blob: a822a6e58290bbedfb0e363bf3bbb601075891fb [file] [log] [blame]
Linus Torvalds96c57ad2014-04-12 15:39:53 -07001#include <linux/ceph/ceph_debug.h>
Sage Weil8f4e91d2009-10-06 11:31:14 -07002#include <linux/in.h>
3
Sage Weil8f4e91d2009-10-06 11:31:14 -07004#include "super.h"
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07005#include "mds_client.h"
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07006#include "ioctl.h"
Sage Weil8f4e91d2009-10-06 11:31:14 -07007
8
9/*
10 * ioctls
11 */
12
13/*
14 * get and set the file layout
15 */
16static long ceph_ioctl_get_layout(struct file *file, void __user *arg)
17{
Al Viro496ad9a2013-01-23 17:07:38 -050018 struct ceph_inode_info *ci = ceph_inode(file_inode(file));
Sage Weil8f4e91d2009-10-06 11:31:14 -070019 struct ceph_ioctl_layout l;
20 int err;
21
Al Viro496ad9a2013-01-23 17:07:38 -050022 err = ceph_do_getattr(file_inode(file), CEPH_STAT_CAP_LAYOUT);
Sage Weil8f4e91d2009-10-06 11:31:14 -070023 if (!err) {
24 l.stripe_unit = ceph_file_layout_su(ci->i_layout);
25 l.stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
26 l.object_size = ceph_file_layout_object_size(ci->i_layout);
27 l.data_pool = le32_to_cpu(ci->i_layout.fl_pg_pool);
Sage Weil3469ac1a2012-05-07 15:33:36 -070028 l.preferred_osd = (s32)-1;
Sage Weil8f4e91d2009-10-06 11:31:14 -070029 if (copy_to_user(arg, &l, sizeof(l)))
30 return -EFAULT;
31 }
32
33 return err;
34}
35
Sage Weile49bf4c2012-05-07 15:34:35 -070036static long __validate_layout(struct ceph_mds_client *mdsc,
37 struct ceph_ioctl_layout *l)
38{
39 int i, err;
40
Sage Weile49bf4c2012-05-07 15:34:35 -070041 /* validate striping parameters */
42 if ((l->object_size & ~PAGE_MASK) ||
43 (l->stripe_unit & ~PAGE_MASK) ||
Sage Weil45f2e082012-08-21 12:11:51 -070044 (l->stripe_unit != 0 &&
45 ((unsigned)l->object_size % (unsigned)l->stripe_unit)))
Sage Weile49bf4c2012-05-07 15:34:35 -070046 return -EINVAL;
47
48 /* make sure it's a valid data pool */
49 mutex_lock(&mdsc->mutex);
50 err = -EINVAL;
51 for (i = 0; i < mdsc->mdsmap->m_num_data_pg_pools; i++)
52 if (mdsc->mdsmap->m_data_pg_pools[i] == l->data_pool) {
53 err = 0;
54 break;
55 }
56 mutex_unlock(&mdsc->mutex);
57 if (err)
58 return err;
59
60 return 0;
61}
62
Sage Weil8f4e91d2009-10-06 11:31:14 -070063static long ceph_ioctl_set_layout(struct file *file, void __user *arg)
64{
Al Viro496ad9a2013-01-23 17:07:38 -050065 struct inode *inode = file_inode(file);
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070066 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
Sage Weil8f4e91d2009-10-06 11:31:14 -070067 struct ceph_mds_request *req;
68 struct ceph_ioctl_layout l;
Al Viro496ad9a2013-01-23 17:07:38 -050069 struct ceph_inode_info *ci = ceph_inode(file_inode(file));
Greg Farnuma35eca92011-08-25 12:43:06 -070070 struct ceph_ioctl_layout nl;
Sage Weile49bf4c2012-05-07 15:34:35 -070071 int err;
