xref: /openbmc/linux/fs/ceph/xattr.c (revision 1f08f2b056ea6c2e12f4e95e88949a882a996208)
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 
14355da1ebSSage Weil static bool ceph_is_valid_xattr(const char *name)
15355da1ebSSage Weil {
1622891907SAlex Elder 	return !strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN) ||
171a756278SSage Weil 	       !strncmp(name, XATTR_SECURITY_PREFIX,
18355da1ebSSage Weil 			XATTR_SECURITY_PREFIX_LEN) ||
19355da1ebSSage Weil 	       !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
20355da1ebSSage Weil 	       !strncmp(name, XATTR_USER_PREFIX, XATTR_USER_PREFIX_LEN);
21355da1ebSSage Weil }
22355da1ebSSage Weil 
23355da1ebSSage Weil /*
24355da1ebSSage Weil  * These define virtual xattrs exposing the recursive directory
25355da1ebSSage Weil  * statistics and layout metadata.
26355da1ebSSage Weil  */
27881a5fa2SAlex Elder struct ceph_vxattr {
28355da1ebSSage Weil 	char *name;
293ce6cd12SAlex Elder 	size_t name_size;	/* strlen(name) + 1 (for '\0') */
30355da1ebSSage Weil 	size_t (*getxattr_cb)(struct ceph_inode_info *ci, char *val,
31355da1ebSSage Weil 			      size_t size);
328860147aSSage Weil 	bool readonly, hidden;
33f36e4472SSage Weil 	bool (*exists_cb)(struct ceph_inode_info *ci);
34355da1ebSSage Weil };
35355da1ebSSage Weil 
3632ab0bd7SSage Weil /* layouts */
3732ab0bd7SSage Weil 
3832ab0bd7SSage Weil static bool ceph_vxattrcb_layout_exists(struct ceph_inode_info *ci)
3932ab0bd7SSage Weil {
4032ab0bd7SSage Weil 	size_t s;
4132ab0bd7SSage Weil 	char *p = (char *)&ci->i_layout;
4232ab0bd7SSage Weil 
4332ab0bd7SSage Weil 	for (s = 0; s < sizeof(ci->i_layout); s++, p++)
4432ab0bd7SSage Weil 		if (*p)
4532ab0bd7SSage Weil 			return true;
4632ab0bd7SSage Weil 	return false;
4732ab0bd7SSage Weil }
4832ab0bd7SSage Weil 
4932ab0bd7SSage Weil static size_t ceph_vxattrcb_layout(struct ceph_inode_info *ci, char *val,
5032ab0bd7SSage Weil 					size_t size)
5132ab0bd7SSage Weil {
5232ab0bd7SSage Weil 	int ret;
5332ab0bd7SSage Weil 	struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
5432ab0bd7SSage Weil 	struct ceph_osd_client *osdc = &fsc->client->osdc;
5532ab0bd7SSage Weil 	s64 pool = ceph_file_layout_pg_pool(ci->i_layout);
5632ab0bd7SSage Weil 	const char *pool_name;
5732ab0bd7SSage Weil 
5832ab0bd7SSage Weil 	dout("ceph_vxattrcb_layout %p\n", &ci->vfs_inode);
5932ab0bd7SSage Weil 	down_read(&osdc->map_sem);
6032ab0bd7SSage Weil 	pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, pool);
6132ab0bd7SSage Weil 	if (pool_name)
6232ab0bd7SSage Weil 		ret = snprintf(val, size,
6332ab0bd7SSage Weil 		"stripe_unit=%lld stripe_count=%lld object_size=%lld pool=%s",
6432ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_su(ci->i_layout),
6532ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
6632ab0bd7SSage Weil 	        (unsigned long long)ceph_file_layout_object_size(ci->i_layout),
6732ab0bd7SSage Weil 		pool_name);
6832ab0bd7SSage Weil 	else
6932ab0bd7SSage Weil 		ret = snprintf(val, size,
7032ab0bd7SSage Weil 		"stripe_unit=%lld stripe_count=%lld object_size=%lld pool=%lld",
7132ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_su(ci->i_layout),
7232ab0bd7SSage Weil 		(unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
7332ab0bd7SSage Weil 	        (unsigned long long)ceph_file_layout_object_size(ci->i_layout),
7432ab0bd7SSage Weil 		(unsigned long long)pool);
7532ab0bd7SSage Weil 
7632ab0bd7SSage Weil 	up_read(&osdc->map_sem);
7732ab0bd7SSage Weil 	return ret;
7832ab0bd7SSage Weil }
7932ab0bd7SSage Weil 
80355da1ebSSage Weil /* directories */
81355da1ebSSage Weil 
82aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_entries(struct ceph_inode_info *ci, char *val,
83355da1ebSSage Weil 					size_t size)
84355da1ebSSage Weil {
85355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_files + ci->i_subdirs);
86355da1ebSSage Weil }
87355da1ebSSage Weil 
88aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_files(struct ceph_inode_info *ci, char *val,
89355da1ebSSage Weil 				      size_t size)
90355da1ebSSage Weil {
91355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_files);
92355da1ebSSage Weil }
93355da1ebSSage Weil 
94aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_subdirs(struct ceph_inode_info *ci, char *val,
95355da1ebSSage Weil 					size_t size)
96355da1ebSSage Weil {
97355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_subdirs);
98355da1ebSSage Weil }
99355da1ebSSage Weil 
100aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rentries(struct ceph_inode_info *ci, char *val,
101355da1ebSSage Weil 					 size_t size)
102355da1ebSSage Weil {
103355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rfiles + ci->i_rsubdirs);
104355da1ebSSage Weil }
105355da1ebSSage Weil 
106aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rfiles(struct ceph_inode_info *ci, char *val,
107355da1ebSSage Weil 				       size_t size)
108355da1ebSSage Weil {
109355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rfiles);
110355da1ebSSage Weil }
111355da1ebSSage Weil 
112aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rsubdirs(struct ceph_inode_info *ci, char *val,
113355da1ebSSage Weil 					 size_t size)
114355da1ebSSage Weil {
115355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rsubdirs);
116355da1ebSSage Weil }
117355da1ebSSage Weil 
118aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rbytes(struct ceph_inode_info *ci, char *val,
119355da1ebSSage Weil 				       size_t size)
120355da1ebSSage Weil {
121355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rbytes);
122355da1ebSSage Weil }
123355da1ebSSage Weil 
124aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rctime(struct ceph_inode_info *ci, char *val,
125355da1ebSSage Weil 				       size_t size)
126355da1ebSSage Weil {
