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 LE |
684 | GO_DISKLESS, /* Disk is being detached, on io-error or admin request. */ |
685 | WAS_IO_ERROR, /* Local disk failed returned IO error */ | |
0c849666 | 686 | FORCE_DETACH, /* Force-detach from local disk, aborting any pending local IO */ |
b411b363 | 687 | RESYNC_AFTER_NEG, /* Resync after online grow after the attach&negotiate finished. */ |
b411b363 PR |
688 | RESIZE_PENDING, /* Size change detected locally, waiting for the response from |
689 | * the peer, if it changed there as well. */ | |
43a5182c | 690 | NEW_CUR_UUID, /* Create new current UUID when thawing IO */ |
0778286a | 691 | AL_SUSPENDED, /* Activity logging is currently suspended. */ |
370a43e7 | 692 | AHEAD_TO_SYNC_SOURCE, /* Ahead -> SyncSource queued */ |
e64a3294 | 693 | B_RS_H_DONE, /* Before resync handler done (already executed) */ |
08b165ba | 694 | DISCARD_MY_DATA, /* discard_my_data flag per volume */ |
380207d0 | 695 | READ_BALANCE_RR, |
b411b363 PR |
696 | }; |
697 | ||
698 | struct drbd_bitmap; /* opaque for drbd_conf */ | |
699 | ||
20ceb2b2 LE |
700 | /* definition of bits in bm_flags to be used in drbd_bm_lock |
701 | * and drbd_bitmap_io and friends. */ | |
702 | enum bm_flag { | |
703 | /* do we need to kfree, or vfree bm_pages? */ | |
704 | BM_P_VMALLOCED = 0x10000, /* internal use only, will be masked out */ | |
705 | ||
706 | /* currently locked for bulk operation */ | |
a220d291 | 707 | BM_LOCKED_MASK = 0xf, |
20ceb2b2 LE |
708 | |
709 | /* in detail, that is: */ | |
710 | BM_DONT_CLEAR = 0x1, | |
711 | BM_DONT_SET = 0x2, | |
712 | BM_DONT_TEST = 0x4, | |
713 | ||
a220d291 LE |
714 | /* so we can mark it locked for bulk operation, |
715 | * and still allow all non-bulk operations */ | |
716 | BM_IS_LOCKED = 0x8, | |
717 | ||
20ceb2b2 | 718 | /* (test bit, count bit) allowed (common case) */ |
a220d291 | 719 | BM_LOCKED_TEST_ALLOWED = BM_DONT_CLEAR | BM_DONT_SET | BM_IS_LOCKED, |
20ceb2b2 LE |
720 | |
721 | /* testing bits, as well as setting new bits allowed, but clearing bits | |
722 | * would be unexpected. Used during bitmap receive. Setting new bits | |
723 | * requires sending of "out-of-sync" information, though. */ | |
a220d291 | 724 | BM_LOCKED_SET_ALLOWED = BM_DONT_CLEAR | BM_IS_LOCKED, |
20ceb2b2 | 725 | |
a220d291 LE |
726 | /* for drbd_bm_write_copy_pages, everything is allowed, |
727 | * only concurrent bulk operations are locked out. */ | |
728 | BM_LOCKED_CHANGE_ALLOWED = BM_IS_LOCKED, | |
20ceb2b2 LE |
729 | }; |
730 | ||
b411b363 PR |
731 | struct drbd_work_queue { |
732 | struct list_head q; | |
b411b363 | 733 | spinlock_t q_lock; /* to protect the list. */ |
8c0785a5 | 734 | wait_queue_head_t q_wait; |
b411b363 PR |
735 | }; |
736 | ||
737 | struct drbd_socket { | |
b411b363 PR |
738 | struct mutex mutex; |
739 | struct socket *socket; | |
740 | /* this way we get our | |
741 | * send/receive buffers off the stack */ | |
5a87d920 | 742 | void *sbuf; |
e6ef8a5c | 743 | void *rbuf; |
b411b363 PR |
744 | }; |
745 | ||
746 | struct drbd_md { | |
747 | u64 md_offset; /* sector offset to 'super' block */ | |
748 | ||
749 | u64 la_size_sect; /* last agreed size, unit sectors */ | |
39a1aa7f | 750 | spinlock_t uuid_lock; |
b411b363 PR |
751 | u64 uuid[UI_SIZE]; |
752 | u64 device_uuid; | |
753 | u32 flags; | |
754 | u32 md_size_sect; | |
755 | ||
756 | s32 al_offset; /* signed relative sector offset to al area */ | |
757 | s32 bm_offset; /* signed relative sector offset to bitmap */ | |
758 | ||
759 | /* u32 al_nr_extents; important for restoring the AL | |
f399002e | 760 | * is stored into ldev->dc.al_extents, which in turn |
b411b363 PR |
761 | * gets applied to act_log->nr_elements |
762 | */ | |
763 | }; | |
764 | ||
b411b363 PR |
765 | struct drbd_backing_dev { |
766 | struct block_device *backing_bdev; | |
767 | struct block_device *md_bdev; | |
b411b363 | 768 | struct drbd_md md; |
daeda1cc | 769 | struct disk_conf *disk_conf; /* RCU, for updates: mdev->tconn->conf_update */ |
b411b363 PR |
770 | sector_t known_size; /* last known size of that backing device */ |
771 | }; | |
772 | ||
773 | struct drbd_md_io { | |
cdfda633 | 774 | unsigned int done; |
b411b363 PR |
775 | int error; |
776 | }; | |
777 | ||
778 | struct bm_io_work { | |
779 | struct drbd_work w; | |
780 | char *why; | |
20ceb2b2 | 781 | enum bm_flag flags; |
b411b363 PR |
782 | int (*io_fn)(struct drbd_conf *mdev); |
783 | void (*done)(struct drbd_conf *mdev, int rv); | |
784 | }; | |
785 | ||
786 | enum write_ordering_e { | |
787 | WO_none, | |
788 | WO_drain_io, | |
789 | WO_bdev_flush, | |
b411b363 PR |
790 | }; |
791 | ||
778f271d | 792 | struct fifo_buffer { |
778f271d PR |
793 | unsigned int head_index; |
794 | unsigned int size; | |
9958c857 PR |
795 | int total; /* sum of all values */ |
796 | int values[0]; | |
778f271d | 797 | }; |
9958c857 | 798 | extern struct fifo_buffer *fifo_alloc(int fifo_size); |
778f271d | 799 | |
01a311a5 PR |
800 | /* flag bits per tconn */ |
801 | enum { | |
802 | NET_CONGESTED, /* The data socket is congested */ | |
427c0434 | 803 | RESOLVE_CONFLICTS, /* Set on one node, cleared on the peer! */ |
e43ef195 | 804 | SEND_PING, /* whether asender should send a ping asap */ |
808e37b8 | 805 | SIGNAL_ASENDER, /* whether asender wants to be interrupted */ |
2a67d8b9 | 806 | GOT_PING_ACK, /* set when we receive a ping_ack packet, ping_wait gets woken */ |
4d0fc3fd | 807 | CONN_WD_ST_CHG_REQ, /* A cluster wide state change on the connection is active */ |
fc3b10a4 PR |
808 | CONN_WD_ST_CHG_OKAY, |
809 | CONN_WD_ST_CHG_FAIL, | |
8169e41b | 810 | CONN_DRY_RUN, /* Expect disconnect after resync handshake. */ |
6936fcb4 | 811 | CREATE_BARRIER, /* next P_DATA is preceded by a P_BARRIER */ |
a1096a6e | 812 | STATE_SENT, /* Do not change state/UUIDs while this is set */ |
6f3465ed LE |
813 | CALLBACK_PENDING, /* Whether we have a call_usermodehelper(, UMH_WAIT_PROC) |
814 | * pending, from drbd worker context. | |
815 | * If set, bdi_write_congested() returns true, | |
816 | * so shrink_page_list() would not recurse into, | |
817 | * and potentially deadlock on, this drbd worker. | |
818 | */ | |
b66623e3 | 819 | DISCONNECT_SENT, |
01a311a5 PR |
820 | }; |
821 | ||
2111438b PR |
822 | struct drbd_tconn { /* is a resource from the config file */ |
823 | char *name; /* Resource name */ | |
543cc10b | 824 | struct list_head all_tconn; /* linked on global drbd_tconns */ |
9dc9fbb3 | 825 | struct kref kref; |
8410da8f PR |
826 | struct idr volumes; /* <tconn, vnr> to mdev mapping */ |
827 | enum drbd_conns cstate; /* Only C_STANDALONE to C_WF_REPORT_PARAMS */ | |
8e0af25f PR |
828 | unsigned susp:1; /* IO suspended by user */ |
829 | unsigned susp_nod:1; /* IO suspended because no data */ | |
