libceph: osdc->req_mempool should be backed by a slab pool
[linux-2.6-block.git] / include / linux / ceph / osd_client.h
CommitLineData
f24e9980
SW
1#ifndef _FS_CEPH_OSD_CLIENT_H
2#define _FS_CEPH_OSD_CLIENT_H
3
4#include <linux/completion.h>
415e49a9 5#include <linux/kref.h>
f24e9980
SW
6#include <linux/mempool.h>
7#include <linux/rbtree.h>
8
6c4a1915
AE
9#include <linux/ceph/types.h>
10#include <linux/ceph/osdmap.h>
11#include <linux/ceph/messenger.h>
12#include <linux/ceph/auth.h>
c885837f 13#include <linux/ceph/pagelist.h>
f24e9980
SW
14
15struct ceph_msg;
16struct ceph_snap_context;
17struct ceph_osd_request;
18struct ceph_osd_client;
4e7a5dcd 19struct ceph_authorizer;
f24e9980
SW
20
21/*
22 * completion callback for async writepages
23 */
24typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *,
25 struct ceph_msg *);
26be8808 26typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool);
f24e9980
SW
27
28/* a given osd we're communicating with */
29struct ceph_osd {
30 atomic_t o_ref;
31 struct ceph_osd_client *o_osdc;
32 int o_osd;
33 int o_incarnation;
34 struct rb_node o_node;
35 struct ceph_connection o_con;
36 struct list_head o_requests;
a40c4f10 37 struct list_head o_linger_requests;
f5a2041b 38 struct list_head o_osd_lru;
6c4a1915 39 struct ceph_auth_handshake o_auth;
f5a2041b 40 unsigned long lru_ttl;
422d2cb8
YS
41 int o_marked_for_keepalive;
42 struct list_head o_keepalive_item;
f24e9980
SW
43};
44
1b83bef2 45
7cc69d42 46#define CEPH_OSD_MAX_OP 3
1b83bef2 47
2ac2b7a6 48enum ceph_osd_data_type {
ec9123c5 49 CEPH_OSD_DATA_TYPE_NONE = 0,
2ac2b7a6 50 CEPH_OSD_DATA_TYPE_PAGES,
9a5e6d09 51 CEPH_OSD_DATA_TYPE_PAGELIST,
2ac2b7a6
AE
52#ifdef CONFIG_BLOCK
53 CEPH_OSD_DATA_TYPE_BIO,
54#endif /* CONFIG_BLOCK */
55};
56
2794a82a 57struct ceph_osd_data {
2ac2b7a6
AE
58 enum ceph_osd_data_type type;
59 union {
2794a82a
AE
60 struct {
61 struct page **pages;
e0c59487 62 u64 length;
2794a82a
AE
63 u32 alignment;
64 bool pages_from_pool;
65 bool own_pages;
66 };
9a5e6d09 67 struct ceph_pagelist *pagelist;
2794a82a 68#ifdef CONFIG_BLOCK
fdce58cc
AE
69 struct {
70 struct bio *bio; /* list of bios */
71 size_t bio_length; /* total in list */
72 };
2794a82a
AE
73#endif /* CONFIG_BLOCK */
74 };
75};
76
79528734
AE
77struct ceph_osd_req_op {
78 u16 op; /* CEPH_OSD_OP_* */
7b25bf5f 79 u32 flags; /* CEPH_OSD_OP_FLAG_* */
de2aa102 80 u32 indata_len; /* request */
7665d85b
YZ
81 u32 outdata_len; /* reply */
82 s32 rval;
83
79528734 84 union {
49719778 85 struct ceph_osd_data raw_data_in;
79528734
AE
86 struct {
87 u64 offset, length;
88 u64 truncate_size;
89 u32 truncate_seq;
5476492f 90 struct ceph_osd_data osd_data;
79528734 91 } extent;
d74b50be 92 struct {
d7d5a007
ID
93 u32 name_len;
94 u32 value_len;
d74b50be
YZ
95 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */
96 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */
97 struct ceph_osd_data osd_data;
98 } xattr;
79528734
AE
99 struct {
100 const char *class_name;
101 const char *method_name;
5476492f 102 struct ceph_osd_data request_info;
04017e29 103 struct ceph_osd_data request_data;
5476492f 104 struct ceph_osd_data response_data;
79528734
AE
105 __u8 class_len;
106 __u8 method_len;
107 __u8 argc;
108 } cls;
109 struct {
110 u64 cookie;
111 u64 ver;
112 u32 prot_ver;
113 u32 timeout;
114 __u8 flag;
115 } watch;
c647b8a8
ID
116 struct {
117 u64 expected_object_size;
118 u64 expected_write_size;
119 } alloc_hint;
79528734
AE
120 };
121};
122
f24e9980
SW
123/* an in-flight request */
124struct ceph_osd_request {
125 u64 r_tid; /* unique for this client */
126 struct rb_node r_node;
422d2cb8 127 struct list_head r_req_lru_item;
f24e9980 128 struct list_head r_osd_item;
a40c4f10 129 struct list_head r_linger_item;
1d0326b1 130 struct list_head r_linger_osd_item;
f24e9980 131 struct ceph_osd *r_osd;
5b191d99 132 struct ceph_pg r_pgid;
d85b7056
SW
133 int r_pg_osds[CEPH_PG_MAX_SIZE];
134 int r_num_pg_osds;
f24e9980
SW
135
136 struct ceph_msg *r_request, *r_reply;
f24e9980
SW
137 int r_flags; /* any additional flags for the osd */
138 u32 r_sent; /* >0 if r_request is sending/sent */
1b83bef2 139
79528734
AE
140 /* request osd ops array */
141 unsigned int r_num_ops;
142 struct ceph_osd_req_op r_ops[CEPH_OSD_MAX_OP];
143
1b83bef2
SW
144 /* these are updated on each send */