Sage Weil8f4e91d2009-10-06 11:31:14 -070072
Sage Weil8f4e91d2009-10-06 11:31:14 -070073 if (copy_from_user(&l, arg, sizeof(l)))
74 return -EFAULT;
75
Greg Farnuma35eca92011-08-25 12:43:06 -070076 /* validate changed params against current layout */
Al Viro496ad9a2013-01-23 17:07:38 -050077 err = ceph_do_getattr(file_inode(file), CEPH_STAT_CAP_LAYOUT);
Sage Weil702aeb12012-05-14 12:34:38 -070078 if (err)
Greg Farnuma35eca92011-08-25 12:43:06 -070079 return err;
80
Sage Weil702aeb12012-05-14 12:34:38 -070081 memset(&nl, 0, sizeof(nl));
Greg Farnuma35eca92011-08-25 12:43:06 -070082 if (l.stripe_count)
83 nl.stripe_count = l.stripe_count;
Sage Weil702aeb12012-05-14 12:34:38 -070084 else
85 nl.stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
Greg Farnuma35eca92011-08-25 12:43:06 -070086 if (l.stripe_unit)
87 nl.stripe_unit = l.stripe_unit;
Sage Weil702aeb12012-05-14 12:34:38 -070088 else
89 nl.stripe_unit = ceph_file_layout_su(ci->i_layout);
Greg Farnuma35eca92011-08-25 12:43:06 -070090 if (l.object_size)
91 nl.object_size = l.object_size;
Sage Weil702aeb12012-05-14 12:34:38 -070092 else
93 nl.object_size = ceph_file_layout_object_size(ci->i_layout);
Greg Farnuma35eca92011-08-25 12:43:06 -070094 if (l.data_pool)
95 nl.data_pool = l.data_pool;
Sage Weil702aeb12012-05-14 12:34:38 -070096 else
97 nl.data_pool = ceph_file_layout_pg_pool(ci->i_layout);
98
99 /* this is obsolete, and always -1 */
100 nl.preferred_osd = le64_to_cpu(-1);
Greg Farnuma35eca92011-08-25 12:43:06 -0700101
Sage Weile49bf4c2012-05-07 15:34:35 -0700102 err = __validate_layout(mdsc, &nl);
103 if (err)
104 return err;
Sage Weil8f4e91d2009-10-06 11:31:14 -0700105
106 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETLAYOUT,
107 USE_AUTH_MDS);
108 if (IS_ERR(req))
109 return PTR_ERR(req);
Sage Weil70b666c2011-05-27 09:24:26 -0700110 req->r_inode = inode;
111 ihold(inode);
Yan, Zheng3bd58142014-04-27 09:17:45 +0800112 req->r_num_caps = 1;
113
Sage Weil8f4e91d2009-10-06 11:31:14 -0700114 req->r_inode_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL;
115
116 req->r_args.setlayout.layout.fl_stripe_unit =
117 cpu_to_le32(l.stripe_unit);
118 req->r_args.setlayout.layout.fl_stripe_count =
119 cpu_to_le32(l.stripe_count);
120 req->r_args.setlayout.layout.fl_object_size =
121 cpu_to_le32(l.object_size);
122 req->r_args.setlayout.layout.fl_pg_pool = cpu_to_le32(l.data_pool);
Sage Weil8f4e91d2009-10-06 11:31:14 -0700123
Sage Weil752c8bd2013-02-05 13:52:29 -0800124 err = ceph_mdsc_do_request(mdsc, NULL, req);
Sage Weil8f4e91d2009-10-06 11:31:14 -0700125 ceph_mdsc_put_request(req);
126 return err;
127}
128
129/*
Greg Farnum571dba52010-09-24 14:56:40 -0700130 * Set a layout policy on a directory inode. All items in the tree
131 * rooted at this inode will inherit this layout on creation,
132 * (It doesn't apply retroactively )
133 * unless a subdirectory has its own layout policy.