830 | unsigned susp_fen:1; /* IO suspended because fence peer handler runs */ | |
8410da8f | 831 | struct mutex cstate_mutex; /* Protects graceful disconnects */ |
2111438b | 832 | |
062e879c | 833 | unsigned long flags; |
44ed167d | 834 | struct net_conf *net_conf; /* content protected by rcu */ |
a0095508 | 835 | struct mutex conf_update; /* mutex for ready-copy-update of net_conf and disk_conf */ |
91fd4dad | 836 | wait_queue_head_t ping_wait; /* Woken upon reception of a ping, and a state change */ |
f399002e | 837 | struct res_opts res_opts; |
e42325a5 | 838 | |
089c075d AG |
839 | struct sockaddr_storage my_addr; |
840 | int my_addr_len; | |
841 | struct sockaddr_storage peer_addr; | |
842 | int peer_addr_len; | |
843 | ||
e42325a5 PR |
844 | struct drbd_socket data; /* data/barrier/cstate/parameter packets */ |
845 | struct drbd_socket meta; /* ping/ack (metadata) packets */ | |
31890f4a PR |
846 | int agreed_pro_version; /* actually used protocol version */ |
847 | unsigned long last_received; /* in jiffies, either socket */ | |
848 | unsigned int ko_count; | |
e6b3ea83 | 849 | |
87eeee41 | 850 | spinlock_t req_lock; |
b6dd1a89 LE |
851 | |
852 | struct list_head transfer_log; /* all requests not yet fully processed */ | |
87eeee41 | 853 | |
a0638456 | 854 | struct crypto_hash *cram_hmac_tfm; |
46e1ce41 | 855 | struct crypto_hash *integrity_tfm; /* checksums we compute, updates protected by tconn->data->mutex */ |
036b17ea | 856 | struct crypto_hash *peer_integrity_tfm; /* checksums we verify, only accessed from receiver thread */ |
f399002e LE |
857 | struct crypto_hash *csums_tfm; |
858 | struct crypto_hash *verify_tfm; | |
a0638456 PR |
859 | void *int_dig_in; |
860 | void *int_dig_vv; | |
861 | ||
b6dd1a89 | 862 | /* receiver side */ |
12038a3a PR |
863 | struct drbd_epoch *current_epoch; |
864 | spinlock_t epoch_lock; | |
865 | unsigned int epochs; | |
4b0007c0 | 866 | enum write_ordering_e write_ordering; |
b379c41e | 867 | atomic_t current_tle_nr; /* transfer log epoch number */ |
b6dd1a89 | 868 | unsigned current_tle_writes; /* writes seen within this tl epoch */ |
4b0007c0 | 869 | |
07be15b1 | 870 | unsigned long last_reconnect_jif; |
e6b3ea83 PR |
871 | struct drbd_thread receiver; |
872 | struct drbd_thread worker; | |
873 | struct drbd_thread asender; | |
80822284 | 874 | cpumask_var_t cpu_mask; |
b6dd1a89 LE |
875 | |
876 | /* sender side */ | |
d5b27b01 | 877 | struct drbd_work_queue sender_work; |
b6dd1a89 LE |
878 | |
879 | struct { | |
880 | /* whether this sender thread | |
881 | * has processed a single write yet. */ | |
882 | bool seen_any_write_yet; | |
883 | ||
884 | /* Which barrier number to send with the next P_BARRIER */ | |
885 | int current_epoch_nr; | |
886 | ||
887 | /* how many write requests have been sent | |
888 | * with req->epoch == current_epoch_nr. | |
889 | * If none, no P_BARRIER will be sent. */ | |
890 | unsigned current_epoch_writes; | |
891 | } send; | |
2111438b PR |
892 | }; |
893 | ||
b411b363 | 894 | struct drbd_conf { |
2111438b PR |
895 | struct drbd_tconn *tconn; |
896 | int vnr; /* volume number within the connection */ | |
81fa2e67 | 897 | struct kref kref; |
2111438b | 898 | |
b411b363 PR |
899 | /* things that are stored as / read from meta data on disk */ |
900 | unsigned long flags; | |
901 | ||
902 | /* configured by drbdsetup */ | |
b411b363 PR |
903 | struct drbd_backing_dev *ldev __protected_by(local); |
904 | ||
905 | sector_t p_size; /* partner's disk size */ | |
906 | struct request_queue *rq_queue; | |
907 | struct block_device *this_bdev; | |
908 | struct gendisk *vdisk; | |
909 | ||
07be15b1 | 910 | unsigned long last_reattach_jif; |
b411b363 PR |
911 | struct drbd_work resync_work, |
912 | unplug_work, | |
e9e6f3ec | 913 | go_diskless, |
c4752ef1 PR |
914 | md_sync_work, |
915 | start_resync_work; | |
b411b363 PR |
916 | struct timer_list resync_timer; |
917 | struct timer_list md_sync_timer; | |
370a43e7 | 918 | struct timer_list start_resync_timer; |
7fde2be9 | 919 | struct timer_list request_timer; |
ee15b038 LE |
920 | #ifdef DRBD_DEBUG_MD_SYNC |
921 | struct { | |
922 | unsigned int line; | |
923 | const char* func; | |
924 | } last_md_mark_dirty; | |
925 | #endif | |
b411b363 PR |
926 | |
927 | /* Used after attach while negotiating new disk state. */ | |
928 | union drbd_state new_state_tmp; | |
929 | ||
da9fbc27 | 930 | union drbd_dev_state state; |
b411b363 PR |
931 | wait_queue_head_t misc_wait; |
932 | wait_queue_head_t state_wait; /* upon each state change. */ | |
933 | unsigned int send_cnt; | |
934 | unsigned int recv_cnt; | |
935 | unsigned int read_cnt; | |
936 | unsigned int writ_cnt; | |
937 | unsigned int al_writ_cnt; | |
938 | unsigned int bm_writ_cnt; | |
939 | atomic_t ap_bio_cnt; /* Requests we need to complete */ | |
940 | atomic_t ap_pending_cnt; /* AP data packets on the wire, ack expected */ | |
941 | atomic_t rs_pending_cnt; /* RS request/data packets on the wire */ | |
d942ae44 | 942 | atomic_t unacked_cnt; /* Need to send replies for */ |
b411b363 | 943 | atomic_t local_cnt; /* Waiting for local completion */ |
b2fb6dbe | 944 | |
dac1389c AG |
945 | /* Interval tree of pending local requests */ |
946 | struct rb_root read_requests; | |
de696716 AG |
947 | struct rb_root write_requests; |
948 | ||
4b0715f0 | 949 | /* blocks to resync in this run [unit BM_BLOCK_SIZE] */ |
b411b363 | 950 | unsigned long rs_total; |
4b0715f0 | 951 | /* number of resync blocks that failed in this run */ |
b411b363 PR |
952 | unsigned long rs_failed; |
953 | /* Syncer's start time [unit jiffies] */ | |
954 | unsigned long rs_start; | |
955 | /* cumulated time in PausedSyncX state [unit jiffies] */ | |
956 | unsigned long rs_paused; | |
1d7734a0 LE |
957 | /* skipped because csum was equal [unit BM_BLOCK_SIZE] */ |
958 | unsigned long rs_same_csum; | |
959 | #define DRBD_SYNC_MARKS 8 | |
960 | #define DRBD_SYNC_MARK_STEP (3*HZ) | |
b411b363 | 961 | /* block not up-to-date at mark [unit BM_BLOCK_SIZE] */ |
1d7734a0 | 962 | unsigned long rs_mark_left[DRBD_SYNC_MARKS]; |
b411b363 | 963 | /* marks's time [unit jiffies] */ |
1d7734a0 LE |
964 | unsigned long rs_mark_time[DRBD_SYNC_MARKS]; |
965 | /* current index into rs_mark_{left,time} */ | |
966 | int rs_last_mark; | |
b411b363 PR |
967 | |
968 | /* where does the admin want us to start? (sector) */ | |
969 | sector_t ov_start_sector; | |
58ffa580 | 970 | sector_t ov_stop_sector; |
b411b363 PR |
971 | /* where are we now? (sector) */ |
972 | sector_t ov_position; | |
973 | /* Start sector of out of sync range (to merge printk reporting). */ | |
974 | sector_t ov_last_oos_start; | |
975 | /* size of out-of-sync range in sectors. */ | |
976 | sector_t ov_last_oos_size; | |