145 __le32 *r_request_osdmap_epoch;
146 __le32 *r_request_flags;
147 __le64 *r_request_pool;
148 void *r_request_pgid;
149 __le32 *r_request_attempts;
d29adb34 150 bool r_paused;
1b83bef2
SW
151 struct ceph_eversion *r_request_reassert_version;
152
153 int r_result;
350b1c32 154 int r_got_reply;
a40c4f10 155 int r_linger;
f24e9980
SW
156
157 struct ceph_osd_client *r_osdc;
415e49a9 158 struct kref r_kref;
f24e9980
SW
159 bool r_mempool;
160 struct completion r_completion, r_safe_completion;
26be8808
AE
161 ceph_osdc_callback_t r_callback;
162 ceph_osdc_unsafe_callback_t r_unsafe_callback;
f24e9980
SW
163 struct ceph_eversion r_reassert_version;
164 struct list_head r_unsafe_item;
165
166 struct inode *r_inode; /* for use by callbacks */
3d14c5d2 167 void *r_priv; /* ditto */
f24e9980 168
3c972c95
ID
169 struct ceph_object_locator r_base_oloc;
170 struct ceph_object_id r_base_oid;
205ee118
ID
171 struct ceph_object_locator r_target_oloc;
172 struct ceph_object_id r_target_oid;
22116525 173
2169aea6 174 u64 r_snapid;
3dd72fc0 175 unsigned long r_stamp; /* send OR check time */
f24e9980 176
f24e9980 177 struct ceph_snap_context *r_snapc; /* snap context for writes */
f24e9980
SW
178};
179
205ee118
ID
180struct ceph_request_redirect {
181 struct ceph_object_locator oloc;
182};
183
a40c4f10
YS
184struct ceph_osd_event {
185 u64 cookie;
186 int one_shot;
187 struct ceph_osd_client *osdc;
188 void (*cb)(u64, u64, u8, void *);
189 void *data;
190 struct rb_node node;
191 struct list_head osd_node;
192 struct kref kref;
a40c4f10
YS
193};
194
195struct ceph_osd_event_work {
196 struct work_struct work;
197 struct ceph_osd_event *event;
198 u64 ver;
199 u64 notify_id;
200 u8 opcode;
201};
202
f24e9980
SW
203struct ceph_osd_client {
204 struct ceph_client *client;
205
206 struct ceph_osdmap *osdmap; /* current map */
207 struct rw_semaphore map_sem;
208 struct completion map_waiters;
209 u64 last_requested_map;
210
211 struct mutex request_mutex;
212 struct rb_root osds; /* osds */
f5a2041b 213 struct list_head osd_lru; /* idle osds */
f24e9980
SW
214 u64 timeout_tid; /* tid of timeout triggering rq */
215 u64 last_tid; /* tid of last request */
216 struct rb_root requests; /* pending requests */
6f6c7006
SW
217 struct list_head req_lru; /* in-flight lru */
218 struct list_head req_unsent; /* unsent/need-resend queue */
219 struct list_head req_notarget; /* map to no osd */
a40c4f10 220 struct list_head req_linger; /* lingering requests */
f24e9980
SW
221 int num_requests;
222 struct delayed_work timeout_work;
f5a2041b 223 struct delayed_work osds_timeout_work;
039934b8 224#ifdef CONFIG_DEBUG_FS
f24e9980 225 struct dentry *debugfs_file;
039934b8 226#endif
f24e9980
SW
227
228 mempool_t *req_mempool;
229
0d59ab81 230 struct ceph_msgpool msgpool_op;
c16e7869 231 struct ceph_msgpool msgpool_op_reply;
a40c4f10
YS
232
233 spinlock_t event_lock;
234 struct rb_root event_tree;
235 u64 event_count;
236
237 struct workqueue_struct *notify_wq;
f24e9980
SW
238};
239
5522ae0b
AE
240extern int ceph_osdc_setup(void);
241extern void ceph_osdc_cleanup(void);
242
f24e9980
SW
243extern int ceph_osdc_init(struct ceph_osd_client *osdc,
244 struct ceph_client *client);
245extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
246
247extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
248 struct ceph_msg *msg);
249extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
250 struct ceph_msg *msg);
251
49719778 252extern void osd_req_op_init(struct ceph_osd_request *osd_req,
144cba14 253 unsigned int which, u16 opcode, u32 flags);
49719778
AE
254
255extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
256 unsigned int which,
257 struct page **pages, u64 length,
258 u32 alignment, bool pages_from_pool,
259 bool own_pages);
260
c99d2d4a
AE
261extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
262 unsigned int which, u16 opcode,
33803f33
AE
263 u64 offset, u64 length,
264 u64 truncate_size, u32 truncate_seq);
c99d2d4a
AE
265extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
266 unsigned int which, u64 length);
a4ce40a9
AE
267
268extern struct ceph_osd_data *osd_req_op_extent_osd_data(
269 struct ceph_osd_request *osd_req,
406e2c9f 270 unsigned int which);
a4ce40a9
AE
271extern struct ceph_osd_data *osd_req_op_cls_response_data(
272 struct ceph_osd_request *osd_req,
