xref: /openbmc/linux/fs/ceph/xattr.c (revision 0bee82fb4b8d49541fe474ed460d2b917f329568)
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 
36355da1ebSSage Weil /* directories */
37355da1ebSSage Weil 
38aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_entries(struct ceph_inode_info *ci, char *val,
39355da1ebSSage Weil 					size_t size)
40355da1ebSSage Weil {
41355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_files + ci->i_subdirs);
42355da1ebSSage Weil }
43355da1ebSSage Weil 
44aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_files(struct ceph_inode_info *ci, char *val,
45355da1ebSSage Weil 				      size_t size)
46355da1ebSSage Weil {
47355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_files);
48355da1ebSSage Weil }
49355da1ebSSage Weil 
50aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_subdirs(struct ceph_inode_info *ci, char *val,
51355da1ebSSage Weil 					size_t size)
52355da1ebSSage Weil {
53355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_subdirs);
54355da1ebSSage Weil }
55355da1ebSSage Weil 
56aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rentries(struct ceph_inode_info *ci, char *val,
57355da1ebSSage Weil 					 size_t size)
58355da1ebSSage Weil {
59355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rfiles + ci->i_rsubdirs);
60355da1ebSSage Weil }
61355da1ebSSage Weil 
62aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rfiles(struct ceph_inode_info *ci, char *val,
63355da1ebSSage Weil 				       size_t size)
64355da1ebSSage Weil {
65355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rfiles);
66355da1ebSSage Weil }
67355da1ebSSage Weil 
68aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rsubdirs(struct ceph_inode_info *ci, char *val,
69355da1ebSSage Weil 					 size_t size)
70355da1ebSSage Weil {
71355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rsubdirs);
72355da1ebSSage Weil }
73355da1ebSSage Weil 
74aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rbytes(struct ceph_inode_info *ci, char *val,
75355da1ebSSage Weil 				       size_t size)
76355da1ebSSage Weil {
77355da1ebSSage Weil 	return snprintf(val, size, "%lld", ci->i_rbytes);
78355da1ebSSage Weil }
79355da1ebSSage Weil 
80aa4066edSAlex Elder static size_t ceph_vxattrcb_dir_rctime(struct ceph_inode_info *ci, char *val,
81355da1ebSSage Weil 				       size_t size)
82355da1ebSSage Weil {
833489b42aSAlex Elder 	return snprintf(val, size, "%ld.09%ld", (long)ci->i_rctime.tv_sec,
84355da1ebSSage Weil 			(long)ci->i_rctime.tv_nsec);
85355da1ebSSage Weil }
86355da1ebSSage Weil 
87eb788084SAlex Elder #define CEPH_XATTR_NAME(_type, _name)	XATTR_CEPH_PREFIX #_type "." #_name
88eb788084SAlex Elder 
89eb788084SAlex Elder #define XATTR_NAME_CEPH(_type, _name)					\
90eb788084SAlex Elder 	{								\
91eb788084SAlex Elder 		.name = CEPH_XATTR_NAME(_type, _name),			\
923ce6cd12SAlex Elder 		.name_size = sizeof (CEPH_XATTR_NAME(_type, _name)), \
93aa4066edSAlex Elder 		.getxattr_cb = ceph_vxattrcb_ ## _type ## _ ## _name, \
94eb788084SAlex Elder 		.readonly = true,				\
958860147aSSage Weil 		.hidden = false,				\
96f36e4472SSage Weil 		.exists_cb = NULL,			\
97eb788084SAlex Elder 	}
98eb788084SAlex Elder 
99881a5fa2SAlex Elder static struct ceph_vxattr ceph_dir_vxattrs[] = {
100eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, entries),
101eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, files),
102eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, subdirs),
103eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rentries),
104eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rfiles),
105eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rsubdirs),
106eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rbytes),
107eb788084SAlex Elder 	XATTR_NAME_CEPH(dir, rctime),
108eb788084SAlex Elder 	{ 0 }	/* Required table terminator */
109355da1ebSSage Weil };
1103ce6cd12SAlex Elder static size_t ceph_dir_vxattrs_name_size;	/* total size of all names */
111355da1ebSSage Weil 
112355da1ebSSage Weil /* files */
113355da1ebSSage Weil 
114aa4066edSAlex Elder static size_t ceph_vxattrcb_file_layout(struct ceph_inode_info *ci, char *val,
115355da1ebSSage Weil 				   size_t size)
116355da1ebSSage Weil {
117b195befdSSage Weil 	int ret;
118b195befdSSage Weil 
119b195befdSSage Weil 	ret = snprintf(val, size,
120355da1ebSSage Weil 		"chunk_bytes=%lld\nstripe_count=%lld\nobject_size=%lld\n",
121355da1ebSSage Weil 		(unsigned long long)ceph_file_layout_su(ci->i_layout),
122355da1ebSSage Weil 		(unsigned long long)ceph_file_layout_stripe_count(ci->i_layout),
123355da1ebSSage Weil 		(unsigned long long)ceph_file_layout_object_size(ci->i_layout));
124b195befdSSage Weil 	return ret;
125355da1ebSSage Weil }
126355da1ebSSage Weil 
127881a5fa2SAlex Elder static struct ceph_vxattr ceph_file_vxattrs[] = {
128eb788084SAlex Elder 	XATTR_NAME_CEPH(file, layout),
129eb788084SAlex Elder 	{ 0 }	/* Required table terminator */
130355da1ebSSage Weil };
1313ce6cd12SAlex Elder static size_t ceph_file_vxattrs_name_size;	/* total size of all names */
132355da1ebSSage Weil 
133881a5fa2SAlex Elder static struct ceph_vxattr *ceph_inode_vxattrs(struct inode *inode)
134355da1ebSSage Weil {
135355da1ebSSage Weil 	if (S_ISDIR(inode->i_mode))
136355da1ebSSage Weil 		return ceph_dir_vxattrs;
137355da1ebSSage Weil 	else if (S_ISREG(inode->i_mode))
138355da1ebSSage Weil 		return ceph_file_vxattrs;
139355da1ebSSage Weil 	return NULL;
140355da1ebSSage Weil }
141355da1ebSSage Weil 
1423ce6cd12SAlex Elder static size_t ceph_vxattrs_name_size(struct ceph_vxattr *vxattrs)
1433ce6cd12SAlex Elder {
1443ce6cd12SAlex Elder 	if (vxattrs == ceph_dir_vxattrs)
1453ce6cd12SAlex Elder 		return ceph_dir_vxattrs_name_size;
1463ce6cd12SAlex Elder 	if (vxattrs == ceph_file_vxattrs)
1473ce6cd12SAlex Elder 		return ceph_file_vxattrs_name_size;
1483ce6cd12SAlex Elder 	BUG();
1493ce6cd12SAlex Elder 
1503ce6cd12SAlex Elder 	return 0;
1513ce6cd12SAlex Elder }
1523ce6cd12SAlex Elder 
1533ce6cd12SAlex Elder /*
1543ce6cd12SAlex Elder  * Compute the aggregate size (including terminating '\0') of all
1553ce6cd12SAlex Elder  * virtual extended attribute names in the given vxattr table.