977 | unsigned long ov_left; /* in bits */ | |
b411b363 | 978 | |
b411b363 PR |
979 | struct drbd_bitmap *bitmap; |
980 | unsigned long bm_resync_fo; /* bit offset for drbd_bm_find_next */ | |
981 | ||
982 | /* Used to track operations of resync... */ | |
983 | struct lru_cache *resync; | |
984 | /* Number of locked elements in resync LRU */ | |
985 | unsigned int resync_locked; | |
986 | /* resync extent number waiting for application requests */ | |
987 | unsigned int resync_wenr; | |
988 | ||
989 | int open_cnt; | |
990 | u64 *p_uuid; | |
4b0007c0 | 991 | |
85719573 PR |
992 | struct list_head active_ee; /* IO in progress (P_DATA gets written to disk) */ |
993 | struct list_head sync_ee; /* IO in progress (P_RS_DATA_REPLY gets written to disk) */ | |
18b75d75 AG |
994 | struct list_head done_ee; /* need to send P_WRITE_ACK */ |
995 | struct list_head read_ee; /* [RS]P_DATA_REQUEST being read */ | |
b411b363 | 996 | struct list_head net_ee; /* zero-copy network send in progress */ |
b411b363 | 997 | |
b411b363 | 998 | int next_barrier_nr; |
b411b363 | 999 | struct list_head resync_reads; |
435f0740 LE |
1000 | atomic_t pp_in_use; /* allocated from page pool */ |
1001 | atomic_t pp_in_use_by_net; /* sendpage()d, still referenced by tcp */ | |
b411b363 PR |
1002 | wait_queue_head_t ee_wait; |
1003 | struct page *md_io_page; /* one page buffer for md_io */ | |
cdfda633 PR |
1004 | struct drbd_md_io md_io; |
1005 | atomic_t md_io_in_use; /* protects the md_io, md_io_page and md_io_tmpp */ | |
b411b363 PR |
1006 | spinlock_t al_lock; |
1007 | wait_queue_head_t al_wait; | |
1008 | struct lru_cache *act_log; /* activity log */ | |
1009 | unsigned int al_tr_number; | |
1010 | int al_tr_cycle; | |
1011 | int al_tr_pos; /* position of the next transaction in the journal */ | |
b411b363 PR |
1012 | wait_queue_head_t seq_wait; |
1013 | atomic_t packet_seq; | |
1014 | unsigned int peer_seq; | |
1015 | spinlock_t peer_seq_lock; | |
1016 | unsigned int minor; | |
1017 | unsigned long comm_bm_set; /* communicated number of set bits. */ | |
b411b363 PR |
1018 | struct bm_io_work bm_io_work; |
1019 | u64 ed_uuid; /* UUID of the exposed data */ | |
8410da8f PR |
1020 | struct mutex own_state_mutex; |
1021 | struct mutex *state_mutex; /* either own_state_mutex or mdev->tconn->cstate_mutex */ | |
b411b363 | 1022 | char congestion_reason; /* Why we where congested... */ |
1d7734a0 LE |
1023 | atomic_t rs_sect_in; /* for incoming resync data rate, SyncTarget */ |
1024 | atomic_t rs_sect_ev; /* for submitted resync data rate, both */ | |
1025 | int rs_last_sect_ev; /* counter to compare with */ | |
1026 | int rs_last_events; /* counter of read or write "events" (unit sectors) | |
1027 | * on the lower level device when we last looked. */ | |
1028 | int c_sync_rate; /* current resync rate after syncer throttle magic */ | |
813472ce | 1029 | struct fifo_buffer *rs_plan_s; /* correction values of resync planer (RCU, tconn->conn_update) */ |
778f271d | 1030 | int rs_in_flight; /* resync sectors in flight (to proxy, in proxy and from proxy) */ |
759fbdfb | 1031 | atomic_t ap_in_flight; /* App sectors in flight (waiting for ack) */ |
99432fcc PR |
1032 | int peer_max_bio_size; |
1033 | int local_max_bio_size; | |
b411b363 PR |
1034 | }; |
1035 | ||
1036 | static inline struct drbd_conf *minor_to_mdev(unsigned int minor) | |
1037 | { | |
81a5d60e | 1038 | return (struct drbd_conf *)idr_find(&minors, minor); |
b411b363 PR |
1039 | } |
1040 | ||
1041 | static inline unsigned int mdev_to_minor(struct drbd_conf *mdev) | |
1042 | { | |
1043 | return mdev->minor; | |
1044 | } | |
1045 | ||
eefc2f7d PR |
1046 | static inline struct drbd_conf *vnr_to_mdev(struct drbd_tconn *tconn, int vnr) |
1047 | { | |
1048 | return (struct drbd_conf *)idr_find(&tconn->volumes, vnr); | |
1049 | } | |
1050 | ||
b411b363 PR |
1051 | /* |
1052 | * function declarations | |
1053 | *************************/ | |
1054 | ||
1055 | /* drbd_main.c */ | |
1056 | ||
e89b591c PR |
1057 | enum dds_flags { |
1058 | DDSF_FORCED = 1, | |
1059 | DDSF_NO_RESYNC = 2, /* Do not run a resync for the new space */ | |
1060 | }; | |
1061 | ||
b411b363 | 1062 | extern void drbd_init_set_defaults(struct drbd_conf *mdev); |
b411b363 PR |
1063 | extern int drbd_thread_start(struct drbd_thread *thi); |
1064 | extern void _drbd_thread_stop(struct drbd_thread *thi, int restart, int wait); | |
392c8801 | 1065 | extern char *drbd_task_to_thread_name(struct drbd_tconn *tconn, struct task_struct *task); |
b411b363 | 1066 | #ifdef CONFIG_SMP |
80822284 PR |
1067 | extern void drbd_thread_current_set_cpu(struct drbd_thread *thi); |
1068 | extern void drbd_calc_cpu_mask(struct drbd_tconn *tconn); | |
b411b363 | 1069 | #else |
80822284 | 1070 | #define drbd_thread_current_set_cpu(A) ({}) |
b411b363 PR |
1071 | #define drbd_calc_cpu_mask(A) ({}) |
1072 | #endif | |
2f5cdd0b | 1073 | extern void tl_release(struct drbd_tconn *, unsigned int barrier_nr, |
b411b363 | 1074 | unsigned int set_size); |
2f5cdd0b | 1075 | extern void tl_clear(struct drbd_tconn *); |
360cc740 | 1076 | extern void drbd_free_sock(struct drbd_tconn *tconn); |
bedbd2a5 PR |
1077 | extern int drbd_send(struct drbd_tconn *tconn, struct socket *sock, |
1078 | void *buf, size_t size, unsigned msg_flags); | |
fb708e40 AG |
1079 | extern int drbd_send_all(struct drbd_tconn *, struct socket *, void *, size_t, |
1080 | unsigned); | |
1081 | ||
d659f2aa | 1082 | extern int __drbd_send_protocol(struct drbd_tconn *tconn, enum drbd_packet cmd); |
dc8228d1 | 1083 | extern int drbd_send_protocol(struct drbd_tconn *tconn); |
b411b363 PR |
1084 | extern int drbd_send_uuids(struct drbd_conf *mdev); |
1085 | extern int drbd_send_uuids_skip_initial_sync(struct drbd_conf *mdev); | |
9c1b7f72 | 1086 | extern void drbd_gen_and_send_sync_uuid(struct drbd_conf *mdev); |
e89b591c | 1087 | extern int drbd_send_sizes(struct drbd_conf *mdev, int trigger_reply, enum dds_flags flags); |
43de7c85 PR |
1088 | extern int drbd_send_state(struct drbd_conf *mdev, union drbd_state s); |
1089 | extern int drbd_send_current_state(struct drbd_conf *mdev); | |
f399002e | 1090 | extern int drbd_send_sync_param(struct drbd_conf *mdev); |
9ed57dcb | 1091 | extern void drbd_send_b_ack(struct drbd_tconn *tconn, u32 barrier_nr, |
d4e67d7c | 1092 | u32 set_size); |
f6ffca9f AG |
1093 | extern int drbd_send_ack(struct drbd_conf *, enum drbd_packet, |
1094 | struct drbd_peer_request *); | |
a9a9994d AG |
1095 | extern void drbd_send_ack_rp(struct drbd_conf *mdev, enum drbd_packet cmd, |
1096 | struct p_block_req *rp); | |
1097 | extern void drbd_send_ack_dp(struct drbd_conf *mdev, enum drbd_packet cmd, | |
1098 | struct p_data *dp, int data_size); | |
