	{
23732ab0bd7SSage Weil 		.name = "ceph.file.layout",
23832ab0bd7SSage Weil 		.name_size = sizeof("ceph.file.layout"),
23932ab0bd7SSage Weil 		.getxattr_cb = ceph_vxattrcb_layout,
24032ab0bd7SSage Weil 		.readonly = false,
24132ab0bd7SSage Weil 		.hidden = false,
24232ab0bd7SSage Weil 		.exists_cb = ceph_vxattrcb_layout_exists,
24332ab0bd7SSage Weil 	},
244695b7119SSage Weil 	XATTR_LAYOUT_FIELD(file, layout, stripe_unit),
245695b7119SSage Weil 	XATTR_LAYOUT_FIELD(file, layout, stripe_count),
246695b7119SSage Weil 	XATTR_LAYOUT_FIELD(file, layout, object_size),
247695b7119SSage Weil 	XATTR_LAYOUT_FIELD(file, layout, pool),
2482c3dd4ffSAlex Elder 	{ .name = NULL, 0 }	/* Required table terminator */
249355da1ebSSage Weil };
2503ce6cd12SAlex Elder static size_t ceph_file_vxattrs_name_size;	/* total size of all names */
251355da1ebSSage Weil 
252881a5fa2SAlex Elder static struct ceph_vxattr *ceph_inode_vxattrs(struct inode *inode)
253355da1ebSSage Weil {
254355da1ebSSage Weil 	if (S_ISDIR(inode->i_mode))
255355da1ebSSage Weil 		return ceph_dir_vxattrs;
256355da1ebSSage Weil 	else if (S_ISREG(inode->i_mode))
257355da1ebSSage Weil 		return ceph_file_vxattrs;
258355da1ebSSage Weil 	return NULL;
259355da1ebSSage Weil }
260355da1ebSSage Weil 
2613ce6cd12SAlex Elder static size_t ceph_vxattrs_name_size(struct ceph_vxattr *vxattrs)
2623ce6cd12SAlex Elder {
2633ce6cd12SAlex Elder 	if (vxattrs == ceph_dir_vxattrs)
2643ce6cd12SAlex Elder 		return ceph_dir_vxattrs_name_size;
2653ce6cd12SAlex Elder 	if (vxattrs == ceph_file_vxattrs)
2663ce6cd12SAlex Elder 		return ceph_file_vxattrs_name_size;
2673ce6cd12SAlex Elder 	BUG();
2683ce6cd12SAlex Elder 
2693ce6cd12SAlex Elder 	return 0;
2703ce6cd12SAlex Elder }
2713ce6cd12SAlex Elder 
2723ce6cd12SAlex Elder /*
2733ce6cd12SAlex Elder  * Compute the aggregate size (including terminating '\0') of all
2743ce6cd12SAlex Elder  * virtual extended attribute names in the given vxattr table.
2753ce6cd12SAlex Elder  */
2763ce6cd12SAlex Elder static size_t __init vxattrs_name_size(struct ceph_vxattr *vxattrs)
2773ce6cd12SAlex Elder {
2783ce6cd12SAlex Elder 	struct ceph_vxattr *vxattr;
2793ce6cd12SAlex Elder 	size_t size = 0;
2803ce6cd12SAlex Elder 
2813ce6cd12SAlex Elder 	for (vxattr = vxattrs; vxattr->name; vxattr++)
2828860147aSSage Weil 		if (!vxattr->hidden)
2833ce6cd12SAlex Elder 			size += vxattr->name_size;
2843ce6cd12SAlex Elder 
2853ce6cd12SAlex Elder 	return size;
2863ce6cd12SAlex Elder }
2873ce6cd12SAlex Elder 
2883ce6cd12SAlex Elder /* Routines called at initialization and exit time */
2893ce6cd12SAlex Elder 
2903ce6cd12SAlex Elder void __init ceph_xattr_init(void)
2913ce6cd12SAlex Elder {
2923ce6cd12SAlex Elder 	ceph_dir_vxattrs_name_size = vxattrs_name_size(ceph_dir_vxattrs);
2933ce6cd12SAlex Elder 	ceph_file_vxattrs_name_size = vxattrs_name_size(ceph_file_vxattrs);
2943ce6cd12SAlex Elder }
2953ce6cd12SAlex Elder 
2963ce6cd12SAlex Elder void ceph_xattr_exit(void)
2973ce6cd12SAlex Elder {
2983ce6cd12SAlex Elder 	ceph_dir_vxattrs_name_size = 0;
2993ce6cd12SAlex Elder 	ceph_file_vxattrs_name_size = 0;
3003ce6cd12SAlex Elder }
3013ce6cd12SAlex Elder 
302881a5fa2SAlex Elder static struct ceph_vxattr *ceph_match_vxattr(struct inode *inode,
303355da1ebSSage Weil 						const char *name)
304355da1ebSSage Weil {
305881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr = ceph_inode_vxattrs(inode);
30606476a69SAlex Elder 
30706476a69SAlex Elder 	if (vxattr) {
30806476a69SAlex Elder 		while (vxattr->name) {
30906476a69SAlex Elder 			if (!strcmp(vxattr->name, name))
310355da1ebSSage Weil 				return vxattr;
311355da1ebSSage Weil 			vxattr++;
31206476a69SAlex Elder 		}
31306476a69SAlex Elder 	}
31406476a69SAlex Elder 
315355da1ebSSage Weil 	return NULL;
316355da1ebSSage Weil }
317355da1ebSSage Weil 
318355da1ebSSage Weil static int __set_xattr(struct ceph_inode_info *ci,
319355da1ebSSage Weil 			   const char *name, int name_len,
320355da1ebSSage Weil 			   const char *val, int val_len,
321355da1ebSSage Weil 			   int dirty,
322355da1ebSSage Weil 			   int should_free_name, int should_free_val,
323355da1ebSSage Weil 			   struct ceph_inode_xattr **newxattr)
324355da1ebSSage Weil {
325355da1ebSSage Weil 	struct rb_node **p;
326355da1ebSSage Weil 	struct rb_node *parent = NULL;
327355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
328355da1ebSSage Weil 	int c;
329355da1ebSSage Weil 	int new = 0;
330355da1ebSSage Weil 
331355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
332355da1ebSSage Weil 	while (*p) {