273 unsigned int which);
274
275extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
406e2c9f 276 unsigned int which,
a4ce40a9
AE
277 struct page **pages, u64 length,
278 u32 alignment, bool pages_from_pool,
279 bool own_pages);
280extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
406e2c9f 281 unsigned int which,
a4ce40a9
AE
282 struct ceph_pagelist *pagelist);
283#ifdef CONFIG_BLOCK
284extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
406e2c9f 285 unsigned int which,
a4ce40a9
AE
286 struct bio *bio, size_t bio_length);
287#endif /* CONFIG_BLOCK */
288
04017e29
AE
289extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
290 unsigned int which,
291 struct ceph_pagelist *pagelist);
6c57b554
AE
292extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
293 unsigned int which,
294 struct page **pages, u64 length,
295 u32 alignment, bool pages_from_pool,
296 bool own_pages);
a4ce40a9 297extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
c99d2d4a 298 unsigned int which,
a4ce40a9
AE
299 struct page **pages, u64 length,
300 u32 alignment, bool pages_from_pool,
301 bool own_pages);
302
c99d2d4a
AE
303extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
304 unsigned int which, u16 opcode,
04017e29 305 const char *class, const char *method);
d74b50be
YZ
306extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
307 u16 opcode, const char *name, const void *value,
308 size_t size, u8 cmp_op, u8 cmp_mode);
c99d2d4a
AE
309extern void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
310 unsigned int which, u16 opcode,
33803f33 311 u64 cookie, u64 version, int flag);
c647b8a8
ID
312extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
313 unsigned int which,
314 u64 expected_object_size,
315 u64 expected_write_size);
33803f33 316
3499e8a5 317extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
3499e8a5 318 struct ceph_snap_context *snapc,
acead002 319 unsigned int num_ops,
3499e8a5 320 bool use_mempool,
54a54007 321 gfp_t gfp_flags);
3499e8a5 322
175face2 323extern void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
68b4476b 324 struct ceph_snap_context *snapc,
4d6b250b 325 u64 snap_id,
af77f26c 326 struct timespec *mtime);
3499e8a5 327
f24e9980
SW
328extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
329 struct ceph_file_layout *layout,
330 struct ceph_vino vino,
acead002 331 u64 offset, u64 *len,
715e4cd4
YZ
332 unsigned int which, int num_ops,
333 int opcode, int flags,
f24e9980 334 struct ceph_snap_context *snapc,
acead002 335 u32 truncate_seq, u64 truncate_size,
153e5167 336 bool use_mempool);
f24e9980 337
a40c4f10
YS
338extern void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
339 struct ceph_osd_request *req);
a40c4f10 340
9e94af20
ID
341extern void ceph_osdc_get_request(struct ceph_osd_request *req);
342extern void ceph_osdc_put_request(struct ceph_osd_request *req);
f24e9980
SW
343
344extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
345 struct ceph_osd_request *req,
346 bool nofail);
c9f9b93d 347extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
f24e9980
SW
348extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
349 struct ceph_osd_request *req);
350extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
351
dd935f44
JD
352extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
353
f24e9980
SW
354extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
355 struct ceph_vino vino,
356 struct ceph_file_layout *layout,
357 u64 off, u64 *plen,
358 u32 truncate_seq, u64 truncate_size,
b7495fc2
SW
359 struct page **pages, int nr_pages,
360 int page_align);
f24e9980
SW
361
362extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
363 struct ceph_vino vino,
364 struct ceph_file_layout *layout,
365 struct ceph_snap_context *sc,
366 u64 off, u64 len,
367 u32 truncate_seq, u64 truncate_size,
368 struct timespec *mtime,
24808826 369 struct page **pages, int nr_pages);
f24e9980 370
a40c4f10
YS
371/* watch/notify events */
372extern int ceph_osdc_create_event(struct ceph_osd_client *osdc,
373 void (*event_cb)(u64, u64, u8, void *),
3c663bbd 374 void *data, struct ceph_osd_event **pevent);
a40c4f10 375extern void ceph_osdc_cancel_event(struct ceph_osd_event *event);
a40c4f10 376extern void ceph_osdc_put_event(struct ceph_osd_event *event);
f24e9980
SW
377#endif
378