Commit | Line | Data |
---|---|---|
b411b363 PR |
1 | /* |
2 | drbd_int.h | |
3 | ||
4 | This file is part of DRBD by Philipp Reisner and Lars Ellenberg. | |
5 | ||
6 | Copyright (C) 2001-2008, LINBIT Information Technologies GmbH. | |
7 | Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>. | |
8 | Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>. | |
9 | ||
10 | drbd is free software; you can redistribute it and/or modify | |
11 | it under the terms of the GNU General Public License as published by | |
12 | the Free Software Foundation; either version 2, or (at your option) | |
13 | any later version. | |
14 | ||
15 | drbd is distributed in the hope that it will be useful, | |
16 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
18 | GNU General Public License for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
21 | along with drbd; see the file COPYING. If not, write to | |
22 | the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. | |
23 | ||
24 | */ | |
25 | ||
26 | #ifndef _DRBD_INT_H | |
27 | #define _DRBD_INT_H | |
28 | ||
29 | #include <linux/compiler.h> | |
30 | #include <linux/types.h> | |
31 | #include <linux/version.h> | |
32 | #include <linux/list.h> | |
33 | #include <linux/sched.h> | |
34 | #include <linux/bitops.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/crypto.h> | |
132cc538 | 37 | #include <linux/ratelimit.h> |
b411b363 PR |
38 | #include <linux/tcp.h> |
39 | #include <linux/mutex.h> | |
40 | #include <linux/major.h> | |
41 | #include <linux/blkdev.h> | |
42 | #include <linux/genhd.h> | |
062e879c | 43 | #include <linux/idr.h> |
b411b363 PR |
44 | #include <net/tcp.h> |
45 | #include <linux/lru_cache.h> | |
70c71606 | 46 | #include <linux/prefetch.h> |
3b98c0c2 | 47 | #include <linux/drbd_genl_api.h> |
b8907339 PR |
48 | #include <linux/drbd.h> |
49 | #include "drbd_state.h" | |
b411b363 PR |
50 | |
51 | #ifdef __CHECKER__ | |
52 | # define __protected_by(x) __attribute__((require_context(x,1,999,"rdwr"))) | |
53 | # define __protected_read_by(x) __attribute__((require_context(x,1,999,"read"))) | |
54 | # define __protected_write_by(x) __attribute__((require_context(x,1,999,"write"))) | |
55 | # define __must_hold(x) __attribute__((context(x,1,1), require_context(x,1,999,"call"))) | |
56 | #else | |
57 | # define __protected_by(x) | |
58 | # define __protected_read_by(x) | |
59 | # define __protected_write_by(x) | |
60 | # define __must_hold(x) | |
61 | #endif | |
62 | ||
63 | #define __no_warn(lock, stmt) do { __acquire(lock); stmt; __release(lock); } while (0) | |
64 | ||
65 | /* module parameter, defined in drbd_main.c */ | |
66 | extern unsigned int minor_count; | |
67 | extern int disable_sendpage; | |
68 | extern int allow_oos; | |
b411b363 PR |
69 | |
70 | #ifdef CONFIG_DRBD_FAULT_INJECTION | |
71 | extern int enable_faults; | |
72 | extern int fault_rate; | |
73 | extern int fault_devs; | |
74 | #endif | |
75 | ||
76 | extern char usermode_helper[]; | |
77 | ||
78 | ||
b411b363 PR |
79 | /* I don't remember why XCPU ... |
80 | * This is used to wake the asender, | |
81 | * and to interrupt sending the sending task | |
82 | * on disconnect. | |
83 | */ | |
84 | #define DRBD_SIG SIGXCPU | |
85 | ||
86 | /* This is used to stop/restart our threads. | |
87 | * Cannot use SIGTERM nor SIGKILL, since these | |
88 | * are sent out by init on runlevel changes | |
89 | * I choose SIGHUP for now. | |
90 | */ | |
91 | #define DRBD_SIGKILL SIGHUP | |
92 | ||
b411b363 PR |
93 | #define ID_IN_SYNC (4711ULL) |
94 | #define ID_OUT_OF_SYNC (4712ULL) | |
b411b363 | 95 | #define ID_SYNCER (-1ULL) |
579b57ed | 96 | |
4a23f264 | 97 | #define UUID_NEW_BM_OFFSET ((u64)0x0001000000000000ULL) |
b411b363 PR |
98 | |
99 | struct drbd_conf; | |
2111438b | 100 | struct drbd_tconn; |
b411b363 PR |
101 | |
102 | ||
103 | /* to shorten dev_warn(DEV, "msg"); and relatives statements */ | |
104 | #define DEV (disk_to_dev(mdev->vdisk)) | |
105 | ||
60ae4966 PR |
106 | #define conn_printk(LEVEL, TCONN, FMT, ARGS...) \ |
107 | printk(LEVEL "d-con %s: " FMT, TCONN->name , ## ARGS) | |
108 | #define conn_alert(TCONN, FMT, ARGS...) conn_printk(KERN_ALERT, TCONN, FMT, ## ARGS) | |
109 | #define conn_crit(TCONN, FMT, ARGS...) conn_printk(KERN_CRIT, TCONN, FMT, ## ARGS) | |
110 | #define conn_err(TCONN, FMT, ARGS...) conn_printk(KERN_ERR, TCONN, FMT, ## ARGS) | |
111 | #define conn_warn(TCONN, FMT, ARGS...) conn_printk(KERN_WARNING, TCONN, FMT, ## ARGS) | |
112 | #define conn_notice(TCONN, FMT, ARGS...) conn_printk(KERN_NOTICE, TCONN, FMT, ## ARGS) | |
113 | #define conn_info(TCONN, FMT, ARGS...) conn_printk(KERN_INFO, TCONN, FMT, ## ARGS) | |
114 | #define conn_dbg(TCONN, FMT, ARGS...) conn_printk(KERN_DEBUG, TCONN, FMT, ## ARGS) | |
115 | ||
b411b363 PR |
116 | #define D_ASSERT(exp) if (!(exp)) \ |
117 | dev_err(DEV, "ASSERT( " #exp " ) in %s:%d\n", __FILE__, __LINE__) | |
118 | ||
841ce241 AG |
119 | /** |
120 | * expect - Make an assertion | |
121 | * | |
122 | * Unlike the assert macro, this macro returns a boolean result. | |
123 | */ | |
124 | #define expect(exp) ({ \ | |
125 | bool _bool = (exp); \ | |
126 | if (!_bool) \ | |
127 | dev_err(DEV, "ASSERTION %s FAILED in %s\n", \ | |
128 | #exp, __func__); \ | |
129 | _bool; \ | |
130 | }) | |
b411b363 PR |
131 | |
132 | /* Defines to control fault insertion */ | |
133 | enum { | |
134 | DRBD_FAULT_MD_WR = 0, /* meta data write */ | |
135 | DRBD_FAULT_MD_RD = 1, /* read */ | |
136 | DRBD_FAULT_RS_WR = 2, /* resync */ | |
137 | DRBD_FAULT_RS_RD = 3, | |
138 | DRBD_FAULT_DT_WR = 4, /* data */ | |
139 | DRBD_FAULT_DT_RD = 5, | |
140 | DRBD_FAULT_DT_RA = 6, /* data read ahead */ | |
141 | DRBD_FAULT_BM_ALLOC = 7, /* bitmap allocation */ | |
142 | DRBD_FAULT_AL_EE = 8, /* alloc ee */ | |
6b4388ac | 143 | DRBD_FAULT_RECEIVE = 9, /* Changes some bytes upon receiving a [rs]data block */ |
b411b363 PR |
144 | |
145 | DRBD_FAULT_MAX, | |
146 | }; | |
147 | ||
b411b363 PR |
148 | extern unsigned int |
149 | _drbd_insert_fault(struct drbd_conf *mdev, unsigned int type); | |
0cf9d27e | 150 | |
b411b363 PR |
151 | static inline int |
152 | drbd_insert_fault(struct drbd_conf *mdev, unsigned int type) { | |
0cf9d27e | 153 | #ifdef CONFIG_DRBD_FAULT_INJECTION |
b411b363 PR |
154 | return fault_rate && |
155 | (enable_faults & (1<<type)) && | |
156 | _drbd_insert_fault(mdev, type); | |
b411b363 | 157 | #else |
0cf9d27e | 158 | return 0; |
b411b363 | 159 | #endif |
0cf9d27e | 160 | } |
b411b363 PR |
161 | |
162 | /* integer division, round _UP_ to the next integer */ | |
163 | #define div_ceil(A, B) ((A)/(B) + ((A)%(B) ? 1 : 0)) | |
164 | /* usual integer division */ | |
165 | #define div_floor(A, B) ((A)/(B)) | |
166 | ||
b411b363 | 167 | extern struct ratelimit_state drbd_ratelimit_state; |
c141ebda PR |
168 | extern struct idr minors; /* RCU, updates: genl_lock() */ |
169 | extern struct list_head drbd_tconns; /* RCU, updates: genl_lock() */ | |
b411b363 PR |
170 | |
171 | /* on the wire */ | |
d8763023 | 172 | enum drbd_packet { |
b411b363 PR |
173 | /* receiver (data socket) */ |
174 | P_DATA = 0x00, | |
175 | P_DATA_REPLY = 0x01, /* Response to P_DATA_REQUEST */ | |
176 | P_RS_DATA_REPLY = 0x02, /* Response to P_RS_DATA_REQUEST */ | |
177 | P_BARRIER = 0x03, | |
178 | P_BITMAP = 0x04, | |
179 | P_BECOME_SYNC_TARGET = 0x05, | |
180 | P_BECOME_SYNC_SOURCE = 0x06, | |
181 | P_UNPLUG_REMOTE = 0x07, /* Used at various times to hint the peer */ | |
182 | P_DATA_REQUEST = 0x08, /* Used to ask for a data block */ | |
183 | P_RS_DATA_REQUEST = 0x09, /* Used to ask for a data block for resync */ | |
184 | P_SYNC_PARAM = 0x0a, | |
185 | P_PROTOCOL = 0x0b, | |
186 | P_UUIDS = 0x0c, | |
187 | P_SIZES = 0x0d, | |
188 | P_STATE = 0x0e, | |
189 | P_SYNC_UUID = 0x0f, | |
190 | P_AUTH_CHALLENGE = 0x10, | |
191 | P_AUTH_RESPONSE = 0x11, | |
192 | P_STATE_CHG_REQ = 0x12, | |
193 | ||
194 | /* asender (meta socket */ | |
195 | P_PING = 0x13, | |
196 | P_PING_ACK = 0x14, | |
197 | P_RECV_ACK = 0x15, /* Used in protocol B */ | |
198 | P_WRITE_ACK = 0x16, /* Used in protocol C */ | |
199 | P_RS_WRITE_ACK = 0x17, /* Is a P_WRITE_ACK, additionally call set_in_sync(). */ | |
d4dabbe2 | 200 | P_SUPERSEDED = 0x18, /* Used in proto C, two-primaries conflict detection */ |
b411b363 PR |
201 | P_NEG_ACK = 0x19, /* Sent if local disk is unusable */ |
202 | P_NEG_DREPLY = 0x1a, /* Local disk is broken... */ | |
203 | P_NEG_RS_DREPLY = 0x1b, /* Local disk is broken... */ | |
204 | P_BARRIER_ACK = 0x1c, | |
205 | P_STATE_CHG_REPLY = 0x1d, | |
206 | ||
207 | /* "new" commands, no longer fitting into the ordering scheme above */ | |
208 | ||
209 | P_OV_REQUEST = 0x1e, /* data socket */ | |
210 | P_OV_REPLY = 0x1f, | |
211 | P_OV_RESULT = 0x20, /* meta socket */ | |
212 | P_CSUM_RS_REQUEST = 0x21, /* data socket */ | |
213 | P_RS_IS_IN_SYNC = 0x22, /* meta socket */ | |
214 | P_SYNC_PARAM89 = 0x23, /* data socket, protocol version 89 replacement for P_SYNC_PARAM */ | |
215 | P_COMPRESSED_BITMAP = 0x24, /* compressed or otherwise encoded bitmap transfer */ | |
0ced55a3 PR |
216 | /* P_CKPT_FENCE_REQ = 0x25, * currently reserved for protocol D */ |
217 | /* P_CKPT_DISABLE_REQ = 0x26, * currently reserved for protocol D */ | |
218 | P_DELAY_PROBE = 0x27, /* is used on BOTH sockets */ | |
73a01a18 | 219 | P_OUT_OF_SYNC = 0x28, /* Mark as out of sync (Outrunning), data socket */ |
d612d309 | 220 | P_RS_CANCEL = 0x29, /* meta: Used to cancel RS_DATA_REQUEST packet by SyncSource */ |
cf29c9d8 PR |
221 | P_CONN_ST_CHG_REQ = 0x2a, /* data sock: Connection wide state request */ |
222 | P_CONN_ST_CHG_REPLY = 0x2b, /* meta sock: Connection side state req reply */ | |
7be8da07 | 223 | P_RETRY_WRITE = 0x2c, /* Protocol C: retry conflicting write request */ |
036b17ea | 224 | P_PROTOCOL_UPDATE = 0x2d, /* data sock: is used in established connections */ |
b411b363 | 225 | |
b411b363 PR |
226 | P_MAY_IGNORE = 0x100, /* Flag to test if (cmd > P_MAY_IGNORE) ... */ |
227 | P_MAX_OPT_CMD = 0x101, | |
228 | ||
229 | /* special command ids for handshake */ | |
230 | ||
e5d6f33a AG |
231 | P_INITIAL_META = 0xfff1, /* First Packet on the MetaSock */ |
232 | P_INITIAL_DATA = 0xfff2, /* First Packet on the Socket */ | |
b411b363 | 233 | |
6038178e | 234 | P_CONNECTION_FEATURES = 0xfffe /* FIXED for the next century! */ |
b411b363 PR |
235 | }; |
236 | ||
d8763023 | 237 | extern const char *cmdname(enum drbd_packet cmd); |
b411b363 PR |
238 | |
239 | /* for sending/receiving the bitmap, | |
240 | * possibly in some encoding scheme */ | |
241 | struct bm_xfer_ctx { | |
242 | /* "const" | |
243 | * stores total bits and long words | |
244 | * of the bitmap, so we don't need to | |
245 | * call the accessor functions over and again. */ | |
246 | unsigned long bm_bits; | |
247 | unsigned long bm_words; | |
248 | /* during xfer, current position within the bitmap */ | |
249 | unsigned long bit_offset; | |
250 | unsigned long word_offset; | |
251 | ||
252 | /* statistics; index: (h->command == P_BITMAP) */ | |
253 | unsigned packets[2]; | |
254 | unsigned bytes[2]; | |
255 | }; | |
256 | ||
257 | extern void INFO_bm_xfer_stats(struct drbd_conf *mdev, | |
258 | const char *direction, struct bm_xfer_ctx *c); | |
259 | ||
260 | static inline void bm_xfer_ctx_bit_to_word_offset(struct bm_xfer_ctx *c) | |
261 | { | |
262 | /* word_offset counts "native long words" (32 or 64 bit), | |
263 | * aligned at 64 bit. | |
264 | * Encoded packet may end at an unaligned bit offset. | |
265 | * In case a fallback clear text packet is transmitted in | |
266 | * between, we adjust this offset back to the last 64bit | |
267 | * aligned "native long word", which makes coding and decoding | |
268 | * the plain text bitmap much more convenient. */ | |
269 | #if BITS_PER_LONG == 64 | |
270 | c->word_offset = c->bit_offset >> 6; | |
271 | #elif BITS_PER_LONG == 32 | |
272 | c->word_offset = c->bit_offset >> 5; | |
273 | c->word_offset &= ~(1UL); | |
274 | #else | |
275 | # error "unsupported BITS_PER_LONG" | |
276 | #endif | |
277 | } | |
278 | ||
279 | #ifndef __packed | |
280 | #define __packed __attribute__((packed)) | |
281 | #endif | |
282 | ||
283 | /* This is the layout for a packet on the wire. | |
284 | * The byteorder is the network byte order. | |
285 | * (except block_id and barrier fields. | |
286 | * these are pointers to local structs | |
287 | * and have no relevance for the partner, | |
288 | * which just echoes them as received.) | |
289 | * | |
290 | * NOTE that the payload starts at a long aligned offset, | |
291 | * regardless of 32 or 64 bit arch! | |
292 | */ | |
0b70a13d | 293 | struct p_header80 { |
b411b363 PR |
294 | u32 magic; |
295 | u16 command; | |
296 | u16 length; /* bytes of data after this header */ | |
b411b363 | 297 | } __packed; |
0b70a13d PR |
298 | |
299 | /* Header for big packets, Used for data packets exceeding 64kB */ | |
300 | struct p_header95 { | |
301 | u16 magic; /* use DRBD_MAGIC_BIG here */ | |
302 | u16 command; | |
b55d84ba | 303 | u32 length; |
0b70a13d PR |
304 | } __packed; |
305 | ||
0c8e36d9 AG |
306 | struct p_header100 { |
307 | u32 magic; | |
308 | u16 volume; | |
309 | u16 command; | |
310 | u32 length; | |
311 | u32 pad; | |
312 | } __packed; | |
313 | ||
52b061a4 AG |
314 | extern unsigned int drbd_header_size(struct drbd_tconn *tconn); |
315 | ||
b411b363 | 316 | /* these defines must not be changed without changing the protocol version */ |
76d2e7ec PR |