1273489b42aSAlex Elder 	return snprintf(val, size, "%ld.09%ld", (long)ci->i_rctime.tv_sec,
128355da1ebSSage Weil 			(long)ci->i_rctime.tv_nsec);
129355da1ebSSage Weil }
130355da1ebSSage Weil 
13132ab0bd7SSage Weil 
132eb788084SAlex Elder #define CEPH_XATTR_NAME(_type, _name)	XATTR_CEPH_PREFIX #_type "." #_name
133eb788084SAlex Elder 
134eb788084SAlex Elder #define XATTR_NAME_CEPH(_type, _name)					\
135eb788084SAlex Elder 	{								\
136eb788084SAlex Elder 		.name = CEPH_XATTR_NAME(_type, _name),			\
1373ce6cd12SAlex Elder 		.name_size = sizeof (CEPH_XATTR_NAME(_type, _name)), \
138aa4066edSAlex Elder 		.getxattr_cb = ceph_vxattrcb_ ## _type ## _ ## _name, \
139eb788084SAlex Elder 		.readonly = true,				\
1408860147aSSage Weil 		.hidden = false,				\
141f36e4472SSage Weil 		.exists_cb = NULL,			\
142eb788084SAlex Elder 	}
143eb788084SAlex Elder 
144881a5fa2SAlex Elder static struct ceph_vxattr ceph_dir_vxattrs[] = {
145*1f08f2b0SSage Weil 	{
146*1f08f2b0SSage Weil 		.name = "ceph.dir.layout",
147*1f08f2b0SSage Weil 		.name_size = sizeof("ceph.dir.layout"),
148*1f08f2b0SSage Weil 		.getxattr_cb = ceph_vxattrcb_layout,
149*1f08f2b0SSage Weil 		.readonly = false,
150*1f08f2b0SSage Weil 		.hidden = false,
151*1f08f2b0SSage Weil 		.exists_cb = ceph_vxattrcb_layout_exists,
152*1f08f2b0SSage Weil 	},
153eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, entries),
154eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, files),
155eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, subdirs),
156eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rentries),
157eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rfiles),
158eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rsubdirs),
159eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rbytes),
160eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rctime),
161eb788084SAlex Elder 	{ 0 }	/* Required table terminator */
162355da1ebSSage Weil };
1633ce6cd12SAlex Elder static size_t ceph_dir_vxattrs_name_size;	/* total size of all names */
164355da1ebSSage Weil 
165355da1ebSSage Weil /* files */
166355da1ebSSage Weil 
167881a5fa2SAlex Elder static struct ceph_vxattr ceph_file_vxattrs[] = {
16832ab0bd7SSage Weil 	{
16932ab0bd7SSage Weil 		.name = "ceph.file.layout",
17032ab0bd7SSage Weil 		.name_size = sizeof("ceph.file.layout"),
17132ab0bd7SSage Weil 		.getxattr_cb = ceph_vxattrcb_layout,
17232ab0bd7SSage Weil 		.readonly = false,
17332ab0bd7SSage Weil 		.hidden = false,
17432ab0bd7SSage Weil 		.exists_cb = ceph_vxattrcb_layout_exists,
17532ab0bd7SSage Weil 	},
176eb788084SAlex Elder 	{ 0 }	/* Required table terminator */
177355da1ebSSage Weil };
1783ce6cd12SAlex Elder static size_t ceph_file_vxattrs_name_size;	/* total size of all names */
179355da1ebSSage Weil 
180881a5fa2SAlex Elder static struct ceph_vxattr *ceph_inode_vxattrs(struct inode *inode)
181355da1ebSSage Weil {
182355da1ebSSage Weil 	if (S_ISDIR(inode->i_mode))
183355da1ebSSage Weil 		return ceph_dir_vxattrs;
184355da1ebSSage Weil 	else if (S_ISREG(inode->i_mode))
185355da1ebSSage Weil 		return ceph_file_vxattrs;
186355da1ebSSage Weil 	return NULL;
187355da1ebSSage Weil }
188355da1ebSSage Weil 
1893ce6cd12SAlex Elder static size_t ceph_vxattrs_name_size(struct ceph_vxattr *vxattrs)
1903ce6cd12SAlex Elder {
1913ce6cd12SAlex Elder 	if (vxattrs == ceph_dir_vxattrs)
1923ce6cd12SAlex Elder 		return ceph_dir_vxattrs_name_size;
1933ce6cd12SAlex Elder 	if (vxattrs == ceph_file_vxattrs)
1943ce6cd12SAlex Elder 		return ceph_file_vxattrs_name_size;
1953ce6cd12SAlex Elder 	BUG();
1963ce6cd12SAlex Elder 
1973ce6cd12SAlex Elder 	return 0;
1983ce6cd12SAlex Elder }
1993ce6cd12SAlex Elder 
2003ce6cd12SAlex Elder /*
2013ce6cd12SAlex Elder  * Compute the aggregate size (including terminating '\0') of all
2023ce6cd12SAlex Elder  * virtual extended attribute names in the given vxattr table.
2033ce6cd12SAlex Elder  */
2043ce6cd12SAlex Elder static size_t __init vxattrs_name_size(struct ceph_vxattr *vxattrs)
2053ce6cd12SAlex Elder {
2063ce6cd12SAlex Elder 	struct ceph_vxattr *vxattr;
2073ce6cd12SAlex Elder 	size_t size = 0;
2083ce6cd12SAlex Elder 
2093ce6cd12SAlex Elder 	for (vxattr = vxattrs; vxattr->name; vxattr++)
2108860147aSSage Weil 		if (!vxattr->hidden)
2113ce6cd12SAlex Elder 			size += vxattr->name_size;
2123ce6cd12SAlex Elder 
2133ce6cd12SAlex Elder 	return size;
2143ce6cd12SAlex Elder }
2153ce6cd12SAlex Elder 
2163ce6cd12SAlex Elder /* Routines called at initialization and exit time */
2173ce6cd12SAlex Elder 
2183ce6cd12SAlex Elder void __init ceph_xattr_init(void)
2193ce6cd12SAlex Elder {
2203ce6cd12SAlex Elder 	ceph_dir_vxattrs_name_size = vxattrs_name_size(ceph_dir_vxattrs);
2213ce6cd12SAlex Elder 	ceph_file_vxattrs_name_size = vxattrs_name_size(ceph_file_vxattrs);
2223ce6cd12SAlex Elder }
2233ce6cd12SAlex Elder 
2243ce6cd12SAlex Elder void ceph_xattr_exit(void)
2253ce6cd12SAlex Elder {
2263ce6cd12SAlex Elder 	ceph_dir_vxattrs_name_size = 0;
2273ce6cd12SAlex Elder 	ceph_file_vxattrs_name_size = 0;
2283ce6cd12SAlex Elder }
2293ce6cd12SAlex Elder 
230881a5fa2SAlex Elder static struct ceph_vxattr *ceph_match_vxattr(struct inode *inode,
231355da1ebSSage Weil 						const char *name)
232355da1ebSSage Weil {
233881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr = ceph_inode_vxattrs(inode);
23406476a69SAlex Elder 
23506476a69SAlex Elder 	if (vxattr) {
23606476a69SAlex Elder 		while (vxattr->name) {
23706476a69SAlex Elder 			if (!strcmp(vxattr->name, name))
238355da1ebSSage Weil 				return vxattr;
239355da1ebSSage Weil 			vxattr++;
24006476a69SAlex Elder 		}
24106476a69SAlex Elder 	}
24206476a69SAlex Elder 
243355da1ebSSage Weil 	return NULL;
244355da1ebSSage Weil }
245355da1ebSSage Weil 
246355da1ebSSage Weil static int __set_xattr(struct ceph_inode_info *ci,