1563ce6cd12SAlex Elder  */
1573ce6cd12SAlex Elder static size_t __init vxattrs_name_size(struct ceph_vxattr *vxattrs)
1583ce6cd12SAlex Elder {
1593ce6cd12SAlex Elder 	struct ceph_vxattr *vxattr;
1603ce6cd12SAlex Elder 	size_t size = 0;
1613ce6cd12SAlex Elder 
1623ce6cd12SAlex Elder 	for (vxattr = vxattrs; vxattr->name; vxattr++)
1638860147aSSage Weil 		if (!vxattr->hidden)
1643ce6cd12SAlex Elder 			size += vxattr->name_size;
1653ce6cd12SAlex Elder 
1663ce6cd12SAlex Elder 	return size;
1673ce6cd12SAlex Elder }
1683ce6cd12SAlex Elder 
1693ce6cd12SAlex Elder /* Routines called at initialization and exit time */
1703ce6cd12SAlex Elder 
1713ce6cd12SAlex Elder void __init ceph_xattr_init(void)
1723ce6cd12SAlex Elder {
1733ce6cd12SAlex Elder 	ceph_dir_vxattrs_name_size = vxattrs_name_size(ceph_dir_vxattrs);
1743ce6cd12SAlex Elder 	ceph_file_vxattrs_name_size = vxattrs_name_size(ceph_file_vxattrs);
1753ce6cd12SAlex Elder }
1763ce6cd12SAlex Elder 
1773ce6cd12SAlex Elder void ceph_xattr_exit(void)
1783ce6cd12SAlex Elder {
1793ce6cd12SAlex Elder 	ceph_dir_vxattrs_name_size = 0;
1803ce6cd12SAlex Elder 	ceph_file_vxattrs_name_size = 0;
1813ce6cd12SAlex Elder }
1823ce6cd12SAlex Elder 
183881a5fa2SAlex Elder static struct ceph_vxattr *ceph_match_vxattr(struct inode *inode,
184355da1ebSSage Weil 						const char *name)
185355da1ebSSage Weil {
186881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr = ceph_inode_vxattrs(inode);
18706476a69SAlex Elder 
18806476a69SAlex Elder 	if (vxattr) {
18906476a69SAlex Elder 		while (vxattr->name) {
19006476a69SAlex Elder 			if (!strcmp(vxattr->name, name))
191355da1ebSSage Weil 				return vxattr;
192355da1ebSSage Weil 			vxattr++;
19306476a69SAlex Elder 		}
19406476a69SAlex Elder 	}
19506476a69SAlex Elder 
196355da1ebSSage Weil 	return NULL;
197355da1ebSSage Weil }
198355da1ebSSage Weil 
199355da1ebSSage Weil static int __set_xattr(struct ceph_inode_info *ci,
200355da1ebSSage Weil 			   const char *name, int name_len,
201355da1ebSSage Weil 			   const char *val, int val_len,
202355da1ebSSage Weil 			   int dirty,
203355da1ebSSage Weil 			   int should_free_name, int should_free_val,
204355da1ebSSage Weil 			   struct ceph_inode_xattr **newxattr)
205355da1ebSSage Weil {
206355da1ebSSage Weil 	struct rb_node **p;
207355da1ebSSage Weil 	struct rb_node *parent = NULL;
208355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
209355da1ebSSage Weil 	int c;
210355da1ebSSage Weil 	int new = 0;
211355da1ebSSage Weil 
212355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
213355da1ebSSage Weil 	while (*p) {
214355da1ebSSage Weil 		parent = *p;
215355da1ebSSage Weil 		xattr = rb_entry(parent, struct ceph_inode_xattr, node);
216355da1ebSSage Weil 		c = strncmp(name, xattr->name, min(name_len, xattr->name_len));
217355da1ebSSage Weil 		if (c < 0)
218355da1ebSSage Weil 			p = &(*p)->rb_left;
219355da1ebSSage Weil 		else if (c > 0)
220355da1ebSSage Weil 			p = &(*p)->rb_right;
221355da1ebSSage Weil 		else {
222355da1ebSSage Weil 			if (name_len == xattr->name_len)
223355da1ebSSage Weil 				break;
224355da1ebSSage Weil 			else if (name_len < xattr->name_len)
225355da1ebSSage Weil 				p = &(*p)->rb_left;
226355da1ebSSage Weil 			else
227355da1ebSSage Weil 				p = &(*p)->rb_right;
228355da1ebSSage Weil 		}
229355da1ebSSage Weil 		xattr = NULL;
230355da1ebSSage Weil 	}
231355da1ebSSage Weil 
232355da1ebSSage Weil 	if (!xattr) {
233355da1ebSSage Weil 		new = 1;
234355da1ebSSage Weil 		xattr = *newxattr;
235355da1ebSSage Weil 		xattr->name = name;
236355da1ebSSage Weil 		xattr->name_len = name_len;
237355da1ebSSage Weil 		xattr->should_free_name = should_free_name;
238355da1ebSSage Weil 
239355da1ebSSage Weil 		ci->i_xattrs.count++;
240355da1ebSSage Weil 		dout("__set_xattr count=%d\n", ci->i_xattrs.count);
241355da1ebSSage Weil 	} else {
242355da1ebSSage Weil 		kfree(*newxattr);
243355da1ebSSage Weil 		*newxattr = NULL;
244355da1ebSSage Weil 		if (xattr->should_free_val)
245355da1ebSSage Weil 			kfree((void *)xattr->val);
246355da1ebSSage Weil 
247355da1ebSSage Weil 		if (should_free_name) {
248355da1ebSSage Weil 			kfree((void *)name);
249355da1ebSSage Weil 			name = xattr->name;
250355da1ebSSage Weil 		}
251355da1ebSSage Weil 		ci->i_xattrs.names_size -= xattr->name_len;
252355da1ebSSage Weil 		ci->i_xattrs.vals_size -= xattr->val_len;
253355da1ebSSage Weil 	}
254355da1ebSSage Weil 	ci->i_xattrs.names_size += name_len;
255355da1ebSSage Weil 	ci->i_xattrs.vals_size += val_len;
256355da1ebSSage Weil 	if (val)
257355da1ebSSage Weil 		xattr->val = val;
258355da1ebSSage Weil 	else
259355da1ebSSage Weil 		xattr->val = "";
260355da1ebSSage Weil 
261355da1ebSSage Weil 	xattr->val_len = val_len;
262355da1ebSSage Weil 	xattr->dirty = dirty;
263355da1ebSSage Weil 	xattr->should_free_val = (val && should_free_val);
264355da1ebSSage Weil 
265355da1ebSSage Weil 	if (new) {
266355da1ebSSage Weil 		rb_link_node(&xattr->node, parent, p);
267355da1ebSSage Weil 		rb_insert_color(&xattr->node, &ci->i_xattrs.index);
268355da1ebSSage Weil 		dout("__set_xattr_val p=%p\n", p);
269355da1ebSSage Weil 	}
270355da1ebSSage Weil 
271355da1ebSSage Weil 	dout("__set_xattr_val added %llx.%llx xattr %p %s=%.*s\n",
272355da1ebSSage Weil 	     ceph_vinop(&ci->vfs_inode), xattr, name, val_len, val);
273355da1ebSSage Weil 
274355da1ebSSage Weil 	return 0;
275355da1ebSSage Weil }
276355da1ebSSage Weil 