317 | #define DP_HARDBARRIER 1 /* depricated */ |
318 | #define DP_RW_SYNC 2 /* equals REQ_SYNC */ | |
b411b363 | 319 | #define DP_MAY_SET_IN_SYNC 4 |
721a9602 | 320 | #define DP_UNPLUG 8 /* not used anymore */ |
76d2e7ec PR |
321 | #define DP_FUA 16 /* equals REQ_FUA */ |
322 | #define DP_FLUSH 32 /* equals REQ_FLUSH */ | |
323 | #define DP_DISCARD 64 /* equals REQ_DISCARD */ | |
303d1448 PR |
324 | #define DP_SEND_RECEIVE_ACK 128 /* This is a proto B write request */ |
325 | #define DP_SEND_WRITE_ACK 256 /* This is a proto C write request */ | |
b411b363 PR |
326 | |
327 | struct p_data { | |
b411b363 PR |
328 | u64 sector; /* 64 bits sector number */ |
329 | u64 block_id; /* to identify the request in protocol B&C */ | |
330 | u32 seq_num; | |
331 | u32 dp_flags; | |
332 | } __packed; | |
333 | ||
334 | /* | |
335 | * commands which share a struct: | |
336 | * p_block_ack: | |
337 | * P_RECV_ACK (proto B), P_WRITE_ACK (proto C), | |
d4dabbe2 | 338 | * P_SUPERSEDED (proto C, two-primaries conflict detection) |
b411b363 PR |
339 | * p_block_req: |
340 | * P_DATA_REQUEST, P_RS_DATA_REQUEST | |
341 | */ | |
342 | struct p_block_ack { | |
b411b363 PR |
343 | u64 sector; |
344 | u64 block_id; | |
345 | u32 blksize; | |
346 | u32 seq_num; | |
347 | } __packed; | |
348 | ||
b411b363 | 349 | struct p_block_req { |
b411b363 PR |
350 | u64 sector; |
351 | u64 block_id; | |
352 | u32 blksize; | |
353 | u32 pad; /* to multiple of 8 Byte */ | |
354 | } __packed; | |
355 | ||
356 | /* | |
357 | * commands with their own struct for additional fields: | |
6038178e | 358 | * P_CONNECTION_FEATURES |
b411b363 PR |
359 | * P_BARRIER |
360 | * P_BARRIER_ACK | |
361 | * P_SYNC_PARAM | |
362 | * ReportParams | |
363 | */ | |
364 | ||
6038178e | 365 | struct p_connection_features { |
b411b363 PR |
366 | u32 protocol_min; |
367 | u32 feature_flags; | |
368 | u32 protocol_max; | |
369 | ||
370 | /* should be more than enough for future enhancements | |
d942ae44 | 371 | * for now, feature_flags and the reserved array shall be zero. |
b411b363 PR |
372 | */ |
373 | ||
374 | u32 _pad; | |
d942ae44 | 375 | u64 reserved[7]; |
b411b363 | 376 | } __packed; |
b411b363 PR |
377 | |
378 | struct p_barrier { | |
b411b363 PR |
379 | u32 barrier; /* barrier number _handle_ only */ |
380 | u32 pad; /* to multiple of 8 Byte */ | |
381 | } __packed; | |
382 | ||
383 | struct p_barrier_ack { | |
b411b363 PR |
384 | u32 barrier; |
385 | u32 set_size; | |
386 | } __packed; | |
387 | ||
388 | struct p_rs_param { | |
6394b935 | 389 | u32 resync_rate; |
b411b363 PR |
390 | |
391 | /* Since protocol version 88 and higher. */ | |
392 | char verify_alg[0]; | |
393 | } __packed; | |
394 | ||
395 | struct p_rs_param_89 { | |
6394b935 | 396 | u32 resync_rate; |
b411b363 PR |
397 | /* protocol version 89: */ |
398 | char verify_alg[SHARED_SECRET_MAX]; | |
399 | char csums_alg[SHARED_SECRET_MAX]; | |
400 | } __packed; | |
401 | ||
8e26f9cc | 402 | struct p_rs_param_95 { |
6394b935 | 403 | u32 resync_rate; |
8e26f9cc PR |
404 | char verify_alg[SHARED_SECRET_MAX]; |
405 | char csums_alg[SHARED_SECRET_MAX]; | |
406 | u32 c_plan_ahead; | |
407 | u32 c_delay_target; | |
408 | u32 c_fill_target; | |
409 | u32 c_max_rate; | |
410 | } __packed; | |
411 | ||
cf14c2e9 | 412 | enum drbd_conn_flags { |
6139f60d | 413 | CF_DISCARD_MY_DATA = 1, |
cf14c2e9 PR |
414 | CF_DRY_RUN = 2, |
415 | }; | |
416 | ||
b411b363 | 417 | struct p_protocol { |
b411b363 PR |
418 | u32 protocol; |
419 | u32 after_sb_0p; | |
420 | u32 after_sb_1p; | |
421 | u32 after_sb_2p; | |
cf14c2e9 | 422 | u32 conn_flags; |
b411b363 PR |
423 | u32 two_primaries; |
424 | ||
425 | /* Since protocol version 87 and higher. */ | |
426 | char integrity_alg[0]; | |
427 | ||
428 | } __packed; | |
429 | ||
430 | struct p_uuids { | |
b411b363 PR |
431 | u64 uuid[UI_EXTENDED_SIZE]; |
432 | } __packed; | |
433 | ||
434 | struct p_rs_uuid { | |
b411b363 PR |
435 | u64 uuid; |
436 | } __packed; | |
437 | ||
438 | struct p_sizes { | |
b411b363 PR |
439 | u64 d_size; /* size of disk */ |
440 | u64 u_size; /* user requested size */ | |
441 | u64 c_size; /* current exported size */ | |
1816a2b4 | 442 | u32 max_bio_size; /* Maximal size of a BIO */ |
e89b591c PR |
443 | u16 queue_order_type; /* not yet implemented in DRBD*/ |
444 | u16 dds_flags; /* use enum dds_flags here. */ | |
b411b363 PR |
445 | } __packed; |
446 | ||
447 | struct p_state { | |
b411b363 PR |
448 | u32 state; |
449 | } __packed; | |
450 | ||
451 | struct p_req_state { | |
b411b363 PR |
452 | u32 mask; |
453 | u32 val; | |
454 | } __packed; | |
455 | ||
456 | struct p_req_state_reply { | |
b411b363 PR |
457 | u32 retcode; |
458 | } __packed; | |
459 | ||
460 | struct p_drbd06_param { | |
461 | u64 size; | |
462 | u32 state; | |
463 | u32 blksize; | |
464 | u32 protocol; | |
465 | u32 version; | |
466 | u32 gen_cnt[5]; | |
467 | u32 bit_map_gen[5]; | |
468 | } __packed; | |
469 | ||
73a01a18 | 470 | struct p_block_desc { |
73a01a18 PR |
471 | u64 sector; |
472 | u32 blksize; | |
473 | u32 pad; /* to multiple of 8 Byte */ | |
474 | } __packed; | |
475 | ||
b411b363 PR |
476 | /* Valid values for the encoding field. |
477 | * Bump proto version when changing this. */ | |
478 | enum drbd_bitmap_code { | |
479 | /* RLE_VLI_Bytes = 0, | |
480 | * and other bit variants had been defined during | |
481 | * algorithm evaluation. */ | |
482 | RLE_VLI_Bits = 2, | |
483 | }; | |
484 | ||
485 | struct p_compressed_bm { | |
b411b363 PR |
486 | /* (encoding & 0x0f): actual encoding, see enum drbd_bitmap_code |
487 | * (encoding & 0x80): polarity (set/unset) of first runlength | |
488 | * ((encoding >> 4) & 0x07): pad_bits, number of trailing zero bits | |
489 | * used to pad up to head.length bytes | |
490 | */ | |
491 | u8 encoding; | |
492 | ||
493 | u8 code[0]; | |
494 | } __packed; | |
495 | ||
0b70a13d | 496 | struct p_delay_probe93 { |
0b70a13d PR |
497 | u32 seq_num; /* sequence number to match the two probe packets */ |
498 | u32 offset; /* usecs the probe got sent after the reference time point */ | |
0ced55a3 PR |
499 | } __packed; |
500 | ||
50d0b1ad AG |
501 | /* |
502 | * Bitmap packets need to fit within a single page on the sender and receiver, | |
503 | * so we are limited to 4 KiB (and not to PAGE_SIZE, which can be bigger). | |
b411b363 | 504 | */ |
e6ef8a5c AG |
505 | #define DRBD_SOCKET_BUFFER_SIZE 4096 |
506 | ||
b411b363 PR |
507 | /**********************************************************************/ |
508 | enum drbd_thread_state { | |
e77a0a5c AG |
509 | NONE, |
510 | RUNNING, | |
511 | EXITING, | |
512 | RESTARTING | |
b411b363 PR |
513 | }; |
514 | ||
515 | struct drbd_thread { | |
516 | spinlock_t t_lock; | |
517 | struct task_struct *task; | |
518 | struct completion stop; | |
519 | enum drbd_thread_state t_state; | |
520 | int (*function) (struct drbd_thread *); | |
392c8801 | 521 | struct drbd_tconn *tconn; |
b411b363 | 522 | int reset_cpu_mask; |
bed879ae | 523 | char name[9]; |
b411b363 PR |
524 | }; |
525 | ||
526 | static inline enum drbd_thread_state get_t_state(struct drbd_thread *thi) | |
527 | { | |
528 | /* THINK testing the t_state seems to be uncritical in all cases | |
529 | * (but thread_{start,stop}), so we can read it *without* the lock. | |
530 | * --lge */ | |
531 | ||
532 | smp_rmb(); | |
533 | return thi->t_state; | |
534 | } | |
535 | ||
b411b363 PR |
536 | struct drbd_work { |
537 | struct list_head list; | |
309a8348 | 538 | int (*cb)(struct drbd_work *, int cancel); |
00d56944 PR |
539 | union { |
540 | struct drbd_conf *mdev; | |
541 | struct drbd_tconn *tconn; | |
542 | }; | |
b411b363 PR |
543 | }; |
544 | ||
ace652ac AG |
545 | #include "drbd_interval.h" |
546 | ||
7be8da07 AG |
547 | extern int drbd_wait_misc(struct drbd_conf *, struct drbd_interval *); |
548 | ||
b411b363 PR |
549 | struct drbd_request { |
550 | struct drbd_work w; | |
b411b363 PR |
551 | |
552 | /* if local IO is not allowed, will be NULL. | |
553 | * if local IO _is_ allowed, holds the locally submitted bio clone, | |
554 | * or, after local IO completion, the ERR_PTR(error). | |
fcefa62e | 555 | * see drbd_request_endio(). */ |
b411b363 PR |
556 | struct bio *private_bio; |
557 | ||
ace652ac | 558 | struct drbd_interval i; |
b411b363 | 559 | |
b6dd1a89 | 560 | /* epoch: used to check on "completion" whether this req was in |
b411b363 | 561 | * the current epoch, and we therefore have to close it, |
b6dd1a89 LE |
562 | * causing a p_barrier packet to be send, starting a new epoch. |
563 | * | |
564 | * This corresponds to "barrier" in struct p_barrier[_ack], | |
565 | * and to "barrier_nr" in struct drbd_epoch (and various | |
566 | * comments/function parameters/local variable names). | |
b411b363 | 567 | */ |
b6dd1a89 | 568 | unsigned int epoch; |
b411b363 | 569 | |
b411b363 PR |
570 | struct list_head tl_requests; /* ring list in the transfer log */ |
571 | struct bio *master_bio; /* master bio pointer */ | |
b411b363 | 572 | unsigned long start_time; |
b406777e LE |
573 | |
574 | /* once it hits 0, we may complete the master_bio */ | |
575 | atomic_t completion_ref; | |
576 | /* once it hits 0, we may destroy this drbd_request object */ | |
577 | struct kref kref; | |
a0d856df LE |
578 | |
579 | unsigned rq_state; /* see comments above _req_mod() */ | |
b411b363 PR |
580 | }; |
581 | ||
b411b363 | 582 | struct drbd_epoch { |
9ed57dcb | 583 | struct drbd_tconn *tconn; |
b411b363 PR |
584 | struct list_head list; |
585 | unsigned int barrier_nr; | |
586 | atomic_t epoch_size; /* increased on every request added. */ | |
587 | atomic_t active; /* increased on every req. added, and dec on every finished. */ | |
588 | unsigned long flags; | |
589 | }; | |
590 | ||
591 | /* drbd_epoch flag bits */ | |
592 | enum { | |
b411b363 | 593 | DE_HAVE_BARRIER_NUMBER, |
b411b363 PR |
594 | }; |
595 | ||
596 | enum epoch_event { | |
597 | EV_PUT, | |
598 | EV_GOT_BARRIER_NR, | |
b411b363 | 599 | EV_BECAME_LAST, |
b411b363 PR |
600 | EV_CLEANUP = 32, /* used as flag */ |
601 | }; | |
602 | ||
b411b363 PR |
603 | struct drbd_wq_barrier { |
604 | struct drbd_work w; | |
605 | struct completion done; | |
606 | }; | |
607 | ||
608 | struct digest_info { | |
609 | int digest_size; | |
610 | void *digest; | |
611 | }; | |
612 | ||
f6ffca9f | 613 | struct drbd_peer_request { |
45bb912b | 614 | struct drbd_work w; |
85719573 | 615 | struct drbd_epoch *epoch; /* for writes */ |
45bb912b LE |
616 | struct page *pages; |
617 | atomic_t pending_bios; | |
010f6e67 | 618 | struct drbd_interval i; |
45bb912b LE |
619 | /* see comments on ee flag bits below */ |
620 | unsigned long flags; | |
85719573 PR |
621 | union { |
622 | u64 block_id; | |
623 | struct digest_info *digest; | |
624 | }; | |
45bb912b LE |
625 | }; |
626 | ||
627 | /* ee flag bits. | |
628 | * While corresponding bios are in flight, the only modification will be | |
629 | * set_bit WAS_ERROR, which has to be atomic. | |
630 | * If no bios are in flight yet, or all have been completed, | |
631 | * non-atomic modification to ee->flags is ok. | |
632 | */ | |
b411b363 PR |
633 | enum { |
634 | __EE_CALL_AL_COMPLETE_IO, | |
b411b363 | 635 | __EE_MAY_SET_IN_SYNC, |
45bb912b | 636 | |
45bb912b LE |
637 | /* In case a barrier failed, |
638 | * we need to resubmit without the barrier flag. */ | |
639 | __EE_RESUBMITTED, | |
640 | ||
6c852bec | 641 | /* we may have several bios per peer request. |
45bb912b LE |
642 | * if any of those fail, we set this flag atomically |
643 | * from the endio callback */ | |
644 | __EE_WAS_ERROR, | |
c36c3ced LE |
645 | |
646 | /* This ee has a pointer to a digest instead of a block id */ | |
647 | __EE_HAS_DIGEST, | |
7be8da07 AG |
648 | |
649 | /* Conflicting local requests need to be restarted after this request */ | |
650 | __EE_RESTART_REQUESTS, | |
303d1448 PR |
651 | |
652 | /* The peer wants a write ACK for this (wire proto C) */ | |
653 | __EE_SEND_WRITE_ACK, | |
302bdeae PR |
654 | |
655 | /* Is set when net_conf had two_primaries set while creating this peer_req */ | |
656 | __EE_IN_INTERVAL_TREE, | |
b411b363 PR |
657 | }; |
658 | #define EE_CALL_AL_COMPLETE_IO (1<<__EE_CALL_AL_COMPLETE_IO) | |
b411b363 | 659 | #define EE_MAY_SET_IN_SYNC (1<<__EE_MAY_SET_IN_SYNC) |
45bb912b LE |
660 | #define EE_RESUBMITTED (1<<__EE_RESUBMITTED) |
661 | #define EE_WAS_ERROR (1<<__EE_WAS_ERROR) | |
c36c3ced | 662 | #define EE_HAS_DIGEST (1<<__EE_HAS_DIGEST) |
7be8da07 | 663 | #define EE_RESTART_REQUESTS (1<<__EE_RESTART_REQUESTS) |
303d1448 | 664 | #define EE_SEND_WRITE_ACK (1<<__EE_SEND_WRITE_ACK) |
302bdeae | 665 | #define EE_IN_INTERVAL_TREE (1<<__EE_IN_INTERVAL_TREE) |
b411b363 | 666 | |
01a311a5 | 667 | /* flag bits per mdev */ |
b411b363 | 668 | enum { |
b411b363 PR |
669 | UNPLUG_REMOTE, /* sending a "UnplugRemote" could help */ |
670 | MD_DIRTY, /* current uuids and flags not yet on disk */ | |
b411b363 | 671 | USE_DEGR_WFC_T, /* degr-wfc-timeout instead of wfc-timeout. */ |
b411b363 PR |
672 | CL_ST_CHG_SUCCESS, |
673 | CL_ST_CHG_FAIL, | |
674 | CRASHED_PRIMARY, /* This node was a crashed primary. | |
675 | * Gets cleared when the state.conn | |
676 | * goes into C_CONNECTED state. */ | |
b411b363 PR |
677 | CONSIDER_RESYNC, |
678 | ||
a8a4e51e | 679 | MD_NO_FUA, /* Users wants us to not use FUA/FLUSH on meta data dev */ |
b411b363 PR |
680 | SUSPEND_IO, /* suspend application io */ |
681 | BITMAP_IO, /* suspend application io; | |
682 | once no more io in flight, start bitmap io */ | |