247355da1ebSSage Weil 			   const char *name, int name_len,
248355da1ebSSage Weil 			   const char *val, int val_len,
249355da1ebSSage Weil 			   int dirty,
250355da1ebSSage Weil 			   int should_free_name, int should_free_val,
251355da1ebSSage Weil 			   struct ceph_inode_xattr **newxattr)
252355da1ebSSage Weil {
253355da1ebSSage Weil 	struct rb_node **p;
254355da1ebSSage Weil 	struct rb_node *parent = NULL;
255355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
256355da1ebSSage Weil 	int c;
257355da1ebSSage Weil 	int new = 0;
258355da1ebSSage Weil 
259355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
260355da1ebSSage Weil 	while (*p) {
261355da1ebSSage Weil 		parent = *p;
262355da1ebSSage Weil 		xattr = rb_entry(parent, struct ceph_inode_xattr, node);
263355da1ebSSage Weil 		c = strncmp(name, xattr->name, min(name_len, xattr->name_len));
264355da1ebSSage Weil 		if (c < 0)
265355da1ebSSage Weil 			p = &(*p)->rb_left;
266355da1ebSSage Weil 		else if (c > 0)
267355da1ebSSage Weil 			p = &(*p)->rb_right;
268355da1ebSSage Weil 		else {
269355da1ebSSage Weil 			if (name_len == xattr->name_len)
270355da1ebSSage Weil 				break;
271355da1ebSSage Weil 			else if (name_len < xattr->name_len)
272355da1ebSSage Weil 				p = &(*p)->rb_left;
273355da1ebSSage Weil 			else
274355da1ebSSage Weil 				p = &(*p)->rb_right;
275355da1ebSSage Weil 		}
276355da1ebSSage Weil 		xattr = NULL;
277355da1ebSSage Weil 	}
278355da1ebSSage Weil 
279355da1ebSSage Weil 	if (!xattr) {
280355da1ebSSage Weil 		new = 1;
281355da1ebSSage Weil 		xattr = *newxattr;
282355da1ebSSage Weil 		xattr->name = name;
283355da1ebSSage Weil 		xattr->name_len = name_len;
284355da1ebSSage Weil 		xattr->should_free_name = should_free_name;
285355da1ebSSage Weil 
286355da1ebSSage Weil 		ci->i_xattrs.count++;
287355da1ebSSage Weil 		dout("__set_xattr count=%d\n", ci->i_xattrs.count);
288355da1ebSSage Weil 	} else {
289355da1ebSSage Weil 		kfree(*newxattr);
290355da1ebSSage Weil 		*newxattr = NULL;
291355da1ebSSage Weil 		if (xattr->should_free_val)
292355da1ebSSage Weil 			kfree((void *)xattr->val);
293355da1ebSSage Weil 
294355da1ebSSage Weil 		if (should_free_name) {
295355da1ebSSage Weil 			kfree((void *)name);
296355da1ebSSage Weil 			name = xattr->name;
297355da1ebSSage Weil 		}
298355da1ebSSage Weil 		ci->i_xattrs.names_size -= xattr->name_len;
299355da1ebSSage Weil 		ci->i_xattrs.vals_size -= xattr->val_len;
300355da1ebSSage Weil 	}
301355da1ebSSage Weil 	ci->i_xattrs.names_size += name_len;
302355da1ebSSage Weil 	ci->i_xattrs.vals_size += val_len;
303355da1ebSSage Weil 	if (val)
304355da1ebSSage Weil 		xattr->val = val;
305355da1ebSSage Weil 	else
306355da1ebSSage Weil 		xattr->val = "";
307355da1ebSSage Weil 
308355da1ebSSage Weil 	xattr->val_len = val_len;
309355da1ebSSage Weil 	xattr->dirty = dirty;
310355da1ebSSage Weil 	xattr->should_free_val = (val && should_free_val);
311355da1ebSSage Weil 
312355da1ebSSage Weil 	if (new) {
313355da1ebSSage Weil 		rb_link_node(&xattr->node, parent, p);
314355da1ebSSage Weil 		rb_insert_color(&xattr->node, &ci->i_xattrs.index);
315355da1ebSSage Weil 		dout("__set_xattr_val p=%p\n", p);
316355da1ebSSage Weil 	}
317355da1ebSSage Weil 
318355da1ebSSage Weil 	dout("__set_xattr_val added %llx.%llx xattr %p %s=%.*s\n",
319355da1ebSSage Weil 	     ceph_vinop(&ci->vfs_inode), xattr, name, val_len, val);
320355da1ebSSage Weil 
321355da1ebSSage Weil 	return 0;
322355da1ebSSage Weil }
323355da1ebSSage Weil 
324355da1ebSSage Weil static struct ceph_inode_xattr *__get_xattr(struct ceph_inode_info *ci,
325355da1ebSSage Weil 			   const char *name)
326355da1ebSSage Weil {
327355da1ebSSage Weil 	struct rb_node **p;
328355da1ebSSage Weil 	struct rb_node *parent = NULL;
329355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
33017db143fSSage Weil 	int name_len = strlen(name);
331355da1ebSSage Weil 	int c;
332355da1ebSSage Weil 
333355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
334355da1ebSSage Weil 	while (*p) {
335355da1ebSSage Weil 		parent = *p;
336355da1ebSSage Weil 		xattr = rb_entry(parent, struct ceph_inode_xattr, node);
337355da1ebSSage Weil 		c = strncmp(name, xattr->name, xattr->name_len);
33817db143fSSage Weil 		if (c == 0 && name_len > xattr->name_len)
33917db143fSSage Weil 			c = 1;
340355da1ebSSage Weil 		if (c < 0)
341355da1ebSSage Weil 			p = &(*p)->rb_left;
342355da1ebSSage Weil 		else if (c > 0)
343355da1ebSSage Weil 			p = &(*p)->rb_right;
344355da1ebSSage Weil 		else {
345355da1ebSSage Weil 			dout("__get_xattr %s: found %.*s\n", name,
346355da1ebSSage Weil 			     xattr->val_len, xattr->val);
347355da1ebSSage Weil 			return xattr;
348355da1ebSSage Weil 		}
349355da1ebSSage Weil 	}
350355da1ebSSage Weil 
351355da1ebSSage Weil 	dout("__get_xattr %s: not found\n", name);
352355da1ebSSage Weil 
353355da1ebSSage Weil 	return NULL;
354355da1ebSSage Weil }
355355da1ebSSage Weil 
356355da1ebSSage Weil static void __free_xattr(struct ceph_inode_xattr *xattr)
357355da1ebSSage Weil {
358355da1ebSSage Weil 	BUG_ON(!xattr);
359355da1ebSSage Weil 
360355da1ebSSage Weil 	if (xattr->should_free_name)
361355da1ebSSage Weil 		kfree((void *)xattr->name);
362355da1ebSSage Weil 	if (xattr->should_free_val)
363355da1ebSSage Weil 		kfree((void *)xattr->val);
364355da1ebSSage Weil 
365355da1ebSSage Weil 	kfree(xattr);
366355da1ebSSage Weil }
367355da1ebSSage Weil 
368355da1ebSSage Weil static int __remove_xattr(struct ceph_inode_info *ci,
369355da1ebSSage Weil 			  struct ceph_inode_xattr *xattr)
370355da1ebSSage Weil {
371355da1ebSSage Weil 	if (!xattr)
372355da1ebSSage Weil 		return -EOPNOTSUPP;
373355da1ebSSage Weil 
374355da1ebSSage Weil 	rb_erase(&xattr->node, &ci->i_xattrs.index);
375355da1ebSSage Weil 
376355da1ebSSage Weil 	if (xattr->should_free_name)
377355da1ebSSage Weil 		kfree((void *)xattr->name);