277355da1ebSSage Weil static struct ceph_inode_xattr *__get_xattr(struct ceph_inode_info *ci,
278355da1ebSSage Weil 			   const char *name)
279355da1ebSSage Weil {
280355da1ebSSage Weil 	struct rb_node **p;
281355da1ebSSage Weil 	struct rb_node *parent = NULL;
282355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
28317db143fSSage Weil 	int name_len = strlen(name);
284355da1ebSSage Weil 	int c;
285355da1ebSSage Weil 
286355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
287355da1ebSSage Weil 	while (*p) {
288355da1ebSSage Weil 		parent = *p;
289355da1ebSSage Weil 		xattr = rb_entry(parent, struct ceph_inode_xattr, node);
290355da1ebSSage Weil 		c = strncmp(name, xattr->name, xattr->name_len);
29117db143fSSage Weil 		if (c == 0 && name_len > xattr->name_len)
29217db143fSSage Weil 			c = 1;
293355da1ebSSage Weil 		if (c < 0)
294355da1ebSSage Weil 			p = &(*p)->rb_left;
295355da1ebSSage Weil 		else if (c > 0)
296355da1ebSSage Weil 			p = &(*p)->rb_right;
297355da1ebSSage Weil 		else {
298355da1ebSSage Weil 			dout("__get_xattr %s: found %.*s\n", name,
299355da1ebSSage Weil 			     xattr->val_len, xattr->val);
300355da1ebSSage Weil 			return xattr;
301355da1ebSSage Weil 		}
302355da1ebSSage Weil 	}
303355da1ebSSage Weil 
304355da1ebSSage Weil 	dout("__get_xattr %s: not found\n", name);
305355da1ebSSage Weil 
306355da1ebSSage Weil 	return NULL;
307355da1ebSSage Weil }
308355da1ebSSage Weil 
309355da1ebSSage Weil static void __free_xattr(struct ceph_inode_xattr *xattr)
310355da1ebSSage Weil {
311355da1ebSSage Weil 	BUG_ON(!xattr);
312355da1ebSSage Weil 
313355da1ebSSage Weil 	if (xattr->should_free_name)
314355da1ebSSage Weil 		kfree((void *)xattr->name);
315355da1ebSSage Weil 	if (xattr->should_free_val)
316355da1ebSSage Weil 		kfree((void *)xattr->val);
317355da1ebSSage Weil 
318355da1ebSSage Weil 	kfree(xattr);
319355da1ebSSage Weil }
320355da1ebSSage Weil 
321355da1ebSSage Weil static int __remove_xattr(struct ceph_inode_info *ci,
322355da1ebSSage Weil 			  struct ceph_inode_xattr *xattr)
323355da1ebSSage Weil {
324355da1ebSSage Weil 	if (!xattr)
325355da1ebSSage Weil 		return -EOPNOTSUPP;
326355da1ebSSage Weil 
327355da1ebSSage Weil 	rb_erase(&xattr->node, &ci->i_xattrs.index);
328355da1ebSSage Weil 
329355da1ebSSage Weil 	if (xattr->should_free_name)
330355da1ebSSage Weil 		kfree((void *)xattr->name);
331355da1ebSSage Weil 	if (xattr->should_free_val)
332355da1ebSSage Weil 		kfree((void *)xattr->val);
333355da1ebSSage Weil 
334355da1ebSSage Weil 	ci->i_xattrs.names_size -= xattr->name_len;
335355da1ebSSage Weil 	ci->i_xattrs.vals_size -= xattr->val_len;
336355da1ebSSage Weil 	ci->i_xattrs.count--;
337355da1ebSSage Weil 	kfree(xattr);
338355da1ebSSage Weil 
339355da1ebSSage Weil 	return 0;
340355da1ebSSage Weil }
341355da1ebSSage Weil 
342355da1ebSSage Weil static int __remove_xattr_by_name(struct ceph_inode_info *ci,
343355da1ebSSage Weil 			   const char *name)
344355da1ebSSage Weil {
345355da1ebSSage Weil 	struct rb_node **p;
346355da1ebSSage Weil 	struct ceph_inode_xattr *xattr;
347355da1ebSSage Weil 	int err;
348355da1ebSSage Weil 
349355da1ebSSage Weil 	p = &ci->i_xattrs.index.rb_node;
350355da1ebSSage Weil 	xattr = __get_xattr(ci, name);
351355da1ebSSage Weil 	err = __remove_xattr(ci, xattr);
352355da1ebSSage Weil 	return err;
353355da1ebSSage Weil }
354355da1ebSSage Weil 
355355da1ebSSage Weil static char *__copy_xattr_names(struct ceph_inode_info *ci,
356355da1ebSSage Weil 				char *dest)
357355da1ebSSage Weil {
358355da1ebSSage Weil 	struct rb_node *p;
359355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
360355da1ebSSage Weil 
361355da1ebSSage Weil 	p = rb_first(&ci->i_xattrs.index);
362355da1ebSSage Weil 	dout("__copy_xattr_names count=%d\n", ci->i_xattrs.count);
363355da1ebSSage Weil 
364355da1ebSSage Weil 	while (p) {
365355da1ebSSage Weil 		xattr = rb_entry(p, struct ceph_inode_xattr, node);
366355da1ebSSage Weil 		memcpy(dest, xattr->name, xattr->name_len);
367355da1ebSSage Weil 		dest[xattr->name_len] = '\0';
368355da1ebSSage Weil 
369355da1ebSSage Weil 		dout("dest=%s %p (%s) (%d/%d)\n", dest, xattr, xattr->name,
370355da1ebSSage Weil 		     xattr->name_len, ci->i_xattrs.names_size);
371355da1ebSSage Weil 
372355da1ebSSage Weil 		dest += xattr->name_len + 1;
373355da1ebSSage Weil 		p = rb_next(p);
374355da1ebSSage Weil 	}
375355da1ebSSage Weil 
376355da1ebSSage Weil 	return dest;
377355da1ebSSage Weil }
378355da1ebSSage Weil 
379355da1ebSSage Weil void __ceph_destroy_xattrs(struct ceph_inode_info *ci)
380355da1ebSSage Weil {
381355da1ebSSage Weil 	struct rb_node *p, *tmp;
382355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
383355da1ebSSage Weil 
384355da1ebSSage Weil 	p = rb_first(&ci->i_xattrs.index);
385355da1ebSSage Weil 
386355da1ebSSage Weil 	dout("__ceph_destroy_xattrs p=%p\n", p);
387355da1ebSSage Weil 
388355da1ebSSage Weil 	while (p) {
389355da1ebSSage Weil 		xattr = rb_entry(p, struct ceph_inode_xattr, node);
390355da1ebSSage Weil 		tmp = p;
391355da1ebSSage Weil 		p = rb_next(tmp);
392355da1ebSSage Weil 		dout("__ceph_destroy_xattrs next p=%p (%.*s)\n", p,
393355da1ebSSage Weil 		     xattr->name_len, xattr->name);
394355da1ebSSage Weil 		rb_erase(tmp, &ci->i_xattrs.index);
395355da1ebSSage Weil 
396355da1ebSSage Weil 		__free_xattr(xattr);
397355da1ebSSage Weil 	}
398355da1ebSSage Weil 
399355da1ebSSage Weil 	ci->i_xattrs.names_size = 0;