683 | BITMAP_IO_QUEUED, /* Started bitmap IO */ | |
82f59cc6 | 684 | GO_DISKLESS, /* Disk is being detached, on io-error or admin request. */ |
edc9f5eb LE |
685 | WAS_IO_ERROR, /* Local disk failed, returned IO error */ |
686 | WAS_READ_ERROR, /* Local disk READ failed (set additionally to the above) */ | |
0c849666 | 687 | FORCE_DETACH, /* Force-detach from local disk, aborting any pending local IO */ |
b411b363 | 688 | RESYNC_AFTER_NEG, /* Resync after online grow after the attach&negotiate finished. */ |
b411b363 PR |
689 | RESIZE_PENDING, /* Size change detected locally, waiting for the response from |
690 | * the peer, if it changed there as well. */ | |
43a5182c | 691 | NEW_CUR_UUID, /* Create new current UUID when thawing IO */ |
0778286a | 692 | AL_SUSPENDED, /* Activity logging is currently suspended. */ |
370a43e7 | 693 | AHEAD_TO_SYNC_SOURCE, /* Ahead -> SyncSource queued */ |
e64a3294 | 694 | B_RS_H_DONE, /* Before resync handler done (already executed) */ |
08b165ba | 695 | DISCARD_MY_DATA, /* discard_my_data flag per volume */ |
380207d0 | 696 | READ_BALANCE_RR, |
b411b363 PR |
697 | }; |
698 | ||
699 | struct drbd_bitmap; /* opaque for drbd_conf */ | |
700 | ||
20ceb2b2 LE |
701 | /* definition of bits in bm_flags to be used in drbd_bm_lock |
702 | * and drbd_bitmap_io and friends. */ | |
703 | enum bm_flag { | |
704 | /* do we need to kfree, or vfree bm_pages? */ | |
705 | BM_P_VMALLOCED = 0x10000, /* internal use only, will be masked out */ | |
706 | ||
707 | /* currently locked for bulk operation */ | |
a220d291 | 708 | BM_LOCKED_MASK = 0xf, |
20ceb2b2 LE |
709 | |
710 | /* in detail, that is: */ | |
711 | BM_DONT_CLEAR = 0x1, | |
712 | BM_DONT_SET = 0x2, | |
713 | BM_DONT_TEST = 0x4, | |
714 | ||
a220d291 LE |
715 | /* so we can mark it locked for bulk operation, |
716 | * and still allow all non-bulk operations */ | |
717 | BM_IS_LOCKED = 0x8, | |
718 | ||
20ceb2b2 | 719 | /* (test bit, count bit) allowed (common case) */ |
a220d291 | 720 | BM_LOCKED_TEST_ALLOWED = BM_DONT_CLEAR | BM_DONT_SET | BM_IS_LOCKED, |
20ceb2b2 LE |
721 | |
722 | /* testing bits, as well as setting new bits allowed, but clearing bits | |
723 | * would be unexpected. Used during bitmap receive. Setting new bits | |
724 | * requires sending of "out-of-sync" information, though. */ | |
a220d291 | 725 | BM_LOCKED_SET_ALLOWED = BM_DONT_CLEAR | BM_IS_LOCKED, |
20ceb2b2 | 726 | |
a220d291 LE |
727 | /* for drbd_bm_write_copy_pages, everything is allowed, |
728 | * only concurrent bulk operations are locked out. */ | |
729 | BM_LOCKED_CHANGE_ALLOWED = BM_IS_LOCKED, | |
20ceb2b2 LE |
730 | }; |
731 | ||
b411b363 PR |
732 | struct drbd_work_queue { |
733 | struct list_head q; | |
b411b363 | 734 | spinlock_t q_lock; /* to protect the list. */ |
8c0785a5 | 735 | wait_queue_head_t q_wait; |
b411b363 PR |
736 | }; |
737 | ||
738 | struct drbd_socket { | |
b411b363 PR |
739 | struct mutex mutex; |
740 | struct socket *socket; | |
741 | /* this way we get our | |
742 | * send/receive buffers off the stack */ | |
5a87d920 | 743 | void *sbuf; |
e6ef8a5c | 744 | void *rbuf; |
b411b363 PR |
745 | }; |
746 | ||
747 | struct drbd_md { | |
748 | u64 md_offset; /* sector offset to 'super' block */ | |
749 | ||
750 | u64 la_size_sect; /* last agreed size, unit sectors */ | |
39a1aa7f | 751 | spinlock_t uuid_lock; |
b411b363 PR |
752 | u64 uuid[UI_SIZE]; |
753 | u64 device_uuid; | |
754 | u32 flags; | |
755 | u32 md_size_sect; | |
756 | ||
757 | s32 al_offset; /* signed relative sector offset to al area */ | |
758 | s32 bm_offset; /* signed relative sector offset to bitmap */ | |
759 | ||
760 | /* u32 al_nr_extents; important for restoring the AL | |
f399002e | 761 | * is stored into ldev->dc.al_extents, which in turn |
b411b363 PR |
762 | * gets applied to act_log->nr_elements |
763 | */ | |
764 | }; | |
765 | ||
b411b363 PR |
766 | struct drbd_backing_dev { |
767 | struct block_device *backing_bdev; | |
768 | struct block_device *md_bdev; | |
b411b363 | 769 | struct drbd_md md; |
daeda1cc | 770 | struct disk_conf *disk_conf; /* RCU, for updates: mdev->tconn->conf_update */ |
b411b363 PR |
771 | sector_t known_size; /* last known size of that backing device */ |
772 | }; | |
773 | ||
774 | struct drbd_md_io { | |
cdfda633 | 775 | unsigned int done; |
b411b363 PR |
776 | int error; |
777 | }; | |
778 | ||
779 | struct bm_io_work { | |
780 | struct drbd_work w; | |
781 | char *why; | |
20ceb2b2 | 782 | enum bm_flag flags; |
b411b363 PR |
783 | int (*io_fn)(struct drbd_conf *mdev); |
784 | void (*done)(struct drbd_conf *mdev, int rv); | |
785 | }; | |
786 | ||
787 | enum write_ordering_e { | |
788 | WO_none, | |
789 | WO_drain_io, | |
790 | WO_bdev_flush, | |
b411b363 PR |
791 | }; |
792 | ||
778f271d | 793 | struct fifo_buffer { |
778f271d PR |
794 | unsigned int head_index; |
795 | unsigned int size; | |
9958c857 PR |
796 | int total; /* sum of all values */ |
797 | int values[0]; | |
778f271d | 798 | }; |
9958c857 | 799 | extern struct fifo_buffer *fifo_alloc(int fifo_size); |
778f271d | 800 | |
01a311a5 PR |
801 | /* flag bits per tconn */ |
802 | enum { | |
803 | NET_CONGESTED, /* The data socket is congested */ | |
427c0434 | 804 | RESOLVE_CONFLICTS, /* Set on one node, cleared on the peer! */ |
e43ef195 | 805 | SEND_PING, /* whether asender should send a ping asap */ |
808e37b8 | 806 | SIGNAL_ASENDER, /* whether asender wants to be interrupted */ |
2a67d8b9 | 807 | GOT_PING_ACK, /* set when we receive a ping_ack packet, ping_wait gets woken */ |
4d0fc3fd | 808 | CONN_WD_ST_CHG_REQ, /* A cluster wide state change on the connection is active */ |
fc3b10a4 PR |
809 | CONN_WD_ST_CHG_OKAY, |
810 | CONN_WD_ST_CHG_FAIL, | |
8169e41b | 811 | CONN_DRY_RUN, /* Expect disconnect after resync handshake. */ |
6936fcb4 | 812 | CREATE_BARRIER, /* next P_DATA is preceded by a P_BARRIER */ |
a1096a6e | 813 | STATE_SENT, /* Do not change state/UUIDs while this is set */ |
6f3465ed LE |
814 | CALLBACK_PENDING, /* Whether we have a call_usermodehelper(, UMH_WAIT_PROC) |
815 | * pending, from drbd worker context. | |
816 | * If set, bdi_write_congested() returns true, | |
817 | * so shrink_page_list() would not recurse into, | |
818 | * and potentially deadlock on, this drbd worker. | |
819 | */ | |
b66623e3 | 820 | DISCONNECT_SENT, |
01a311a5 PR |
821 | }; |
822 | ||
2111438b PR |
823 | struct drbd_tconn { /* is a resource from the config file */ |
824 | char *name; /* Resource name */ | |
543cc10b | 825 | struct list_head all_tconn; /* linked on global drbd_tconns */ |
9dc9fbb3 | 826 | struct kref kref; |
8410da8f PR |
827 | struct idr volumes; /* <tconn, vnr> to mdev mapping */ |
828 | enum drbd_conns cstate; /* Only C_STANDALONE to C_WF_REPORT_PARAMS */ | |
8e0af25f PR |
829 | unsigned susp:1; /* IO suspended by user */ |
830 | unsigned susp_nod:1; /* IO suspended because no data */ | |
831 | unsigned susp_fen:1; /* IO suspended because fence peer handler runs */ | |
8410da8f | 832 | struct mutex cstate_mutex; /* Protects graceful disconnects */ |
2111438b | 833 | |
062e879c | 834 | unsigned long flags; |
44ed167d | 835 | struct net_conf *net_conf; /* content protected by rcu */ |
a0095508 | 836 | struct mutex conf_update; /* mutex for ready-copy-update of net_conf and disk_conf */ |
91fd4dad | 837 | wait_queue_head_t ping_wait; /* Woken upon reception of a ping, and a state change */ |
f399002e | 838 | struct res_opts res_opts; |
e42325a5 | 839 | |
089c075d AG |
840 | struct sockaddr_storage my_addr; |
841 | int my_addr_len; | |
842 | struct sockaddr_storage peer_addr; | |
843 | int peer_addr_len; | |
844 | ||
e42325a5 PR |
845 | struct drbd_socket data; /* data/barrier/cstate/parameter packets */ |
846 | struct drbd_socket meta; /* ping/ack (metadata) packets */ | |
31890f4a PR |
847 | int agreed_pro_version; /* actually used protocol version */ |
848 | unsigned long last_received; /* in jiffies, either socket */ | |
849 | unsigned int ko_count; | |
e6b3ea83 | 850 | |
87eeee41 | 851 | spinlock_t req_lock; |
b6dd1a89 LE |
852 | |
853 | struct list_head transfer_log; /* all requests not yet fully processed */ | |
87eeee41 | 854 | |
a0638456 | 855 | struct crypto_hash *cram_hmac_tfm; |
46e1ce41 | 856 | struct crypto_hash *integrity_tfm; /* checksums we compute, updates protected by tconn->data->mutex */ |
036b17ea | 857 | struct crypto_hash *peer_integrity_tfm; /* checksums we verify, only accessed from receiver thread */ |
f399002e LE |
858 | struct crypto_hash *csums_tfm; |
859 | struct crypto_hash *verify_tfm; | |
a0638456 PR |
860 | void *int_dig_in; |
861 | void *int_dig_vv; | |
862 | ||
b6dd1a89 | 863 | /* receiver side */ |
12038a3a PR |
864 | struct drbd_epoch *current_epoch; |
865 | spinlock_t epoch_lock; | |
866 | unsigned int epochs; | |
4b0007c0 | 867 | enum write_ordering_e write_ordering; |
b379c41e | 868 | atomic_t current_tle_nr; /* transfer log epoch number */ |
b6dd1a89 | 869 | unsigned current_tle_writes; /* writes seen within this tl epoch */ |
4b0007c0 | 870 | |
07be15b1 | 871 | unsigned long last_reconnect_jif; |
e6b3ea83 PR |
872 | struct drbd_thread receiver; |
873 | struct drbd_thread worker; | |
874 | struct drbd_thread asender; | |
80822284 | 875 | cpumask_var_t cpu_mask; |
b6dd1a89 LE |
876 | |
877 | /* sender side */ | |
d5b27b01 | 878 | struct drbd_work_queue sender_work; |
b6dd1a89 LE |
879 | |
880 | struct { | |
881 | /* whether this sender thread | |
882 | * has processed a single write yet. */ | |
883 | bool seen_any_write_yet; | |
884 | ||
885 | /* Which barrier number to send with the next P_BARRIER */ | |
886 | int current_epoch_nr; | |
887 | ||
888 | /* how many write requests have been sent | |
889 | * with req->epoch == current_epoch_nr. | |
890 | * If none, no P_BARRIER will be sent. */ | |
891 | unsigned current_epoch_writes; | |
892 | } send; | |
2111438b PR |
893 | }; |
894 | ||
b411b363 | 895 | struct drbd_conf { |
2111438b PR |
896 | struct drbd_tconn *tconn; |
897 | int vnr; /* volume number within the connection */ | |
81fa2e67 | 898 | struct kref kref; |
2111438b | 899 | |
b411b363 PR |
900 | /* things that are stored as / read from meta data on disk */ |
901 | unsigned long flags; | |
902 | ||
903 | /* configured by drbdsetup */ | |
b411b363 PR |
904 | struct drbd_backing_dev *ldev __protected_by(local); |
905 | ||
906 | sector_t p_size; /* partner's disk size */ | |
907 | struct request_queue *rq_queue; | |
908 | struct block_device *this_bdev; | |
909 | struct gendisk *vdisk; | |
910 | ||
07be15b1 | 911 | unsigned long last_reattach_jif; |
b411b363 PR |
912 | struct drbd_work resync_work, |
913 | unplug_work, | |
e9e6f3ec | 914 | go_diskless, |
c4752ef1 PR |
915 | md_sync_work, |
916 | start_resync_work; | |
b411b363 PR |
917 | struct timer_list resync_timer; |
918 | struct timer_list md_sync_timer; | |
370a43e7 | 919 | struct timer_list start_resync_timer; |
7fde2be9 | 920 | struct timer_list request_timer; |
ee15b038 LE |
921 | #ifdef DRBD_DEBUG_MD_SYNC |
922 | struct { | |
923 | unsigned int line; | |
924 | const char* func; | |
925 | } last_md_mark_dirty; | |
926 | #endif | |
b411b363 PR |
927 | |
928 | /* Used after attach while negotiating new disk state. */ | |
929 | union drbd_state new_state_tmp; | |
930 | ||
da9fbc27 | 931 | union drbd_dev_state state; |
b411b363 PR |
932 | wait_queue_head_t misc_wait; |
933 | wait_queue_head_t state_wait; /* upon each state change. */ | |
934 | unsigned int send_cnt; | |
935 | unsigned int recv_cnt; | |
936 | unsigned int read_cnt; | |
937 | unsigned int writ_cnt; | |
938 | unsigned int al_writ_cnt; | |
939 | unsigned int bm_writ_cnt; | |
940 | atomic_t ap_bio_cnt; /* Requests we need to complete */ | |
941 | atomic_t ap_pending_cnt; /* AP data packets on the wire, ack expected */ | |
942 | atomic_t rs_pending_cnt; /* RS request/data packets on the wire */ | |
d942ae44 | 943 | atomic_t unacked_cnt; /* Need to send replies for */ |
b411b363 | 944 | atomic_t local_cnt; /* Waiting for local completion */ |
b2fb6dbe | 945 | |
dac1389c AG |
946 | /* Interval tree of pending local requests */ |
947 | struct rb_root read_requests; | |
de696716 AG |
948 | struct rb_root write_requests; |
949 | ||
4b0715f0 | 950 | /* blocks to resync in this run [unit BM_BLOCK_SIZE] */ |
b411b363 | 951 | unsigned long rs_total; |
4b0715f0 | 952 | /* number of resync blocks that failed in this run */ |
b411b363 PR |
953 | unsigned long rs_failed; |
954 | /* Syncer's start time [unit jiffies] */ | |
955 | unsigned long rs_start; | |
956 | /* cumulated time in PausedSyncX state [unit jiffies] */ | |
957 | unsigned long rs_paused; | |
1d7734a0 LE |
958 | /* skipped because csum was equal [unit BM_BLOCK_SIZE] */ |
959 | unsigned long rs_same_csum; | |
960 | #define DRBD_SYNC_MARKS 8 | |
961 | #define DRBD_SYNC_MARK_STEP (3*HZ) | |
b411b363 | 962 | /* block not up-to-date at mark [unit BM_BLOCK_SIZE] */ |
1d7734a0 | 963 | unsigned long rs_mark_left[DRBD_SYNC_MARKS]; |
b411b363 | 964 | /* marks's time [unit jiffies] */ |
1d7734a0 LE |
965 | unsigned long rs_mark_time[DRBD_SYNC_MARKS]; |
966 | /* current index into rs_mark_{left,time} */ | |
967 | int rs_last_mark; | |
b411b363 PR |
968 | |
969 | /* where does the admin want us to start? (sector) */ | |
970 | sector_t ov_start_sector; | |
58ffa580 | 971 | sector_t ov_stop_sector; |
b411b363 PR |
972 | /* where are we now? (sector) */ |
973 | sector_t ov_position; | |
974 | /* Start sector of out of sync range (to merge printk reporting). */ | |
