blob: cedfb1a8434a11a0ba0b32348a0687e0ff9e7836 [file] [log] [blame]
Sage Weilf24e9982009-10-06 11:31:10 -07001#ifndef _FS_CEPH_OSD_CLIENT_H
2#define _FS_CEPH_OSD_CLIENT_H
3
4#include <linux/completion.h>
Sage Weil415e49a2009-12-07 13:37:03 -08005#include <linux/kref.h>
Sage Weilf24e9982009-10-06 11:31:10 -07006#include <linux/mempool.h>
7#include <linux/rbtree.h>
8
Alex Elder6c4a1912012-05-16 15:16:38 -05009#include <linux/ceph/types.h>
10#include <linux/ceph/osdmap.h>
11#include <linux/ceph/messenger.h>
12#include <linux/ceph/auth.h>
Sage Weilf24e9982009-10-06 11:31:10 -070013
Stratos Psomadakis224736d2011-11-10 15:45:37 +020014/*
15 * Maximum object name size
16 * (must be at least as big as RBD_MAX_MD_NAME_LEN -- currently 100)
17 */
18#define MAX_OBJ_NAME_SIZE 100
19
Sage Weilf24e9982009-10-06 11:31:10 -070020struct ceph_msg;
21struct ceph_snap_context;
22struct ceph_osd_request;
23struct ceph_osd_client;
Sage Weil4e7a5dc2009-11-18 16:19:57 -080024struct ceph_authorizer;
Yehuda Sadeh68b44762010-04-06 15:01:27 -070025struct ceph_pagelist;
Sage Weilf24e9982009-10-06 11:31:10 -070026
27/*
28 * completion callback for async writepages
29 */
30typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *,
31 struct ceph_msg *);
32
33/* a given osd we're communicating with */
34struct ceph_osd {
35 atomic_t o_ref;
36 struct ceph_osd_client *o_osdc;
37 int o_osd;
38 int o_incarnation;
39 struct rb_node o_node;
40 struct ceph_connection o_con;
41 struct list_head o_requests;
Yehuda Sadeha40c4f12011-03-21 15:07:16 -070042 struct list_head o_linger_requests;
Yehuda Sadehf5a20412010-02-03 11:00:26 -080043 struct list_head o_osd_lru;
Alex Elder6c4a1912012-05-16 15:16:38 -050044 struct ceph_auth_handshake o_auth;
Yehuda Sadehf5a20412010-02-03 11:00:26 -080045 unsigned long lru_ttl;
Yehuda Sadeh422d2cb2010-02-26 15:32:31 -080046 int o_marked_for_keepalive;
47 struct list_head o_keepalive_item;
Sage Weilf24e9982009-10-06 11:31:10 -070048};
49
50/* an in-flight request */
51struct ceph_osd_request {
52 u64 r_tid; /* unique for this client */
53 struct rb_node r_node;
Yehuda Sadeh422d2cb2010-02-26 15:32:31 -080054 struct list_head r_req_lru_item;
Sage Weilf24e9982009-10-06 11:31:10 -070055 struct list_head r_osd_item;
Yehuda Sadeha40c4f12011-03-21 15:07:16 -070056 struct list_head r_linger_item;
57 struct list_head r_linger_osd;
Sage Weilf24e9982009-10-06 11:31:10 -070058 struct ceph_osd *r_osd;
Sage Weil7740a422010-01-08 15:58:25 -080059 struct ceph_pg r_pgid;
Sage Weild85b7052010-05-10 10:24:48 -070060 int r_pg_osds[CEPH_PG_MAX_SIZE];
61 int r_num_pg_osds;
Sage Weilf24e9982009-10-06 11:31:10 -070062
Yehuda Sadeh0d59ab82010-01-13 17:03:23 -080063 struct ceph_connection *r_con_filling_msg;
Sage Weil350b1c32009-12-22 10:45:45 -080064
Sage Weilf24e9982009-10-06 11:31:10 -070065 struct ceph_msg *r_request, *r_reply;
66 int r_result;
67 int r_flags; /* any additional flags for the osd */
68 u32 r_sent; /* >0 if r_request is sending/sent */
Sage Weil350b1c32009-12-22 10:45:45 -080069 int r_got_reply;
Yehuda Sadeha40c4f12011-03-21 15:07:16 -070070 int r_linger;
Sage Weilf24e9982009-10-06 11:31:10 -070071
72 struct ceph_osd_client *r_osdc;
Sage Weil415e49a2009-12-07 13:37:03 -080073 struct kref r_kref;
Sage Weilf24e9982009-10-06 11:31:10 -070074 bool r_mempool;
75 struct completion r_completion, r_safe_completion;
76 ceph_osdc_callback_t r_callback, r_safe_callback;
77 struct ceph_eversion r_reassert_version;
78 struct list_head r_unsafe_item;
79
80 struct inode *r_inode; /* for use by callbacks */
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070081 void *r_priv; /* ditto */
Sage Weilf24e9982009-10-06 11:31:10 -070082
Stratos Psomadakis224736d2011-11-10 15:45:37 +020083 char r_oid[MAX_OBJ_NAME_SIZE]; /* object name */
Sage Weilf24e9982009-10-06 11:31:10 -070084 int r_oid_len;
Sage Weil3dd72fc2010-03-22 14:42:30 -070085 unsigned long r_stamp; /* send OR check time */
Sage Weilf24e9982009-10-06 11:31:10 -070086
87 struct ceph_file_layout r_file_layout;
88 struct ceph_snap_context *r_snapc; /* snap context for writes */
89 unsigned r_num_pages; /* size of page array (follows) */