400355da1ebSSage Weil 	ci->i_xattrs.vals_size = 0;
401355da1ebSSage Weil 	ci->i_xattrs.index_version = 0;
402355da1ebSSage Weil 	ci->i_xattrs.count = 0;
403355da1ebSSage Weil 	ci->i_xattrs.index = RB_ROOT;
404355da1ebSSage Weil }
405355da1ebSSage Weil 
406355da1ebSSage Weil static int __build_xattrs(struct inode *inode)
407be655596SSage Weil 	__releases(ci->i_ceph_lock)
408be655596SSage Weil 	__acquires(ci->i_ceph_lock)
409355da1ebSSage Weil {
410355da1ebSSage Weil 	u32 namelen;
411355da1ebSSage Weil 	u32 numattr = 0;
412355da1ebSSage Weil 	void *p, *end;
413355da1ebSSage Weil 	u32 len;
414355da1ebSSage Weil 	const char *name, *val;
415355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
416355da1ebSSage Weil 	int xattr_version;
417355da1ebSSage Weil 	struct ceph_inode_xattr **xattrs = NULL;
41863ff78b2SSage Weil 	int err = 0;
419355da1ebSSage Weil 	int i;
420355da1ebSSage Weil 
421355da1ebSSage Weil 	dout("__build_xattrs() len=%d\n",
422355da1ebSSage Weil 	     ci->i_xattrs.blob ? (int)ci->i_xattrs.blob->vec.iov_len : 0);
423355da1ebSSage Weil 
424355da1ebSSage Weil 	if (ci->i_xattrs.index_version >= ci->i_xattrs.version)
425355da1ebSSage Weil 		return 0; /* already built */
426355da1ebSSage Weil 
427355da1ebSSage Weil 	__ceph_destroy_xattrs(ci);
428355da1ebSSage Weil 
429355da1ebSSage Weil start:
430355da1ebSSage Weil 	/* updated internal xattr rb tree */
431355da1ebSSage Weil 	if (ci->i_xattrs.blob && ci->i_xattrs.blob->vec.iov_len > 4) {
432355da1ebSSage Weil 		p = ci->i_xattrs.blob->vec.iov_base;
433355da1ebSSage Weil 		end = p + ci->i_xattrs.blob->vec.iov_len;
434355da1ebSSage Weil 		ceph_decode_32_safe(&p, end, numattr, bad);
435355da1ebSSage Weil 		xattr_version = ci->i_xattrs.version;
436be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
437355da1ebSSage Weil 
438355da1ebSSage Weil 		xattrs = kcalloc(numattr, sizeof(struct ceph_xattr *),
439355da1ebSSage Weil 				 GFP_NOFS);
440355da1ebSSage Weil 		err = -ENOMEM;
441355da1ebSSage Weil 		if (!xattrs)
442355da1ebSSage Weil 			goto bad_lock;
443355da1ebSSage Weil 		memset(xattrs, 0, numattr*sizeof(struct ceph_xattr *));
444355da1ebSSage Weil 		for (i = 0; i < numattr; i++) {
445355da1ebSSage Weil 			xattrs[i] = kmalloc(sizeof(struct ceph_inode_xattr),
446355da1ebSSage Weil 					    GFP_NOFS);
447355da1ebSSage Weil 			if (!xattrs[i])
448355da1ebSSage Weil 				goto bad_lock;
449355da1ebSSage Weil 		}
450355da1ebSSage Weil 
451be655596SSage Weil 		spin_lock(&ci->i_ceph_lock);
452355da1ebSSage Weil 		if (ci->i_xattrs.version != xattr_version) {
453355da1ebSSage Weil 			/* lost a race, retry */
454355da1ebSSage Weil 			for (i = 0; i < numattr; i++)
455355da1ebSSage Weil 				kfree(xattrs[i]);
456355da1ebSSage Weil 			kfree(xattrs);
45721ec6ffaSAlan Cox 			xattrs = NULL;
458355da1ebSSage Weil 			goto start;
459355da1ebSSage Weil 		}
460355da1ebSSage Weil 		err = -EIO;
461355da1ebSSage Weil 		while (numattr--) {
462355da1ebSSage Weil 			ceph_decode_32_safe(&p, end, len, bad);
463355da1ebSSage Weil 			namelen = len;
464355da1ebSSage Weil 			name = p;
465355da1ebSSage Weil 			p += len;
466355da1ebSSage Weil 			ceph_decode_32_safe(&p, end, len, bad);
467355da1ebSSage Weil 			val = p;
468355da1ebSSage Weil 			p += len;
469355da1ebSSage Weil 
470355da1ebSSage Weil 			err = __set_xattr(ci, name, namelen, val, len,
471355da1ebSSage Weil 					  0, 0, 0, &xattrs[numattr]);
472355da1ebSSage Weil 
473355da1ebSSage Weil 			if (err < 0)
474355da1ebSSage Weil 				goto bad;
475355da1ebSSage Weil 		}
476355da1ebSSage Weil 		kfree(xattrs);
477355da1ebSSage Weil 	}
478355da1ebSSage Weil 	ci->i_xattrs.index_version = ci->i_xattrs.version;
479355da1ebSSage Weil 	ci->i_xattrs.dirty = false;
480355da1ebSSage Weil 
481355da1ebSSage Weil 	return err;
482355da1ebSSage Weil bad_lock:
483be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
484355da1ebSSage Weil bad:
485355da1ebSSage Weil 	if (xattrs) {
486355da1ebSSage Weil 		for (i = 0; i < numattr; i++)
487355da1ebSSage Weil 			kfree(xattrs[i]);
488355da1ebSSage Weil 		kfree(xattrs);
489355da1ebSSage Weil 	}
490355da1ebSSage Weil 	ci->i_xattrs.names_size = 0;
491355da1ebSSage Weil 	return err;
492355da1ebSSage Weil }
493355da1ebSSage Weil 
494355da1ebSSage Weil static int __get_required_blob_size(struct ceph_inode_info *ci, int name_size,
495355da1ebSSage Weil 				    int val_size)
496355da1ebSSage Weil {
497355da1ebSSage Weil 	/*
498355da1ebSSage Weil 	 * 4 bytes for the length, and additional 4 bytes per each xattr name,
499355da1ebSSage Weil 	 * 4 bytes per each value
500355da1ebSSage Weil 	 */
501355da1ebSSage Weil 	int size = 4 + ci->i_xattrs.count*(4 + 4) +
502355da1ebSSage Weil 			     ci->i_xattrs.names_size +
503355da1ebSSage Weil 			     ci->i_xattrs.vals_size;
504355da1ebSSage Weil 	dout("__get_required_blob_size c=%d names.size=%d vals.size=%d\n",
505355da1ebSSage Weil 	     ci->i_xattrs.count, ci->i_xattrs.names_size,
506355da1ebSSage Weil 	     ci->i_xattrs.vals_size);
507355da1ebSSage Weil 
508355da1ebSSage Weil 	if (name_size)
509355da1ebSSage Weil 		size += 4 + 4 + name_size + val_size;
510355da1ebSSage Weil 
511355da1ebSSage Weil 	return size;
512355da1ebSSage Weil }
513355da1ebSSage Weil 
514355da1ebSSage Weil /*
515355da1ebSSage Weil  * If there are dirty xattrs, reencode xattrs into the prealloc_blob
516355da1ebSSage Weil  * and swap into place.