975 | sector_t ov_last_oos_start; | |
976 | /* size of out-of-sync range in sectors. */ | |
977 | sector_t ov_last_oos_size; | |
978 | unsigned long ov_left; /* in bits */ | |
b411b363 | 979 | |
b411b363 PR |
980 | struct drbd_bitmap *bitmap; |
981 | unsigned long bm_resync_fo; /* bit offset for drbd_bm_find_next */ | |
982 | ||
983 | /* Used to track operations of resync... */ | |
984 | struct lru_cache *resync; | |
985 | /* Number of locked elements in resync LRU */ | |
986 | unsigned int resync_locked; | |
987 | /* resync extent number waiting for application requests */ | |
988 | unsigned int resync_wenr; | |
989 | ||
990 | int open_cnt; | |
991 | u64 *p_uuid; | |
4b0007c0 | 992 | |
85719573 PR |
993 | struct list_head active_ee; /* IO in progress (P_DATA gets written to disk) */ |
994 | struct list_head sync_ee; /* IO in progress (P_RS_DATA_REPLY gets written to disk) */ | |
18b75d75 AG |
995 | struct list_head done_ee; /* need to send P_WRITE_ACK */ |
996 | struct list_head read_ee; /* [RS]P_DATA_REQUEST being read */ | |
b411b363 | 997 | struct list_head net_ee; /* zero-copy network send in progress */ |
b411b363 | 998 | |
b411b363 | 999 | int next_barrier_nr; |
b411b363 | 1000 | struct list_head resync_reads; |
435f0740 LE |
1001 | atomic_t pp_in_use; /* allocated from page pool */ |
1002 | atomic_t pp_in_use_by_net; /* sendpage()d, still referenced by tcp */ | |
b411b363 PR |
1003 | wait_queue_head_t ee_wait; |
1004 | struct page *md_io_page; /* one page buffer for md_io */ | |
cdfda633 PR |
1005 | struct drbd_md_io md_io; |
1006 | atomic_t md_io_in_use; /* protects the md_io, md_io_page and md_io_tmpp */ | |
b411b363 PR |
1007 | spinlock_t al_lock; |
1008 | wait_queue_head_t al_wait; | |
1009 | struct lru_cache *act_log; /* activity log */ | |
1010 | unsigned int al_tr_number; | |
1011 | int al_tr_cycle; | |
1012 | int al_tr_pos; /* position of the next transaction in the journal */ | |
b411b363 PR |
1013 | wait_queue_head_t seq_wait; |
1014 | atomic_t packet_seq; | |
1015 | unsigned int peer_seq; | |
1016 | spinlock_t peer_seq_lock; | |
1017 | unsigned int minor; | |
1018 | unsigned long comm_bm_set; /* communicated number of set bits. */ | |
b411b363 PR |
1019 | struct bm_io_work bm_io_work; |
1020 | u64 ed_uuid; /* UUID of the exposed data */ | |
8410da8f PR |
1021 | struct mutex own_state_mutex; |
1022 | struct mutex *state_mutex; /* either own_state_mutex or mdev->tconn->cstate_mutex */ | |
b411b363 | 1023 | char congestion_reason; /* Why we where congested... */ |
1d7734a0 LE |
1024 | atomic_t rs_sect_in; /* for incoming resync data rate, SyncTarget */ |
1025 | atomic_t rs_sect_ev; /* for submitted resync data rate, both */ | |
1026 | int rs_last_sect_ev; /* counter to compare with */ | |
1027 | int rs_last_events; /* counter of read or write "events" (unit sectors) | |
1028 | * on the lower level device when we last looked. */ | |
1029 | int c_sync_rate; /* current resync rate after syncer throttle magic */ | |
813472ce | 1030 | struct fifo_buffer *rs_plan_s; /* correction values of resync planer (RCU, tconn->conn_update) */ |
778f271d | 1031 | int rs_in_flight; /* resync sectors in flight (to proxy, in proxy and from proxy) */ |
759fbdfb | 1032 | atomic_t ap_in_flight; /* App sectors in flight (waiting for ack) */ |
99432fcc PR |
1033 | int peer_max_bio_size; |
1034 | int local_max_bio_size; | |
b411b363 PR |
1035 | }; |
1036 | ||
1037 | static inline struct drbd_conf *minor_to_mdev(unsigned int minor) | |
1038 | { | |
81a5d60e | 1039 | return (struct drbd_conf *)idr_find(&minors, minor); |
b411b363 PR |
1040 | } |
1041 | ||
1042 | static inline unsigned int mdev_to_minor(struct drbd_conf *mdev) | |
1043 | { | |
1044 | return mdev->minor; | |
1045 | } | |
1046 | ||
eefc2f7d PR |
1047 | static inline struct drbd_conf *vnr_to_mdev(struct drbd_tconn *tconn, int vnr) |
1048 | { | |
1049 | return (struct drbd_conf *)idr_find(&tconn->volumes, vnr); | |
1050 | } | |
1051 | ||
b411b363 PR |
1052 | /* |
1053 | * function declarations | |
1054 | *************************/ | |
1055 | ||
1056 | /* drbd_main.c */ | |
1057 | ||
e89b591c PR |
1058 | enum dds_flags { |
1059 | DDSF_FORCED = 1, | |
1060 | DDSF_NO_RESYNC = 2, /* Do not run a resync for the new space */ | |
1061 | }; | |
1062 | ||
b411b363 | 1063 | extern void drbd_init_set_defaults(struct drbd_conf *mdev); |
b411b363 PR |
1064 | extern int drbd_thread_start(struct drbd_thread *thi); |
1065 | extern void _drbd_thread_stop(struct drbd_thread *thi, int restart, int wait); | |
392c8801 | 1066 | extern char *drbd_task_to_thread_name(struct drbd_tconn *tconn, struct task_struct *task); |
b411b363 | 1067 | #ifdef CONFIG_SMP |
80822284 PR |
1068 | extern void drbd_thread_current_set_cpu(struct drbd_thread *thi); |
1069 | extern void drbd_calc_cpu_mask(struct drbd_tconn *tconn); | |
b411b363 | 1070 | #else |
80822284 | 1071 | #define drbd_thread_current_set_cpu(A) ({}) |
b411b363 PR |
1072 | #define drbd_calc_cpu_mask(A) ({}) |
1073 | #endif | |
2f5cdd0b | 1074 | extern void tl_release(struct drbd_tconn *, unsigned int barrier_nr, |
b411b363 | 1075 | unsigned int set_size); |
2f5cdd0b | 1076 | extern void tl_clear(struct drbd_tconn *); |
360cc740 | 1077 | extern void drbd_free_sock(struct drbd_tconn *tconn); |
bedbd2a5 PR |
1078 | extern int drbd_send(struct drbd_tconn *tconn, struct socket *sock, |
1079 | void *buf, size_t size, unsigned msg_flags); | |
fb708e40 AG |
1080 | extern int drbd_send_all(struct drbd_tconn *, struct socket *, void *, size_t, |
1081 | unsigned); | |
1082 | ||
d659f2aa | 1083 | extern int __drbd_send_protocol(struct drbd_tconn *tconn, enum drbd_packet cmd); |
dc8228d1 | 1084 | extern int drbd_send_protocol(struct drbd_tconn *tconn); |
b411b363 PR |
1085 | extern int drbd_send_uuids(struct drbd_conf *mdev); |
1086 | extern int drbd_send_uuids_skip_initial_sync(struct drbd_conf *mdev); | |
9c1b7f72 | 1087 | extern void drbd_gen_and_send_sync_uuid(struct drbd_conf *mdev); |
e89b591c | 1088 | extern int drbd_send_sizes(struct drbd_conf *mdev, int trigger_reply, enum dds_flags flags); |
43de7c85 PR |
1089 | extern int drbd_send_state(struct drbd_conf *mdev, union drbd_state s); |
1090 | extern int drbd_send_current_state(struct drbd_conf *mdev); | |
f399002e | 1091 | extern int drbd_send_sync_param(struct drbd_conf *mdev); |
9ed57dcb | 1092 | extern void drbd_send_b_ack(struct drbd_tconn *tconn, u32 barrier_nr, |
d4e67d7c | 1093 | u32 set_size); |
f6ffca9f AG |
1094 | extern int drbd_send_ack(struct drbd_conf *, enum drbd_packet, |
1095 | struct drbd_peer_request *); | |
a9a9994d AG |
1096 | extern void drbd_send_ack_rp(struct drbd_conf *mdev, enum drbd_packet cmd, |
1097 | struct p_block_req *rp); | |
1098 | extern void drbd_send_ack_dp(struct drbd_conf *mdev, enum drbd_packet cmd, | |
1099 | struct p_data *dp, int data_size); | |
d8763023 | 1100 | extern int drbd_send_ack_ex(struct drbd_conf *mdev, enum drbd_packet cmd, |
b411b363 | 1101 | sector_t sector, int blksize, u64 block_id); |
8f7bed77 | 1102 | extern int drbd_send_out_of_sync(struct drbd_conf *, struct drbd_request *); |
f6ffca9f AG |
1103 | extern int drbd_send_block(struct drbd_conf *, enum drbd_packet, |
1104 | struct drbd_peer_request *); | |
b411b363 | 1105 | extern int drbd_send_dblock(struct drbd_conf *mdev, struct drbd_request *req); |
b411b363 PR |
1106 | extern int drbd_send_drequest(struct drbd_conf *mdev, int cmd, |
1107 | sector_t sector, int size, u64 block_id); | |
d8763023 AG |
1108 | extern int drbd_send_drequest_csum(struct drbd_conf *mdev, sector_t sector, |
1109 | int size, void *digest, int digest_size, | |
1110 | enum drbd_packet cmd); | |
b411b363 PR |
1111 | extern int drbd_send_ov_request(struct drbd_conf *mdev,sector_t sector,int size); |
1112 | ||
1113 | extern int drbd_send_bitmap(struct drbd_conf *mdev); | |
2f4e7abe | 1114 | extern void drbd_send_sr_reply(struct drbd_conf *mdev, enum drbd_state_rv retcode); |
9f5bdc33 | 1115 | extern void conn_send_sr_reply(struct drbd_tconn *tconn, enum drbd_state_rv retcode); |
b411b363 PR |
1116 | extern void drbd_free_bc(struct drbd_backing_dev *ldev); |
1117 | extern void drbd_mdev_cleanup(struct drbd_conf *mdev); | |
62b0da3a | 1118 | void drbd_print_uuids(struct drbd_conf *mdev, const char *text); |
b411b363 | 1119 | |
19fffd7b | 1120 | extern void conn_md_sync(struct drbd_tconn *tconn); |
b411b363 PR |
1121 | extern void drbd_md_sync(struct drbd_conf *mdev); |
1122 | extern int drbd_md_read(struct drbd_conf *mdev, struct drbd_backing_dev *bdev); | |
b411b363 PR |
1123 | extern void drbd_uuid_set(struct drbd_conf *mdev, int idx, u64 val) __must_hold(local); |
1124 | extern void _drbd_uuid_set(struct drbd_conf *mdev, int idx, u64 val) __must_hold(local); | |
1125 | extern void drbd_uuid_new_current(struct drbd_conf *mdev) __must_hold(local); | |
b411b363 | 1126 | extern void drbd_uuid_set_bm(struct drbd_conf *mdev, u64 val) __must_hold(local); |
39a1aa7f PR |
1127 | extern void drbd_uuid_move_history(struct drbd_conf *mdev) __must_hold(local); |
1128 | extern void __drbd_uuid_set(struct drbd_conf *mdev, int idx, u64 val) __must_hold(local); | |
b411b363 PR |
1129 | extern void drbd_md_set_flag(struct drbd_conf *mdev, int flags) __must_hold(local); |
1130 | extern void drbd_md_clear_flag(struct drbd_conf *mdev, int flags)__must_hold(local); | |
1131 | extern int drbd_md_test_flag(struct drbd_backing_dev *, int); | |
ee15b038 | 1132 | #ifndef DRBD_DEBUG_MD_SYNC |
b411b363 | 1133 | extern void drbd_md_mark_dirty(struct drbd_conf *mdev); |
ee15b038 LE |
1134 | #else |
1135 | #define drbd_md_mark_dirty(m) drbd_md_mark_dirty_(m, __LINE__ , __func__ ) | |
1136 | extern void drbd_md_mark_dirty_(struct drbd_conf *mdev, | |
1137 | unsigned int line, const char *func); | |
1138 | #endif | |
b411b363 PR |
1139 | extern void drbd_queue_bitmap_io(struct drbd_conf *mdev, |
1140 | int (*io_fn)(struct drbd_conf *), | |
1141 | void (*done)(struct drbd_conf *, int), | |
20ceb2b2 LE |
1142 | char *why, enum bm_flag flags); |
1143 | extern int drbd_bitmap_io(struct drbd_conf *mdev, | |
1144 | int (*io_fn)(struct drbd_conf *), | |
1145 | char *why, enum bm_flag flags); | |
edc9f5eb LE |
1146 | extern int drbd_bitmap_io_from_worker(struct drbd_conf *mdev, |
1147 | int (*io_fn)(struct drbd_conf *), | |
1148 | char *why, enum bm_flag flags); | |
b411b363 PR |
1149 | extern int drbd_bmio_set_n_write(struct drbd_conf *mdev); |
1150 | extern int drbd_bmio_clear_n_write(struct drbd_conf *mdev); | |
e9e6f3ec | 1151 | extern void drbd_go_diskless(struct drbd_conf *mdev); |
82f59cc6 | 1152 | extern void drbd_ldev_destroy(struct drbd_conf *mdev); |
b411b363 | 1153 | |
b411b363 PR |
1154 | /* Meta data layout |
1155 | We reserve a 128MB Block (4k aligned) | |
1156 | * either at the end of the backing device | |
3ad2f3fb | 1157 | * or on a separate meta data device. */ |
b411b363 | 1158 | |
b411b363 | 1159 | /* The following numbers are sectors */ |
7ad651b5 LE |
1160 | /* Allows up to about 3.8TB, so if you want more, |
1161 | * you need to use the "flexible" meta data format. */ | |
1162 | #define MD_RESERVED_SECT (128LU << 11) /* 128 MB, unit sectors */ | |
1163 | #define MD_AL_OFFSET 8 /* 8 Sectors after start of meta area */ | |
1164 | #define MD_AL_SECTORS 64 /* = 32 kB on disk activity log ring buffer */ | |
1165 | #define MD_BM_OFFSET (MD_AL_OFFSET + MD_AL_SECTORS) | |
1166 | ||
1167 | /* we do all meta data IO in 4k blocks */ | |
1168 | #define MD_BLOCK_SHIFT 12 | |
1169 | #define MD_BLOCK_SIZE (1<<MD_BLOCK_SHIFT) | |
1170 | ||
1171 | /* One activity log extent represents 4M of storage */ | |
1172 | #define AL_EXTENT_SHIFT 22 | |
b411b363 PR |
1173 | #define AL_EXTENT_SIZE (1<<AL_EXTENT_SHIFT) |
1174 | ||
7ad651b5 LE |
1175 | /* We could make these currently hardcoded constants configurable |
1176 | * variables at create-md time (or even re-configurable at runtime?). | |
1177 | * Which will require some more changes to the DRBD "super block" | |
1178 | * and attach code. | |
1179 | * | |
1180 | * updates per transaction: | |
1181 | * This many changes to the active set can be logged with one transaction. | |
1182 | * This number is arbitrary. | |
1183 | * context per transaction: | |
1184 | * This many context extent numbers are logged with each transaction. | |
1185 | * This number is resulting from the transaction block size (4k), the layout | |
1186 | * of the transaction header, and the number of updates per transaction. | |
1187 | * See drbd_actlog.c:struct al_transaction_on_disk | |
1188 | * */ | |
1189 | #define AL_UPDATES_PER_TRANSACTION 64 // arbitrary | |
1190 | #define AL_CONTEXT_PER_TRANSACTION 919 // (4096 - 36 - 6*64)/4 | |
1191 | ||
b411b363 PR |
1192 | #if BITS_PER_LONG == 32 |
1193 | #define LN2_BPL 5 | |
1194 | #define cpu_to_lel(A) cpu_to_le32(A) | |
1195 | #define lel_to_cpu(A) le32_to_cpu(A) | |
1196 | #elif BITS_PER_LONG == 64 | |
1197 | #define LN2_BPL 6 | |
1198 | #define cpu_to_lel(A) cpu_to_le64(A) | |
1199 | #define lel_to_cpu(A) le64_to_cpu(A) | |
1200 | #else | |
1201 | #error "LN2 of BITS_PER_LONG unknown!" | |
1202 | #endif | |
1203 | ||
1204 | /* resync bitmap */ | |
1205 | /* 16MB sized 'bitmap extent' to track syncer usage */ | |
1206 | struct bm_extent { | |
1207 | int rs_left; /* number of bits set (out of sync) in this extent. */ | |
1208 | int rs_failed; /* number of failed resync requests in this extent. */ | |
1209 | unsigned long flags; | |