Sage Weilb7495fc2010-11-09 12:43:12 -080090 unsigned r_page_alignment; /* io offset in first page */
Sage Weilf24e9982009-10-06 11:31:10 -070091 struct page **r_pages; /* pages for data payload */
92 int r_pages_from_pool;
93 int r_own_pages; /* if true, i own page list */
Yehuda Sadeh68b44762010-04-06 15:01:27 -070094#ifdef CONFIG_BLOCK
95 struct bio *r_bio; /* instead of pages */
96#endif
97
98 struct ceph_pagelist *r_trail; /* trailing part of the data */
Sage Weilf24e9982009-10-06 11:31:10 -070099};
100
Yehuda Sadeha40c4f12011-03-21 15:07:16 -0700101struct ceph_osd_event {
102 u64 cookie;
103 int one_shot;
104 struct ceph_osd_client *osdc;
105 void (*cb)(u64, u64, u8, void *);
106 void *data;
107 struct rb_node node;
108 struct list_head osd_node;
109 struct kref kref;
110 struct completion completion;
111};
112
113struct ceph_osd_event_work {
114 struct work_struct work;
115 struct ceph_osd_event *event;
116 u64 ver;
117 u64 notify_id;
118 u8 opcode;
119};
120
Sage Weilf24e9982009-10-06 11:31:10 -0700121struct ceph_osd_client {
122 struct ceph_client *client;
123
124 struct ceph_osdmap *osdmap; /* current map */
125 struct rw_semaphore map_sem;
126 struct completion map_waiters;
127 u64 last_requested_map;
128
129 struct mutex request_mutex;
130 struct rb_root osds; /* osds */
Yehuda Sadehf5a20412010-02-03 11:00:26 -0800131 struct list_head osd_lru; /* idle osds */
Sage Weilf24e9982009-10-06 11:31:10 -0700132 u64 timeout_tid; /* tid of timeout triggering rq */
133 u64 last_tid; /* tid of last request */
134 struct rb_root requests; /* pending requests */
Sage Weil6f6c7002011-01-17 20:34:08 -0800135 struct list_head req_lru; /* in-flight lru */
136 struct list_head req_unsent; /* unsent/need-resend queue */
137 struct list_head req_notarget; /* map to no osd */
Yehuda Sadeha40c4f12011-03-21 15:07:16 -0700138 struct list_head req_linger; /* lingering requests */
Sage Weilf24e9982009-10-06 11:31:10 -0700139 int num_requests;
140 struct delayed_work timeout_work;
Yehuda Sadehf5a20412010-02-03 11:00:26 -0800141 struct delayed_work osds_timeout_work;
Sage Weil039934b2009-11-12 15:05:52 -0800142#ifdef CONFIG_DEBUG_FS
Sage Weilf24e9982009-10-06 11:31:10 -0700143 struct dentry *debugfs_file;
Sage Weil039934b2009-11-12 15:05:52 -0800144#endif
Sage Weilf24e9982009-10-06 11:31:10 -0700145
146 mempool_t *req_mempool;
147
Yehuda Sadeh0d59ab82010-01-13 17:03:23 -0800148 struct ceph_msgpool msgpool_op;
Sage Weilc16e7862010-03-01 13:02:00 -0800149 struct ceph_msgpool msgpool_op_reply;
Yehuda Sadeha40c4f12011-03-21 15:07:16 -0700150
151 spinlock_t event_lock;
152 struct rb_root event_tree;
153 u64 event_count;
154
155 struct workqueue_struct *notify_wq;
Sage Weilf24e9982009-10-06 11:31:10 -0700156};
157
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700158struct ceph_osd_req_op {
159 u16 op; /* CEPH_OSD_OP_* */
160 u32 flags; /* CEPH_OSD_FLAG_* */
161 union {
162 struct {
163 u64 offset, length;
164 u64 truncate_size;
165 u32 truncate_seq;
166 } extent;
167 struct {
168 const char *name;
169 u32 name_len;
170 const char *val;
171 u32 value_len;
172 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */
173 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */
174 } xattr;
175 struct {
Yehuda Sadehae1533b2010-05-18 16:38:08 -0700176 const char *class_name;
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700177 __u8 class_len;
Yehuda Sadehae1533b2010-05-18 16:38:08 -0700178 const char *method_name;
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700179 __u8 method_len;
180 __u8 argc;
Yehuda Sadehae1533b2010-05-18 16:38:08 -0700181 const char *indata;
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700182 u32 indata_len;
183 } cls;
184 struct {
185 u64 cookie, count;
186 } pgls;
Yehuda Sadehae1533b2010-05-18 16:38:08 -0700187 struct {
188 u64 snapid;
189 } snap;
Yehuda Sadeha40c4f12011-03-21 15:07:16 -0700190 struct {
191 u64 cookie;
192 u64 ver;
193 __u8 flag;
194 u32 prot_ver;
195 u32 timeout;
196 } watch;
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700197 };
198 u32 payload_len;
199};
200
Sage Weilf24e9982009-10-06 11:31:10 -0700201extern int ceph_osdc_init(struct ceph_osd_client *osdc,