d8763023 | 1099 | extern int drbd_send_ack_ex(struct drbd_conf *mdev, enum drbd_packet cmd, |
b411b363 | 1100 | sector_t sector, int blksize, u64 block_id); |
8f7bed77 | 1101 | extern int drbd_send_out_of_sync(struct drbd_conf *, struct drbd_request *); |
f6ffca9f AG |
1102 | extern int drbd_send_block(struct drbd_conf *, enum drbd_packet, |
1103 | struct drbd_peer_request *); | |
b411b363 | 1104 | extern int drbd_send_dblock(struct drbd_conf *mdev, struct drbd_request *req); |
b411b363 PR |
1105 | extern int drbd_send_drequest(struct drbd_conf *mdev, int cmd, |
1106 | sector_t sector, int size, u64 block_id); | |
d8763023 AG |
1107 | extern int drbd_send_drequest_csum(struct drbd_conf *mdev, sector_t sector, |
1108 | int size, void *digest, int digest_size, | |
1109 | enum drbd_packet cmd); | |
b411b363 PR |
1110 | extern int drbd_send_ov_request(struct drbd_conf *mdev,sector_t sector,int size); |
1111 | ||
1112 | extern int drbd_send_bitmap(struct drbd_conf *mdev); | |
2f4e7abe | 1113 | extern void drbd_send_sr_reply(struct drbd_conf *mdev, enum drbd_state_rv retcode); |
9f5bdc33 | 1114 | extern void conn_send_sr_reply(struct drbd_tconn *tconn, enum drbd_state_rv retcode); |
b411b363 PR |
1115 | extern void drbd_free_bc(struct drbd_backing_dev *ldev); |
1116 | extern void drbd_mdev_cleanup(struct drbd_conf *mdev); | |
62b0da3a | 1117 | void drbd_print_uuids(struct drbd_conf *mdev, const char *text); |
b411b363 | 1118 | |
19fffd7b | 1119 | extern void conn_md_sync(struct drbd_tconn *tconn); |
b411b363 PR |
1120 | extern void drbd_md_sync(struct drbd_conf *mdev); |
1121 | extern int drbd_md_read(struct drbd_conf *mdev, struct drbd_backing_dev *bdev); | |
b411b363 PR |
1122 | extern void drbd_uuid_set(struct drbd_conf *mdev, int idx, u64 val) __must_hold(local); |
1123 | extern void _drbd_uuid_set(struct drbd_conf *mdev, int idx, u64 val) __must_hold(local); | |
1124 | extern void drbd_uuid_new_current(struct drbd_conf *mdev) __must_hold(local); | |
b411b363 | 1125 | extern void drbd_uuid_set_bm(struct drbd_conf *mdev, u64 val) __must_hold(local); |
39a1aa7f PR |
1126 | extern void drbd_uuid_move_history(struct drbd_conf *mdev) __must_hold(local); |
1127 | extern void __drbd_uuid_set(struct drbd_conf *mdev, int idx, u64 val) __must_hold(local); | |
b411b363 PR |
1128 | extern void drbd_md_set_flag(struct drbd_conf *mdev, int flags) __must_hold(local); |
1129 | extern void drbd_md_clear_flag(struct drbd_conf *mdev, int flags)__must_hold(local); | |
1130 | extern int drbd_md_test_flag(struct drbd_backing_dev *, int); | |
ee15b038 | 1131 | #ifndef DRBD_DEBUG_MD_SYNC |
b411b363 | 1132 | extern void drbd_md_mark_dirty(struct drbd_conf *mdev); |
ee15b038 LE |
1133 | #else |
1134 | #define drbd_md_mark_dirty(m) drbd_md_mark_dirty_(m, __LINE__ , __func__ ) | |
1135 | extern void drbd_md_mark_dirty_(struct drbd_conf *mdev, | |
1136 | unsigned int line, const char *func); | |
1137 | #endif | |
b411b363 PR |
1138 | extern void drbd_queue_bitmap_io(struct drbd_conf *mdev, |
1139 | int (*io_fn)(struct drbd_conf *), | |
1140 | void (*done)(struct drbd_conf *, int), | |
20ceb2b2 LE |
1141 | char *why, enum bm_flag flags); |
1142 | extern int drbd_bitmap_io(struct drbd_conf *mdev, | |
1143 | int (*io_fn)(struct drbd_conf *), | |
1144 | char *why, enum bm_flag flags); | |
b411b363 PR |
1145 | extern int drbd_bmio_set_n_write(struct drbd_conf *mdev); |
1146 | extern int drbd_bmio_clear_n_write(struct drbd_conf *mdev); | |
e9e6f3ec | 1147 | extern void drbd_go_diskless(struct drbd_conf *mdev); |
82f59cc6 | 1148 | extern void drbd_ldev_destroy(struct drbd_conf *mdev); |
b411b363 | 1149 | |
b411b363 PR |
1150 | /* Meta data layout |
1151 | We reserve a 128MB Block (4k aligned) | |
1152 | * either at the end of the backing device | |
3ad2f3fb | 1153 | * or on a separate meta data device. */ |
b411b363 | 1154 | |
b411b363 | 1155 | /* The following numbers are sectors */ |
7ad651b5 LE |
1156 | /* Allows up to about 3.8TB, so if you want more, |
1157 | * you need to use the "flexible" meta data format. */ | |
1158 | #define MD_RESERVED_SECT (128LU << 11) /* 128 MB, unit sectors */ | |
1159 | #define MD_AL_OFFSET 8 /* 8 Sectors after start of meta area */ | |
1160 | #define MD_AL_SECTORS 64 /* = 32 kB on disk activity log ring buffer */ | |
1161 | #define MD_BM_OFFSET (MD_AL_OFFSET + MD_AL_SECTORS) | |
1162 | ||
1163 | /* we do all meta data IO in 4k blocks */ | |
1164 | #define MD_BLOCK_SHIFT 12 | |
1165 | #define MD_BLOCK_SIZE (1<<MD_BLOCK_SHIFT) | |
1166 | ||
1167 | /* One activity log extent represents 4M of storage */ | |
1168 | #define AL_EXTENT_SHIFT 22 | |
b411b363 PR |
1169 | #define AL_EXTENT_SIZE (1<<AL_EXTENT_SHIFT) |
1170 | ||
7ad651b5 LE |
1171 | /* We could make these currently hardcoded constants configurable |
1172 | * variables at create-md time (or even re-configurable at runtime?). | |
1173 | * Which will require some more changes to the DRBD "super block" | |
1174 | * and attach code. | |
1175 | * | |
1176 | * updates per transaction: | |
1177 | * This many changes to the active set can be logged with one transaction. | |
1178 | * This number is arbitrary. | |
1179 | * context per transaction: | |
1180 | * This many context extent numbers are logged with each transaction. | |
1181 | * This number is resulting from the transaction block size (4k), the layout | |
1182 | * of the transaction header, and the number of updates per transaction. | |
1183 | * See drbd_actlog.c:struct al_transaction_on_disk | |
1184 | * */ | |
1185 | #define AL_UPDATES_PER_TRANSACTION 64 // arbitrary | |
1186 | #define AL_CONTEXT_PER_TRANSACTION 919 // (4096 - 36 - 6*64)/4 | |
1187 | ||
b411b363 PR |
1188 | #if BITS_PER_LONG == 32 |
1189 | #define LN2_BPL 5 | |
1190 | #define cpu_to_lel(A) cpu_to_le32(A) | |
1191 | #define lel_to_cpu(A) le32_to_cpu(A) | |
1192 | #elif BITS_PER_LONG == 64 | |
1193 | #define LN2_BPL 6 | |
1194 | #define cpu_to_lel(A) cpu_to_le64(A) | |
1195 | #define lel_to_cpu(A) le64_to_cpu(A) | |
1196 | #else | |
1197 | #error "LN2 of BITS_PER_LONG unknown!" | |
1198 | #endif | |
1199 | ||
1200 | /* resync bitmap */ | |
1201 | /* 16MB sized 'bitmap extent' to track syncer usage */ | |
1202 | struct bm_extent { | |
1203 | int rs_left; /* number of bits set (out of sync) in this extent. */ | |
1204 | int rs_failed; /* number of failed resync requests in this extent. */ | |
1205 | unsigned long flags; | |
1206 | struct lc_element lce; | |
1207 | }; | |
1208 | ||
1209 | #define BME_NO_WRITES 0 /* bm_extent.flags: no more requests on this one! */ | |
1210 | #define BME_LOCKED 1 /* bm_extent.flags: syncer active on this one. */ | |
e3555d85 | 1211 | #define BME_PRIORITY 2 /* finish resync IO on this extent ASAP! App IO waiting! */ |
b411b363 PR |
1212 | |