1210 | struct lc_element lce; | |
1211 | }; | |
1212 | ||
1213 | #define BME_NO_WRITES 0 /* bm_extent.flags: no more requests on this one! */ | |
1214 | #define BME_LOCKED 1 /* bm_extent.flags: syncer active on this one. */ | |
e3555d85 | 1215 | #define BME_PRIORITY 2 /* finish resync IO on this extent ASAP! App IO waiting! */ |
b411b363 PR |
1216 | |
1217 | /* drbd_bitmap.c */ | |
1218 | /* | |
1219 | * We need to store one bit for a block. | |
1220 | * Example: 1GB disk @ 4096 byte blocks ==> we need 32 KB bitmap. | |
1221 | * Bit 0 ==> local node thinks this block is binary identical on both nodes | |
1222 | * Bit 1 ==> local node thinks this block needs to be synced. | |
1223 | */ | |
1224 | ||
8e26f9cc PR |
1225 | #define SLEEP_TIME (HZ/10) |
1226 | ||
45dfffeb LE |
1227 | /* We do bitmap IO in units of 4k blocks. |
1228 | * We also still have a hardcoded 4k per bit relation. */ | |
1229 | #define BM_BLOCK_SHIFT 12 /* 4k per bit */ | |
b411b363 | 1230 | #define BM_BLOCK_SIZE (1<<BM_BLOCK_SHIFT) |
45dfffeb LE |
1231 | /* mostly arbitrarily set the represented size of one bitmap extent, |
1232 | * aka resync extent, to 16 MiB (which is also 512 Byte worth of bitmap | |
1233 | * at 4k per bit resolution) */ | |
1234 | #define BM_EXT_SHIFT 24 /* 16 MiB per resync extent */ | |
b411b363 PR |
1235 | #define BM_EXT_SIZE (1<<BM_EXT_SHIFT) |
1236 | ||
1237 | #if (BM_EXT_SHIFT != 24) || (BM_BLOCK_SHIFT != 12) | |
1238 | #error "HAVE YOU FIXED drbdmeta AS WELL??" | |
1239 | #endif | |
1240 | ||
1241 | /* thus many _storage_ sectors are described by one bit */ | |
1242 | #define BM_SECT_TO_BIT(x) ((x)>>(BM_BLOCK_SHIFT-9)) | |
1243 | #define BM_BIT_TO_SECT(x) ((sector_t)(x)<<(BM_BLOCK_SHIFT-9)) | |
1244 | #define BM_SECT_PER_BIT BM_BIT_TO_SECT(1) | |
1245 | ||
1246 | /* bit to represented kilo byte conversion */ | |
1247 | #define Bit2KB(bits) ((bits)<<(BM_BLOCK_SHIFT-10)) | |
1248 | ||
1249 | /* in which _bitmap_ extent (resp. sector) the bit for a certain | |
1250 | * _storage_ sector is located in */ | |
1251 | #define BM_SECT_TO_EXT(x) ((x)>>(BM_EXT_SHIFT-9)) | |
1252 | ||
1253 | /* how much _storage_ sectors we have per bitmap sector */ | |
1254 | #define BM_EXT_TO_SECT(x) ((sector_t)(x) << (BM_EXT_SHIFT-9)) | |
1255 | #define BM_SECT_PER_EXT BM_EXT_TO_SECT(1) | |
1256 | ||
1257 | /* in one sector of the bitmap, we have this many activity_log extents. */ | |
1258 | #define AL_EXT_PER_BM_SECT (1 << (BM_EXT_SHIFT - AL_EXTENT_SHIFT)) | |
1259 | #define BM_WORDS_PER_AL_EXT (1 << (AL_EXTENT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) | |
1260 | ||
1261 | #define BM_BLOCKS_PER_BM_EXT_B (BM_EXT_SHIFT - BM_BLOCK_SHIFT) | |
1262 | #define BM_BLOCKS_PER_BM_EXT_MASK ((1<<BM_BLOCKS_PER_BM_EXT_B) - 1) | |
1263 | ||
1264 | /* the extent in "PER_EXTENT" below is an activity log extent | |
1265 | * we need that many (long words/bytes) to store the bitmap | |
1266 | * of one AL_EXTENT_SIZE chunk of storage. | |
1267 | * we can store the bitmap for that many AL_EXTENTS within | |
1268 | * one sector of the _on_disk_ bitmap: | |
1269 | * bit 0 bit 37 bit 38 bit (512*8)-1 | |
1270 | * ...|........|........|.. // ..|........| | |
1271 | * sect. 0 `296 `304 ^(512*8*8)-1 | |
1272 | * | |
1273 | #define BM_WORDS_PER_EXT ( (AL_EXT_SIZE/BM_BLOCK_SIZE) / BITS_PER_LONG ) | |
1274 | #define BM_BYTES_PER_EXT ( (AL_EXT_SIZE/BM_BLOCK_SIZE) / 8 ) // 128 | |
1275 | #define BM_EXT_PER_SECT ( 512 / BM_BYTES_PER_EXTENT ) // 4 | |
1276 | */ | |
1277 | ||
1278 | #define DRBD_MAX_SECTORS_32 (0xffffffffLU) | |
1279 | #define DRBD_MAX_SECTORS_BM \ | |
1280 | ((MD_RESERVED_SECT - MD_BM_OFFSET) * (1LL<<(BM_EXT_SHIFT-9))) | |
1281 | #if DRBD_MAX_SECTORS_BM < DRBD_MAX_SECTORS_32 | |
1282 | #define DRBD_MAX_SECTORS DRBD_MAX_SECTORS_BM | |
1283 | #define DRBD_MAX_SECTORS_FLEX DRBD_MAX_SECTORS_BM | |
36bfc7e2 | 1284 | #elif !defined(CONFIG_LBDAF) && BITS_PER_LONG == 32 |
b411b363 PR |
1285 | #define DRBD_MAX_SECTORS DRBD_MAX_SECTORS_32 |
1286 | #define DRBD_MAX_SECTORS_FLEX DRBD_MAX_SECTORS_32 | |
1287 | #else | |
1288 | #define DRBD_MAX_SECTORS DRBD_MAX_SECTORS_BM | |
1289 | /* 16 TB in units of sectors */ | |
1290 | #if BITS_PER_LONG == 32 | |
1291 | /* adjust by one page worth of bitmap, | |
1292 | * so we won't wrap around in drbd_bm_find_next_bit. | |
1293 | * you should use 64bit OS for that much storage, anyways. */ | |
1294 | #define DRBD_MAX_SECTORS_FLEX BM_BIT_TO_SECT(0xffff7fff) | |
1295 | #else | |
4b0715f0 LE |
1296 | /* we allow up to 1 PiB now on 64bit architecture with "flexible" meta data */ |
1297 | #define DRBD_MAX_SECTORS_FLEX (1UL << 51) | |
1298 | /* corresponds to (1UL << 38) bits right now. */ | |
b411b363 PR |
1299 | #endif |
1300 | #endif | |
1301 | ||
23361cf3 LE |
1302 | /* BIO_MAX_SIZE is 256 * PAGE_CACHE_SIZE, |
1303 | * so for typical PAGE_CACHE_SIZE of 4k, that is (1<<20) Byte. | |
1304 | * Since we may live in a mixed-platform cluster, | |
1305 | * we limit us to a platform agnostic constant here for now. | |
1306 | * A followup commit may allow even bigger BIO sizes, | |
1307 | * once we thought that through. */ | |
1308 | #define DRBD_MAX_BIO_SIZE (1 << 20) | |
1309 | #if DRBD_MAX_BIO_SIZE > BIO_MAX_SIZE | |
1310 | #error Architecture not supported: DRBD_MAX_BIO_SIZE > BIO_MAX_SIZE | |
1311 | #endif | |
99432fcc | 1312 | #define DRBD_MAX_BIO_SIZE_SAFE (1 << 12) /* Works always = 4k */ |
b411b363 | 1313 | |
2ffca4f3 PR |
1314 | #define DRBD_MAX_SIZE_H80_PACKET (1 << 15) /* Header 80 only allows packets up to 32KiB data */ |
1315 | #define DRBD_MAX_BIO_SIZE_P95 (1 << 17) /* Protocol 95 to 99 allows bios up to 128KiB */ | |
b411b363 PR |
1316 | |
1317 | extern int drbd_bm_init(struct drbd_conf *mdev); | |
02d9a94b | 1318 | extern int drbd_bm_resize(struct drbd_conf *mdev, sector_t sectors, int set_new_bits); |
b411b363 PR |
1319 | extern void drbd_bm_cleanup(struct drbd_conf *mdev); |
1320 | extern void drbd_bm_set_all(struct drbd_conf *mdev); | |
1321 | extern void drbd_bm_clear_all(struct drbd_conf *mdev); | |
4b0715f0 | 1322 | /* set/clear/test only a few bits at a time */ |
b411b363 PR |
1323 | extern int drbd_bm_set_bits( |
1324 | struct drbd_conf *mdev, unsigned long s, unsigned long e); | |
1325 | extern int drbd_bm_clear_bits( | |
1326 | struct drbd_conf *mdev, unsigned long s, unsigned long e); | |
4b0715f0 LE |
1327 | extern int drbd_bm_count_bits( |
1328 | struct drbd_conf *mdev, const unsigned long s, const unsigned long e); | |
1329 | /* bm_set_bits variant for use while holding drbd_bm_lock, | |
1330 | * may process the whole bitmap in one go */ | |
b411b363 PR |
1331 | extern void _drbd_bm_set_bits(struct drbd_conf *mdev, |
1332 | const unsigned long s, const unsigned long e); | |
1333 | extern int drbd_bm_test_bit(struct drbd_conf *mdev, unsigned long bitnr); | |
1334 | extern int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr); | |
19f843aa | 1335 | extern int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local); |
b411b363 | 1336 | extern int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local); |
45dfffeb | 1337 | extern void drbd_bm_mark_for_writeout(struct drbd_conf *mdev, int page_nr); |
b411b363 | 1338 | extern int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local); |
45dfffeb | 1339 | extern int drbd_bm_write_hinted(struct drbd_conf *mdev) __must_hold(local); |
fef45d29 | 1340 | extern int drbd_bm_write_all(struct drbd_conf *mdev) __must_hold(local); |
a220d291 | 1341 | extern int drbd_bm_write_copy_pages(struct drbd_conf *mdev) __must_hold(local); |
b411b363 PR |
1342 | extern size_t drbd_bm_words(struct drbd_conf *mdev); |
1343 | extern unsigned long drbd_bm_bits(struct drbd_conf *mdev); | |
1344 | extern sector_t drbd_bm_capacity(struct drbd_conf *mdev); | |
4b0715f0 LE |
1345 | |
1346 | #define DRBD_END_OF_BITMAP (~(unsigned long)0) | |
b411b363 PR |
1347 | extern unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo); |
1348 | /* bm_find_next variants for use while you hold drbd_bm_lock() */ | |
1349 | extern unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo); | |
1350 | extern unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo); | |
0778286a | 1351 | extern unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev); |
b411b363 PR |
1352 | extern unsigned long drbd_bm_total_weight(struct drbd_conf *mdev); |
1353 | extern int drbd_bm_rs_done(struct drbd_conf *mdev); | |
1354 | /* for receive_bitmap */ | |
1355 | extern void drbd_bm_merge_lel(struct drbd_conf *mdev, size_t offset, | |
1356 | size_t number, unsigned long *buffer); | |
19f843aa | 1357 | /* for _drbd_send_bitmap */ |
b411b363 PR |
1358 | extern void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset, |
1359 | size_t number, unsigned long *buffer); | |
1360 | ||
20ceb2b2 | 1361 | extern void drbd_bm_lock(struct drbd_conf *mdev, char *why, enum bm_flag flags); |
b411b363 | 1362 | extern void drbd_bm_unlock(struct drbd_conf *mdev); |
b411b363 PR |
1363 | /* drbd_main.c */ |
1364 | ||
1365 | extern struct kmem_cache *drbd_request_cache; | |
6c852bec | 1366 | extern struct kmem_cache *drbd_ee_cache; /* peer requests */ |
b411b363 PR |
1367 | extern struct kmem_cache *drbd_bm_ext_cache; /* bitmap extents */ |
1368 | extern struct kmem_cache *drbd_al_ext_cache; /* activity log extents */ | |
1369 | extern mempool_t *drbd_request_mempool; | |
1370 | extern mempool_t *drbd_ee_mempool; | |
1371 | ||
35abf594 LE |
1372 | /* drbd's page pool, used to buffer data received from the peer, |
1373 | * or data requested by the peer. | |
1374 | * | |
1375 | * This does not have an emergency reserve. | |
1376 | * | |
1377 | * When allocating from this pool, it first takes pages from the pool. | |
1378 | * Only if the pool is depleted will try to allocate from the system. | |
1379 | * | |
1380 | * The assumption is that pages taken from this pool will be processed, | |
1381 | * and given back, "quickly", and then can be recycled, so we can avoid | |
1382 | * frequent calls to alloc_page(), and still will be able to make progress even | |
1383 | * under memory pressure. | |
1384 | */ | |
1385 | extern struct page *drbd_pp_pool; | |
b411b363 PR |
1386 | extern spinlock_t drbd_pp_lock; |
1387 | extern int drbd_pp_vacant; | |
1388 | extern wait_queue_head_t drbd_pp_wait; | |
1389 | ||
35abf594 LE |
1390 | /* We also need a standard (emergency-reserve backed) page pool |
1391 | * for meta data IO (activity log, bitmap). | |
1392 | * We can keep it global, as long as it is used as "N pages at a time". | |
1393 | * 128 should be plenty, currently we probably can get away with as few as 1. | |
1394 | */ | |
1395 | #define DRBD_MIN_POOL_PAGES 128 | |
1396 | extern mempool_t *drbd_md_io_page_pool; | |
1397 | ||
da4a75d2 LE |
1398 | /* We also need to make sure we get a bio |
1399 | * when we need it for housekeeping purposes */ | |
1400 | extern struct bio_set *drbd_md_io_bio_set; | |
1401 | /* to allocate from that set */ | |
1402 | extern struct bio *bio_alloc_drbd(gfp_t gfp_mask); | |
1403 | ||
b411b363 PR |
1404 | extern rwlock_t global_state_lock; |
1405 | ||
80883197 | 1406 | extern int conn_lowest_minor(struct drbd_tconn *tconn); |
774b3055 | 1407 | enum drbd_ret_code conn_new_minor(struct drbd_tconn *tconn, unsigned int minor, int vnr); |
81fa2e67 | 1408 | extern void drbd_minor_destroy(struct kref *kref); |
b411b363 | 1409 | |
afbbfa88 AG |
1410 | extern int set_resource_options(struct drbd_tconn *tconn, struct res_opts *res_opts); |
1411 | extern struct drbd_tconn *conn_create(const char *name, struct res_opts *res_opts); | |
9dc9fbb3 | 1412 | extern void conn_destroy(struct kref *kref); |
0ace9dfa | 1413 | struct drbd_tconn *conn_get_by_name(const char *name); |
089c075d AG |
1414 | extern struct drbd_tconn *conn_get_by_addrs(void *my_addr, int my_addr_len, |
1415 | void *peer_addr, int peer_addr_len); | |
91fd4dad | 1416 | extern void conn_free_crypto(struct drbd_tconn *tconn); |
2111438b | 1417 | |
b411b363 PR |
1418 | extern int proc_details; |
1419 | ||
1420 | /* drbd_req */ | |
5df69ece | 1421 | extern void __drbd_make_request(struct drbd_conf *, struct bio *, unsigned long); |
2f58dcfc | 1422 | extern int drbd_make_request(struct request_queue *q, struct bio *bio); |
b411b363 PR |
1423 | extern int drbd_read_remote(struct drbd_conf *mdev, struct drbd_request *req); |
1424 | extern int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec); | |
1425 | extern int is_valid_ar_handle(struct drbd_request *, sector_t); | |
1426 | ||
1427 | ||
1428 | /* drbd_nl.c */ | |
8432b314 | 1429 | extern int drbd_msg_put_info(const char *info); |
b411b363 PR |
1430 | extern void drbd_suspend_io(struct drbd_conf *mdev); |
1431 | extern void drbd_resume_io(struct drbd_conf *mdev); | |
1432 | extern char *ppsize(char *buf, unsigned long long size); | |
ef5e44a6 | 1433 | extern sector_t drbd_new_dev_size(struct drbd_conf *, struct drbd_backing_dev *, sector_t, int); |
b411b363 | 1434 | enum determine_dev_size { dev_size_error = -1, unchanged = 0, shrunk = 1, grew = 2 }; |
24c4830c | 1435 | extern enum determine_dev_size drbd_determine_dev_size(struct drbd_conf *, enum dds_flags) __must_hold(local); |
b411b363 | 1436 | extern void resync_after_online_grow(struct drbd_conf *); |
99432fcc | 1437 | extern void drbd_reconsider_max_bio_size(struct drbd_conf *mdev); |
bf885f8a AG |
1438 | extern enum drbd_state_rv drbd_set_role(struct drbd_conf *mdev, |
1439 | enum drbd_role new_role, | |
1440 | int force); | |