333355da1ebSSage Weil 		parent = *p;
334355da1ebSSage Weil 		xattr = rb_entry(parent, struct ceph_inode_xattr, node);
335355da1ebSSage Weil 		c = strncmp(name, xattr->name, min(name_len, xattr->name_len));
336355da1ebSSage Weil 		if (c < 0)
337355da1ebSSage Weil 			p = &(*p)->rb_left;
338355da1ebSSage Weil 		else if (c > 0)
339355da1ebSSage Weil 			p = &(*p)->rb_right;
340355da1ebSSage Weil 		else {
341355da1ebSSage Weil 			if (name_len == xattr->name_len)
342355da1ebSSage Weil 				break;
343355da1ebSSage Weil 			else if (name_len < xattr->name_len)
344355da1ebSSage Weil 				p = &(*p)->rb_left;
345355da1ebSSage Weil 			else
346355da1ebSSage Weil 				p = &(*p)->rb_right;
347355da1ebSSage Weil 		}
348355da1ebSSage Weil 		xattr = NULL;
349355da1ebSSage Weil 	}
350355da1ebSSage Weil 
351355da1ebSSage Weil 	if (!xattr) {
352355da1ebSSage Weil 		new = 1;
353355da1ebSSage Weil 		xattr = *newxattr;
354355da1ebSSage Weil 		xattr->name = name;
355355da1ebSSage Weil 		xattr->name_len = name_len;
356355da1ebSSage Weil 		xattr->should_free_name = should_free_name;
357355da1ebSSage Weil 
358355da1ebSSage Weil 		ci->i_xattrs.count++;
359355da1ebSSage Weil 		dout("__set_xattr count=%d\n", ci->i_xattrs.count);
360355da1ebSSage Weil 	} else {
361355da1ebSSage Weil 		kfree(*newxattr);
362355da1ebSSage Weil 		*newxattr = NULL;
363355da1ebSSage Weil 		if (xattr->should_free_val)
364355da1ebSSage Weil 			kfree((void *)xattr->val);
365355da1ebSSage Weil 
366355da1ebSSage Weil 		if (should_free_name) {
367355da1ebSSage Weil 			kfree((void *)name);
368355da1ebSSage Weil 			name = xattr->name;
369355da1ebSSage Weil 		}
370355da1ebSSage Weil 		ci->i_xattrs.names_size -= xattr->name_len;
371355da1ebSSage Weil 		ci->i_xattrs.vals_size -= xattr->val_len;
372355da1ebSSage Weil 	}
373355da1ebSSage Weil 	ci->i_xattrs.names_size += name_len;
374355da1ebSSage Weil 	ci->i_xattrs.vals_size += val_len;
375355da1ebSSage Weil 	if (val)
376355da1ebSSage Weil 		xattr->val = val;
377355da1ebSSage Weil 	else
378355da1ebSSage Weil 		xattr->val = "";
379355da1ebSSage Weil 
380355da1ebSSage Weil 	xattr->val_len = val_len;
381355da1ebSSage Weil 	xattr->dirty = dirty;
382355da1ebSSage Weil 	xattr->should_free_val = (val && should_free_val);
383355da1ebSSage Weil 
384355da1ebSSage Weil 	if (new) {
385355da1ebSSage Weil 		rb_link_node(&xattr->node, parent, p);
386355da1ebSSage Weil 		rb_insert_color(&xattr->node, &ci->i_xattrs.index);
387355da1ebSSage Weil 		dout("__set_xattr_val p=%p\n", p);
388355da1ebSSage Weil 	}
389355da1ebSSage Weil 
390355da1ebSSage Weil 	dout("__set_xattr_val added %llx.%llx xattr %p %s=%.*s\n",
391355da1ebSSage Weil 	     ceph_vinop(&ci->vfs_inode), xattr, name, val_len, val);
392355da1ebSSage Weil 
393355da1ebSSage Weil 	return 0;
394355da1ebSSage Weil }
395355da1ebSSage Weil 
396355da1ebSSage Weil static struct ceph_inode_xattr *__get_xattr(struct ceph_inode_info *ci,
397355da1ebSSage Weil 			   const char *name)
398355da1ebSSage Weil {
399355da1ebSSage Weil 	struct rb_node **p;
400355da1ebSSage Weil 	struct rb_node *parent = NULL;
401355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
40217db143fSSage Weil 	int name_len = strlen(name);
403355da1ebSSage Weil 	int c;
404355da1ebSSage Weil 
405355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
406355da1ebSSage Weil 	while (*p) {
407355da1ebSSage Weil 		parent = *p;
408355da1ebSSage Weil 		xattr = rb_entry(parent, struct ceph_inode_xattr, node);
409355da1ebSSage Weil 		c = strncmp(name, xattr->name, xattr->name_len);
41017db143fSSage Weil 		if (c == 0 && name_len > xattr->name_len)
41117db143fSSage Weil 			c = 1;
412355da1ebSSage Weil 		if (c < 0)
413355da1ebSSage Weil 			p = &(*p)->rb_left;
414355da1ebSSage Weil 		else if (c > 0)
415355da1ebSSage Weil 			p = &(*p)->rb_right;
416355da1ebSSage Weil 		else {
417355da1ebSSage Weil 			dout("__get_xattr %s: found %.*s\n", name,
418355da1ebSSage Weil 			     xattr->val_len, xattr->val);
419355da1ebSSage Weil 			return xattr;
420355da1ebSSage Weil 		}
421355da1ebSSage Weil 	}
422355da1ebSSage Weil 
423355da1ebSSage Weil 	dout("__get_xattr %s: not found\n", name);
424355da1ebSSage Weil 
425355da1ebSSage Weil 	return NULL;
426355da1ebSSage Weil }
427355da1ebSSage Weil 
428355da1ebSSage Weil static void __free_xattr(struct ceph_inode_xattr *xattr)
429355da1ebSSage Weil {
430355da1ebSSage Weil 	BUG_ON(!xattr);
431355da1ebSSage Weil 
432355da1ebSSage Weil 	if (xattr->should_free_name)
433355da1ebSSage Weil 		kfree((void *)xattr->name);