517355da1ebSSage Weil  */
518355da1ebSSage Weil void __ceph_build_xattrs_blob(struct ceph_inode_info *ci)
519355da1ebSSage Weil {
520355da1ebSSage Weil 	struct rb_node *p;
521355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
522355da1ebSSage Weil 	void *dest;
523355da1ebSSage Weil 
524355da1ebSSage Weil 	dout("__build_xattrs_blob %p\n", &ci->vfs_inode);
525355da1ebSSage Weil 	if (ci->i_xattrs.dirty) {
526355da1ebSSage Weil 		int need = __get_required_blob_size(ci, 0, 0);
527355da1ebSSage Weil 
528355da1ebSSage Weil 		BUG_ON(need > ci->i_xattrs.prealloc_blob->alloc_len);
529355da1ebSSage Weil 
530355da1ebSSage Weil 		p = rb_first(&ci->i_xattrs.index);
531355da1ebSSage Weil 		dest = ci->i_xattrs.prealloc_blob->vec.iov_base;
532355da1ebSSage Weil 
533355da1ebSSage Weil 		ceph_encode_32(&dest, ci->i_xattrs.count);
534355da1ebSSage Weil 		while (p) {
535355da1ebSSage Weil 			xattr = rb_entry(p, struct ceph_inode_xattr, node);
536355da1ebSSage Weil 
537355da1ebSSage Weil 			ceph_encode_32(&dest, xattr->name_len);
538355da1ebSSage Weil 			memcpy(dest, xattr->name, xattr->name_len);
539355da1ebSSage Weil 			dest += xattr->name_len;
540355da1ebSSage Weil 			ceph_encode_32(&dest, xattr->val_len);
541355da1ebSSage Weil 			memcpy(dest, xattr->val, xattr->val_len);
542355da1ebSSage Weil 			dest += xattr->val_len;
543355da1ebSSage Weil 
544355da1ebSSage Weil 			p = rb_next(p);
545355da1ebSSage Weil 		}
546355da1ebSSage Weil 
547355da1ebSSage Weil 		/* adjust buffer len; it may be larger than we need */
548355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob->vec.iov_len =
549355da1ebSSage Weil 			dest - ci->i_xattrs.prealloc_blob->vec.iov_base;
550355da1ebSSage Weil 
551b6c1d5b8SSage Weil 		if (ci->i_xattrs.blob)
552355da1ebSSage Weil 			ceph_buffer_put(ci->i_xattrs.blob);
553355da1ebSSage Weil 		ci->i_xattrs.blob = ci->i_xattrs.prealloc_blob;
554355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob = NULL;
555355da1ebSSage Weil 		ci->i_xattrs.dirty = false;
5564a625be4SSage Weil 		ci->i_xattrs.version++;
557355da1ebSSage Weil 	}
558355da1ebSSage Weil }
559355da1ebSSage Weil 
560355da1ebSSage Weil ssize_t ceph_getxattr(struct dentry *dentry, const char *name, void *value,
561355da1ebSSage Weil 		      size_t size)
562355da1ebSSage Weil {
563355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
564355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
565355da1ebSSage Weil 	int err;
566355da1ebSSage Weil 	struct ceph_inode_xattr *xattr;
567881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr = NULL;
568355da1ebSSage Weil 
569355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
570355da1ebSSage Weil 		return -ENODATA;
571355da1ebSSage Weil 
572be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
573355da1ebSSage Weil 	dout("getxattr %p ver=%lld index_ver=%lld\n", inode,
574355da1ebSSage Weil 	     ci->i_xattrs.version, ci->i_xattrs.index_version);
575355da1ebSSage Weil 
576*0bee82fbSSage Weil 	/* let's see if a virtual xattr was requested */
577*0bee82fbSSage Weil 	vxattr = ceph_match_vxattr(inode, name);
578*0bee82fbSSage Weil 	if (vxattr && !(vxattr->exists_cb && !vxattr->exists_cb(ci))) {
579*0bee82fbSSage Weil 		err = vxattr->getxattr_cb(ci, value, size);
580*0bee82fbSSage Weil 		goto out;
581*0bee82fbSSage Weil 	}
582*0bee82fbSSage Weil 
583355da1ebSSage Weil 	if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
584355da1ebSSage Weil 	    (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
585355da1ebSSage Weil 		goto get_xattr;
586355da1ebSSage Weil 	} else {
587be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
588355da1ebSSage Weil 		/* get xattrs from mds (if we don't already have them) */
589355da1ebSSage Weil 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
590355da1ebSSage Weil 		if (err)
591355da1ebSSage Weil 			return err;
592355da1ebSSage Weil 	}
593355da1ebSSage Weil 
594be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
595355da1ebSSage Weil 
596355da1ebSSage Weil 	err = __build_xattrs(inode);
597355da1ebSSage Weil 	if (err < 0)
598355da1ebSSage Weil 		goto out;
599355da1ebSSage Weil 
600355da1ebSSage Weil get_xattr:
601355da1ebSSage Weil 	err = -ENODATA;  /* == ENOATTR */
602355da1ebSSage Weil 	xattr = __get_xattr(ci, name);
603*0bee82fbSSage Weil 	if (!xattr)
604355da1ebSSage Weil 		goto out;
605355da1ebSSage Weil 
606355da1ebSSage Weil 	err = -ERANGE;
607355da1ebSSage Weil 	if (size && size < xattr->val_len)
608355da1ebSSage Weil 		goto out;
609355da1ebSSage Weil 
610355da1ebSSage Weil 	err = xattr->val_len;
611355da1ebSSage Weil 	if (size == 0)
612355da1ebSSage Weil 		goto out;
613355da1ebSSage Weil 
614355da1ebSSage Weil 	memcpy(value, xattr->val, xattr->val_len);
615355da1ebSSage Weil 
616355da1ebSSage Weil out:
617be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
618355da1ebSSage Weil 	return err;
619355da1ebSSage Weil }
620355da1ebSSage Weil 
621355da1ebSSage Weil ssize_t ceph_listxattr(struct dentry *dentry, char *names, size_t size)
622355da1ebSSage Weil {