cb703454 PR |
1441 | extern bool conn_try_outdate_peer(struct drbd_tconn *tconn); |
1442 | extern void conn_try_outdate_peer_async(struct drbd_tconn *tconn); | |
b411b363 PR |
1443 | extern int drbd_khelper(struct drbd_conf *mdev, char *cmd); |
1444 | ||
1445 | /* drbd_worker.c */ | |
1446 | extern int drbd_worker(struct drbd_thread *thi); | |
95f8efd0 AG |
1447 | enum drbd_ret_code drbd_resync_after_valid(struct drbd_conf *mdev, int o_minor); |
1448 | void drbd_resync_after_changed(struct drbd_conf *mdev); | |
b411b363 PR |
1449 | extern void drbd_start_resync(struct drbd_conf *mdev, enum drbd_conns side); |
1450 | extern void resume_next_sg(struct drbd_conf *mdev); | |
1451 | extern void suspend_other_sg(struct drbd_conf *mdev); | |
1452 | extern int drbd_resync_finished(struct drbd_conf *mdev); | |
1453 | /* maybe rather drbd_main.c ? */ | |
cdfda633 PR |
1454 | extern void *drbd_md_get_buffer(struct drbd_conf *mdev); |
1455 | extern void drbd_md_put_buffer(struct drbd_conf *mdev); | |
b411b363 PR |
1456 | extern int drbd_md_sync_page_io(struct drbd_conf *mdev, |
1457 | struct drbd_backing_dev *bdev, sector_t sector, int rw); | |
8f7bed77 | 1458 | extern void drbd_ov_out_of_sync_found(struct drbd_conf *, sector_t, int); |
e34b677d LE |
1459 | extern void wait_until_done_or_force_detached(struct drbd_conf *mdev, |
1460 | struct drbd_backing_dev *bdev, unsigned int *done); | |
9bd28d3c | 1461 | extern void drbd_rs_controller_reset(struct drbd_conf *mdev); |
b411b363 | 1462 | |
8f7bed77 | 1463 | static inline void ov_out_of_sync_print(struct drbd_conf *mdev) |
b411b363 PR |
1464 | { |
1465 | if (mdev->ov_last_oos_size) { | |
1466 | dev_err(DEV, "Out of sync: start=%llu, size=%lu (sectors)\n", | |
1467 | (unsigned long long)mdev->ov_last_oos_start, | |
1468 | (unsigned long)mdev->ov_last_oos_size); | |
1469 | } | |
1470 | mdev->ov_last_oos_size=0; | |
1471 | } | |
1472 | ||
1473 | ||
45bb912b | 1474 | extern void drbd_csum_bio(struct drbd_conf *, struct crypto_hash *, struct bio *, void *); |
f6ffca9f AG |
1475 | extern void drbd_csum_ee(struct drbd_conf *, struct crypto_hash *, |
1476 | struct drbd_peer_request *, void *); | |
b411b363 | 1477 | /* worker callbacks */ |
99920dc5 AG |
1478 | extern int w_e_end_data_req(struct drbd_work *, int); |
1479 | extern int w_e_end_rsdata_req(struct drbd_work *, int); | |
1480 | extern int w_e_end_csum_rs_req(struct drbd_work *, int); | |
1481 | extern int w_e_end_ov_reply(struct drbd_work *, int); | |
1482 | extern int w_e_end_ov_req(struct drbd_work *, int); | |
1483 | extern int w_ov_finished(struct drbd_work *, int); | |
1484 | extern int w_resync_timer(struct drbd_work *, int); | |
1485 | extern int w_send_write_hint(struct drbd_work *, int); | |
1486 | extern int w_make_resync_request(struct drbd_work *, int); | |
1487 | extern int w_send_dblock(struct drbd_work *, int); | |
99920dc5 AG |
1488 | extern int w_send_read_req(struct drbd_work *, int); |
1489 | extern int w_prev_work_done(struct drbd_work *, int); | |
1490 | extern int w_e_reissue(struct drbd_work *, int); | |
1491 | extern int w_restart_disk_io(struct drbd_work *, int); | |
8f7bed77 | 1492 | extern int w_send_out_of_sync(struct drbd_work *, int); |
99920dc5 | 1493 | extern int w_start_resync(struct drbd_work *, int); |
b411b363 PR |
1494 | |
1495 | extern void resync_timer_fn(unsigned long data); | |
370a43e7 | 1496 | extern void start_resync_timer_fn(unsigned long data); |
b411b363 PR |
1497 | |
1498 | /* drbd_receiver.c */ | |
e3555d85 | 1499 | extern int drbd_rs_should_slow_down(struct drbd_conf *mdev, sector_t sector); |
fbe29dec AG |
1500 | extern int drbd_submit_peer_request(struct drbd_conf *, |
1501 | struct drbd_peer_request *, const unsigned, | |
1502 | const int); | |
7721f567 | 1503 | extern int drbd_free_peer_reqs(struct drbd_conf *, struct list_head *); |
0db55363 AG |
1504 | extern struct drbd_peer_request *drbd_alloc_peer_req(struct drbd_conf *, u64, |
1505 | sector_t, unsigned int, | |
1506 | gfp_t) __must_hold(local); | |
3967deb1 AG |
1507 | extern void __drbd_free_peer_req(struct drbd_conf *, struct drbd_peer_request *, |
1508 | int); | |
1509 | #define drbd_free_peer_req(m,e) __drbd_free_peer_req(m, e, 0) | |
1510 | #define drbd_free_net_peer_req(m,e) __drbd_free_peer_req(m, e, 1) | |
c37c8ecf | 1511 | extern struct page *drbd_alloc_pages(struct drbd_conf *, unsigned int, bool); |
b411b363 PR |
1512 | extern void drbd_set_recv_tcq(struct drbd_conf *mdev, int tcq_enabled); |
1513 | extern void _drbd_clear_done_ee(struct drbd_conf *mdev, struct list_head *to_be_freed); | |
0e29d163 | 1514 | extern void conn_flush_workqueue(struct drbd_tconn *tconn); |
c141ebda | 1515 | extern int drbd_connected(struct drbd_conf *mdev); |
0e29d163 PR |
1516 | static inline void drbd_flush_workqueue(struct drbd_conf *mdev) |
1517 | { | |
1518 | conn_flush_workqueue(mdev->tconn); | |
1519 | } | |
b411b363 | 1520 | |
ed439848 LE |
1521 | /* Yes, there is kernel_setsockopt, but only since 2.6.18. |
1522 | * So we have our own copy of it here. */ | |
b411b363 | 1523 | static inline int drbd_setsockopt(struct socket *sock, int level, int optname, |
ed439848 | 1524 | char *optval, int optlen) |
b411b363 | 1525 | { |
ed439848 LE |
1526 | mm_segment_t oldfs = get_fs(); |
1527 | char __user *uoptval; | |
b411b363 | 1528 | int err; |
ed439848 LE |
1529 | |
1530 | uoptval = (char __user __force *)optval; | |
1531 | ||
1532 | set_fs(KERNEL_DS); | |
b411b363 | 1533 | if (level == SOL_SOCKET) |
ed439848 | 1534 | err = sock_setsockopt(sock, level, optname, uoptval, optlen); |
b411b363 | 1535 | else |
ed439848 | 1536 | err = sock->ops->setsockopt(sock, level, optname, uoptval, |
b411b363 | 1537 | optlen); |
ed439848 | 1538 | set_fs(oldfs); |
b411b363 PR |
1539 | return err; |
1540 | } | |
1541 | ||
1542 | static inline void drbd_tcp_cork(struct socket *sock) | |
1543 | { | |
ed439848 | 1544 | int val = 1; |
b411b363 | 1545 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_CORK, |
ed439848 | 1546 | (char*)&val, sizeof(val)); |
b411b363 PR |
1547 | } |
1548 | ||
1549 | static inline void drbd_tcp_uncork(struct socket *sock) | |
1550 | { | |
ed439848 | 1551 | int val = 0; |
b411b363 | 1552 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_CORK, |
ed439848 | 1553 | (char*)&val, sizeof(val)); |
b411b363 PR |
1554 | } |
1555 | ||
1556 | static inline void drbd_tcp_nodelay(struct socket *sock) | |
1557 | { | |
ed439848 | 1558 | int val = 1; |
b411b363 | 1559 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_NODELAY, |
ed439848 | 1560 | (char*)&val, sizeof(val)); |
b411b363 PR |
1561 | } |
1562 | ||
1563 | static inline void drbd_tcp_quickack(struct socket *sock) | |
1564 | { | |
ed439848 | 1565 | int val = 2; |
b411b363 | 1566 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_QUICKACK, |
ed439848 | 1567 | (char*)&val, sizeof(val)); |
b411b363 PR |
1568 | } |
1569 | ||
4b0007c0 | 1570 | void drbd_bump_write_ordering(struct drbd_tconn *tconn, enum write_ordering_e wo); |
b411b363 PR |
1571 | |
1572 | /* drbd_proc.c */ | |
1573 | extern struct proc_dir_entry *drbd_proc; | |
7d4e9d09 | 1574 | extern const struct file_operations drbd_proc_fops; |
b411b363 PR |
1575 | extern const char *drbd_conn_str(enum drbd_conns s); |
1576 | extern const char *drbd_role_str(enum drbd_role s); | |
1577 | ||
1578 | /* drbd_actlog.c */ | |
181286ad LE |
1579 | extern void drbd_al_begin_io(struct drbd_conf *mdev, struct drbd_interval *i); |
1580 | extern void drbd_al_complete_io(struct drbd_conf *mdev, struct drbd_interval *i); | |
b411b363 PR |
1581 | extern void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector); |
1582 | extern int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector); | |
1583 | extern int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector); | |
1584 | extern void drbd_rs_cancel_all(struct drbd_conf *mdev); | |
1585 | extern int drbd_rs_del_all(struct drbd_conf *mdev); | |
1586 | extern void drbd_rs_failed_io(struct drbd_conf *mdev, | |
1587 | sector_t sector, int size); | |
ea5442af | 1588 | extern void drbd_advance_rs_marks(struct drbd_conf *mdev, unsigned long still_to_go); |
b411b363 PR |
1589 | extern void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, |
1590 | int size, const char *file, const unsigned int line); | |
1591 | #define drbd_set_in_sync(mdev, sector, size) \ | |
1592 | __drbd_set_in_sync(mdev, sector, size, __FILE__, __LINE__) | |
73a01a18 | 1593 | extern int __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, |
b411b363 PR |
1594 | int size, const char *file, const unsigned int line); |
1595 | #define drbd_set_out_of_sync(mdev, sector, size) \ | |
1596 | __drbd_set_out_of_sync(mdev, sector, size, __FILE__, __LINE__) | |
b411b363 PR |
1597 | extern void drbd_al_shrink(struct drbd_conf *mdev); |
1598 | ||
b411b363 | 1599 | /* drbd_nl.c */ |
3b98c0c2 LE |
1600 | /* state info broadcast */ |
1601 | struct sib_info { | |
1602 | enum drbd_state_info_bcast_reason sib_reason; | |
1603 | union { | |
1604 | struct { | |
1605 | char *helper_name; | |
1606 | unsigned helper_exit_code; | |
1607 | }; | |
1608 | struct { | |
1609 | union drbd_state os; | |
1610 | union drbd_state ns; | |
1611 | }; | |
1612 | }; | |
1613 | }; | |
1614 | void drbd_bcast_event(struct drbd_conf *mdev, const struct sib_info *sib); | |
b411b363 | 1615 | |
b411b363 PR |
1616 | /* |
1617 | * inline helper functions | |
1618 | *************************/ | |
1619 | ||
45bb912b LE |
1620 | /* see also page_chain_add and friends in drbd_receiver.c */ |
1621 | static inline struct page *page_chain_next(struct page *page) | |
1622 | { | |
1623 | return (struct page *)page_private(page); | |
1624 | } | |
1625 | #define page_chain_for_each(page) \ | |
1626 | for (; page && ({ prefetch(page_chain_next(page)); 1; }); \ | |
1627 | page = page_chain_next(page)) | |
1628 | #define page_chain_for_each_safe(page, n) \ | |
1629 | for (; page && ({ n = page_chain_next(page); 1; }); page = n) | |
1630 | ||
45bb912b | 1631 | |
045417f7 | 1632 | static inline int drbd_peer_req_has_active_page(struct drbd_peer_request *peer_req) |
45bb912b | 1633 | { |
db830c46 | 1634 | struct page *page = peer_req->pages; |
45bb912b LE |
1635 | page_chain_for_each(page) { |
1636 | if (page_count(page) > 1) | |
1637 | return 1; | |
1638 | } | |
1639 | return 0; | |
1640 | } | |
1641 | ||
bf885f8a AG |
1642 | static inline enum drbd_state_rv |
1643 | _drbd_set_state(struct drbd_conf *mdev, union drbd_state ns, | |
1644 | enum chg_state_flags flags, struct completion *done) | |
b411b363 | 1645 | { |
bf885f8a | 1646 | enum drbd_state_rv rv; |
b411b363 PR |
1647 | |
1648 | read_lock(&global_state_lock); | |
1649 | rv = __drbd_set_state(mdev, ns, flags, done); | |
1650 | read_unlock(&global_state_lock); | |
1651 | ||
1652 | return rv; | |
1653 | } | |
1654 | ||
78bae59b PR |
1655 | static inline union drbd_state drbd_read_state(struct drbd_conf *mdev) |
1656 | { | |
1657 | union drbd_state rv; | |
1658 | ||
da9fbc27 | 1659 | rv.i = mdev->state.i; |
8e0af25f PR |
1660 | rv.susp = mdev->tconn->susp; |
1661 | rv.susp_nod = mdev->tconn->susp_nod; | |
1662 | rv.susp_fen = mdev->tconn->susp_fen; | |
78bae59b PR |
1663 | |
1664 | return rv; | |
1665 | } | |
1666 | ||
0c849666 | 1667 | enum drbd_force_detach_flags { |
edc9f5eb LE |
1668 | DRBD_READ_ERROR, |
1669 | DRBD_WRITE_ERROR, | |
0c849666 LE |
1670 | DRBD_META_IO_ERROR, |
1671 | DRBD_FORCE_DETACH, | |
1672 | }; | |
1673 | ||
b411b363 | 1674 | #define __drbd_chk_io_error(m,f) __drbd_chk_io_error_(m,f, __func__) |
0c849666 | 1675 | static inline void __drbd_chk_io_error_(struct drbd_conf *mdev, |
edc9f5eb | 1676 | enum drbd_force_detach_flags df, |
0c849666 | 1677 | const char *where) |
b411b363 | 1678 | { |
daeda1cc PR |
1679 | enum drbd_io_error_p ep; |
1680 | ||
1681 | rcu_read_lock(); | |
1682 | ep = rcu_dereference(mdev->ldev->disk_conf)->on_io_error; | |
1683 | rcu_read_unlock(); | |
1684 | switch (ep) { | |
1685 | case EP_PASS_ON: /* FIXME would this be better named "Ignore"? */ | |
edc9f5eb | 1686 | if (df == DRBD_READ_ERROR || df == DRBD_WRITE_ERROR) { |
7383506c | 1687 | if (__ratelimit(&drbd_ratelimit_state)) |
82f59cc6 | 1688 | dev_err(DEV, "Local IO failed in %s.\n", where); |
d2e17807 PR |
1689 | if (mdev->state.disk > D_INCONSISTENT) |
1690 | _drbd_set_state(_NS(mdev, disk, D_INCONSISTENT), CS_HARD, NULL); | |
b411b363 PR |
1691 | break; |
1692 | } | |
edc9f5eb | 1693 | /* NOTE fall through for DRBD_META_IO_ERROR or DRBD_FORCE_DETACH */ |
b411b363 PR |
1694 | case EP_DETACH: |
1695 | case EP_CALL_HELPER: | |
edc9f5eb LE |
1696 | /* Remember whether we saw a READ or WRITE error. |
1697 | * | |
1698 | * Recovery of the affected area for WRITE failure is covered | |
1699 | * by the activity log. | |
1700 | * READ errors may fall outside that area though. Certain READ | |
1701 | * errors can be "healed" by writing good data to the affected | |
1702 | * blocks, which triggers block re-allocation in lower layers. | |
1703 | * | |
1704 | * If we can not write the bitmap after a READ error, | |
1705 | * we may need to trigger a full sync (see w_go_diskless()). | |
1706 | * | |
1707 | * Force-detach is not really an IO error, but rather a | |
1708 | * desperate measure to try to deal with a completely | |
1709 | * unresponsive lower level IO stack. | |
1710 | * Still it should be treated as a WRITE error. | |
1711 | * | |
1712 | * Meta IO error is always WRITE error: | |