434355da1ebSSage Weil 	if (xattr->should_free_val)
435355da1ebSSage Weil 		kfree((void *)xattr->val);
436355da1ebSSage Weil 
437355da1ebSSage Weil 	kfree(xattr);
438355da1ebSSage Weil }
439355da1ebSSage Weil 
440355da1ebSSage Weil static int __remove_xattr(struct ceph_inode_info *ci,
441355da1ebSSage Weil 			  struct ceph_inode_xattr *xattr)
442355da1ebSSage Weil {
443355da1ebSSage Weil 	if (!xattr)
444355da1ebSSage Weil 		return -EOPNOTSUPP;
445355da1ebSSage Weil 
446355da1ebSSage Weil 	rb_erase(&xattr->node, &ci->i_xattrs.index);
447355da1ebSSage Weil 
448355da1ebSSage Weil 	if (xattr->should_free_name)
449355da1ebSSage Weil 		kfree((void *)xattr->name);
450355da1ebSSage Weil 	if (xattr->should_free_val)
451355da1ebSSage Weil 		kfree((void *)xattr->val);
452355da1ebSSage Weil 
453355da1ebSSage Weil 	ci->i_xattrs.names_size -= xattr->name_len;
454355da1ebSSage Weil 	ci->i_xattrs.vals_size -= xattr->val_len;
455355da1ebSSage Weil 	ci->i_xattrs.count--;
456355da1ebSSage Weil 	kfree(xattr);
457355da1ebSSage Weil 
458355da1ebSSage Weil 	return 0;
459355da1ebSSage Weil }
460355da1ebSSage Weil 
461355da1ebSSage Weil static int __remove_xattr_by_name(struct ceph_inode_info *ci,
462355da1ebSSage Weil 			   const char *name)
463355da1ebSSage Weil {
464355da1ebSSage Weil 	struct rb_node **p;
465355da1ebSSage Weil 	struct ceph_inode_xattr *xattr;
466355da1ebSSage Weil 	int err;
467355da1ebSSage Weil 
468355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
469355da1ebSSage Weil 	xattr = __get_xattr(ci, name);
470355da1ebSSage Weil 	err = __remove_xattr(ci, xattr);
471355da1ebSSage Weil 	return err;
472355da1ebSSage Weil }
473355da1ebSSage Weil 
474355da1ebSSage Weil static char *__copy_xattr_names(struct ceph_inode_info *ci,
475355da1ebSSage Weil 				char *dest)
476355da1ebSSage Weil {
477355da1ebSSage Weil 	struct rb_node *p;
478355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
479355da1ebSSage Weil 
480355da1ebSSage Weil 	p = rb_first(&ci->i_xattrs.index);
481355da1ebSSage Weil 	dout("__copy_xattr_names count=%d\n", ci->i_xattrs.count);
482355da1ebSSage Weil 
483355da1ebSSage Weil 	while (p) {
484355da1ebSSage Weil 		xattr = rb_entry(p, struct ceph_inode_xattr, node);
485355da1ebSSage Weil 		memcpy(dest, xattr->name, xattr->name_len);
486355da1ebSSage Weil 		dest[xattr->name_len] = '\0';
487355da1ebSSage Weil 
488355da1ebSSage Weil 		dout("dest=%s %p (%s) (%d/%d)\n", dest, xattr, xattr->name,
489355da1ebSSage Weil 		     xattr->name_len, ci->i_xattrs.names_size);
490355da1ebSSage Weil 
491355da1ebSSage Weil 		dest += xattr->name_len + 1;
492355da1ebSSage Weil 		p = rb_next(p);
493355da1ebSSage Weil 	}
494355da1ebSSage Weil 
495355da1ebSSage Weil 	return dest;
496355da1ebSSage Weil }
497355da1ebSSage Weil 
498355da1ebSSage Weil void __ceph_destroy_xattrs(struct ceph_inode_info *ci)
499355da1ebSSage Weil {
500355da1ebSSage Weil 	struct rb_node *p, *tmp;
501355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
502355da1ebSSage Weil 
503355da1ebSSage Weil 	p = rb_first(&ci->i_xattrs.index);
504355da1ebSSage Weil 
505355da1ebSSage Weil 	dout("__ceph_destroy_xattrs p=%p\n", p);
506355da1ebSSage Weil 
507355da1ebSSage Weil 	while (p) {
508355da1ebSSage Weil 		xattr = rb_entry(p, struct ceph_inode_xattr, node);
509355da1ebSSage Weil 		tmp = p;
510355da1ebSSage Weil 		p = rb_next(tmp);
511355da1ebSSage Weil 		dout("__ceph_destroy_xattrs next p=%p (%.*s)\n", p,
512355da1ebSSage Weil 		     xattr->name_len, xattr->name);
513355da1ebSSage Weil 		rb_erase(tmp, &ci->i_xattrs.index);
514355da1ebSSage Weil 
515355da1ebSSage Weil 		__free_xattr(xattr);
516355da1ebSSage Weil 	}
517355da1ebSSage Weil 
518355da1ebSSage Weil 	ci->i_xattrs.names_size = 0;
519355da1ebSSage Weil 	ci->i_xattrs.vals_size = 0;
520355da1ebSSage Weil 	ci->i_xattrs.index_version = 0;
521355da1ebSSage Weil 	ci->i_xattrs.count = 0;
522355da1ebSSage Weil 	ci->i_xattrs.index = RB_ROOT;
523355da1ebSSage Weil }
524355da1ebSSage Weil 
525355da1ebSSage Weil static int __build_xattrs(struct inode *inode)
526be655596SSage Weil 	__releases(ci->i_ceph_lock)
527be655596SSage Weil 	__acquires(ci->i_ceph_lock)
528355da1ebSSage Weil {
529355da1ebSSage Weil 	u32 namelen;
530355da1ebSSage Weil 	u32 numattr = 0;
531355da1ebSSage Weil 	void *p, *end;
532355da1ebSSage Weil 	u32 len;
533355da1ebSSage Weil 	const char *name, *val;
534355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
535355da1ebSSage Weil 	int xattr_version;
536355da1ebSSage Weil 	struct ceph_inode_xattr **xattrs = NULL;
53763ff78b2SSage Weil 	int err = 0;