623355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
624355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
625881a5fa2SAlex Elder 	struct ceph_vxattr *vxattrs = ceph_inode_vxattrs(inode);
626355da1ebSSage Weil 	u32 vir_namelen = 0;
627355da1ebSSage Weil 	u32 namelen;
628355da1ebSSage Weil 	int err;
629355da1ebSSage Weil 	u32 len;
630355da1ebSSage Weil 	int i;
631355da1ebSSage Weil 
632be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
633355da1ebSSage Weil 	dout("listxattr %p ver=%lld index_ver=%lld\n", inode,
634355da1ebSSage Weil 	     ci->i_xattrs.version, ci->i_xattrs.index_version);
635355da1ebSSage Weil 
636355da1ebSSage Weil 	if (__ceph_caps_issued_mask(ci, CEPH_CAP_XATTR_SHARED, 1) &&
637bddfa3ccSHenry C Chang 	    (ci->i_xattrs.index_version >= ci->i_xattrs.version)) {
638355da1ebSSage Weil 		goto list_xattr;
639355da1ebSSage Weil 	} else {
640be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
641355da1ebSSage Weil 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_XATTR);
642355da1ebSSage Weil 		if (err)
643355da1ebSSage Weil 			return err;
644355da1ebSSage Weil 	}
645355da1ebSSage Weil 
646be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
647355da1ebSSage Weil 
648355da1ebSSage Weil 	err = __build_xattrs(inode);
649355da1ebSSage Weil 	if (err < 0)
650355da1ebSSage Weil 		goto out;
651355da1ebSSage Weil 
652355da1ebSSage Weil list_xattr:
6533ce6cd12SAlex Elder 	/*
6543ce6cd12SAlex Elder 	 * Start with virtual dir xattr names (if any) (including
6553ce6cd12SAlex Elder 	 * terminating '\0' characters for each).
6563ce6cd12SAlex Elder 	 */
6573ce6cd12SAlex Elder 	vir_namelen = ceph_vxattrs_name_size(vxattrs);
6583ce6cd12SAlex Elder 
659355da1ebSSage Weil 	/* adding 1 byte per each variable due to the null termination */
660355da1ebSSage Weil 	namelen = vir_namelen + ci->i_xattrs.names_size + ci->i_xattrs.count;
661355da1ebSSage Weil 	err = -ERANGE;
662355da1ebSSage Weil 	if (size && namelen > size)
663355da1ebSSage Weil 		goto out;
664355da1ebSSage Weil 
665355da1ebSSage Weil 	err = namelen;
666355da1ebSSage Weil 	if (size == 0)
667355da1ebSSage Weil 		goto out;
668355da1ebSSage Weil 
669355da1ebSSage Weil 	names = __copy_xattr_names(ci, names);
670355da1ebSSage Weil 
671355da1ebSSage Weil 	/* virtual xattr names, too */
672355da1ebSSage Weil 	if (vxattrs)
673355da1ebSSage Weil 		for (i = 0; vxattrs[i].name; i++) {
674355da1ebSSage Weil 			len = sprintf(names, "%s", vxattrs[i].name);
675355da1ebSSage Weil 			names += len + 1;
676355da1ebSSage Weil 		}
677355da1ebSSage Weil 
678355da1ebSSage Weil out:
679be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
680355da1ebSSage Weil 	return err;
681355da1ebSSage Weil }
682355da1ebSSage Weil 
683355da1ebSSage Weil static int ceph_sync_setxattr(struct dentry *dentry, const char *name,
684355da1ebSSage Weil 			      const char *value, size_t size, int flags)
685355da1ebSSage Weil {
6863d14c5d2SYehuda Sadeh 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
687355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
688355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
6895f21c96dSSage Weil 	struct inode *parent_inode;
690355da1ebSSage Weil 	struct ceph_mds_request *req;
6913d14c5d2SYehuda Sadeh 	struct ceph_mds_client *mdsc = fsc->mdsc;
692355da1ebSSage Weil 	int err;
693355da1ebSSage Weil 	int i, nr_pages;
694355da1ebSSage Weil 	struct page **pages = NULL;
695355da1ebSSage Weil 	void *kaddr;
696355da1ebSSage Weil 
697355da1ebSSage Weil 	/* copy value into some pages */
698355da1ebSSage Weil 	nr_pages = calc_pages_for(0, size);
699355da1ebSSage Weil 	if (nr_pages) {
700355da1ebSSage Weil 		pages = kmalloc(sizeof(pages[0])*nr_pages, GFP_NOFS);
701355da1ebSSage Weil 		if (!pages)
702355da1ebSSage Weil 			return -ENOMEM;
703355da1ebSSage Weil 		err = -ENOMEM;
704355da1ebSSage Weil 		for (i = 0; i < nr_pages; i++) {
70531459fe4SYehuda Sadeh 			pages[i] = __page_cache_alloc(GFP_NOFS);
706355da1ebSSage Weil 			if (!pages[i]) {
707355da1ebSSage Weil 				nr_pages = i;
708355da1ebSSage Weil 				goto out;
709355da1ebSSage Weil 			}
710355da1ebSSage Weil 			kaddr = kmap(pages[i]);
711355da1ebSSage Weil 			memcpy(kaddr, value + i*PAGE_CACHE_SIZE,
712355da1ebSSage Weil 			       min(PAGE_CACHE_SIZE, size-i*PAGE_CACHE_SIZE));
713355da1ebSSage Weil 		}
714355da1ebSSage Weil 	}
715355da1ebSSage Weil 
716355da1ebSSage Weil 	dout("setxattr value=%.*s\n", (int)size, value);
717355da1ebSSage Weil 
718355da1ebSSage Weil 	/* do request */
719355da1ebSSage Weil 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETXATTR,
720355da1ebSSage Weil 				       USE_AUTH_MDS);
72160d87733SJulia Lawall 	if (IS_ERR(req)) {
72260d87733SJulia Lawall 		err = PTR_ERR(req);
72360d87733SJulia Lawall 		goto out;
72460d87733SJulia Lawall 	}
72570b666c3SSage Weil 	req->r_inode = inode;
72670b666c3SSage Weil 	ihold(inode);
727355da1ebSSage Weil 	req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
728355da1ebSSage Weil 	req->r_num_caps = 1;
729355da1ebSSage Weil 	req->r_args.setxattr.flags = cpu_to_le32(flags);