1213 | /* drbd_bitmap.c */ | |
1214 | /* | |
1215 | * We need to store one bit for a block. | |
1216 | * Example: 1GB disk @ 4096 byte blocks ==> we need 32 KB bitmap. | |
1217 | * Bit 0 ==> local node thinks this block is binary identical on both nodes | |
1218 | * Bit 1 ==> local node thinks this block needs to be synced. | |
1219 | */ | |
1220 | ||
8e26f9cc PR |
1221 | #define SLEEP_TIME (HZ/10) |
1222 | ||
45dfffeb LE |
1223 | /* We do bitmap IO in units of 4k blocks. |
1224 | * We also still have a hardcoded 4k per bit relation. */ | |
1225 | #define BM_BLOCK_SHIFT 12 /* 4k per bit */ | |
b411b363 | 1226 | #define BM_BLOCK_SIZE (1<<BM_BLOCK_SHIFT) |
45dfffeb LE |
1227 | /* mostly arbitrarily set the represented size of one bitmap extent, |
1228 | * aka resync extent, to 16 MiB (which is also 512 Byte worth of bitmap | |
1229 | * at 4k per bit resolution) */ | |
1230 | #define BM_EXT_SHIFT 24 /* 16 MiB per resync extent */ | |
b411b363 PR |
1231 | #define BM_EXT_SIZE (1<<BM_EXT_SHIFT) |
1232 | ||
1233 | #if (BM_EXT_SHIFT != 24) || (BM_BLOCK_SHIFT != 12) | |
1234 | #error "HAVE YOU FIXED drbdmeta AS WELL??" | |
1235 | #endif | |
1236 | ||
1237 | /* thus many _storage_ sectors are described by one bit */ | |
1238 | #define BM_SECT_TO_BIT(x) ((x)>>(BM_BLOCK_SHIFT-9)) | |
1239 | #define BM_BIT_TO_SECT(x) ((sector_t)(x)<<(BM_BLOCK_SHIFT-9)) | |
1240 | #define BM_SECT_PER_BIT BM_BIT_TO_SECT(1) | |
1241 | ||
1242 | /* bit to represented kilo byte conversion */ | |
1243 | #define Bit2KB(bits) ((bits)<<(BM_BLOCK_SHIFT-10)) | |
1244 | ||
1245 | /* in which _bitmap_ extent (resp. sector) the bit for a certain | |
1246 | * _storage_ sector is located in */ | |
1247 | #define BM_SECT_TO_EXT(x) ((x)>>(BM_EXT_SHIFT-9)) | |
1248 | ||
1249 | /* how much _storage_ sectors we have per bitmap sector */ | |
1250 | #define BM_EXT_TO_SECT(x) ((sector_t)(x) << (BM_EXT_SHIFT-9)) | |
1251 | #define BM_SECT_PER_EXT BM_EXT_TO_SECT(1) | |
1252 | ||
1253 | /* in one sector of the bitmap, we have this many activity_log extents. */ | |
1254 | #define AL_EXT_PER_BM_SECT (1 << (BM_EXT_SHIFT - AL_EXTENT_SHIFT)) | |
1255 | #define BM_WORDS_PER_AL_EXT (1 << (AL_EXTENT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) | |
1256 | ||
1257 | #define BM_BLOCKS_PER_BM_EXT_B (BM_EXT_SHIFT - BM_BLOCK_SHIFT) | |
1258 | #define BM_BLOCKS_PER_BM_EXT_MASK ((1<<BM_BLOCKS_PER_BM_EXT_B) - 1) | |
1259 | ||
1260 | /* the extent in "PER_EXTENT" below is an activity log extent | |
1261 | * we need that many (long words/bytes) to store the bitmap | |
1262 | * of one AL_EXTENT_SIZE chunk of storage. | |
1263 | * we can store the bitmap for that many AL_EXTENTS within | |
1264 | * one sector of the _on_disk_ bitmap: | |
1265 | * bit 0 bit 37 bit 38 bit (512*8)-1 | |
1266 | * ...|........|........|.. // ..|........| | |
1267 | * sect. 0 `296 `304 ^(512*8*8)-1 | |
1268 | * | |
1269 | #define BM_WORDS_PER_EXT ( (AL_EXT_SIZE/BM_BLOCK_SIZE) / BITS_PER_LONG ) | |
1270 | #define BM_BYTES_PER_EXT ( (AL_EXT_SIZE/BM_BLOCK_SIZE) / 8 ) // 128 | |
1271 | #define BM_EXT_PER_SECT ( 512 / BM_BYTES_PER_EXTENT ) // 4 | |
1272 | */ | |
1273 | ||
1274 | #define DRBD_MAX_SECTORS_32 (0xffffffffLU) | |
1275 | #define DRBD_MAX_SECTORS_BM \ | |
1276 | ((MD_RESERVED_SECT - MD_BM_OFFSET) * (1LL<<(BM_EXT_SHIFT-9))) | |
1277 | #if DRBD_MAX_SECTORS_BM < DRBD_MAX_SECTORS_32 | |
1278 | #define DRBD_MAX_SECTORS DRBD_MAX_SECTORS_BM | |
1279 | #define DRBD_MAX_SECTORS_FLEX DRBD_MAX_SECTORS_BM | |
36bfc7e2 | 1280 | #elif !defined(CONFIG_LBDAF) && BITS_PER_LONG == 32 |
b411b363 PR |
1281 | #define DRBD_MAX_SECTORS DRBD_MAX_SECTORS_32 |
1282 | #define DRBD_MAX_SECTORS_FLEX DRBD_MAX_SECTORS_32 | |
1283 | #else | |
1284 | #define DRBD_MAX_SECTORS DRBD_MAX_SECTORS_BM | |
1285 | /* 16 TB in units of sectors */ | |
1286 | #if BITS_PER_LONG == 32 | |
1287 | /* adjust by one page worth of bitmap, | |
1288 | * so we won't wrap around in drbd_bm_find_next_bit. | |
1289 | * you should use 64bit OS for that much storage, anyways. */ | |
1290 | #define DRBD_MAX_SECTORS_FLEX BM_BIT_TO_SECT(0xffff7fff) | |
1291 | #else | |
4b0715f0 LE |
1292 | /* we allow up to 1 PiB now on 64bit architecture with "flexible" meta data */ |
1293 | #define DRBD_MAX_SECTORS_FLEX (1UL << 51) | |
1294 | /* corresponds to (1UL << 38) bits right now. */ | |
b411b363 PR |
1295 | #endif |
1296 | #endif | |
1297 | ||
23361cf3 LE |
1298 | /* BIO_MAX_SIZE is 256 * PAGE_CACHE_SIZE, |
1299 | * so for typical PAGE_CACHE_SIZE of 4k, that is (1<<20) Byte. | |
1300 | * Since we may live in a mixed-platform cluster, | |
1301 | * we limit us to a platform agnostic constant here for now. | |
1302 | * A followup commit may allow even bigger BIO sizes, | |
1303 | * once we thought that through. */ | |
1304 | #define DRBD_MAX_BIO_SIZE (1 << 20) | |
1305 | #if DRBD_MAX_BIO_SIZE > BIO_MAX_SIZE | |
1306 | #error Architecture not supported: DRBD_MAX_BIO_SIZE > BIO_MAX_SIZE | |
1307 | #endif | |
99432fcc | 1308 | #define DRBD_MAX_BIO_SIZE_SAFE (1 << 12) /* Works always = 4k */ |
b411b363 | 1309 | |
2ffca4f3 PR |
1310 | #define DRBD_MAX_SIZE_H80_PACKET (1 << 15) /* Header 80 only allows packets up to 32KiB data */ |
1311 | #define DRBD_MAX_BIO_SIZE_P95 (1 << 17) /* Protocol 95 to 99 allows bios up to 128KiB */ | |
b411b363 PR |
1312 | |
1313 | extern int drbd_bm_init(struct drbd_conf *mdev); | |
02d9a94b | 1314 | extern int drbd_bm_resize(struct drbd_conf *mdev, sector_t sectors, int set_new_bits); |
b411b363 PR |
1315 | extern void drbd_bm_cleanup(struct drbd_conf *mdev); |
1316 | extern void drbd_bm_set_all(struct drbd_conf *mdev); | |
1317 | extern void drbd_bm_clear_all(struct drbd_conf *mdev); | |
4b0715f0 | 1318 | /* set/clear/test only a few bits at a time */ |
b411b363 PR |
1319 | extern int drbd_bm_set_bits( |
1320 | struct drbd_conf *mdev, unsigned long s, unsigned long e); | |
1321 | extern int drbd_bm_clear_bits( | |
1322 | struct drbd_conf *mdev, unsigned long s, unsigned long e); | |
4b0715f0 LE |
1323 | extern int drbd_bm_count_bits( |
1324 | struct drbd_conf *mdev, const unsigned long s, const unsigned long e); | |
1325 | /* bm_set_bits variant for use while holding drbd_bm_lock, | |
1326 | * may process the whole bitmap in one go */ | |
b411b363 PR |
1327 | extern void _drbd_bm_set_bits(struct drbd_conf *mdev, |
1328 | const unsigned long s, const unsigned long e); | |
1329 | extern int drbd_bm_test_bit(struct drbd_conf *mdev, unsigned long bitnr); | |
1330 | extern int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr); | |
19f843aa | 1331 | extern int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local); |
b411b363 | 1332 | extern int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local); |
45dfffeb | 1333 | extern void drbd_bm_mark_for_writeout(struct drbd_conf *mdev, int page_nr); |