538355da1ebSSage Weil 	int i;
539355da1ebSSage Weil 
540355da1ebSSage Weil 	dout("__build_xattrs() len=%d\n",
541355da1ebSSage Weil 	     ci->i_xattrs.blob ? (int)ci->i_xattrs.blob->vec.iov_len : 0);
542355da1ebSSage Weil 
543355da1ebSSage Weil 	if (ci->i_xattrs.index_version >= ci->i_xattrs.version)
544355da1ebSSage Weil 		return 0; /* already built */
545355da1ebSSage Weil 
546355da1ebSSage Weil 	__ceph_destroy_xattrs(ci);
547355da1ebSSage Weil 
548355da1ebSSage Weil start:
549355da1ebSSage Weil 	/* updated internal xattr rb tree */
550355da1ebSSage Weil 	if (ci->i_xattrs.blob && ci->i_xattrs.blob->vec.iov_len > 4) {
551355da1ebSSage Weil 		p = ci->i_xattrs.blob->vec.iov_base;
552355da1ebSSage Weil 		end = p + ci->i_xattrs.blob->vec.iov_len;
553355da1ebSSage Weil 		ceph_decode_32_safe(&p, end, numattr, bad);
554355da1ebSSage Weil 		xattr_version = ci->i_xattrs.version;
555be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
556355da1ebSSage Weil 
557355da1ebSSage Weil 		xattrs = kcalloc(numattr, sizeof(struct ceph_xattr *),
558355da1ebSSage Weil 				 GFP_NOFS);
559355da1ebSSage Weil 		err = -ENOMEM;
560355da1ebSSage Weil 		if (!xattrs)
561355da1ebSSage Weil 			goto bad_lock;
562355da1ebSSage Weil 		memset(xattrs, 0, numattr*sizeof(struct ceph_xattr *));
563355da1ebSSage Weil 		for (i = 0; i < numattr; i++) {
564355da1ebSSage Weil 			xattrs[i] = kmalloc(sizeof(struct ceph_inode_xattr),
565355da1ebSSage Weil 					    GFP_NOFS);
566355da1ebSSage Weil 			if (!xattrs[i])
567355da1ebSSage Weil 				goto bad_lock;
568355da1ebSSage Weil 		}
569355da1ebSSage Weil 
570be655596SSage Weil 		spin_lock(&ci->i_ceph_lock);
571355da1ebSSage Weil 		if (ci->i_xattrs.version != xattr_version) {
572355da1ebSSage Weil 			/* lost a race, retry */
573355da1ebSSage Weil 			for (i = 0; i < numattr; i++)
574355da1ebSSage Weil 				kfree(xattrs[i]);
575355da1ebSSage Weil 			kfree(xattrs);
57621ec6ffaSAlan Cox 			xattrs = NULL;
577355da1ebSSage Weil 			goto start;
578355da1ebSSage Weil 		}
579355da1ebSSage Weil 		err = -EIO;
580355da1ebSSage Weil 		while (numattr--) {
581355da1ebSSage Weil 			ceph_decode_32_safe(&p, end, len, bad);
582355da1ebSSage Weil 			namelen = len;
583355da1ebSSage Weil 			name = p;
584355da1ebSSage Weil 			p += len;
585355da1ebSSage Weil 			ceph_decode_32_safe(&p, end, len, bad);
586355da1ebSSage Weil 			val = p;
587355da1ebSSage Weil 			p += len;
588355da1ebSSage Weil 
589355da1ebSSage Weil 			err = __set_xattr(ci, name, namelen, val, len,
590355da1ebSSage Weil 					  0, 0, 0, &xattrs[numattr]);
591355da1ebSSage Weil 
592355da1ebSSage Weil 			if (err < 0)
593355da1ebSSage Weil 				goto bad;
594355da1ebSSage Weil 		}
595355da1ebSSage Weil 		kfree(xattrs);
596355da1ebSSage Weil 	}
597355da1ebSSage Weil 	ci->i_xattrs.index_version = ci->i_xattrs.version;
598355da1ebSSage Weil 	ci->i_xattrs.dirty = false;
599355da1ebSSage Weil 
600355da1ebSSage Weil 	return err;
601355da1ebSSage Weil bad_lock:
602be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
603355da1ebSSage Weil bad:
604355da1ebSSage Weil 	if (xattrs) {
605355da1ebSSage Weil 		for (i = 0; i < numattr; i++)
606355da1ebSSage Weil 			kfree(xattrs[i]);
607355da1ebSSage Weil 		kfree(xattrs);
608355da1ebSSage Weil 	}
609355da1ebSSage Weil 	ci->i_xattrs.names_size = 0;
610355da1ebSSage Weil 	return err;
611355da1ebSSage Weil }
612355da1ebSSage Weil 
613355da1ebSSage Weil static int __get_required_blob_size(struct ceph_inode_info *ci, int name_size,
614355da1ebSSage Weil 				    int val_size)
615355da1ebSSage Weil {
616355da1ebSSage Weil 	/*
617355da1ebSSage Weil 	 * 4 bytes for the length, and additional 4 bytes per each xattr name,
618355da1ebSSage Weil 	 * 4 bytes per each value
619355da1ebSSage Weil 	 */
620355da1ebSSage Weil 	int size = 4 + ci->i_xattrs.count*(4 + 4) +
621355da1ebSSage Weil 			     ci->i_xattrs.names_size +
622355da1ebSSage Weil 			     ci->i_xattrs.vals_size;
623355da1ebSSage Weil 	dout("__get_required_blob_size c=%d names.size=%d vals.size=%d\n",
624355da1ebSSage Weil 	     ci->i_xattrs.count, ci->i_xattrs.names_size,
625355da1ebSSage Weil 	     ci->i_xattrs.vals_size);
626355da1ebSSage Weil 
627355da1ebSSage Weil 	if (name_size)
628355da1ebSSage Weil 		size += 4 + 4 + name_size + val_size;
629355da1ebSSage Weil 
630355da1ebSSage Weil 	return size;
631355da1ebSSage Weil }
632355da1ebSSage Weil 
633355da1ebSSage Weil /*
634355da1ebSSage Weil  * If there are dirty xattrs, reencode xattrs into the prealloc_blob
635355da1ebSSage Weil  * and swap into place.