730355da1ebSSage Weil 	req->r_path2 = kstrdup(name, GFP_NOFS);
731355da1ebSSage Weil 
732355da1ebSSage Weil 	req->r_pages = pages;
733355da1ebSSage Weil 	req->r_num_pages = nr_pages;
734355da1ebSSage Weil 	req->r_data_len = size;
735355da1ebSSage Weil 
736355da1ebSSage Weil 	dout("xattr.ver (before): %lld\n", ci->i_xattrs.version);
7375f21c96dSSage Weil 	parent_inode = ceph_get_dentry_parent_inode(dentry);
738355da1ebSSage Weil 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
7395f21c96dSSage Weil 	iput(parent_inode);
740355da1ebSSage Weil 	ceph_mdsc_put_request(req);
741355da1ebSSage Weil 	dout("xattr.ver (after): %lld\n", ci->i_xattrs.version);
742355da1ebSSage Weil 
743355da1ebSSage Weil out:
744355da1ebSSage Weil 	if (pages) {
745355da1ebSSage Weil 		for (i = 0; i < nr_pages; i++)
746355da1ebSSage Weil 			__free_page(pages[i]);
747355da1ebSSage Weil 		kfree(pages);
748355da1ebSSage Weil 	}
749355da1ebSSage Weil 	return err;
750355da1ebSSage Weil }
751355da1ebSSage Weil 
752355da1ebSSage Weil int ceph_setxattr(struct dentry *dentry, const char *name,
753355da1ebSSage Weil 		  const void *value, size_t size, int flags)
754355da1ebSSage Weil {
755355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
756881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr;
757355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
75818fa8b3fSAlex Elder 	int issued;
759355da1ebSSage Weil 	int err;
76018fa8b3fSAlex Elder 	int dirty;
761355da1ebSSage Weil 	int name_len = strlen(name);
762355da1ebSSage Weil 	int val_len = size;
763355da1ebSSage Weil 	char *newname = NULL;
764355da1ebSSage Weil 	char *newval = NULL;
765355da1ebSSage Weil 	struct ceph_inode_xattr *xattr = NULL;
766355da1ebSSage Weil 	int required_blob_size;
767355da1ebSSage Weil 
768355da1ebSSage Weil 	if (ceph_snap(inode) != CEPH_NOSNAP)
769355da1ebSSage Weil 		return -EROFS;
770355da1ebSSage Weil 
771355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
772355da1ebSSage Weil 		return -EOPNOTSUPP;
773355da1ebSSage Weil 
77406476a69SAlex Elder 	vxattr = ceph_match_vxattr(inode, name);
775355da1ebSSage Weil 	if (vxattr && vxattr->readonly)
776355da1ebSSage Weil 		return -EOPNOTSUPP;
777355da1ebSSage Weil 
7783adf654dSSage Weil 	/* pass any unhandled ceph.* xattrs through to the MDS */
7793adf654dSSage Weil 	if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
7803adf654dSSage Weil 		goto do_sync_unlocked;
7813adf654dSSage Weil 
782355da1ebSSage Weil 	/* preallocate memory for xattr name, value, index node */
783355da1ebSSage Weil 	err = -ENOMEM;
78461413c2fSJulia Lawall 	newname = kmemdup(name, name_len + 1, GFP_NOFS);
785355da1ebSSage Weil 	if (!newname)
786355da1ebSSage Weil 		goto out;
787355da1ebSSage Weil 
788355da1ebSSage Weil 	if (val_len) {
789b829c195SAlex Elder 		newval = kmemdup(value, val_len, GFP_NOFS);
790355da1ebSSage Weil 		if (!newval)
791355da1ebSSage Weil 			goto out;
792355da1ebSSage Weil 	}
793355da1ebSSage Weil 
794355da1ebSSage Weil 	xattr = kmalloc(sizeof(struct ceph_inode_xattr), GFP_NOFS);
795355da1ebSSage Weil 	if (!xattr)
796355da1ebSSage Weil 		goto out;
797355da1ebSSage Weil 
798be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
799355da1ebSSage Weil retry:
800355da1ebSSage Weil 	issued = __ceph_caps_issued(ci, NULL);
80118fa8b3fSAlex Elder 	dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
802355da1ebSSage Weil 	if (!(issued & CEPH_CAP_XATTR_EXCL))
803355da1ebSSage Weil 		goto do_sync;
804355da1ebSSage Weil 	__build_xattrs(inode);
805355da1ebSSage Weil 
806355da1ebSSage Weil 	required_blob_size = __get_required_blob_size(ci, name_len, val_len);
807355da1ebSSage Weil 
808355da1ebSSage Weil 	if (!ci->i_xattrs.prealloc_blob ||
809355da1ebSSage Weil 	    required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
81018fa8b3fSAlex Elder 		struct ceph_buffer *blob;
811355da1ebSSage Weil 
812be655596SSage Weil 		spin_unlock(&ci->i_ceph_lock);
813355da1ebSSage Weil 		dout(" preaallocating new blob size=%d\n", required_blob_size);
814b6c1d5b8SSage Weil 		blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
815355da1ebSSage Weil 		if (!blob)
816355da1ebSSage Weil 			goto out;
817be655596SSage Weil 		spin_lock(&ci->i_ceph_lock);
818b6c1d5b8SSage Weil 		if (ci->i_xattrs.prealloc_blob)
819355da1ebSSage Weil 			ceph_buffer_put(ci->i_xattrs.prealloc_blob);
820355da1ebSSage Weil 		ci->i_xattrs.prealloc_blob = blob;
821355da1ebSSage Weil 		goto retry;
822355da1ebSSage Weil 	}
823355da1ebSSage Weil 
824355da1ebSSage Weil 	err = __set_xattr(ci, newname, name_len, newval,
825355da1ebSSage Weil 			  val_len, 1, 1, 1, &xattr);
82618fa8b3fSAlex Elder 
827fca65b4aSSage Weil 	dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
828355da1ebSSage Weil 	ci->i_xattrs.dirty = true;
829355da1ebSSage Weil 	inode->i_ctime = CURRENT_TIME;
83018fa8b3fSAlex Elder 
831be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
832fca65b4aSSage Weil 	if (dirty)
833fca65b4aSSage Weil 		__mark_inode_dirty(inode, dirty);
834355da1ebSSage Weil 	return err;
835355da1ebSSage Weil 