134 */
135static long ceph_ioctl_set_layout_policy (struct file *file, void __user *arg)
136{
Al Viro496ad9a2013-01-23 17:07:38 -0500137 struct inode *inode = file_inode(file);
Greg Farnum571dba52010-09-24 14:56:40 -0700138 struct ceph_mds_request *req;
139 struct ceph_ioctl_layout l;
Sage Weile49bf4c2012-05-07 15:34:35 -0700140 int err;
Greg Farnum571dba52010-09-24 14:56:40 -0700141 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
142
143 /* copy and validate */
144 if (copy_from_user(&l, arg, sizeof(l)))
145 return -EFAULT;
146
Sage Weile49bf4c2012-05-07 15:34:35 -0700147 err = __validate_layout(mdsc, &l);
148 if (err)
149 return err;
Greg Farnum571dba52010-09-24 14:56:40 -0700150
151 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETDIRLAYOUT,
152 USE_AUTH_MDS);
153
154 if (IS_ERR(req))
155 return PTR_ERR(req);
Sage Weil70b666c2011-05-27 09:24:26 -0700156 req->r_inode = inode;
157 ihold(inode);
Yan, Zheng3bd58142014-04-27 09:17:45 +0800158 req->r_num_caps = 1;
Greg Farnum571dba52010-09-24 14:56:40 -0700159
160 req->r_args.setlayout.layout.fl_stripe_unit =
161 cpu_to_le32(l.stripe_unit);
162 req->r_args.setlayout.layout.fl_stripe_count =
163 cpu_to_le32(l.stripe_count);
164 req->r_args.setlayout.layout.fl_object_size =
165 cpu_to_le32(l.object_size);
166 req->r_args.setlayout.layout.fl_pg_pool =
167 cpu_to_le32(l.data_pool);
Greg Farnum571dba52010-09-24 14:56:40 -0700168
169 err = ceph_mdsc_do_request(mdsc, inode, req);
170 ceph_mdsc_put_request(req);
171 return err;
172}
173
174/*
Sage Weil8f4e91d2009-10-06 11:31:14 -0700175 * Return object name, size/offset information, and location (OSD
176 * number, network address) for a given file offset.
177 */
178static long ceph_ioctl_get_dataloc(struct file *file, void __user *arg)
179{
180 struct ceph_ioctl_dataloc dl;
Al Viro496ad9a2013-01-23 17:07:38 -0500181 struct inode *inode = file_inode(file);
Sage Weil8f4e91d2009-10-06 11:31:14 -0700182 struct ceph_inode_info *ci = ceph_inode(inode);
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700183 struct ceph_osd_client *osdc =
184 &ceph_sb_to_client(inode->i_sb)->client->osdc;
Ilya Dryomov7c13cb62014-01-27 17:40:19 +0200185 struct ceph_object_locator oloc;
186 struct ceph_object_id oid;
Sage Weil8f4e91d2009-10-06 11:31:14 -0700187 u64 len = 1, olen;
188 u64 tmp;
Sage Weil51042122009-11-04 11:39:12 -0800189 struct ceph_pg pgid;
Sage Weil457712a2012-09-24 21:04:57 -0700190 int r;
Sage Weil8f4e91d2009-10-06 11:31:14 -0700191
192 /* copy and validate */
193 if (copy_from_user(&dl, arg, sizeof(dl)))
194 return -EFAULT;
195
196 down_read(&osdc->map_sem);
Alex Eldere8afad62012-11-14 09:38:19 -0600197 r = ceph_calc_file_object_mapping(&ci->i_layout, dl.file_offset, len,
Sage Weil457712a2012-09-24 21:04:57 -0700198 &dl.object_no, &dl.object_offset,
199 &olen);
majianpeng494ddd12013-07-16 19:36:21 +0800200 if (r < 0) {
201 up_read(&osdc->map_sem);
Sage Weil457712a2012-09-24 21:04:57 -0700202 return -EIO;
majianpeng494ddd12013-07-16 19:36:21 +0800203 }
Sage Weil8f4e91d2009-10-06 11:31:14 -0700204 dl.file_offset -= dl.object_offset;
205 dl.object_size = ceph_file_layout_object_size(ci->i_layout);
206 dl.block_size = ceph_file_layout_su(ci->i_layout);
207
208 /* block_offset = object_offset % block_size */