636355da1ebSSage Weil  */
637355da1ebSSage Weil void __ceph_build_xattrs_blob(struct ceph_inode_info *ci)
638355da1ebSSage Weil {
639355da1ebSSage Weil 	struct rb_node *p;
640355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
641355da1ebSSage Weil 	void *dest;
642355da1ebSSage Weil 
643355da1ebSSage Weil 	dout("__build_xattrs_blob %p\n", &ci->vfs_inode);
644355da1ebSSage Weil 	if (ci->i_xattrs.dirty) {
645355da1ebSSage Weil 		int need = __get_required_blob_size(ci, 0, 0);
646355da1ebSSage Weil 
647355da1ebSSage Weil 		BUG_ON(need > ci->i_xattrs.prealloc_blob->alloc_len);
648355da1ebSSage Weil 
649355da1ebSSage Weil 		p = rb_first(&ci->i_xattrs.index);
650355da1ebSSage Weil 		dest = ci->i_xattrs.prealloc_blob->vec.iov_base;
651355da1ebSSage Weil 
652355da1ebSSage Weil 		ceph_encode_32(&dest, ci->i_xattrs.count);
653355da1ebSSage Weil 		while (p) {
654355da1ebSSage Weil 			xattr = rb_entry(p, struct ceph_inode_xattr, node);
655355da1ebSSage Weil 
656355da1ebSSage Weil 			ceph_encode_32(&dest, xattr->name_len);
657355da1ebSSage Weil 			memcpy(dest, xattr->name, xattr->name_len);
658355da1ebSSage Weil 			dest += xattr->name_len;
659355da1ebSSage Weil 			ceph_encode_32(&dest, xattr->val_len);
660355da1ebSSage Weil 			memcpy(dest, xattr->val, xattr->val_len);
661355da1ebSSage Weil 			dest += xattr->val_len;
662355da1ebSSage Weil 
663355da1ebSSage Weil 			p = rb_next(p);
664355da1ebSSage Weil 		}
665355da1ebSSage Weil 
666355da1ebSSage Weil 		/* adjust buffer len; it may be larger than we need */
667355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob->vec.iov_len =
668355da1ebSSage Weil 			dest - ci->i_xattrs.prealloc_blob->vec.iov_base;
669355da1ebSSage Weil 
670b6c1d5b8SSage Weil 		if (ci->i_xattrs.blob)
671355da1ebSSage Weil 			ceph_buffer_put(ci->i_xattrs.blob);
672355da1ebSSage Weil 		ci->i_xattrs.blob = ci->i_xattrs.prealloc_blob;
673355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob = NULL;
674355da1ebSSage Weil 		ci->i_xattrs.dirty = false;
6754a625be4SSage Weil 		ci->i_xattrs.version++;
676355da1ebSSage Weil 	}
677355da1ebSSage Weil }
678355da1ebSSage Weil 
679*7221fe4cSGuangliang Zhao ssize_t __ceph_getxattr(struct inode *inode, const char *name, void *value,
680355da1ebSSage Weil 		      size_t size)
681355da1ebSSage Weil {
682355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
683355da1ebSSage Weil 	int err;
684355da1ebSSage Weil 	struct ceph_inode_xattr *xattr;
685881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr = NULL;
686355da1ebSSage Weil 
687355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
688355da1ebSSage Weil 		return -ENODATA;
689355da1ebSSage Weil 
6900bee82fbSSage Weil 	/* let's see if a virtual xattr was requested */
6910bee82fbSSage Weil 	vxattr = ceph_match_vxattr(inode, name);
6920bee82fbSSage Weil 	if (vxattr && !(vxattr->exists_cb && !vxattr->exists_cb(ci))) {
6930bee82fbSSage Weil 		err = vxattr->getxattr_cb(ci, value, size);
694a1dc1937Smajianpeng 		return err;
6950bee82fbSSage Weil 	}
6960bee82fbSSage Weil 
697a1dc1937Smajianpeng 	spin_lock(&ci->i_ceph_lock);
698a1dc1937Smajianpeng 	dout("getxattr %p ver=%lld index_ver=%lld\n", inode,
699a1dc1937Smajianpeng 	     ci->i_xattrs.version, ci->i_xattrs.index_version);
700a1dc1937Smajianpeng 
701355da1ebSSage Weil 	if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
702355da1ebSSage Weil 	    (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
703355da1ebSSage Weil 		goto get_xattr;
704355da1ebSSage Weil 	} else {
705be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
706355da1ebSSage Weil 		/* get xattrs from mds (if we don't already have them) */
707355da1ebSSage Weil 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
708355da1ebSSage Weil 		if (err)
709355da1ebSSage Weil 			return err;
710355da1ebSSage Weil 	}
711355da1ebSSage Weil 
712be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
713355da1ebSSage Weil 
714355da1ebSSage Weil 	err = __build_xattrs(inode);
715355da1ebSSage Weil 	if (err < 0)
716355da1ebSSage Weil 		goto out;
717355da1ebSSage Weil 
718355da1ebSSage Weil get_xattr:
719355da1ebSSage Weil 	err = -ENODATA;  /* == ENOATTR */
720355da1ebSSage Weil 	xattr = __get_xattr(ci, name);
7210bee82fbSSage Weil 	if (!xattr)
722355da1ebSSage Weil 		goto out;
723355da1ebSSage Weil 
724355da1ebSSage Weil 	err = -ERANGE;
725355da1ebSSage Weil 	if (size && size < xattr->val_len)
726355da1ebSSage Weil 		goto out;
727355da1ebSSage Weil 
728355da1ebSSage Weil 	err = xattr->val_len;
729355da1ebSSage Weil 	if (size == 0)
730355da1ebSSage Weil 		goto out;
731355da1ebSSage Weil 
732355da1ebSSage Weil 	memcpy(value, xattr->val, xattr->val_len);
733355da1ebSSage Weil 
734355da1ebSSage Weil out:
735be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
736355da1ebSSage Weil 	return err;
737355da1ebSSage Weil }
738355da1ebSSage Weil 
739*7221fe4cSGuangliang Zhao ssize_t ceph_getxattr(struct dentry *dentry, const char *name, void *value,
740*7221fe4cSGuangliang Zhao 		      size_t size)
741*7221fe4cSGuangliang Zhao {
742*7221fe4cSGuangliang Zhao 	if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
743*7221fe4cSGuangliang Zhao 		return generic_getxattr(dentry, name, value, size);
744*7221fe4cSGuangliang Zhao 
745*7221fe4cSGuangliang Zhao 	return __ceph_getxattr(dentry->d_inode, name, value, size);
746*7221fe4cSGuangliang Zhao }
747*7221fe4cSGuangliang Zhao 
748355da1ebSSage Weil ssize_t ceph_listxattr(struct dentry *dentry, char *names, size_t size)
749355da1ebSSage Weil {
750355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
751355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
752881a5fa2SAlex Elder 	struct ceph_vxattr *vxattrs = ceph_inode_vxattrs(inode);
753355da1ebSSage Weil 	u32 vir_namelen = 0;
754355da1ebSSage Weil 	u32 namelen;
755355da1ebSSage Weil 	int err;
756355da1ebSSage Weil 	u32 len;
757355da1ebSSage Weil 	int i;
758355da1ebSSage Weil 
759be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
760355da1ebSSage Weil 	dout("listxattr %p ver=%lld index_ver=%lld\n", inode,
761355da1ebSSage Weil 	     ci->i_xattrs.version, ci->i_xattrs.index_version);
762355da1ebSSage Weil 
763355da1ebSSage Weil 	if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
764bddfa3ccSHenry C Chang 	    (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
765355da1ebSSage Weil 		goto list_xattr;
766355da1ebSSage Weil 	} else {
767be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
768355da1ebSSage Weil 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
769355da1ebSSage Weil 		if (err)
770355da1ebSSage Weil 			return err;
771355da1ebSSage Weil 	}
772355da1ebSSage Weil 
773be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
774355da1ebSSage Weil 
775355da1ebSSage Weil 	err = __build_xattrs(inode);
776355da1ebSSage Weil 	if (err < 0)
777355da1ebSSage Weil 		goto out;
778355da1ebSSage Weil 
779355da1ebSSage Weil list_xattr:
7803ce6cd12SAlex Elder 	/*
7813ce6cd12SAlex Elder 	 * Start with virtual dir xattr names (if any) (including
7823ce6cd12SAlex Elder 	 * terminating '\0' characters for each).