b411b363 | 1334 | extern int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local); |
45dfffeb | 1335 | extern int drbd_bm_write_hinted(struct drbd_conf *mdev) __must_hold(local); |
fef45d29 | 1336 | extern int drbd_bm_write_all(struct drbd_conf *mdev) __must_hold(local); |
a220d291 | 1337 | extern int drbd_bm_write_copy_pages(struct drbd_conf *mdev) __must_hold(local); |
b411b363 PR |
1338 | extern size_t drbd_bm_words(struct drbd_conf *mdev); |
1339 | extern unsigned long drbd_bm_bits(struct drbd_conf *mdev); | |
1340 | extern sector_t drbd_bm_capacity(struct drbd_conf *mdev); | |
4b0715f0 LE |
1341 | |
1342 | #define DRBD_END_OF_BITMAP (~(unsigned long)0) | |
b411b363 PR |
1343 | extern unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo); |
1344 | /* bm_find_next variants for use while you hold drbd_bm_lock() */ | |
1345 | extern unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo); | |
1346 | extern unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo); | |
0778286a | 1347 | extern unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev); |
b411b363 PR |
1348 | extern unsigned long drbd_bm_total_weight(struct drbd_conf *mdev); |
1349 | extern int drbd_bm_rs_done(struct drbd_conf *mdev); | |
1350 | /* for receive_bitmap */ | |
1351 | extern void drbd_bm_merge_lel(struct drbd_conf *mdev, size_t offset, | |
1352 | size_t number, unsigned long *buffer); | |
19f843aa | 1353 | /* for _drbd_send_bitmap */ |
b411b363 PR |
1354 | extern void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset, |
1355 | size_t number, unsigned long *buffer); | |
1356 | ||
20ceb2b2 | 1357 | extern void drbd_bm_lock(struct drbd_conf *mdev, char *why, enum bm_flag flags); |
b411b363 | 1358 | extern void drbd_bm_unlock(struct drbd_conf *mdev); |
b411b363 PR |
1359 | /* drbd_main.c */ |
1360 | ||
1361 | extern struct kmem_cache *drbd_request_cache; | |
6c852bec | 1362 | extern struct kmem_cache *drbd_ee_cache; /* peer requests */ |
b411b363 PR |
1363 | extern struct kmem_cache *drbd_bm_ext_cache; /* bitmap extents */ |
1364 | extern struct kmem_cache *drbd_al_ext_cache; /* activity log extents */ | |
1365 | extern mempool_t *drbd_request_mempool; | |
1366 | extern mempool_t *drbd_ee_mempool; | |
1367 | ||
35abf594 LE |
1368 | /* drbd's page pool, used to buffer data received from the peer, |
1369 | * or data requested by the peer. | |
1370 | * | |
1371 | * This does not have an emergency reserve. | |
1372 | * | |
1373 | * When allocating from this pool, it first takes pages from the pool. | |
1374 | * Only if the pool is depleted will try to allocate from the system. | |
1375 | * | |
1376 | * The assumption is that pages taken from this pool will be processed, | |
1377 | * and given back, "quickly", and then can be recycled, so we can avoid | |
1378 | * frequent calls to alloc_page(), and still will be able to make progress even | |
1379 | * under memory pressure. | |
1380 | */ | |
1381 | extern struct page *drbd_pp_pool; | |
b411b363 PR |
1382 | extern spinlock_t drbd_pp_lock; |
1383 | extern int drbd_pp_vacant; | |
1384 | extern wait_queue_head_t drbd_pp_wait; | |
1385 | ||
35abf594 LE |
1386 | /* We also need a standard (emergency-reserve backed) page pool |
1387 | * for meta data IO (activity log, bitmap). | |
1388 | * We can keep it global, as long as it is used as "N pages at a time". | |
1389 | * 128 should be plenty, currently we probably can get away with as few as 1. | |
1390 | */ | |
1391 | #define DRBD_MIN_POOL_PAGES 128 | |
1392 | extern mempool_t *drbd_md_io_page_pool; | |
1393 | ||
da4a75d2 LE |
1394 | /* We also need to make sure we get a bio |
1395 | * when we need it for housekeeping purposes */ | |
1396 | extern struct bio_set *drbd_md_io_bio_set; | |
1397 | /* to allocate from that set */ | |
1398 | extern struct bio *bio_alloc_drbd(gfp_t gfp_mask); | |
1399 | ||
b411b363 PR |
1400 | extern rwlock_t global_state_lock; |
1401 | ||
80883197 | 1402 | extern int conn_lowest_minor(struct drbd_tconn *tconn); |
774b3055 | 1403 | enum drbd_ret_code conn_new_minor(struct drbd_tconn *tconn, unsigned int minor, int vnr); |
81fa2e67 | 1404 | extern void drbd_minor_destroy(struct kref *kref); |
b411b363 | 1405 | |
afbbfa88 AG |
1406 | extern int set_resource_options(struct drbd_tconn *tconn, struct res_opts *res_opts); |
1407 | extern struct drbd_tconn *conn_create(const char *name, struct res_opts *res_opts); | |
9dc9fbb3 | 1408 | extern void conn_destroy(struct kref *kref); |
0ace9dfa | 1409 | struct drbd_tconn *conn_get_by_name(const char *name); |
089c075d AG |
1410 | extern struct drbd_tconn *conn_get_by_addrs(void *my_addr, int my_addr_len, |
1411 | void *peer_addr, int peer_addr_len); | |
91fd4dad | 1412 | extern void conn_free_crypto(struct drbd_tconn *tconn); |
2111438b | 1413 | |
b411b363 PR |
1414 | extern int proc_details; |
1415 | ||
1416 | /* drbd_req */ | |
5df69ece | 1417 | extern void __drbd_make_request(struct drbd_conf *, struct bio *, unsigned long); |
2f58dcfc | 1418 | extern int drbd_make_request(struct request_queue *q, struct bio *bio); |
b411b363 PR |
1419 | extern int drbd_read_remote(struct drbd_conf *mdev, struct drbd_request *req); |
1420 | extern int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec); | |
1421 | extern int is_valid_ar_handle(struct drbd_request *, sector_t); | |
1422 | ||
1423 | ||
1424 | /* drbd_nl.c */ | |
8432b314 | 1425 | extern int drbd_msg_put_info(const char *info); |
b411b363 PR |
1426 | extern void drbd_suspend_io(struct drbd_conf *mdev); |
1427 | extern void drbd_resume_io(struct drbd_conf *mdev); | |
1428 | extern char *ppsize(char *buf, unsigned long long size); | |
ef5e44a6 | 1429 | extern sector_t drbd_new_dev_size(struct drbd_conf *, struct drbd_backing_dev *, sector_t, int); |
b411b363 | 1430 | enum determine_dev_size { dev_size_error = -1, unchanged = 0, shrunk = 1, grew = 2 }; |
24c4830c | 1431 | extern enum determine_dev_size drbd_determine_dev_size(struct drbd_conf *, enum dds_flags) __must_hold(local); |
b411b363 | 1432 | extern void resync_after_online_grow(struct drbd_conf *); |
99432fcc | 1433 | extern void drbd_reconsider_max_bio_size(struct drbd_conf *mdev); |
bf885f8a AG |
1434 | extern enum drbd_state_rv drbd_set_role(struct drbd_conf *mdev, |
1435 | enum drbd_role new_role, | |
1436 | int force); | |
cb703454 PR |
1437 | extern bool conn_try_outdate_peer(struct drbd_tconn *tconn); |
1438 | extern void conn_try_outdate_peer_async(struct drbd_tconn *tconn); | |
b411b363 PR |
1439 | extern int drbd_khelper(struct drbd_conf *mdev, char *cmd); |
1440 | ||
1441 | /* drbd_worker.c */ | |
1442 | extern int drbd_worker(struct drbd_thread *thi); | |
95f8efd0 AG |