378355da1ebSSage Weil 	if (xattr->should_free_val)
379355da1ebSSage Weil 		kfree((void *)xattr->val);
380355da1ebSSage Weil 
381355da1ebSSage Weil 	ci->i_xattrs.names_size -= xattr->name_len;
382355da1ebSSage Weil 	ci->i_xattrs.vals_size -= xattr->val_len;
383355da1ebSSage Weil 	ci->i_xattrs.count--;
384355da1ebSSage Weil 	kfree(xattr);
385355da1ebSSage Weil 
386355da1ebSSage Weil 	return 0;
387355da1ebSSage Weil }
388355da1ebSSage Weil 
389355da1ebSSage Weil static int __remove_xattr_by_name(struct ceph_inode_info *ci,
390355da1ebSSage Weil 			   const char *name)
391355da1ebSSage Weil {
392355da1ebSSage Weil 	struct rb_node **p;
393355da1ebSSage Weil 	struct ceph_inode_xattr *xattr;
394355da1ebSSage Weil 	int err;
395355da1ebSSage Weil 
396355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
397355da1ebSSage Weil 	xattr = __get_xattr(ci, name);
398355da1ebSSage Weil 	err = __remove_xattr(ci, xattr);
399355da1ebSSage Weil 	return err;
400355da1ebSSage Weil }
401355da1ebSSage Weil 
402355da1ebSSage Weil static char *__copy_xattr_names(struct ceph_inode_info *ci,
403355da1ebSSage Weil 				char *dest)
404355da1ebSSage Weil {
405355da1ebSSage Weil 	struct rb_node *p;
406355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
407355da1ebSSage Weil 
408355da1ebSSage Weil 	p = rb_first(&ci->i_xattrs.index);
409355da1ebSSage Weil 	dout("__copy_xattr_names count=%d\n", ci->i_xattrs.count);
410355da1ebSSage Weil 
411355da1ebSSage Weil 	while (p) {
412355da1ebSSage Weil 		xattr = rb_entry(p, struct ceph_inode_xattr, node);
413355da1ebSSage Weil 		memcpy(dest, xattr->name, xattr->name_len);
414355da1ebSSage Weil 		dest[xattr->name_len] = '\0';
415355da1ebSSage Weil 
416355da1ebSSage Weil 		dout("dest=%s %p (%s) (%d/%d)\n", dest, xattr, xattr->name,
417355da1ebSSage Weil 		     xattr->name_len, ci->i_xattrs.names_size);
418355da1ebSSage Weil 
419355da1ebSSage Weil 		dest += xattr->name_len + 1;
420355da1ebSSage Weil 		p = rb_next(p);
421355da1ebSSage Weil 	}
422355da1ebSSage Weil 
423355da1ebSSage Weil 	return dest;
424355da1ebSSage Weil }
425355da1ebSSage Weil 
426355da1ebSSage Weil void __ceph_destroy_xattrs(struct ceph_inode_info *ci)
427355da1ebSSage Weil {
428355da1ebSSage Weil 	struct rb_node *p, *tmp;
429355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
430355da1ebSSage Weil 
431355da1ebSSage Weil 	p = rb_first(&ci->i_xattrs.index);
432355da1ebSSage Weil 
433355da1ebSSage Weil 	dout("__ceph_destroy_xattrs p=%p\n", p);
434355da1ebSSage Weil 
435355da1ebSSage Weil 	while (p) {
436355da1ebSSage Weil 		xattr = rb_entry(p, struct ceph_inode_xattr, node);
437355da1ebSSage Weil 		tmp = p;
438355da1ebSSage Weil 		p = rb_next(tmp);
439355da1ebSSage Weil 		dout("__ceph_destroy_xattrs next p=%p (%.*s)\n", p,
440355da1ebSSage Weil 		     xattr->name_len, xattr->name);
441355da1ebSSage Weil 		rb_erase(tmp, &ci->i_xattrs.index);
442355da1ebSSage Weil 
443355da1ebSSage Weil 		__free_xattr(xattr);
444355da1ebSSage Weil 	}
445355da1ebSSage Weil 
446355da1ebSSage Weil 	ci->i_xattrs.names_size = 0;
447355da1ebSSage Weil 	ci->i_xattrs.vals_size = 0;
448355da1ebSSage Weil 	ci->i_xattrs.index_version = 0;
449355da1ebSSage Weil 	ci->i_xattrs.count = 0;
450355da1ebSSage Weil 	ci->i_xattrs.index = RB_ROOT;
451355da1ebSSage Weil }
452355da1ebSSage Weil 
453355da1ebSSage Weil static int __build_xattrs(struct inode *inode)
454be655596SSage Weil 	__releases(ci->i_ceph_lock)
455be655596SSage Weil 	__acquires(ci->i_ceph_lock)
456355da1ebSSage Weil {
457355da1ebSSage Weil 	u32 namelen;
458355da1ebSSage Weil 	u32 numattr = 0;
459355da1ebSSage Weil 	void *p, *end;
460355da1ebSSage Weil 	u32 len;
461355da1ebSSage Weil 	const char *name, *val;
462355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
463355da1ebSSage Weil 	int xattr_version;
464355da1ebSSage Weil 	struct ceph_inode_xattr **xattrs = NULL;
46563ff78b2SSage Weil 	int err = 0;
466355da1ebSSage Weil 	int i;
467355da1ebSSage Weil 
468355da1ebSSage Weil 	dout("__build_xattrs() len=%d\n",
469355da1ebSSage Weil 	     ci->i_xattrs.blob ? (int)ci->i_xattrs.blob->vec.iov_len : 0);
470355da1ebSSage Weil 
471355da1ebSSage Weil 	if (ci->i_xattrs.index_version >= ci->i_xattrs.version)
472355da1ebSSage Weil 		return 0; /* already built */
473355da1ebSSage Weil 
474355da1ebSSage Weil 	__ceph_destroy_xattrs(ci);
475355da1ebSSage Weil 
476355da1ebSSage Weil start:
477355da1ebSSage Weil 	/* updated internal xattr rb tree */
478355da1ebSSage Weil 	if (ci->i_xattrs.blob && ci->i_xattrs.blob->vec.iov_len > 4) {
479355da1ebSSage Weil 		p = ci->i_xattrs.blob->vec.iov_base;
480355da1ebSSage Weil 		end = p + ci->i_xattrs.blob->vec.iov_len;
481355da1ebSSage Weil 		ceph_decode_32_safe(&p, end, numattr, bad);
482355da1ebSSage Weil 		xattr_version = ci->i_xattrs.version;
483be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
484355da1ebSSage Weil 
485355da1ebSSage Weil 		xattrs = kcalloc(numattr, sizeof(struct ceph_xattr *),
486355da1ebSSage Weil 				 GFP_NOFS);
487355da1ebSSage Weil 		err = -ENOMEM;
488355da1ebSSage Weil 		if (!xattrs)
489355da1ebSSage Weil 			goto bad_lock;
490355da1ebSSage Weil 		memset(xattrs, 0, numattr*sizeof(struct ceph_xattr *));
491355da1ebSSage Weil 		for (i = 0; i < numattr; i++) {
492355da1ebSSage Weil 			xattrs[i] = kmalloc(sizeof(struct ceph_inode_xattr),
493355da1ebSSage Weil 					    GFP_NOFS);
494355da1ebSSage Weil 			if (!xattrs[i])
495355da1ebSSage Weil 				goto bad_lock;
496355da1ebSSage Weil 		}
497355da1ebSSage Weil 
498be655596SSage Weil 		spin_lock(&ci->i_ceph_lock);
499355da1ebSSage Weil 		if (ci->i_xattrs.version != xattr_version) {
500355da1ebSSage Weil 			/* lost a race, retry */
501355da1ebSSage Weil 			for (i = 0; i < numattr; i++)
502355da1ebSSage Weil 				kfree(xattrs[i]);