7833ce6cd12SAlex Elder 	 */
7843ce6cd12SAlex Elder 	vir_namelen = ceph_vxattrs_name_size(vxattrs);
7853ce6cd12SAlex Elder 
786355da1ebSSage Weil 	/* adding 1 byte per each variable due to the null termination */
787b65917ddSSage Weil 	namelen = ci->i_xattrs.names_size + ci->i_xattrs.count;
788355da1ebSSage Weil 	err = -ERANGE;
789b65917ddSSage Weil 	if (size && vir_namelen + namelen > size)
790355da1ebSSage Weil 		goto out;
791355da1ebSSage Weil 
792b65917ddSSage Weil 	err = namelen + vir_namelen;
793355da1ebSSage Weil 	if (size == 0)
794355da1ebSSage Weil 		goto out;
795355da1ebSSage Weil 
796355da1ebSSage Weil 	names = __copy_xattr_names(ci, names);
797355da1ebSSage Weil 
798355da1ebSSage Weil 	/* virtual xattr names, too */
799b65917ddSSage Weil 	err = namelen;
800b65917ddSSage Weil 	if (vxattrs) {
801355da1ebSSage Weil 		for (i = 0; vxattrs[i].name; i++) {
802b65917ddSSage Weil 			if (!vxattrs[i].hidden &&
803b65917ddSSage Weil 			    !(vxattrs[i].exists_cb &&
804b65917ddSSage Weil 			      !vxattrs[i].exists_cb(ci))) {
805355da1ebSSage Weil 				len = sprintf(names, "%s", vxattrs[i].name);
806355da1ebSSage Weil 				names += len + 1;
807b65917ddSSage Weil 				err += len + 1;
808b65917ddSSage Weil 			}
809b65917ddSSage Weil 		}
810355da1ebSSage Weil 	}
811355da1ebSSage Weil 
812355da1ebSSage Weil out:
813be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
814355da1ebSSage Weil 	return err;
815355da1ebSSage Weil }
816355da1ebSSage Weil 
817355da1ebSSage Weil static int ceph_sync_setxattr(struct dentry *dentry, const char *name,
818355da1ebSSage Weil 			      const char *value, size_t size, int flags)
819355da1ebSSage Weil {
8203d14c5d2SYehuda Sadeh 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
821355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
822355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
8235f21c96dSSage Weil 	struct inode *parent_inode;
824355da1ebSSage Weil 	struct ceph_mds_request *req;
8253d14c5d2SYehuda Sadeh 	struct ceph_mds_client *mdsc = fsc->mdsc;
826355da1ebSSage Weil 	int err;
827355da1ebSSage Weil 	int i, nr_pages;
828355da1ebSSage Weil 	struct page **pages = NULL;
829355da1ebSSage Weil 	void *kaddr;
830355da1ebSSage Weil 
831355da1ebSSage Weil 	/* copy value into some pages */
832355da1ebSSage Weil 	nr_pages = calc_pages_for(0, size);
833355da1ebSSage Weil 	if (nr_pages) {
834355da1ebSSage Weil 		pages = kmalloc(sizeof(pages[0])*nr_pages, GFP_NOFS);
835355da1ebSSage Weil 		if (!pages)
836355da1ebSSage Weil 			return -ENOMEM;
837355da1ebSSage Weil 		err = -ENOMEM;
838355da1ebSSage Weil 		for (i = 0; i < nr_pages; i++) {
83931459fe4SYehuda Sadeh 			pages[i] = __page_cache_alloc(GFP_NOFS);
840355da1ebSSage Weil 			if (!pages[i]) {
841355da1ebSSage Weil 				nr_pages = i;
842355da1ebSSage Weil 				goto out;
843355da1ebSSage Weil 			}
844355da1ebSSage Weil 			kaddr = kmap(pages[i]);
845355da1ebSSage Weil 			memcpy(kaddr, value + i*PAGE_CACHE_SIZE,
846355da1ebSSage Weil 			       min(PAGE_CACHE_SIZE, size-i*PAGE_CACHE_SIZE));
847355da1ebSSage Weil 		}
848355da1ebSSage Weil 	}
849355da1ebSSage Weil 
850355da1ebSSage Weil 	dout("setxattr value=%.*s\n", (int)size, value);
851355da1ebSSage Weil 
852355da1ebSSage Weil 	/* do request */
853355da1ebSSage Weil 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETXATTR,
854355da1ebSSage Weil 				       USE_AUTH_MDS);
85560d87733SJulia Lawall 	if (IS_ERR(req)) {
85660d87733SJulia Lawall 		err = PTR_ERR(req);
85760d87733SJulia Lawall 		goto out;
85860d87733SJulia Lawall 	}
85970b666c3SSage Weil 	req->r_inode = inode;
86070b666c3SSage Weil 	ihold(inode);
861355da1ebSSage Weil 	req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
862355da1ebSSage Weil 	req->r_num_caps = 1;
863355da1ebSSage Weil 	req->r_args.setxattr.flags = cpu_to_le32(flags);
864355da1ebSSage Weil 	req->r_path2 = kstrdup(name, GFP_NOFS);
865355da1ebSSage Weil 
866355da1ebSSage Weil 	req->r_pages = pages;
867355da1ebSSage Weil 	req->r_num_pages = nr_pages;
868355da1ebSSage Weil 	req->r_data_len = size;
869355da1ebSSage Weil 
870355da1ebSSage Weil 	dout("xattr.ver (before): %lld\n", ci->i_xattrs.version);
8715f21c96dSSage Weil 	parent_inode = ceph_get_dentry_parent_inode(dentry);
872355da1ebSSage Weil 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
8735f21c96dSSage Weil 	iput(parent_inode);
874355da1ebSSage Weil 	ceph_mdsc_put_request(req);
875355da1ebSSage Weil 	dout("xattr.ver (after): %lld\n", ci->i_xattrs.version);
876355da1ebSSage Weil 
877355da1ebSSage Weil out:
878355da1ebSSage Weil 	if (pages) {
879355da1ebSSage Weil 		for (i = 0; i < nr_pages; i++)
880355da1ebSSage Weil 			__free_page(pages[i]);
881355da1ebSSage Weil 		kfree(pages);
882355da1ebSSage Weil 	}
883355da1ebSSage Weil 	return err;
884355da1ebSSage Weil }
885355da1ebSSage Weil 
886*7221fe4cSGuangliang Zhao int __ceph_setxattr(struct dentry *dentry, const char *name,
887355da1ebSSage Weil 			const void *value, size_t size, int flags)
888355da1ebSSage Weil {
889355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