1713 | * we read meta data only once during attach, | |
1714 | * which will fail in case of errors. | |
1715 | */ | |
82f59cc6 | 1716 | set_bit(WAS_IO_ERROR, &mdev->flags); |
edc9f5eb LE |
1717 | if (df == DRBD_READ_ERROR) |
1718 | set_bit(WAS_READ_ERROR, &mdev->flags); | |
1719 | if (df == DRBD_FORCE_DETACH) | |
0c849666 | 1720 | set_bit(FORCE_DETACH, &mdev->flags); |
b411b363 PR |
1721 | if (mdev->state.disk > D_FAILED) { |
1722 | _drbd_set_state(_NS(mdev, disk, D_FAILED), CS_HARD, NULL); | |
82f59cc6 LE |
1723 | dev_err(DEV, |
1724 | "Local IO failed in %s. Detaching...\n", where); | |
b411b363 PR |
1725 | } |
1726 | break; | |
1727 | } | |
1728 | } | |
1729 | ||
1730 | /** | |
1731 | * drbd_chk_io_error: Handle the on_io_error setting, should be called from all io completion handlers | |
1732 | * @mdev: DRBD device. | |
1733 | * @error: Error code passed to the IO completion callback | |
1734 | * @forcedetach: Force detach. I.e. the error happened while accessing the meta data | |
1735 | * | |
1736 | * See also drbd_main.c:after_state_ch() if (os.disk > D_FAILED && ns.disk == D_FAILED) | |
1737 | */ | |
1738 | #define drbd_chk_io_error(m,e,f) drbd_chk_io_error_(m,e,f, __func__) | |
1739 | static inline void drbd_chk_io_error_(struct drbd_conf *mdev, | |
0c849666 | 1740 | int error, enum drbd_force_detach_flags forcedetach, const char *where) |
b411b363 PR |
1741 | { |
1742 | if (error) { | |
1743 | unsigned long flags; | |
87eeee41 | 1744 | spin_lock_irqsave(&mdev->tconn->req_lock, flags); |
b411b363 | 1745 | __drbd_chk_io_error_(mdev, forcedetach, where); |
87eeee41 | 1746 | spin_unlock_irqrestore(&mdev->tconn->req_lock, flags); |
b411b363 PR |
1747 | } |
1748 | } | |
1749 | ||
1750 | ||
1751 | /** | |
1752 | * drbd_md_first_sector() - Returns the first sector number of the meta data area | |
1753 | * @bdev: Meta data block device. | |
1754 | * | |
1755 | * BTW, for internal meta data, this happens to be the maximum capacity | |
1756 | * we could agree upon with our peer node. | |
1757 | */ | |
daeda1cc | 1758 | static inline sector_t _drbd_md_first_sector(int meta_dev_idx, struct drbd_backing_dev *bdev) |
b411b363 | 1759 | { |
daeda1cc | 1760 | switch (meta_dev_idx) { |
b411b363 PR |
1761 | case DRBD_MD_INDEX_INTERNAL: |
1762 | case DRBD_MD_INDEX_FLEX_INT: | |
1763 | return bdev->md.md_offset + bdev->md.bm_offset; | |
1764 | case DRBD_MD_INDEX_FLEX_EXT: | |
1765 | default: | |
1766 | return bdev->md.md_offset; | |
1767 | } | |
1768 | } | |
1769 | ||
daeda1cc PR |
1770 | static inline sector_t drbd_md_first_sector(struct drbd_backing_dev *bdev) |
1771 | { | |
1772 | int meta_dev_idx; | |
1773 | ||
1774 | rcu_read_lock(); | |
1775 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1776 | rcu_read_unlock(); | |
1777 | ||
1778 | return _drbd_md_first_sector(meta_dev_idx, bdev); | |
1779 | } | |
1780 | ||
b411b363 PR |
1781 | /** |
1782 | * drbd_md_last_sector() - Return the last sector number of the meta data area | |
1783 | * @bdev: Meta data block device. | |
1784 | */ | |
1785 | static inline sector_t drbd_md_last_sector(struct drbd_backing_dev *bdev) | |
1786 | { | |
daeda1cc PR |
1787 | int meta_dev_idx; |
1788 | ||
1789 | rcu_read_lock(); | |
1790 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1791 | rcu_read_unlock(); | |
1792 | ||
1793 | switch (meta_dev_idx) { | |
b411b363 PR |
1794 | case DRBD_MD_INDEX_INTERNAL: |
1795 | case DRBD_MD_INDEX_FLEX_INT: | |
1796 | return bdev->md.md_offset + MD_AL_OFFSET - 1; | |
1797 | case DRBD_MD_INDEX_FLEX_EXT: | |
1798 | default: | |
1799 | return bdev->md.md_offset + bdev->md.md_size_sect; | |
1800 | } | |
1801 | } | |
1802 | ||
1803 | /* Returns the number of 512 byte sectors of the device */ | |
1804 | static inline sector_t drbd_get_capacity(struct block_device *bdev) | |
1805 | { | |
1806 | /* return bdev ? get_capacity(bdev->bd_disk) : 0; */ | |
77304d2a | 1807 | return bdev ? i_size_read(bdev->bd_inode) >> 9 : 0; |
b411b363 PR |
1808 | } |
1809 | ||
1810 | /** | |
1811 | * drbd_get_max_capacity() - Returns the capacity we announce to out peer | |
1812 | * @bdev: Meta data block device. | |
1813 | * | |
1814 | * returns the capacity we announce to out peer. we clip ourselves at the | |
1815 | * various MAX_SECTORS, because if we don't, current implementation will | |
1816 | * oops sooner or later | |
1817 | */ | |
1818 | static inline sector_t drbd_get_max_capacity(struct drbd_backing_dev *bdev) | |
1819 | { | |
1820 | sector_t s; | |
daeda1cc PR |
1821 | int meta_dev_idx; |
1822 | ||
1823 | rcu_read_lock(); | |
1824 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1825 | rcu_read_unlock(); | |
1826 | ||
1827 | switch (meta_dev_idx) { | |
b411b363 PR |
1828 | case DRBD_MD_INDEX_INTERNAL: |
1829 | case DRBD_MD_INDEX_FLEX_INT: | |
1830 | s = drbd_get_capacity(bdev->backing_bdev) | |
1831 | ? min_t(sector_t, DRBD_MAX_SECTORS_FLEX, | |
daeda1cc | 1832 | _drbd_md_first_sector(meta_dev_idx, bdev)) |
b411b363 PR |
1833 | : 0; |
1834 | break; | |
1835 | case DRBD_MD_INDEX_FLEX_EXT: | |
1836 | s = min_t(sector_t, DRBD_MAX_SECTORS_FLEX, | |
1837 | drbd_get_capacity(bdev->backing_bdev)); | |
1838 | /* clip at maximum size the meta device can support */ | |
1839 | s = min_t(sector_t, s, | |
1840 | BM_EXT_TO_SECT(bdev->md.md_size_sect | |
1841 | - bdev->md.bm_offset)); | |
1842 | break; | |
1843 | default: | |
1844 | s = min_t(sector_t, DRBD_MAX_SECTORS, | |
1845 | drbd_get_capacity(bdev->backing_bdev)); | |
1846 | } | |
1847 | return s; | |
1848 | } | |
1849 | ||
1850 | /** | |
1851 | * drbd_md_ss__() - Return the sector number of our meta data super block | |
1852 | * @mdev: DRBD device. | |
1853 | * @bdev: Meta data block device. | |
1854 | */ | |
1855 | static inline sector_t drbd_md_ss__(struct drbd_conf *mdev, | |
1856 | struct drbd_backing_dev *bdev) | |
1857 | { | |
daeda1cc PR |
1858 | int meta_dev_idx; |
1859 | ||
1860 | rcu_read_lock(); | |
1861 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1862 | rcu_read_unlock(); | |
1863 | ||
1864 | switch (meta_dev_idx) { | |
b411b363 | 1865 | default: /* external, some index */ |
daeda1cc | 1866 | return MD_RESERVED_SECT * meta_dev_idx; |
b411b363 PR |
1867 | case DRBD_MD_INDEX_INTERNAL: |
1868 | /* with drbd08, internal meta data is always "flexible" */ | |
1869 | case DRBD_MD_INDEX_FLEX_INT: | |
1870 | /* sizeof(struct md_on_disk_07) == 4k | |
1871 | * position: last 4k aligned block of 4k size */ | |
1872 | if (!bdev->backing_bdev) { | |
1873 | if (__ratelimit(&drbd_ratelimit_state)) { | |
1874 | dev_err(DEV, "bdev->backing_bdev==NULL\n"); | |
1875 | dump_stack(); | |
1876 | } | |
1877 | return 0; | |
1878 | } | |
1879 | return (drbd_get_capacity(bdev->backing_bdev) & ~7ULL) | |
1880 | - MD_AL_OFFSET; | |
1881 | case DRBD_MD_INDEX_FLEX_EXT: | |
1882 | return 0; | |
1883 | } | |
1884 | } | |
1885 | ||
b411b363 PR |
1886 | static inline void |
1887 | drbd_queue_work_front(struct drbd_work_queue *q, struct drbd_work *w) | |
1888 | { | |
1889 | unsigned long flags; | |
1890 | spin_lock_irqsave(&q->q_lock, flags); | |
1891 | list_add(&w->list, &q->q); | |
b411b363 | 1892 | spin_unlock_irqrestore(&q->q_lock, flags); |
8c0785a5 | 1893 | wake_up(&q->q_wait); |
b411b363 PR |
1894 | } |
1895 | ||
1896 | static inline void | |
1897 | drbd_queue_work(struct drbd_work_queue *q, struct drbd_work *w) | |
1898 | { | |
1899 | unsigned long flags; | |
1900 | spin_lock_irqsave(&q->q_lock, flags); | |
1901 | list_add_tail(&w->list, &q->q); | |
b411b363 | 1902 | spin_unlock_irqrestore(&q->q_lock, flags); |
8c0785a5 | 1903 | wake_up(&q->q_wait); |
b411b363 PR |
1904 | } |
1905 | ||
0625ac19 | 1906 | static inline void wake_asender(struct drbd_tconn *tconn) |
b411b363 | 1907 | { |
0625ac19 PR |
1908 | if (test_bit(SIGNAL_ASENDER, &tconn->flags)) |
1909 | force_sig(DRBD_SIG, tconn->asender.task); | |
b411b363 PR |
1910 | } |
1911 | ||
0625ac19 | 1912 | static inline void request_ping(struct drbd_tconn *tconn) |
b411b363 | 1913 | { |
0625ac19 PR |
1914 | set_bit(SEND_PING, &tconn->flags); |
1915 | wake_asender(tconn); | |
b411b363 PR |
1916 | } |
1917 | ||
dba58587 AG |
1918 | extern void *conn_prepare_command(struct drbd_tconn *, struct drbd_socket *); |
1919 | extern void *drbd_prepare_command(struct drbd_conf *, struct drbd_socket *); | |
1920 | extern int conn_send_command(struct drbd_tconn *, struct drbd_socket *, | |
1921 | enum drbd_packet, unsigned int, void *, | |
1922 | unsigned int); | |
1923 | extern int drbd_send_command(struct drbd_conf *, struct drbd_socket *, | |
1924 | enum drbd_packet, unsigned int, void *, | |
1925 | unsigned int); | |
1926 | ||
e307f352 AG |
1927 | extern int drbd_send_ping(struct drbd_tconn *tconn); |
1928 | extern int drbd_send_ping_ack(struct drbd_tconn *tconn); | |
9f5bdc33 AG |
1929 | extern int drbd_send_state_req(struct drbd_conf *, union drbd_state, union drbd_state); |
1930 | extern int conn_send_state_req(struct drbd_tconn *, union drbd_state, union drbd_state); | |
cf29c9d8 | 1931 | |
b411b363 PR |
1932 | static inline void drbd_thread_stop(struct drbd_thread *thi) |
1933 | { | |
81e84650 | 1934 | _drbd_thread_stop(thi, false, true); |
b411b363 PR |
1935 | } |
1936 | ||
1937 | static inline void drbd_thread_stop_nowait(struct drbd_thread *thi) | |
1938 | { | |
81e84650 | 1939 | _drbd_thread_stop(thi, false, false); |
b411b363 PR |
1940 | } |
1941 | ||
1942 | static inline void drbd_thread_restart_nowait(struct drbd_thread *thi) | |
1943 | { | |
81e84650 | 1944 | _drbd_thread_stop(thi, true, false); |
b411b363 PR |
1945 | } |
1946 | ||
1947 | /* counts how many answer packets packets we expect from our peer, | |
1948 | * for either explicit application requests, | |
1949 | * or implicit barrier packets as necessary. | |
1950 | * increased: | |
1951 | * w_send_barrier | |
8554df1c | 1952 | * _req_mod(req, QUEUE_FOR_NET_WRITE or QUEUE_FOR_NET_READ); |
b411b363 PR |
1953 | * it is much easier and equally valid to count what we queue for the |
1954 | * worker, even before it actually was queued or send. | |
1955 | * (drbd_make_request_common; recovery path on read io-error) | |
1956 | * decreased: | |
1957 | * got_BarrierAck (respective tl_clear, tl_clear_barrier) | |
8554df1c | 1958 | * _req_mod(req, DATA_RECEIVED) |
b411b363 | 1959 | * [from receive_DataReply] |
8554df1c | 1960 | * _req_mod(req, WRITE_ACKED_BY_PEER or RECV_ACKED_BY_PEER or NEG_ACKED) |
b411b363 PR |
1961 | * [from got_BlockAck (P_WRITE_ACK, P_RECV_ACK)] |
1962 | * for some reason it is NOT decreased in got_NegAck, | |
1963 | * but in the resulting cleanup code from report_params. | |
1964 | * we should try to remember the reason for that... | |
8554df1c AG |
1965 | * _req_mod(req, SEND_FAILED or SEND_CANCELED) |
1966 | * _req_mod(req, CONNECTION_LOST_WHILE_PENDING) | |
b411b363 PR |
1967 | * [from tl_clear_barrier] |
1968 | */ | |
1969 | static inline void inc_ap_pending(struct drbd_conf *mdev) | |
1970 | { | |
1971 | atomic_inc(&mdev->ap_pending_cnt); | |
1972 | } | |
1973 | ||
49559d87 PR |
1974 | #define ERR_IF_CNT_IS_NEGATIVE(which, func, line) \ |
1975 | if (atomic_read(&mdev->which) < 0) \ | |
b411b363 | 1976 | dev_err(DEV, "in %s:%d: " #which " = %d < 0 !\n", \ |
49559d87 PR |
1977 | func, line, \ |
1978 | atomic_read(&mdev->which)) | |
b411b363 | 1979 | |
49559d87 PR |
1980 | #define dec_ap_pending(mdev) _dec_ap_pending(mdev, __FUNCTION__, __LINE__) |
1981 | static inline void _dec_ap_pending(struct drbd_conf *mdev, const char *func, int line) | |
1982 | { | |
1983 | if (atomic_dec_and_test(&mdev->ap_pending_cnt)) | |
1984 | wake_up(&mdev->misc_wait); | |
1985 | ERR_IF_CNT_IS_NEGATIVE(ap_pending_cnt, func, line); | |
1986 | } | |
b411b363 PR |
1987 | |
1988 | /* counts how many resync-related answers we still expect from the peer | |
1989 | * increase decrease | |
1990 | * C_SYNC_TARGET sends P_RS_DATA_REQUEST (and expects P_RS_DATA_REPLY) | |
25985edc | 1991 | * C_SYNC_SOURCE sends P_RS_DATA_REPLY (and expects P_WRITE_ACK with ID_SYNCER) |
b411b363 PR |
1992 | * (or P_NEG_ACK with ID_SYNCER) |
1993 | */ | |
1994 | static inline void inc_rs_pending(struct drbd_conf *mdev) | |
1995 | { | |
1996 | atomic_inc(&mdev->rs_pending_cnt); | |
1997 | } | |
1998 | ||
49559d87 PR |
1999 | #define dec_rs_pending(mdev) _dec_rs_pending(mdev, __FUNCTION__, __LINE__) |
2000 | static inline void _dec_rs_pending(struct drbd_conf *mdev, const char *func, int line) | |
2001 | { | |
2002 | atomic_dec(&mdev->rs_pending_cnt); | |
2003 | ERR_IF_CNT_IS_NEGATIVE(rs_pending_cnt, func, line); | |
2004 | } | |
b411b363 PR |
2005 | |
2006 | /* counts how many answers we still need to send to the peer. | |
2007 | * increased on | |
2008 | * receive_Data unless protocol A; | |
2009 | * we need to send a P_RECV_ACK (proto B) | |
2010 | * or P_WRITE_ACK (proto C) | |
2011 | * receive_RSDataReply (recv_resync_read) we need to send a P_WRITE_ACK | |
2012 | * receive_DataRequest (receive_RSDataRequest) we need to send back P_DATA | |
2013 | * receive_Barrier_* we need to send a P_BARRIER_ACK | |
2014 | */ | |
2015 | static inline void inc_unacked(struct drbd_conf *mdev) | |
2016 | { | |
2017 | atomic_inc(&mdev->unacked_cnt); | |
2018 | } | |
2019 | ||
49559d87 PR |
2020 | #define dec_unacked(mdev) _dec_unacked(mdev, __FUNCTION__, __LINE__) |
2021 | static inline void _dec_unacked(struct drbd_conf *mdev, const char *func, int line) | |
2022 | { | |
2023 | atomic_dec(&mdev->unacked_cnt); | |
2024 | ERR_IF_CNT_IS_NEGATIVE(unacked_cnt, func, line); | |