202 struct ceph_client *client);
203extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
204
205extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
206 struct ceph_msg *msg);
207extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
208 struct ceph_msg *msg);
209
Yehuda Sadeh3499e8a2010-04-06 14:51:47 -0700210extern void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
211 struct ceph_file_layout *layout,
212 u64 snapid,
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700213 u64 off, u64 *plen, u64 *bno,
214 struct ceph_osd_request *req,
215 struct ceph_osd_req_op *op);
Yehuda Sadeh3499e8a2010-04-06 14:51:47 -0700216
217extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
218 int flags,
219 struct ceph_snap_context *snapc,
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700220 struct ceph_osd_req_op *ops,
Yehuda Sadeh3499e8a2010-04-06 14:51:47 -0700221 bool use_mempool,
222 gfp_t gfp_flags,
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700223 struct page **pages,
224 struct bio *bio);
Yehuda Sadeh3499e8a2010-04-06 14:51:47 -0700225
226extern void ceph_osdc_build_request(struct ceph_osd_request *req,
Yehuda Sadeh68b44762010-04-06 15:01:27 -0700227 u64 off, u64 *plen,
228 struct ceph_osd_req_op *src_ops,
229 struct ceph_snap_context *snapc,
230 struct timespec *mtime,
231 const char *oid,
232 int oid_len);
Yehuda Sadeh3499e8a2010-04-06 14:51:47 -0700233
Sage Weilf24e9982009-10-06 11:31:10 -0700234extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
235 struct ceph_file_layout *layout,
236 struct ceph_vino vino,
237 u64 offset, u64 *len, int op, int flags,
238 struct ceph_snap_context *snapc,
239 int do_sync, u32 truncate_seq,
240 u64 truncate_size,
241 struct timespec *mtime,
Sage Weilb7495fc2010-11-09 12:43:12 -0800242 bool use_mempool, int num_reply,
243 int page_align);
Sage Weilf24e9982009-10-06 11:31:10 -0700244
Yehuda Sadeha40c4f12011-03-21 15:07:16 -0700245extern void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
246 struct ceph_osd_request *req);
247extern void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
248 struct ceph_osd_request *req);
249
Sage Weilf24e9982009-10-06 11:31:10 -0700250static inline void ceph_osdc_get_request(struct ceph_osd_request *req)
251{
Sage Weil415e49a2009-12-07 13:37:03 -0800252 kref_get(&req->r_kref);
Sage Weilf24e9982009-10-06 11:31:10 -0700253}
Sage Weil415e49a2009-12-07 13:37:03 -0800254extern void ceph_osdc_release_request(struct kref *kref);
255static inline void ceph_osdc_put_request(struct ceph_osd_request *req)
256{
257 kref_put(&req->r_kref, ceph_osdc_release_request);
258}
Sage Weilf24e9982009-10-06 11:31:10 -0700259
260extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
261 struct ceph_osd_request *req,
262 bool nofail);
263extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
264 struct ceph_osd_request *req);
265extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
266
267extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
268 struct ceph_vino vino,
269 struct ceph_file_layout *layout,
270 u64 off, u64 *plen,
271 u32 truncate_seq, u64 truncate_size,
Sage Weilb7495fc2010-11-09 12:43:12 -0800272 struct page **pages, int nr_pages,
273 int page_align);
Sage Weilf24e9982009-10-06 11:31:10 -0700274
275extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
276 struct ceph_vino vino,
277 struct ceph_file_layout *layout,
278 struct ceph_snap_context *sc,
279 u64 off, u64 len,
280 u32 truncate_seq, u64 truncate_size,
281 struct timespec *mtime,
282 struct page **pages, int nr_pages,
283 int flags, int do_sync, bool nofail);
284
Yehuda Sadeha40c4f12011-03-21 15:07:16 -0700285/* watch/notify events */
286extern int ceph_osdc_create_event(struct ceph_osd_client *osdc,
287 void (*event_cb)(u64, u64, u8, void *),
288 int one_shot, void *data,
289 struct ceph_osd_event **pevent);
290extern void ceph_osdc_cancel_event(struct ceph_osd_event *event);
291extern int ceph_osdc_wait_event(struct ceph_osd_event *event,
292 unsigned long timeout);
293extern void ceph_osdc_put_event(struct ceph_osd_event *event);
Sage Weilf24e9982009-10-06 11:31:10 -0700294#endif
295