890881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr;
891355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
89218fa8b3fSAlex Elder 	int issued;
893355da1ebSSage Weil 	int err;
89418fa8b3fSAlex Elder 	int dirty;
895355da1ebSSage Weil 	int name_len = strlen(name);
896355da1ebSSage Weil 	int val_len = size;
897355da1ebSSage Weil 	char *newname = NULL;
898355da1ebSSage Weil 	char *newval = NULL;
899355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
900355da1ebSSage Weil 	int required_blob_size;
901355da1ebSSage Weil 
902355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
903355da1ebSSage Weil 		return -EOPNOTSUPP;
904355da1ebSSage Weil 
90506476a69SAlex Elder 	vxattr = ceph_match_vxattr(inode, name);
906355da1ebSSage Weil 	if (vxattr && vxattr->readonly)
907355da1ebSSage Weil 		return -EOPNOTSUPP;
908355da1ebSSage Weil 
9093adf654dSSage Weil 	/* pass any unhandled ceph.* xattrs through to the MDS */
9103adf654dSSage Weil 	if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
9113adf654dSSage Weil 		goto do_sync_unlocked;
9123adf654dSSage Weil 
913355da1ebSSage Weil 	/* preallocate memory for xattr name, value, index node */
914355da1ebSSage Weil 	err = -ENOMEM;
91561413c2fSJulia Lawall 	newname = kmemdup(name, name_len + 1, GFP_NOFS);
916355da1ebSSage Weil 	if (!newname)
917355da1ebSSage Weil 		goto out;
918355da1ebSSage Weil 
919355da1ebSSage Weil 	if (val_len) {
920b829c195SAlex Elder 		newval = kmemdup(value, val_len, GFP_NOFS);
921355da1ebSSage Weil 		if (!newval)
922355da1ebSSage Weil 			goto out;
923355da1ebSSage Weil 	}
924355da1ebSSage Weil 
925355da1ebSSage Weil 	xattr = kmalloc(sizeof(struct ceph_inode_xattr), GFP_NOFS);
926355da1ebSSage Weil 	if (!xattr)
927355da1ebSSage Weil 		goto out;
928355da1ebSSage Weil 
929be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
930355da1ebSSage Weil retry:
931355da1ebSSage Weil 	issued = __ceph_caps_issued(ci, NULL);
93218fa8b3fSAlex Elder 	dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
933355da1ebSSage Weil 	if (!(issued & CEPH_CAP_XATTR_EXCL))
934355da1ebSSage Weil 		goto do_sync;
935355da1ebSSage Weil 	__build_xattrs(inode);
936355da1ebSSage Weil 
937355da1ebSSage Weil 	required_blob_size = __get_required_blob_size(ci, name_len, val_len);
938355da1ebSSage Weil 
939355da1ebSSage Weil 	if (!ci->i_xattrs.prealloc_blob ||
940355da1ebSSage Weil 	    required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
94118fa8b3fSAlex Elder 		struct ceph_buffer *blob;
942355da1ebSSage Weil 
943be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
944355da1ebSSage Weil 		dout(" preaallocating new blob size=%d\n", required_blob_size);
945b6c1d5b8SSage Weil 		blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
946355da1ebSSage Weil 		if (!blob)
947355da1ebSSage Weil 			goto out;
948be655596SSage Weil 		spin_lock(&ci->i_ceph_lock);
949b6c1d5b8SSage Weil 		if (ci->i_xattrs.prealloc_blob)
950355da1ebSSage Weil 			ceph_buffer_put(ci->i_xattrs.prealloc_blob);
951355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob = blob;
952355da1ebSSage Weil 		goto retry;
953355da1ebSSage Weil 	}
954355da1ebSSage Weil 
955355da1ebSSage Weil 	err = __set_xattr(ci, newname, name_len, newval,
956355da1ebSSage Weil 			  val_len, 1, 1, 1, &xattr);
95718fa8b3fSAlex Elder 
958fca65b4aSSage Weil 	dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
959355da1ebSSage Weil 	ci->i_xattrs.dirty = true;
960355da1ebSSage Weil 	inode->i_ctime = CURRENT_TIME;
96118fa8b3fSAlex Elder 
962be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
963fca65b4aSSage Weil 	if (dirty)
964fca65b4aSSage Weil 		__mark_inode_dirty(inode, dirty);
965355da1ebSSage Weil 	return err;
966355da1ebSSage Weil 
967355da1ebSSage Weil do_sync:
968be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
9693adf654dSSage Weil do_sync_unlocked:
970355da1ebSSage Weil 	err = ceph_sync_setxattr(dentry, name, value, size, flags);
971355da1ebSSage Weil out:
972355da1ebSSage Weil 	kfree(newname);
973355da1ebSSage Weil 	kfree(newval);
974355da1ebSSage Weil 	kfree(xattr);
975355da1ebSSage Weil 	return err;
976355da1ebSSage Weil }
977355da1ebSSage Weil 
978*7221fe4cSGuangliang Zhao int ceph_setxattr(struct dentry *dentry, const char *name,
979*7221fe4cSGuangliang Zhao 		  const void *value, size_t size, int flags)
980*7221fe4cSGuangliang Zhao {
981*7221fe4cSGuangliang Zhao 	if (ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
982*7221fe4cSGuangliang Zhao 		return -EROFS;
983*7221fe4cSGuangliang Zhao 
984*7221fe4cSGuangliang Zhao 	if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
985*7221fe4cSGuangliang Zhao 		return generic_setxattr(dentry, name, value, size, flags);
986*7221fe4cSGuangliang Zhao 
987*7221fe4cSGuangliang Zhao 	return __ceph_setxattr(dentry, name, value, size, flags);
988*7221fe4cSGuangliang Zhao }
989*7221fe4cSGuangliang Zhao 
990355da1ebSSage Weil static int ceph_send_removexattr(struct dentry *dentry, const char *name)
991355da1ebSSage Weil {
9923d14c5d2SYehuda Sadeh 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