2025 | } | |
b411b363 | 2026 | |
49559d87 PR |
2027 | #define sub_unacked(mdev, n) _sub_unacked(mdev, n, __FUNCTION__, __LINE__) |
2028 | static inline void _sub_unacked(struct drbd_conf *mdev, int n, const char *func, int line) | |
2029 | { | |
2030 | atomic_sub(n, &mdev->unacked_cnt); | |
2031 | ERR_IF_CNT_IS_NEGATIVE(unacked_cnt, func, line); | |
2032 | } | |
b411b363 | 2033 | |
b411b363 PR |
2034 | /** |
2035 | * get_ldev() - Increase the ref count on mdev->ldev. Returns 0 if there is no ldev | |
2036 | * @M: DRBD device. | |
2037 | * | |
2038 | * You have to call put_ldev() when finished working with mdev->ldev. | |
2039 | */ | |
2040 | #define get_ldev(M) __cond_lock(local, _get_ldev_if_state(M,D_INCONSISTENT)) | |
2041 | #define get_ldev_if_state(M,MINS) __cond_lock(local, _get_ldev_if_state(M,MINS)) | |
2042 | ||
2043 | static inline void put_ldev(struct drbd_conf *mdev) | |
2044 | { | |
1d7734a0 | 2045 | int i = atomic_dec_return(&mdev->local_cnt); |
9a0d9d03 LE |
2046 | |
2047 | /* This may be called from some endio handler, | |
2048 | * so we must not sleep here. */ | |
2049 | ||
b411b363 | 2050 | __release(local); |
1d7734a0 | 2051 | D_ASSERT(i >= 0); |
e9e6f3ec | 2052 | if (i == 0) { |
82f59cc6 LE |
2053 | if (mdev->state.disk == D_DISKLESS) |
2054 | /* even internal references gone, safe to destroy */ | |
2055 | drbd_ldev_destroy(mdev); | |
e9e6f3ec | 2056 | if (mdev->state.disk == D_FAILED) |
82f59cc6 | 2057 | /* all application IO references gone. */ |
e9e6f3ec | 2058 | drbd_go_diskless(mdev); |
b411b363 | 2059 | wake_up(&mdev->misc_wait); |
e9e6f3ec | 2060 | } |
b411b363 PR |
2061 | } |
2062 | ||
2063 | #ifndef __CHECKER__ | |
2064 | static inline int _get_ldev_if_state(struct drbd_conf *mdev, enum drbd_disk_state mins) | |
2065 | { | |
2066 | int io_allowed; | |
2067 | ||
82f59cc6 LE |
2068 | /* never get a reference while D_DISKLESS */ |
2069 | if (mdev->state.disk == D_DISKLESS) | |
2070 | return 0; | |
2071 | ||
b411b363 PR |
2072 | atomic_inc(&mdev->local_cnt); |
2073 | io_allowed = (mdev->state.disk >= mins); | |
2074 | if (!io_allowed) | |
2075 | put_ldev(mdev); | |
2076 | return io_allowed; | |
2077 | } | |
2078 | #else | |
2079 | extern int _get_ldev_if_state(struct drbd_conf *mdev, enum drbd_disk_state mins); | |
2080 | #endif | |
2081 | ||
2082 | /* you must have an "get_ldev" reference */ | |
2083 | static inline void drbd_get_syncer_progress(struct drbd_conf *mdev, | |
2084 | unsigned long *bits_left, unsigned int *per_mil_done) | |
2085 | { | |
4b0715f0 LE |
2086 | /* this is to break it at compile time when we change that, in case we |
2087 | * want to support more than (1<<32) bits on a 32bit arch. */ | |
b411b363 PR |
2088 | typecheck(unsigned long, mdev->rs_total); |
2089 | ||
2090 | /* note: both rs_total and rs_left are in bits, i.e. in | |
2091 | * units of BM_BLOCK_SIZE. | |
2092 | * for the percentage, we don't care. */ | |
2093 | ||
439d5953 LE |
2094 | if (mdev->state.conn == C_VERIFY_S || mdev->state.conn == C_VERIFY_T) |
2095 | *bits_left = mdev->ov_left; | |
2096 | else | |
2097 | *bits_left = drbd_bm_total_weight(mdev) - mdev->rs_failed; | |
b411b363 PR |
2098 | /* >> 10 to prevent overflow, |
2099 | * +1 to prevent division by zero */ | |
2100 | if (*bits_left > mdev->rs_total) { | |
2101 | /* doh. maybe a logic bug somewhere. | |
2102 | * may also be just a race condition | |
2103 | * between this and a disconnect during sync. | |
2104 | * for now, just prevent in-kernel buffer overflow. | |
2105 | */ | |
2106 | smp_rmb(); | |
2107 | dev_warn(DEV, "cs:%s rs_left=%lu > rs_total=%lu (rs_failed %lu)\n", | |
2108 | drbd_conn_str(mdev->state.conn), | |
2109 | *bits_left, mdev->rs_total, mdev->rs_failed); | |
2110 | *per_mil_done = 0; | |
2111 | } else { | |
4b0715f0 LE |
2112 | /* Make sure the division happens in long context. |
2113 | * We allow up to one petabyte storage right now, | |
2114 | * at a granularity of 4k per bit that is 2**38 bits. | |
2115 | * After shift right and multiplication by 1000, | |
2116 | * this should still fit easily into a 32bit long, | |
2117 | * so we don't need a 64bit division on 32bit arch. | |
2118 | * Note: currently we don't support such large bitmaps on 32bit | |
2119 | * arch anyways, but no harm done to be prepared for it here. | |
2120 | */ | |
3b7cd457 | 2121 | unsigned int shift = mdev->rs_total > UINT_MAX ? 16 : 10; |
4b0715f0 LE |
2122 | unsigned long left = *bits_left >> shift; |
2123 | unsigned long total = 1UL + (mdev->rs_total >> shift); | |
2124 | unsigned long tmp = 1000UL - left * 1000UL/total; | |
b411b363 PR |
2125 | *per_mil_done = tmp; |
2126 | } | |
2127 | } | |
2128 | ||
2129 | ||
2130 | /* this throttles on-the-fly application requests | |
2131 | * according to max_buffers settings; | |
2132 | * maybe re-implement using semaphores? */ | |
2133 | static inline int drbd_get_max_buffers(struct drbd_conf *mdev) | |
2134 | { | |
44ed167d PR |
2135 | struct net_conf *nc; |
2136 | int mxb; | |
2137 | ||
2138 | rcu_read_lock(); | |
2139 | nc = rcu_dereference(mdev->tconn->net_conf); | |
2140 | mxb = nc ? nc->max_buffers : 1000000; /* arbitrary limit on open requests */ | |
2141 | rcu_read_unlock(); | |
2142 | ||
b411b363 PR |
2143 | return mxb; |
2144 | } | |
2145 | ||
3719094e | 2146 | static inline int drbd_state_is_stable(struct drbd_conf *mdev) |
b411b363 | 2147 | { |
da9fbc27 | 2148 | union drbd_dev_state s = mdev->state; |
b411b363 PR |
2149 | |
2150 | /* DO NOT add a default clause, we want the compiler to warn us | |
2151 | * for any newly introduced state we may have forgotten to add here */ | |
2152 | ||
2153 | switch ((enum drbd_conns)s.conn) { | |
2154 | /* new io only accepted when there is no connection, ... */ | |
2155 | case C_STANDALONE: | |
2156 | case C_WF_CONNECTION: | |
2157 | /* ... or there is a well established connection. */ | |
2158 | case C_CONNECTED: | |
2159 | case C_SYNC_SOURCE: | |
2160 | case C_SYNC_TARGET: | |
2161 | case C_VERIFY_S: | |
2162 | case C_VERIFY_T: | |
2163 | case C_PAUSED_SYNC_S: | |
2164 | case C_PAUSED_SYNC_T: | |
67531718 PR |
2165 | case C_AHEAD: |
2166 | case C_BEHIND: | |
3719094e | 2167 | /* transitional states, IO allowed */ |
b411b363 PR |
2168 | case C_DISCONNECTING: |
2169 | case C_UNCONNECTED: | |
2170 | case C_TIMEOUT: | |
2171 | case C_BROKEN_PIPE: | |
2172 | case C_NETWORK_FAILURE: | |
2173 | case C_PROTOCOL_ERROR: | |
2174 | case C_TEAR_DOWN: | |
2175 | case C_WF_REPORT_PARAMS: | |
2176 | case C_STARTING_SYNC_S: | |
2177 | case C_STARTING_SYNC_T: | |
3719094e PR |
2178 | break; |
2179 | ||
2180 | /* Allow IO in BM exchange states with new protocols */ | |
b411b363 | 2181 | case C_WF_BITMAP_S: |
31890f4a | 2182 | if (mdev->tconn->agreed_pro_version < 96) |
3719094e PR |
2183 | return 0; |
2184 | break; | |
2185 | ||
2186 | /* no new io accepted in these states */ | |
b411b363 PR |
2187 | case C_WF_BITMAP_T: |
2188 | case C_WF_SYNC_UUID: | |
2189 | case C_MASK: | |
2190 | /* not "stable" */ | |
2191 | return 0; | |
2192 | } | |
2193 | ||
2194 | switch ((enum drbd_disk_state)s.disk) { | |
2195 | case D_DISKLESS: | |
2196 | case D_INCONSISTENT: | |
2197 | case D_OUTDATED: | |
2198 | case D_CONSISTENT: | |
2199 | case D_UP_TO_DATE: | |
cdfda633 | 2200 | case D_FAILED: |
b411b363 PR |
2201 | /* disk state is stable as well. */ |
2202 | break; | |
2203 | ||
d942ae44 | 2204 | /* no new io accepted during transitional states */ |
b411b363 | 2205 | case D_ATTACHING: |
b411b363 PR |
2206 | case D_NEGOTIATING: |
2207 | case D_UNKNOWN: | |
2208 | case D_MASK: | |
2209 | /* not "stable" */ | |
2210 | return 0; | |
2211 | } | |
2212 | ||
2213 | return 1; | |
2214 | } | |
2215 | ||
2aebfabb | 2216 | static inline int drbd_suspended(struct drbd_conf *mdev) |
fb22c402 | 2217 | { |
8e0af25f PR |
2218 | struct drbd_tconn *tconn = mdev->tconn; |
2219 | ||
2220 | return tconn->susp || tconn->susp_fen || tconn->susp_nod; | |
fb22c402 PR |
2221 | } |
2222 | ||
1b881ef7 | 2223 | static inline bool may_inc_ap_bio(struct drbd_conf *mdev) |
b411b363 PR |
2224 | { |
2225 | int mxb = drbd_get_max_buffers(mdev); | |
2226 | ||
2aebfabb | 2227 | if (drbd_suspended(mdev)) |
1b881ef7 | 2228 | return false; |
b411b363 | 2229 | if (test_bit(SUSPEND_IO, &mdev->flags)) |
1b881ef7 | 2230 | return false; |
b411b363 PR |
2231 | |
2232 | /* to avoid potential deadlock or bitmap corruption, | |
2233 | * in various places, we only allow new application io | |
2234 | * to start during "stable" states. */ | |
2235 | ||
2236 | /* no new io accepted when attaching or detaching the disk */ | |
3719094e | 2237 | if (!drbd_state_is_stable(mdev)) |
1b881ef7 | 2238 | return false; |
b411b363 PR |
2239 | |
2240 | /* since some older kernels don't have atomic_add_unless, | |
2241 | * and we are within the spinlock anyways, we have this workaround. */ | |
2242 | if (atomic_read(&mdev->ap_bio_cnt) > mxb) | |
1b881ef7 | 2243 | return false; |
b411b363 | 2244 | if (test_bit(BITMAP_IO, &mdev->flags)) |
1b881ef7 AG |
2245 | return false; |
2246 | return true; | |
b411b363 PR |
2247 | } |
2248 | ||
23361cf3 | 2249 | static inline bool inc_ap_bio_cond(struct drbd_conf *mdev) |
b411b363 | 2250 | { |
1b881ef7 | 2251 | bool rv = false; |
8869d683 | 2252 | |
87eeee41 | 2253 | spin_lock_irq(&mdev->tconn->req_lock); |
1b881ef7 | 2254 | rv = may_inc_ap_bio(mdev); |
8869d683 | 2255 | if (rv) |
23361cf3 | 2256 | atomic_inc(&mdev->ap_bio_cnt); |
87eeee41 | 2257 | spin_unlock_irq(&mdev->tconn->req_lock); |
8869d683 PR |
2258 | |
2259 | return rv; | |
2260 | } | |
b411b363 | 2261 | |
23361cf3 | 2262 | static inline void inc_ap_bio(struct drbd_conf *mdev) |
8869d683 | 2263 | { |
b411b363 PR |
2264 | /* we wait here |
2265 | * as long as the device is suspended | |
2266 | * until the bitmap is no longer on the fly during connection | |
d942ae44 | 2267 | * handshake as long as we would exceed the max_buffer limit. |
b411b363 PR |
2268 | * |
2269 | * to avoid races with the reconnect code, | |
2270 | * we need to atomic_inc within the spinlock. */ | |
2271 | ||
23361cf3 | 2272 | wait_event(mdev->misc_wait, inc_ap_bio_cond(mdev)); |
b411b363 PR |
2273 | } |
2274 | ||
2275 | static inline void dec_ap_bio(struct drbd_conf *mdev) | |
2276 | { | |
2277 | int mxb = drbd_get_max_buffers(mdev); | |
2278 | int ap_bio = atomic_dec_return(&mdev->ap_bio_cnt); | |
2279 | ||
2280 | D_ASSERT(ap_bio >= 0); | |
970fbde1 LE |
2281 | |
2282 | if (ap_bio == 0 && test_bit(BITMAP_IO, &mdev->flags)) { | |
2283 | if (!test_and_set_bit(BITMAP_IO_QUEUED, &mdev->flags)) | |
2284 | drbd_queue_work(&mdev->tconn->sender_work, &mdev->bm_io_work.w); | |
2285 | } | |
2286 | ||
b411b363 PR |
2287 | /* this currently does wake_up for every dec_ap_bio! |
2288 | * maybe rather introduce some type of hysteresis? | |
2289 | * e.g. (ap_bio == mxb/2 || ap_bio == 0) ? */ | |
2290 | if (ap_bio < mxb) | |
2291 | wake_up(&mdev->misc_wait); | |
b411b363 PR |
2292 | } |
2293 | ||
58ffa580 LE |
2294 | static inline bool verify_can_do_stop_sector(struct drbd_conf *mdev) |
2295 | { | |
2296 | return mdev->tconn->agreed_pro_version >= 97 && | |
2297 | mdev->tconn->agreed_pro_version != 100; | |
2298 | } | |
2299 | ||
62b0da3a | 2300 | static inline int drbd_set_ed_uuid(struct drbd_conf *mdev, u64 val) |
b411b363 | 2301 | { |
62b0da3a | 2302 | int changed = mdev->ed_uuid != val; |
b411b363 | 2303 | mdev->ed_uuid = val; |
62b0da3a | 2304 | return changed; |
b411b363 PR |
2305 | } |
2306 | ||
b411b363 PR |
2307 | static inline int drbd_queue_order_type(struct drbd_conf *mdev) |
2308 | { | |
2309 | /* sorry, we currently have no working implementation | |
2310 | * of distributed TCQ stuff */ | |
2311 | #ifndef QUEUE_ORDERED_NONE | |
2312 | #define QUEUE_ORDERED_NONE 0 | |
2313 | #endif | |
2314 | return QUEUE_ORDERED_NONE; | |
2315 | } | |
2316 | ||
b411b363 PR |
2317 | static inline void drbd_md_flush(struct drbd_conf *mdev) |
2318 | { | |
2319 | int r; | |
2320 | ||
fd0017c1 PR |
2321 | if (mdev->ldev == NULL) { |
2322 | dev_warn(DEV, "mdev->ldev == NULL in drbd_md_flush\n"); | |
2323 | return; | |
2324 | } | |
2325 | ||
a8a4e51e | 2326 | if (test_bit(MD_NO_FUA, &mdev->flags)) |
b411b363 PR |
2327 | return; |
2328 | ||
8747d30a | 2329 | r = blkdev_issue_flush(mdev->ldev->md_bdev, GFP_NOIO, NULL); |
b411b363 | 2330 | if (r) { |
a8a4e51e | 2331 | set_bit(MD_NO_FUA, &mdev->flags); |
b411b363 PR |
2332 | dev_err(DEV, "meta data flush failed with status %d, disabling md-flushes\n", r); |
2333 | } | |
2334 | } | |
2335 | ||
2336 | #endif | |
c1fd29a1 PR |
2337 | |
2338 | /* This is defined in drivers/md/md.h as well. Should go into wait.h */ | |
2339 | #define __wait_event_lock_irq(wq, condition, lock, cmd) \ | |
2340 | do { \ | |
2341 | wait_queue_t __wait; \ | |
2342 | init_waitqueue_entry(&__wait, current); \ | |
2343 | \ | |
2344 | add_wait_queue(&wq, &__wait); \ | |
2345 | for (;;) { \ | |
2346 | set_current_state(TASK_UNINTERRUPTIBLE); \ | |
2347 | if (condition) \ | |
2348 | break; \ | |
2349 | spin_unlock_irq(&lock); \ | |
2350 | cmd; \ | |
2351 | schedule(); \ | |
2352 | spin_lock_irq(&lock); \ | |
2353 | } \ | |
2354 | current->state = TASK_RUNNING; \ | |
2355 | remove_wait_queue(&wq, &__wait); \ | |
2356 | } while (0) | |
2357 | ||
2358 | #define wait_event_lock_irq(wq, condition, lock, cmd) \ | |
2359 | do { \ | |
2360 | if (condition) \ | |
2361 | break; \ | |
2362 | __wait_event_lock_irq(wq, condition, lock, cmd); \ | |
2363 | } while (0) |