503355da1ebSSage Weil 			kfree(xattrs);
50421ec6ffaSAlan Cox 			xattrs = NULL;
505355da1ebSSage Weil 			goto start;
506355da1ebSSage Weil 		}
507355da1ebSSage Weil 		err = -EIO;
508355da1ebSSage Weil 		while (numattr--) {
509355da1ebSSage Weil 			ceph_decode_32_safe(&p, end, len, bad);
510355da1ebSSage Weil 			namelen = len;
511355da1ebSSage Weil 			name = p;
512355da1ebSSage Weil 			p += len;
513355da1ebSSage Weil 			ceph_decode_32_safe(&p, end, len, bad);
514355da1ebSSage Weil 			val = p;
515355da1ebSSage Weil 			p += len;
516355da1ebSSage Weil 
517355da1ebSSage Weil 			err = __set_xattr(ci, name, namelen, val, len,
518355da1ebSSage Weil 					  0, 0, 0, &xattrs[numattr]);
519355da1ebSSage Weil 
520355da1ebSSage Weil 			if (err < 0)
521355da1ebSSage Weil 				goto bad;
522355da1ebSSage Weil 		}
523355da1ebSSage Weil 		kfree(xattrs);
524355da1ebSSage Weil 	}
525355da1ebSSage Weil 	ci->i_xattrs.index_version = ci->i_xattrs.version;
526355da1ebSSage Weil 	ci->i_xattrs.dirty = false;
527355da1ebSSage Weil 
528355da1ebSSage Weil 	return err;
529355da1ebSSage Weil bad_lock:
530be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
531355da1ebSSage Weil bad:
532355da1ebSSage Weil 	if (xattrs) {
533355da1ebSSage Weil 		for (i = 0; i < numattr; i++)
534355da1ebSSage Weil 			kfree(xattrs[i]);
535355da1ebSSage Weil 		kfree(xattrs);
536355da1ebSSage Weil 	}
537355da1ebSSage Weil 	ci->i_xattrs.names_size = 0;
538355da1ebSSage Weil 	return err;
539355da1ebSSage Weil }
540355da1ebSSage Weil 
541355da1ebSSage Weil static int __get_required_blob_size(struct ceph_inode_info *ci, int name_size,
542355da1ebSSage Weil 				    int val_size)
543355da1ebSSage Weil {
544355da1ebSSage Weil 	/*
545355da1ebSSage Weil 	 * 4 bytes for the length, and additional 4 bytes per each xattr name,
546355da1ebSSage Weil 	 * 4 bytes per each value
547355da1ebSSage Weil 	 */
548355da1ebSSage Weil 	int size = 4 + ci->i_xattrs.count*(4 + 4) +
549355da1ebSSage Weil 			     ci->i_xattrs.names_size +
550355da1ebSSage Weil 			     ci->i_xattrs.vals_size;
551355da1ebSSage Weil 	dout("__get_required_blob_size c=%d names.size=%d vals.size=%d\n",
552355da1ebSSage Weil 	     ci->i_xattrs.count, ci->i_xattrs.names_size,
553355da1ebSSage Weil 	     ci->i_xattrs.vals_size);
554355da1ebSSage Weil 
555355da1ebSSage Weil 	if (name_size)
556355da1ebSSage Weil 		size += 4 + 4 + name_size + val_size;
557355da1ebSSage Weil 
558355da1ebSSage Weil 	return size;
559355da1ebSSage Weil }
560355da1ebSSage Weil 
561355da1ebSSage Weil /*
562355da1ebSSage Weil  * If there are dirty xattrs, reencode xattrs into the prealloc_blob
563355da1ebSSage Weil  * and swap into place.
564355da1ebSSage Weil  */
565355da1ebSSage Weil void __ceph_build_xattrs_blob(struct ceph_inode_info *ci)
566355da1ebSSage Weil {
567355da1ebSSage Weil 	struct rb_node *p;
568355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
569355da1ebSSage Weil 	void *dest;
570355da1ebSSage Weil 
571355da1ebSSage Weil 	dout("__build_xattrs_blob %p\n", &ci->vfs_inode);
572355da1ebSSage Weil 	if (ci->i_xattrs.dirty) {
573355da1ebSSage Weil 		int need = __get_required_blob_size(ci, 0, 0);
574355da1ebSSage Weil 
575355da1ebSSage Weil 		BUG_ON(need > ci->i_xattrs.prealloc_blob->alloc_len);
576355da1ebSSage Weil 
577355da1ebSSage Weil 		p = rb_first(&ci->i_xattrs.index);
578355da1ebSSage Weil 		dest = ci->i_xattrs.prealloc_blob->vec.iov_base;
579355da1ebSSage Weil 
580355da1ebSSage Weil 		ceph_encode_32(&dest, ci->i_xattrs.count);
581355da1ebSSage Weil 		while (p) {
582355da1ebSSage Weil 			xattr = rb_entry(p, struct ceph_inode_xattr, node);
583355da1ebSSage Weil 
584355da1ebSSage Weil 			ceph_encode_32(&dest, xattr->name_len);
585355da1ebSSage Weil 			memcpy(dest, xattr->name, xattr->name_len);
586355da1ebSSage Weil 			dest += xattr->name_len;
587355da1ebSSage Weil 			ceph_encode_32(&dest, xattr->val_len);
588355da1ebSSage Weil 			memcpy(dest, xattr->val, xattr->val_len);
589355da1ebSSage Weil 			dest += xattr->val_len;
590355da1ebSSage Weil 
591355da1ebSSage Weil 			p = rb_next(p);
592355da1ebSSage Weil 		}
593355da1ebSSage Weil 
594355da1ebSSage Weil 		/* adjust buffer len; it may be larger than we need */
595355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob->vec.iov_len =
596355da1ebSSage Weil 			dest - ci->i_xattrs.prealloc_blob->vec.iov_base;
597355da1ebSSage Weil 
598b6c1d5b8SSage Weil 		if (ci->i_xattrs.blob)
599355da1ebSSage Weil 			ceph_buffer_put(ci->i_xattrs.blob);
600355da1ebSSage Weil 		ci->i_xattrs.blob = ci->i_xattrs.prealloc_blob;
601355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob = NULL;
602355da1ebSSage Weil 		ci->i_xattrs.dirty = false;
6034a625be4SSage Weil 		ci->i_xattrs.version++;
604355da1ebSSage Weil 	}
605355da1ebSSage Weil }
606355da1ebSSage Weil 
607355da1ebSSage Weil ssize_t ceph_getxattr(struct dentry *dentry, const char *name, void *value,
608355da1ebSSage Weil 		      size_t size)
609355da1ebSSage Weil {
610355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
611355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
612355da1ebSSage Weil 	int err;
613355da1ebSSage Weil 	struct ceph_inode_xattr *xattr;
614881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr = NULL;
615355da1ebSSage Weil 
616355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
617355da1ebSSage Weil 		return -ENODATA;
618355da1ebSSage Weil 
619be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
620355da1ebSSage Weil 	dout("getxattr %p ver=%lld index_ver=%lld\n", inode,
621355da1ebSSage Weil 	     ci->i_xattrs.version, ci->i_xattrs.index_version);
622355da1ebSSage Weil 
6230bee82fbSSage Weil 	/* let's see if a virtual xattr was requested */
6240bee82fbSSage Weil 	vxattr = ceph_match_vxattr(inode, name);
6250bee82fbSSage Weil 	if (vxattr && !(vxattr->exists_cb && !vxattr->exists_cb(ci))) {