9933d14c5d2SYehuda Sadeh 	struct ceph_mds_client *mdsc = fsc->mdsc;
994355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
9955f21c96dSSage Weil 	struct inode *parent_inode;
996355da1ebSSage Weil 	struct ceph_mds_request *req;
997355da1ebSSage Weil 	int err;
998355da1ebSSage Weil 
999355da1ebSSage Weil 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RMXATTR,
1000355da1ebSSage Weil 				       USE_AUTH_MDS);
1001355da1ebSSage Weil 	if (IS_ERR(req))
1002355da1ebSSage Weil 		return PTR_ERR(req);
100370b666c3SSage Weil 	req->r_inode = inode;
100470b666c3SSage Weil 	ihold(inode);
1005355da1ebSSage Weil 	req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
1006355da1ebSSage Weil 	req->r_num_caps = 1;
1007355da1ebSSage Weil 	req->r_path2 = kstrdup(name, GFP_NOFS);
1008355da1ebSSage Weil 
10095f21c96dSSage Weil 	parent_inode = ceph_get_dentry_parent_inode(dentry);
1010355da1ebSSage Weil 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
10115f21c96dSSage Weil 	iput(parent_inode);
1012355da1ebSSage Weil 	ceph_mdsc_put_request(req);
1013355da1ebSSage Weil 	return err;
1014355da1ebSSage Weil }
1015355da1ebSSage Weil 
1016*7221fe4cSGuangliang Zhao int __ceph_removexattr(struct dentry *dentry, const char *name)
1017355da1ebSSage Weil {
1018355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
1019881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr;
1020355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
1021355da1ebSSage Weil 	int issued;
1022355da1ebSSage Weil 	int err;
102383eb26afSAlex Elder 	int required_blob_size;
1024fca65b4aSSage Weil 	int dirty;
1025355da1ebSSage Weil 
1026355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
1027355da1ebSSage Weil 		return -EOPNOTSUPP;
1028355da1ebSSage Weil 
102906476a69SAlex Elder 	vxattr = ceph_match_vxattr(inode, name);
1030355da1ebSSage Weil 	if (vxattr && vxattr->readonly)
1031355da1ebSSage Weil 		return -EOPNOTSUPP;
1032355da1ebSSage Weil 
1033d421acb1SSage Weil 	/* pass any unhandled ceph.* xattrs through to the MDS */
1034d421acb1SSage Weil 	if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
1035d421acb1SSage Weil 		goto do_sync_unlocked;
1036d421acb1SSage Weil 
103783eb26afSAlex Elder 	err = -ENOMEM;
1038be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
103983eb26afSAlex Elder retry:
1040355da1ebSSage Weil 	issued = __ceph_caps_issued(ci, NULL);
1041355da1ebSSage Weil 	dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
1042355da1ebSSage Weil 
1043355da1ebSSage Weil 	if (!(issued & CEPH_CAP_XATTR_EXCL))
1044355da1ebSSage Weil 		goto do_sync;
104518fa8b3fSAlex Elder 	__build_xattrs(inode);
1046355da1ebSSage Weil 
104783eb26afSAlex Elder 	required_blob_size = __get_required_blob_size(ci, 0, 0);
104883eb26afSAlex Elder 
104983eb26afSAlex Elder 	if (!ci->i_xattrs.prealloc_blob ||
105083eb26afSAlex Elder 	    required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
105183eb26afSAlex Elder 		struct ceph_buffer *blob;
105283eb26afSAlex Elder 
105383eb26afSAlex Elder 		spin_unlock(&ci->i_ceph_lock);
105483eb26afSAlex Elder 		dout(" preaallocating new blob size=%d\n", required_blob_size);
105583eb26afSAlex Elder 		blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
105683eb26afSAlex Elder 		if (!blob)
105783eb26afSAlex Elder 			goto out;
105883eb26afSAlex Elder 		spin_lock(&ci->i_ceph_lock);
105983eb26afSAlex Elder 		if (ci->i_xattrs.prealloc_blob)
106083eb26afSAlex Elder 			ceph_buffer_put(ci->i_xattrs.prealloc_blob);
106183eb26afSAlex Elder 		ci->i_xattrs.prealloc_blob = blob;
106283eb26afSAlex Elder 		goto retry;
106383eb26afSAlex Elder 	}
106483eb26afSAlex Elder 
1065355da1ebSSage Weil 	err = __remove_xattr_by_name(ceph_inode(inode), name);
106618fa8b3fSAlex Elder 
1067fca65b4aSSage Weil 	dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
1068355da1ebSSage Weil 	ci->i_xattrs.dirty = true;
1069355da1ebSSage Weil 	inode->i_ctime = CURRENT_TIME;
1070be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
1071fca65b4aSSage Weil 	if (dirty)
1072fca65b4aSSage Weil 		__mark_inode_dirty(inode, dirty);
1073355da1ebSSage Weil 	return err;
1074355da1ebSSage Weil do_sync:
1075be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
1076d421acb1SSage Weil do_sync_unlocked:
1077355da1ebSSage Weil 	err = ceph_send_removexattr(dentry, name);
107883eb26afSAlex Elder out:
1079355da1ebSSage Weil 	return err;
1080355da1ebSSage Weil }
1081355da1ebSSage Weil 
1082*7221fe4cSGuangliang Zhao int ceph_removexattr(struct dentry *dentry, const char *name)
1083*7221fe4cSGuangliang Zhao {
1084*7221fe4cSGuangliang Zhao 	if (ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
1085*7221fe4cSGuangliang Zhao 		return -EROFS;
1086*7221fe4cSGuangliang Zhao 
1087*7221fe4cSGuangliang Zhao 	if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
1088*7221fe4cSGuangliang Zhao 		return generic_removexattr(dentry, name);
1089*7221fe4cSGuangliang Zhao 
1090*7221fe4cSGuangliang Zhao 	return __ceph_removexattr(dentry, name);
1091*7221fe4cSGuangliang Zhao }
1092