1443 | enum drbd_ret_code drbd_resync_after_valid(struct drbd_conf *mdev, int o_minor); |
1444 | void drbd_resync_after_changed(struct drbd_conf *mdev); | |
b411b363 PR |
1445 | extern void drbd_start_resync(struct drbd_conf *mdev, enum drbd_conns side); |
1446 | extern void resume_next_sg(struct drbd_conf *mdev); | |
1447 | extern void suspend_other_sg(struct drbd_conf *mdev); | |
1448 | extern int drbd_resync_finished(struct drbd_conf *mdev); | |
1449 | /* maybe rather drbd_main.c ? */ | |
cdfda633 PR |
1450 | extern void *drbd_md_get_buffer(struct drbd_conf *mdev); |
1451 | extern void drbd_md_put_buffer(struct drbd_conf *mdev); | |
b411b363 PR |
1452 | extern int drbd_md_sync_page_io(struct drbd_conf *mdev, |
1453 | struct drbd_backing_dev *bdev, sector_t sector, int rw); | |
8f7bed77 | 1454 | extern void drbd_ov_out_of_sync_found(struct drbd_conf *, sector_t, int); |
e34b677d LE |
1455 | extern void wait_until_done_or_force_detached(struct drbd_conf *mdev, |
1456 | struct drbd_backing_dev *bdev, unsigned int *done); | |
9bd28d3c | 1457 | extern void drbd_rs_controller_reset(struct drbd_conf *mdev); |
b411b363 | 1458 | |
8f7bed77 | 1459 | static inline void ov_out_of_sync_print(struct drbd_conf *mdev) |
b411b363 PR |
1460 | { |
1461 | if (mdev->ov_last_oos_size) { | |
1462 | dev_err(DEV, "Out of sync: start=%llu, size=%lu (sectors)\n", | |
1463 | (unsigned long long)mdev->ov_last_oos_start, | |
1464 | (unsigned long)mdev->ov_last_oos_size); | |
1465 | } | |
1466 | mdev->ov_last_oos_size=0; | |
1467 | } | |
1468 | ||
1469 | ||
45bb912b | 1470 | extern void drbd_csum_bio(struct drbd_conf *, struct crypto_hash *, struct bio *, void *); |
f6ffca9f AG |
1471 | extern void drbd_csum_ee(struct drbd_conf *, struct crypto_hash *, |
1472 | struct drbd_peer_request *, void *); | |
b411b363 | 1473 | /* worker callbacks */ |
99920dc5 AG |
1474 | extern int w_e_end_data_req(struct drbd_work *, int); |
1475 | extern int w_e_end_rsdata_req(struct drbd_work *, int); | |
1476 | extern int w_e_end_csum_rs_req(struct drbd_work *, int); | |
1477 | extern int w_e_end_ov_reply(struct drbd_work *, int); | |
1478 | extern int w_e_end_ov_req(struct drbd_work *, int); | |
1479 | extern int w_ov_finished(struct drbd_work *, int); | |
1480 | extern int w_resync_timer(struct drbd_work *, int); | |
1481 | extern int w_send_write_hint(struct drbd_work *, int); | |
1482 | extern int w_make_resync_request(struct drbd_work *, int); | |
1483 | extern int w_send_dblock(struct drbd_work *, int); | |
99920dc5 AG |
1484 | extern int w_send_read_req(struct drbd_work *, int); |
1485 | extern int w_prev_work_done(struct drbd_work *, int); | |
1486 | extern int w_e_reissue(struct drbd_work *, int); | |
1487 | extern int w_restart_disk_io(struct drbd_work *, int); | |
8f7bed77 | 1488 | extern int w_send_out_of_sync(struct drbd_work *, int); |
99920dc5 | 1489 | extern int w_start_resync(struct drbd_work *, int); |
b411b363 PR |
1490 | |
1491 | extern void resync_timer_fn(unsigned long data); | |
370a43e7 | 1492 | extern void start_resync_timer_fn(unsigned long data); |
b411b363 PR |
1493 | |
1494 | /* drbd_receiver.c */ | |
e3555d85 | 1495 | extern int drbd_rs_should_slow_down(struct drbd_conf *mdev, sector_t sector); |
fbe29dec AG |
1496 | extern int drbd_submit_peer_request(struct drbd_conf *, |
1497 | struct drbd_peer_request *, const unsigned, | |
1498 | const int); | |
7721f567 | 1499 | extern int drbd_free_peer_reqs(struct drbd_conf *, struct list_head *); |
0db55363 AG |
1500 | extern struct drbd_peer_request *drbd_alloc_peer_req(struct drbd_conf *, u64, |
1501 | sector_t, unsigned int, | |
1502 | gfp_t) __must_hold(local); | |
3967deb1 AG |
1503 | extern void __drbd_free_peer_req(struct drbd_conf *, struct drbd_peer_request *, |
1504 | int); | |
1505 | #define drbd_free_peer_req(m,e) __drbd_free_peer_req(m, e, 0) | |
1506 | #define drbd_free_net_peer_req(m,e) __drbd_free_peer_req(m, e, 1) | |
c37c8ecf | 1507 | extern struct page *drbd_alloc_pages(struct drbd_conf *, unsigned int, bool); |
b411b363 PR |
1508 | extern void drbd_set_recv_tcq(struct drbd_conf *mdev, int tcq_enabled); |
1509 | extern void _drbd_clear_done_ee(struct drbd_conf *mdev, struct list_head *to_be_freed); | |
0e29d163 | 1510 | extern void conn_flush_workqueue(struct drbd_tconn *tconn); |
c141ebda | 1511 | extern int drbd_connected(struct drbd_conf *mdev); |
0e29d163 PR |
1512 | static inline void drbd_flush_workqueue(struct drbd_conf *mdev) |
1513 | { | |
1514 | conn_flush_workqueue(mdev->tconn); | |
1515 | } | |
b411b363 | 1516 | |
ed439848 LE |
1517 | /* Yes, there is kernel_setsockopt, but only since 2.6.18. |
1518 | * So we have our own copy of it here. */ | |
b411b363 | 1519 | static inline int drbd_setsockopt(struct socket *sock, int level, int optname, |
ed439848 | 1520 | char *optval, int optlen) |
b411b363 | 1521 | { |
ed439848 LE |
1522 | mm_segment_t oldfs = get_fs(); |
1523 | char __user *uoptval; | |
b411b363 | 1524 | int err; |
ed439848 LE |
1525 | |
1526 | uoptval = (char __user __force *)optval; | |
1527 | ||
1528 | set_fs(KERNEL_DS); | |
b411b363 | 1529 | if (level == SOL_SOCKET) |
ed439848 | 1530 | err = sock_setsockopt(sock, level, optname, uoptval, optlen); |
b411b363 | 1531 | else |
ed439848 | 1532 | err = sock->ops->setsockopt(sock, level, optname, uoptval, |
b411b363 | 1533 | optlen); |
ed439848 | 1534 | set_fs(oldfs); |
b411b363 PR |
1535 | return err; |
1536 | } | |
1537 | ||
1538 | static inline void drbd_tcp_cork(struct socket *sock) | |
1539 | { | |
ed439848 | 1540 | int val = 1; |
b411b363 | 1541 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_CORK, |
ed439848 | 1542 | (char*)&val, sizeof(val)); |
b411b363 PR |
1543 | } |
1544 | ||
1545 | static inline void drbd_tcp_uncork(struct socket *sock) | |
1546 | { | |
ed439848 | 1547 | int val = 0; |
b411b363 | 1548 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_CORK, |
ed439848 | 1549 | (char*)&val, sizeof(val)); |
b411b363 PR |
1550 | } |
1551 | ||
1552 | static inline void drbd_tcp_nodelay(struct socket *sock) | |
1553 | { | |
ed439848 | 1554 | int val = 1; |
b411b363 | 1555 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_NODELAY, |
ed439848 | 1556 | (char*)&val, sizeof(val)); |
b411b363 PR |
1557 | } |
1558 | ||
1559 | static inline void drbd_tcp_quickack(struct socket *sock) | |
1560 | { | |
ed439848 | 1561 | int val = 2; |
b411b363 | 1562 | (void) drbd_setsockopt(sock, SOL_TCP, TCP_QUICKACK, |
ed439848 | 1563 | (char*)&val, sizeof(val)); |
b411b363 PR |
1564 | } |
1565 | ||
4b0007c0 | 1566 | void drbd_bump_write_ordering(struct drbd_tconn *tconn, enum write_ordering_e wo); |
b411b363 PR |
1567 | |
1568 | /* drbd_proc.c */ | |
1569 | extern struct proc_dir_entry *drbd_proc; | |
7d4e9d09 | 1570 | extern const struct file_operations drbd_proc_fops; |