6260bee82fbSSage Weil 		err = vxattr->getxattr_cb(ci, value, size);
6270bee82fbSSage Weil 		goto out;
6280bee82fbSSage Weil 	}
6290bee82fbSSage Weil 
630355da1ebSSage Weil 	if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
631355da1ebSSage Weil 	    (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
632355da1ebSSage Weil 		goto get_xattr;
633355da1ebSSage Weil 	} else {
634be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
635355da1ebSSage Weil 		/* get xattrs from mds (if we don't already have them) */
636355da1ebSSage Weil 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
637355da1ebSSage Weil 		if (err)
638355da1ebSSage Weil 			return err;
639355da1ebSSage Weil 	}
640355da1ebSSage Weil 
641be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
642355da1ebSSage Weil 
643355da1ebSSage Weil 	err = __build_xattrs(inode);
644355da1ebSSage Weil 	if (err < 0)
645355da1ebSSage Weil 		goto out;
646355da1ebSSage Weil 
647355da1ebSSage Weil get_xattr:
648355da1ebSSage Weil 	err = -ENODATA;  /* == ENOATTR */
649355da1ebSSage Weil 	xattr = __get_xattr(ci, name);
6500bee82fbSSage Weil 	if (!xattr)
651355da1ebSSage Weil 		goto out;
652355da1ebSSage Weil 
653355da1ebSSage Weil 	err = -ERANGE;
654355da1ebSSage Weil 	if (size && size < xattr->val_len)
655355da1ebSSage Weil 		goto out;
656355da1ebSSage Weil 
657355da1ebSSage Weil 	err = xattr->val_len;
658355da1ebSSage Weil 	if (size == 0)
659355da1ebSSage Weil 		goto out;
660355da1ebSSage Weil 
661355da1ebSSage Weil 	memcpy(value, xattr->val, xattr->val_len);
662355da1ebSSage Weil 
663355da1ebSSage Weil out:
664be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
665355da1ebSSage Weil 	return err;
666355da1ebSSage Weil }
667355da1ebSSage Weil 
668355da1ebSSage Weil ssize_t ceph_listxattr(struct dentry *dentry, char *names, size_t size)
669355da1ebSSage Weil {
670355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
671355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
672881a5fa2SAlex Elder 	struct ceph_vxattr *vxattrs = ceph_inode_vxattrs(inode);
673355da1ebSSage Weil 	u32 vir_namelen = 0;
674355da1ebSSage Weil 	u32 namelen;
675355da1ebSSage Weil 	int err;
676355da1ebSSage Weil 	u32 len;
677355da1ebSSage Weil 	int i;
678355da1ebSSage Weil 
679be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
680355da1ebSSage Weil 	dout("listxattr %p ver=%lld index_ver=%lld\n", inode,
681355da1ebSSage Weil 	     ci->i_xattrs.version, ci->i_xattrs.index_version);
682355da1ebSSage Weil 
683355da1ebSSage Weil 	if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
684bddfa3ccSHenry C Chang 	    (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
685355da1ebSSage Weil 		goto list_xattr;
686355da1ebSSage Weil 	} else {
687be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
688355da1ebSSage Weil 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
689355da1ebSSage Weil 		if (err)
690355da1ebSSage Weil 			return err;
691355da1ebSSage Weil 	}
692355da1ebSSage Weil 
693be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
694355da1ebSSage Weil 
695355da1ebSSage Weil 	err = __build_xattrs(inode);
696355da1ebSSage Weil 	if (err < 0)
697355da1ebSSage Weil 		goto out;
698355da1ebSSage Weil 
699355da1ebSSage Weil list_xattr:
7003ce6cd12SAlex Elder 	/*
7013ce6cd12SAlex Elder 	 * Start with virtual dir xattr names (if any) (including
7023ce6cd12SAlex Elder 	 * terminating '\0' characters for each).
7033ce6cd12SAlex Elder 	 */
7043ce6cd12SAlex Elder 	vir_namelen = ceph_vxattrs_name_size(vxattrs);
7053ce6cd12SAlex Elder 
706355da1ebSSage Weil 	/* adding 1 byte per each variable due to the null termination */
707b65917ddSSage Weil 	namelen = ci->i_xattrs.names_size + ci->i_xattrs.count;
708355da1ebSSage Weil 	err = -ERANGE;
709b65917ddSSage Weil 	if (size && vir_namelen + namelen > size)
710355da1ebSSage Weil 		goto out;
711355da1ebSSage Weil 
712b65917ddSSage Weil 	err = namelen + vir_namelen;
713355da1ebSSage Weil 	if (size == 0)
714355da1ebSSage Weil 		goto out;
715355da1ebSSage Weil 
716355da1ebSSage Weil 	names = __copy_xattr_names(ci, names);
717355da1ebSSage Weil 
718355da1ebSSage Weil 	/* virtual xattr names, too */
719b65917ddSSage Weil 	err = namelen;
720b65917ddSSage Weil 	if (vxattrs) {
721355da1ebSSage Weil 		for (i = 0; vxattrs[i].name; i++) {
722b65917ddSSage Weil 			if (!vxattrs[i].hidden &&
723b65917ddSSage Weil 			    !(vxattrs[i].exists_cb &&
724b65917ddSSage Weil 			      !vxattrs[i].exists_cb(ci))) {
725355da1ebSSage Weil 				len = sprintf(names, "%s", vxattrs[i].name);
726355da1ebSSage Weil 				names += len + 1;
727b65917ddSSage Weil 				err += len + 1;
728b65917ddSSage Weil 			}
729b65917ddSSage Weil 		}
730355da1ebSSage Weil 	}
731355da1ebSSage Weil 
732355da1ebSSage Weil out:
733be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
734355da1ebSSage Weil 	return err;
735355da1ebSSage Weil }
736355da1ebSSage Weil 
737355da1ebSSage Weil static int ceph_sync_setxattr(struct dentry *dentry, const char *name,
738355da1ebSSage Weil 			      const char *value, size_t size, int flags)
739355da1ebSSage Weil {
7403d14c5d2SYehuda Sadeh 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
741355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
742355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
7435f21c96dSSage Weil 	struct inode *parent_inode;
744355da1ebSSage Weil 	struct ceph_mds_request *req;
7453d14c5d2SYehuda Sadeh 	struct ceph_mds_client *mdsc = fsc->mdsc;
746355da1ebSSage Weil 	int err;
747355da1ebSSage Weil 	int i, nr_pages;
748355da1ebSSage Weil 	struct page **pages = NULL;
749355da1ebSSage Weil 	void *kaddr;
750355da1ebSSage Weil 
751355da1ebSSage Weil 	/* copy value into some pages */
752355da1ebSSage Weil 	nr_pages = calc_pages_for(0, size);
753355da1ebSSage Weil 	if (nr_pages) {