209 tmp = dl.object_offset;
210 dl.block_offset = do_div(tmp, dl.block_size);
211
212 snprintf(dl.object_name, sizeof(dl.object_name), "%llx.%08llx",
213 ceph_ino(inode), dl.object_no);
Alex Elder41766f82013-03-01 18:00:15 -0600214
Ilya Dryomov7c13cb62014-01-27 17:40:19 +0200215 oloc.pool = ceph_file_layout_pg_pool(ci->i_layout);
216 ceph_oid_set_name(&oid, dl.object_name);
217
218 r = ceph_oloc_oid_to_pg(osdc->osdmap, &oloc, &oid, &pgid);
majianpeng2fbcbff2013-08-02 18:14:48 +0800219 if (r < 0) {
220 up_read(&osdc->map_sem);
221 return r;
222 }
Sage Weil8f4e91d2009-10-06 11:31:14 -0700223
Sage Weil8f4e91d2009-10-06 11:31:14 -0700224 dl.osd = ceph_calc_pg_primary(osdc->osdmap, pgid);
225 if (dl.osd >= 0) {
226 struct ceph_entity_addr *a =
227 ceph_osd_addr(osdc->osdmap, dl.osd);
228 if (a)
229 memcpy(&dl.osd_addr, &a->in_addr, sizeof(dl.osd_addr));
230 } else {
231 memset(&dl.osd_addr, 0, sizeof(dl.osd_addr));
232 }
233 up_read(&osdc->map_sem);
234
235 /* send result back to user */
236 if (copy_to_user(arg, &dl, sizeof(dl)))
237 return -EFAULT;
238
239 return 0;
240}
241
Sage Weil8c6e9222010-04-16 09:53:43 -0700242static long ceph_ioctl_lazyio(struct file *file)
243{
244 struct ceph_file_info *fi = file->private_data;
Al Viro496ad9a2013-01-23 17:07:38 -0500245 struct inode *inode = file_inode(file);
Sage Weil8c6e9222010-04-16 09:53:43 -0700246 struct ceph_inode_info *ci = ceph_inode(inode);
247
248 if ((fi->fmode & CEPH_FILE_MODE_LAZY) == 0) {
Sage Weilbe655592011-11-30 09:47:09 -0800249 spin_lock(&ci->i_ceph_lock);
Sage Weil8c6e9222010-04-16 09:53:43 -0700250 ci->i_nr_by_mode[fi->fmode]--;
251 fi->fmode |= CEPH_FILE_MODE_LAZY;
252 ci->i_nr_by_mode[fi->fmode]++;
Sage Weilbe655592011-11-30 09:47:09 -0800253 spin_unlock(&ci->i_ceph_lock);
Sage Weil8c6e9222010-04-16 09:53:43 -0700254 dout("ioctl_layzio: file %p marked lazy\n", file);
255
256 ceph_check_caps(ci, 0, NULL);
257 } else {
258 dout("ioctl_layzio: file %p already lazy\n", file);
259 }
260 return 0;
261}
262
Sage Weil4918b6d2011-07-26 11:26:07 -0700263static long ceph_ioctl_syncio(struct file *file)
264{
265 struct ceph_file_info *fi = file->private_data;
266
267 fi->flags |= CEPH_F_SYNC;
268 return 0;
269}
270
Sage Weil8f4e91d2009-10-06 11:31:14 -0700271long ceph_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
272{
273 dout("ioctl file %p cmd %u arg %lu\n", file, cmd, arg);
274 switch (cmd) {
275 case CEPH_IOC_GET_LAYOUT:
276 return ceph_ioctl_get_layout(file, (void __user *)arg);
277
278 case CEPH_IOC_SET_LAYOUT:
279 return ceph_ioctl_set_layout(file, (void __user *)arg);
280
Greg Farnum571dba52010-09-24 14:56:40 -0700281 case CEPH_IOC_SET_LAYOUT_POLICY:
282 return ceph_ioctl_set_layout_policy(file, (void __user *)arg);
283
Sage Weil8f4e91d2009-10-06 11:31:14 -0700284 case CEPH_IOC_GET_DATALOC:
285 return ceph_ioctl_get_dataloc(file, (void __user *)arg);
Sage Weil8c6e9222010-04-16 09:53:43 -0700286
287 case CEPH_IOC_LAZYIO:
288 return ceph_ioctl_lazyio(file);
Sage Weil4918b6d2011-07-26 11:26:07 -0700289
290 case CEPH_IOC_SYNCIO:
291 return ceph_ioctl_syncio(file);
Sage Weil8f4e91d2009-10-06 11:31:14 -0700292 }
Greg Farnum571dba52010-09-24 14:56:40 -0700293
Sage Weil8f4e91d2009-10-06 11:31:14 -0700294 return -ENOTTY;
295}