836355da1ebSSage Weil do_sync:
837be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
8383adf654dSSage Weil do_sync_unlocked:
839355da1ebSSage Weil 	err = ceph_sync_setxattr(dentry, name, value, size, flags);
840355da1ebSSage Weil out:
841355da1ebSSage Weil 	kfree(newname);
842355da1ebSSage Weil 	kfree(newval);
843355da1ebSSage Weil 	kfree(xattr);
844355da1ebSSage Weil 	return err;
845355da1ebSSage Weil }
846355da1ebSSage Weil 
847355da1ebSSage Weil static int ceph_send_removexattr(struct dentry *dentry, const char *name)
848355da1ebSSage Weil {
8493d14c5d2SYehuda Sadeh 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
8503d14c5d2SYehuda Sadeh 	struct ceph_mds_client *mdsc = fsc->mdsc;
851355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
8525f21c96dSSage Weil 	struct inode *parent_inode;
853355da1ebSSage Weil 	struct ceph_mds_request *req;
854355da1ebSSage Weil 	int err;
855355da1ebSSage Weil 
856355da1ebSSage Weil 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RMXATTR,
857355da1ebSSage Weil 				       USE_AUTH_MDS);
858355da1ebSSage Weil 	if (IS_ERR(req))
859355da1ebSSage Weil 		return PTR_ERR(req);
86070b666c3SSage Weil 	req->r_inode = inode;
86170b666c3SSage Weil 	ihold(inode);
862355da1ebSSage Weil 	req->r_inode_drop = CEPH_CAP_XATTR_SHARED;
863355da1ebSSage Weil 	req->r_num_caps = 1;
864355da1ebSSage Weil 	req->r_path2 = kstrdup(name, GFP_NOFS);
865355da1ebSSage Weil 
8665f21c96dSSage Weil 	parent_inode = ceph_get_dentry_parent_inode(dentry);
867355da1ebSSage Weil 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
8685f21c96dSSage Weil 	iput(parent_inode);
869355da1ebSSage Weil 	ceph_mdsc_put_request(req);
870355da1ebSSage Weil 	return err;
871355da1ebSSage Weil }
872355da1ebSSage Weil 
873355da1ebSSage Weil int ceph_removexattr(struct dentry *dentry, const char *name)
874355da1ebSSage Weil {
875355da1ebSSage Weil 	struct inode *inode = dentry->d_inode;
876881a5fa2SAlex Elder 	struct ceph_vxattr *vxattr;
877355da1ebSSage Weil 	struct ceph_inode_info *ci = ceph_inode(inode);
878355da1ebSSage Weil 	int issued;
879355da1ebSSage Weil 	int err;
88083eb26afSAlex Elder 	int required_blob_size;
881fca65b4aSSage Weil 	int dirty;
882355da1ebSSage Weil 
883355da1ebSSage Weil 	if (ceph_snap(inode) != CEPH_NOSNAP)
884355da1ebSSage Weil 		return -EROFS;
885355da1ebSSage Weil 
886355da1ebSSage Weil 	if (!ceph_is_valid_xattr(name))
887355da1ebSSage Weil 		return -EOPNOTSUPP;
888355da1ebSSage Weil 
88906476a69SAlex Elder 	vxattr = ceph_match_vxattr(inode, name);
890355da1ebSSage Weil 	if (vxattr && vxattr->readonly)
891355da1ebSSage Weil 		return -EOPNOTSUPP;
892355da1ebSSage Weil 
893d421acb1SSage Weil 	/* pass any unhandled ceph.* xattrs through to the MDS */
894d421acb1SSage Weil 	if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
895d421acb1SSage Weil 		goto do_sync_unlocked;
896d421acb1SSage Weil 
89783eb26afSAlex Elder 	err = -ENOMEM;
898be655596SSage Weil 	spin_lock(&ci->i_ceph_lock);
89983eb26afSAlex Elder retry:
900355da1ebSSage Weil 	issued = __ceph_caps_issued(ci, NULL);
901355da1ebSSage Weil 	dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
902355da1ebSSage Weil 
903355da1ebSSage Weil 	if (!(issued & CEPH_CAP_XATTR_EXCL))
904355da1ebSSage Weil 		goto do_sync;
90518fa8b3fSAlex Elder 	__build_xattrs(inode);
906355da1ebSSage Weil 
90783eb26afSAlex Elder 	required_blob_size = __get_required_blob_size(ci, 0, 0);
90883eb26afSAlex Elder 
90983eb26afSAlex Elder 	if (!ci->i_xattrs.prealloc_blob ||
91083eb26afSAlex Elder 	    required_blob_size > ci->i_xattrs.prealloc_blob->alloc_len) {
91183eb26afSAlex Elder 		struct ceph_buffer *blob;
91283eb26afSAlex Elder 
91383eb26afSAlex Elder 		spin_unlock(&ci->i_ceph_lock);
91483eb26afSAlex Elder 		dout(" preaallocating new blob size=%d\n", required_blob_size);
91583eb26afSAlex Elder 		blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
91683eb26afSAlex Elder 		if (!blob)
91783eb26afSAlex Elder 			goto out;
91883eb26afSAlex Elder 		spin_lock(&ci->i_ceph_lock);
91983eb26afSAlex Elder 		if (ci->i_xattrs.prealloc_blob)
92083eb26afSAlex Elder 			ceph_buffer_put(ci->i_xattrs.prealloc_blob);
92183eb26afSAlex Elder 		ci->i_xattrs.prealloc_blob = blob;
92283eb26afSAlex Elder 		goto retry;
92383eb26afSAlex Elder 	}
92483eb26afSAlex Elder 
925355da1ebSSage Weil 	err = __remove_xattr_by_name(ceph_inode(inode), name);
92618fa8b3fSAlex Elder 
927fca65b4aSSage Weil 	dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL);
928355da1ebSSage Weil 	ci->i_xattrs.dirty = true;
929355da1ebSSage Weil 	inode->i_ctime = CURRENT_TIME;
930be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
931fca65b4aSSage Weil 	if (dirty)
932fca65b4aSSage Weil 		__mark_inode_dirty(inode, dirty);
933355da1ebSSage Weil 	return err;
934355da1ebSSage Weil do_sync:
935be655596SSage Weil 	spin_unlock(&ci->i_ceph_lock);
936d421acb1SSage Weil do_sync_unlocked:
937355da1ebSSage Weil 	err = ceph_send_removexattr(dentry, name);
93883eb26afSAlex Elder out:
939355da1ebSSage Weil 	return err;
940355da1ebSSage Weil }
941355da1ebSSage Weil 
942