754355da1ebSSage Weil 		pages = kmalloc(sizeof(pages[0])*nr_pages, GFP_NOFS);
755355da1ebSSage Weil 		if (!pages)
756355da1ebSSage Weil 			return -ENOMEM;
757355da1ebSSage Weil 		err = -ENOMEM;
758355da1ebSSage Weil 		for (i = 0; i < nr_pages; i++) {
75931459fe4SYehuda Sadeh 			pages[i] = __page_cache_alloc(GFP_NOFS);
760355da1ebSSage Weil 			if (!pages[i]) {
761355da1ebSSage Weil 				nr_pages = i;
762355da1ebSSage Weil 				goto out;
763355da1ebSSage Weil 			}
764355da1ebSSage Weil 			kaddr = kmap(pages[i]);
765355da1ebSSage Weil 			memcpy(kaddr, value + i*PAGE_CACHE_SIZE,
766355da1ebSSage Weil 			       min(PAGE_CACHE_SIZE, size-i*PAGE_CACHE_SIZE));
767355da1ebSSage Weil 		}
768355da1ebSSage Weil 	}
769355da1ebSSage Weil 
770355da1ebSSage Weil 	dout("setxattr value=%.*s\n", (int)size, value);
771355da1ebSSage Weil 
772355da1ebSSage Weil 	/* do request */
773355da1ebSSage Weil 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETXATTR,
774355da1ebSSage Weil 				       USE_AUTH_MDS);
77560d87733SJulia Lawall 	if (IS_ERR(req)) {
77660d87733SJulia Lawall 		err = PTR_ERR(req);
77760d87733SJulia Lawall 		goto out;
77860d87733SJulia Lawall 	}
77970b666c3SSage Weil 	req->r_inode = inode;
78070b666c3SSage Weil 	ihold(inode);
781355da1ebSSage Weil 	req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
782355da1ebSSage Weil 	req->r_num_caps = 1;
783355da1ebSSage Weil 	req->r_args.setxattr.flags = cpu_to_le32(flags);
784355da1ebSSage Weil 	req->r_path2 = kstrdup(name, GFP_NOFS);
785355da1ebSSage Weil 
786355da1ebSSage Weil 	req->r_pages = pages;
787355da1ebSSage Weil 	req->r_num_pages = nr_pages;
788355da1ebSSage Weil 	req->r_data_len = size;
789355da1ebSSage Weil 
790355da1ebSSage Weil 	dout("xattr.ver (before): %lld\n", ci->i_xattrs.version);
7915f21c96dSSage Weil 	parent_inode = ceph_get_dentry_parent_inode(dentry);
792355da1ebSSage Weil 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
7935f21c96dSSage Weil 	iput(parent_inode);
794355da1ebSSage Weil 	ceph_mdsc_put_request(req);
795355da1ebSSage Weil 	dout("xattr.ver (after): %lld\n", ci->i_xattrs.version);
796355da1ebSSage Weil 
797355da1ebSSage Weil out:
798355da1ebSSage Weil 	if (pages) {
799355da1ebSSage Weil 		for (i = 0; i < nr_pages; i++)
800355da1ebSSage Weil 			__free_page(pages[i]);
801355da1ebSSage Weil 		kfree(pages);
802355da1ebSSage Weil 	}
803355da1ebSSage Weil 	return err;
804355da1ebSSage Weil }
805355da1ebSSage Weil 
806355da1ebSSage Weil int ceph_setxattr(struct dentry *dentry, const char *name,
807355da1ebSSage Weil 		  const void *value, size_t size, int flags)
808355da1ebSSage Weil {
809355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
810881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr;
811355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
81218fa8b3fSAlex Elder 	int issued;
813355da1ebSSage Weil 	int err;
81418fa8b3fSAlex Elder 	int dirty;
815355da1ebSSage Weil 	int name_len = strlen(name);
816355da1ebSSage Weil 	int val_len = size;
817355da1ebSSage Weil 	char *newname = NULL;
818355da1ebSSage Weil 	char *newval = NULL;
819355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
820355da1ebSSage Weil 	int required_blob_size;
821355da1ebSSage Weil 
822355da1ebSSage Weil 	if (ceph_snap(inode) != CEPH_NOSNAP)
823355da1ebSSage Weil 		return -EROFS;
824355da1ebSSage Weil 
825355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
826355da1ebSSage Weil 		return -EOPNOTSUPP;
827355da1ebSSage Weil 
82806476a69SAlex Elder 	vxattr = ceph_match_vxattr(inode, name);
829355da1ebSSage Weil 	if (vxattr && vxattr->readonly)
830355da1ebSSage Weil 		return -EOPNOTSUPP;
831355da1ebSSage Weil 
8323adf654dSSage Weil 	/* pass any unhandled ceph.* xattrs through to the MDS */
8333adf654dSSage Weil 	if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
8343adf654dSSage Weil 		goto do_sync_unlocked;
8353adf654dSSage Weil 
836355da1ebSSage Weil 	/* preallocate memory for xattr name, value, index node */
837355da1ebSSage Weil 	err = -ENOMEM;
83861413c2fSJulia Lawall 	newname = kmemdup(name, name_len + 1, GFP_NOFS);
839355da1ebSSage Weil 	if (!newname)
840355da1ebSSage Weil 		goto out;
841355da1ebSSage Weil 
842355da1ebSSage Weil 	if (val_len) {
843b829c195SAlex Elder 		newval = kmemdup(value, val_len, GFP_NOFS);
844355da1ebSSage Weil 		if (!newval)
845355da1ebSSage Weil 			goto out;
846355da1ebSSage Weil 	}
847355da1ebSSage Weil 
848355da1ebSSage Weil 	xattr = kmalloc(sizeof(struct ceph_inode_xattr), GFP_NOFS);
849355da1ebSSage Weil 	if (!xattr)
850355da1ebSSage Weil 		goto out;
851355da1ebSSage Weil 
852be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
853355da1ebSSage Weil retry:
854355da1ebSSage Weil 	issued = __ceph_caps_issued(ci, NULL);
85518fa8b3fSAlex Elder 	dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
856355da1ebSSage Weil 	if (!(issued & CEPH_CAP_XATTR_EXCL))
857355da1ebSSage Weil 		goto do_sync;
858355da1ebSSage Weil 	__build_xattrs(inode);
859355da1ebSSage Weil 
860355da1ebSSage Weil 	required_blob_size = __get_required_blob_size(ci, name_len, val_len);
861355da1ebSSage Weil 
862355da1ebSSage Weil 	if (!ci->i_xattrs.prealloc_blob ||
863355da1ebSSage Weil 	    required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
86418fa8b3fSAlex Elder 		struct ceph_buffer *blob;
865355da1ebSSage Weil 
866be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
867355da1ebSSage Weil 		dout(" preaallocating new blob size=%d\n", required_blob_size);
868b6c1d5b8SSage Weil 		blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
869355da1ebSSage Weil 		if (!blob)
870355da1ebSSage Weil 			goto out;
871be655596SSage Weil 		spin_lock(&ci->i_ceph_lock);
872b6c1d5b8SSage Weil 		if (ci->i_xattrs.prealloc_blob)
873355da1ebSSage Weil 			ceph_buffer_put(ci->i_xattrs.prealloc_blob);
874355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob = blob;
875355da1ebSSage Weil 		goto retry;
876355da1ebSSage Weil 	}
877355da1ebSSage Weil 
878355da1ebSSage Weil 	err = __set_xattr(ci, newname, name_len, newval,
879355da1ebSSage Weil 			  val_len, 1, 1, 1, &xattr);
88018fa8b3fSAlex Elder 
881fca65b4aSSage Weil 	dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
882355da1ebSSage Weil 	ci->i_xattrs.dirty = true;
883355da1ebSSage Weil 	inode->i_ctime = CURRENT_TIME;
88418fa8b3fSAlex Elder 
885be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
886fca65b4aSSage Weil 	if (dirty)
887fca65b4aSSage Weil 		__mark_inode_dirty(inode, dirty);
888355da1ebSSage Weil 	return err;
889355da1ebSSage Weil 
890355da1ebSSage Weil do_sync:
891be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
8923adf654dSSage Weil do_sync_unlocked:
893355da1ebSSage Weil 	err = ceph_sync_setxattr(dentry, name, value, size, flags);
894355da1ebSSage Weil out:
895355da1ebSSage Weil 	kfree(newname);
896355da1ebSSage Weil 	kfree(newval);
897355da1ebSSage Weil 	kfree(xattr);
898355da1ebSSage Weil 	return err;
899355da1ebSSage Weil }
900355da1ebSSage Weil 
901355da1ebSSage Weil static int ceph_send_removexattr(struct dentry *dentry, const char *name)
902355da1ebSSage Weil {
9033d14c5d2SYehuda Sadeh 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
9043d14c5d2SYehuda Sadeh 	struct ceph_mds_client *mdsc = fsc->mdsc;
905355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
9065f21c96dSSage Weil 	struct inode *parent_inode;
907355da1ebSSage Weil 	struct ceph_mds_request *req;
908355da1ebSSage Weil 	int err;
909355da1ebSSage Weil 
910355da1ebSSage Weil 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RMXATTR,
911355da1ebSSage Weil 				       USE_AUTH_MDS);
912355da1ebSSage Weil 	if (IS_ERR(req))
913355da1ebSSage Weil 		return PTR_ERR(req);
91470b666c3SSage Weil 	req->r_inode = inode;
91570b666c3SSage Weil 	ihold(inode);
916355da1ebSSage Weil 	req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
917355da1ebSSage Weil 	req->r_num_caps = 1;
918355da1ebSSage Weil 	req->r_path2 = kstrdup(name, GFP_NOFS);
919355da1ebSSage Weil 
9205f21c96dSSage Weil 	parent_inode = ceph_get_dentry_parent_inode(dentry);
921355da1ebSSage Weil 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
9225f21c96dSSage Weil 	iput(parent_inode);
923355da1ebSSage Weil 	ceph_mdsc_put_request(req);
924355da1ebSSage Weil 	return err;
925355da1ebSSage Weil }
926355da1ebSSage Weil 
927355da1ebSSage Weil int ceph_removexattr(struct dentry *dentry, const char *name)
928355da1ebSSage Weil {
929355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
930881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr;
931355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
932355da1ebSSage Weil 	int issued;
933355da1ebSSage Weil 	int err;
93483eb26afSAlex Elder 	int required_blob_size;
935fca65b4aSSage Weil 	int dirty;
936355da1ebSSage Weil 
937355da1ebSSage Weil 	if (ceph_snap(inode) != CEPH_NOSNAP)
938355da1ebSSage Weil 		return -EROFS;
939355da1ebSSage Weil 
940355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
941355da1ebSSage Weil 		return -EOPNOTSUPP;
942355da1ebSSage Weil 
94306476a69SAlex Elder 	vxattr = ceph_match_vxattr(inode, name);
944355da1ebSSage Weil 	if (vxattr && vxattr->readonly)
945355da1ebSSage Weil 		return -EOPNOTSUPP;
946355da1ebSSage Weil 
947d421acb1SSage Weil 	/* pass any unhandled ceph.* xattrs through to the MDS */
948d421acb1SSage Weil 	if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
949d421acb1SSage Weil 		goto do_sync_unlocked;
950d421acb1SSage Weil 
95183eb26afSAlex Elder 	err = -ENOMEM;
952be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
95383eb26afSAlex Elder retry:
954355da1ebSSage Weil 	issued = __ceph_caps_issued(ci, NULL);
955355da1ebSSage Weil 	dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
956355da1ebSSage Weil 
957355da1ebSSage Weil 	if (!(issued & CEPH_CAP_XATTR_EXCL))
958355da1ebSSage Weil 		goto do_sync;
95918fa8b3fSAlex Elder 	__build_xattrs(inode);
960355da1ebSSage Weil 
96183eb26afSAlex Elder 	required_blob_size = __get_required_blob_size(ci, 0, 0);
96283eb26afSAlex Elder 
96383eb26afSAlex Elder 	if (!ci->i_xattrs.prealloc_blob ||
96483eb26afSAlex Elder 	    required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
96583eb26afSAlex Elder 		struct ceph_buffer *blob;
96683eb26afSAlex Elder 
96783eb26afSAlex Elder 		spin_unlock(&ci->i_ceph_lock);
96883eb26afSAlex Elder 		dout(" preaallocating new blob size=%d\n", required_blob_size);
96983eb26afSAlex Elder 		blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
97083eb26afSAlex Elder 		if (!blob)
97183eb26afSAlex Elder 			goto out;
97283eb26afSAlex Elder 		spin_lock(&ci->i_ceph_lock);
97383eb26afSAlex Elder 		if (ci->i_xattrs.prealloc_blob)
97483eb26afSAlex Elder 			ceph_buffer_put(ci->i_xattrs.prealloc_blob);
97583eb26afSAlex Elder 		ci->i_xattrs.prealloc_blob = blob;
97683eb26afSAlex Elder 		goto retry;
97783eb26afSAlex Elder 	}
97883eb26afSAlex Elder 
979355da1ebSSage Weil 	err = __remove_xattr_by_name(ceph_inode(inode), name);
98018fa8b3fSAlex Elder 
981fca65b4aSSage Weil 	dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
982355da1ebSSage Weil 	ci->i_xattrs.dirty = true;
983355da1ebSSage Weil 	inode->i_ctime = CURRENT_TIME;
984be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
985fca65b4aSSage Weil 	if (dirty)
986fca65b4aSSage Weil 		__mark_inode_dirty(inode, dirty);
987355da1ebSSage Weil 	return err;
988355da1ebSSage Weil do_sync:
989be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
990d421acb1SSage Weil do_sync_unlocked:
991355da1ebSSage Weil 	err = ceph_send_removexattr(dentry, name);
99283eb26afSAlex Elder out:
993355da1ebSSage Weil 	return err;
994355da1ebSSage Weil }
995355da1ebSSage Weil 
996