b411b363 PR |
1571 | extern const char *drbd_conn_str(enum drbd_conns s); |
1572 | extern const char *drbd_role_str(enum drbd_role s); | |
1573 | ||
1574 | /* drbd_actlog.c */ | |
181286ad LE |
1575 | extern void drbd_al_begin_io(struct drbd_conf *mdev, struct drbd_interval *i); |
1576 | extern void drbd_al_complete_io(struct drbd_conf *mdev, struct drbd_interval *i); | |
b411b363 PR |
1577 | extern void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector); |
1578 | extern int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector); | |
1579 | extern int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector); | |
1580 | extern void drbd_rs_cancel_all(struct drbd_conf *mdev); | |
1581 | extern int drbd_rs_del_all(struct drbd_conf *mdev); | |
1582 | extern void drbd_rs_failed_io(struct drbd_conf *mdev, | |
1583 | sector_t sector, int size); | |
ea5442af | 1584 | extern void drbd_advance_rs_marks(struct drbd_conf *mdev, unsigned long still_to_go); |
b411b363 PR |
1585 | extern void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, |
1586 | int size, const char *file, const unsigned int line); | |
1587 | #define drbd_set_in_sync(mdev, sector, size) \ | |
1588 | __drbd_set_in_sync(mdev, sector, size, __FILE__, __LINE__) | |
73a01a18 | 1589 | extern int __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, |
b411b363 PR |
1590 | int size, const char *file, const unsigned int line); |
1591 | #define drbd_set_out_of_sync(mdev, sector, size) \ | |
1592 | __drbd_set_out_of_sync(mdev, sector, size, __FILE__, __LINE__) | |
b411b363 PR |
1593 | extern void drbd_al_shrink(struct drbd_conf *mdev); |
1594 | ||
b411b363 | 1595 | /* drbd_nl.c */ |
3b98c0c2 LE |
1596 | /* state info broadcast */ |
1597 | struct sib_info { | |
1598 | enum drbd_state_info_bcast_reason sib_reason; | |
1599 | union { | |
1600 | struct { | |
1601 | char *helper_name; | |
1602 | unsigned helper_exit_code; | |
1603 | }; | |
1604 | struct { | |
1605 | union drbd_state os; | |
1606 | union drbd_state ns; | |
1607 | }; | |
1608 | }; | |
1609 | }; | |
1610 | void drbd_bcast_event(struct drbd_conf *mdev, const struct sib_info *sib); | |
b411b363 | 1611 | |
b411b363 PR |
1612 | /* |
1613 | * inline helper functions | |
1614 | *************************/ | |
1615 | ||
45bb912b LE |
1616 | /* see also page_chain_add and friends in drbd_receiver.c */ |
1617 | static inline struct page *page_chain_next(struct page *page) | |
1618 | { | |
1619 | return (struct page *)page_private(page); | |
1620 | } | |
1621 | #define page_chain_for_each(page) \ | |
1622 | for (; page && ({ prefetch(page_chain_next(page)); 1; }); \ | |
1623 | page = page_chain_next(page)) | |
1624 | #define page_chain_for_each_safe(page, n) \ | |
1625 | for (; page && ({ n = page_chain_next(page); 1; }); page = n) | |
1626 | ||
45bb912b | 1627 | |
045417f7 | 1628 | static inline int drbd_peer_req_has_active_page(struct drbd_peer_request *peer_req) |
45bb912b | 1629 | { |
db830c46 | 1630 | struct page *page = peer_req->pages; |
45bb912b LE |
1631 | page_chain_for_each(page) { |
1632 | if (page_count(page) > 1) | |
1633 | return 1; | |
1634 | } | |
1635 | return 0; | |
1636 | } | |
1637 | ||
bf885f8a AG |
1638 | static inline enum drbd_state_rv |
1639 | _drbd_set_state(struct drbd_conf *mdev, union drbd_state ns, | |
1640 | enum chg_state_flags flags, struct completion *done) | |
b411b363 | 1641 | { |
bf885f8a | 1642 | enum drbd_state_rv rv; |
b411b363 PR |
1643 | |
1644 | read_lock(&global_state_lock); | |
1645 | rv = __drbd_set_state(mdev, ns, flags, done); | |
1646 | read_unlock(&global_state_lock); | |
1647 | ||
1648 | return rv; | |
1649 | } | |
1650 | ||
78bae59b PR |
1651 | static inline union drbd_state drbd_read_state(struct drbd_conf *mdev) |
1652 | { | |
1653 | union drbd_state rv; | |
1654 | ||
da9fbc27 | 1655 | rv.i = mdev->state.i; |
8e0af25f PR |
1656 | rv.susp = mdev->tconn->susp; |
1657 | rv.susp_nod = mdev->tconn->susp_nod; | |
1658 | rv.susp_fen = mdev->tconn->susp_fen; | |
78bae59b PR |
1659 | |
1660 | return rv; | |
1661 | } | |
1662 | ||
0c849666 LE |
1663 | enum drbd_force_detach_flags { |
1664 | DRBD_IO_ERROR, | |
1665 | DRBD_META_IO_ERROR, | |
1666 | DRBD_FORCE_DETACH, | |
1667 | }; | |
1668 | ||
b411b363 | 1669 | #define __drbd_chk_io_error(m,f) __drbd_chk_io_error_(m,f, __func__) |
0c849666 LE |
1670 | static inline void __drbd_chk_io_error_(struct drbd_conf *mdev, |
1671 | enum drbd_force_detach_flags forcedetach, | |
1672 | const char *where) | |
b411b363 | 1673 | { |
daeda1cc PR |
1674 | enum drbd_io_error_p ep; |
1675 | ||
1676 | rcu_read_lock(); | |
1677 | ep = rcu_dereference(mdev->ldev->disk_conf)->on_io_error; | |
1678 | rcu_read_unlock(); | |
1679 | switch (ep) { | |
1680 | case EP_PASS_ON: /* FIXME would this be better named "Ignore"? */ | |
0c849666 | 1681 | if (forcedetach == DRBD_IO_ERROR) { |
7383506c | 1682 | if (__ratelimit(&drbd_ratelimit_state)) |
82f59cc6 | 1683 | dev_err(DEV, "Local IO failed in %s.\n", where); |
d2e17807 PR |
1684 | if (mdev->state.disk > D_INCONSISTENT) |
1685 | _drbd_set_state(_NS(mdev, disk, D_INCONSISTENT), CS_HARD, NULL); | |
b411b363 PR |
1686 | break; |
1687 | } | |
1688 | /* NOTE fall through to detach case if forcedetach set */ | |
1689 | case EP_DETACH: | |
1690 | case EP_CALL_HELPER: | |
82f59cc6 | 1691 | set_bit(WAS_IO_ERROR, &mdev->flags); |
0c849666 LE |
1692 | if (forcedetach == DRBD_FORCE_DETACH) |
1693 | set_bit(FORCE_DETACH, &mdev->flags); | |
b411b363 PR |
1694 | if (mdev->state.disk > D_FAILED) { |
1695 | _drbd_set_state(_NS(mdev, disk, D_FAILED), CS_HARD, NULL); | |
82f59cc6 LE |
1696 | dev_err(DEV, |
1697 | "Local IO failed in %s. Detaching...\n", where); | |
b411b363 PR |
1698 | } |
1699 | break; | |
1700 | } | |
1701 | } | |
1702 | ||
1703 | /** | |
1704 | * drbd_chk_io_error: Handle the on_io_error setting, should be called from all io completion handlers | |
1705 | * @mdev: DRBD device. | |
1706 | * @error: Error code passed to the IO completion callback | |
1707 | * @forcedetach: Force detach. I.e. the error happened while accessing the meta data | |
1708 | * | |
1709 | * See also drbd_main.c:after_state_ch() if (os.disk > D_FAILED && ns.disk == D_FAILED) | |
1710 | */ | |
1711 | #define drbd_chk_io_error(m,e,f) drbd_chk_io_error_(m,e,f, __func__) | |
1712 | static inline void drbd_chk_io_error_(struct drbd_conf *mdev, | |
0c849666 | 1713 | int error, enum drbd_force_detach_flags forcedetach, const char *where) |
b411b363 PR |
1714 | { |
1715 | if (error) { | |
1716 | unsigned long flags; | |
87eeee41 | 1717 | spin_lock_irqsave(&mdev->tconn->req_lock, flags); |
b411b363 | 1718 | __drbd_chk_io_error_(mdev, forcedetach, where); |
87eeee41 | 1719 | spin_unlock_irqrestore(&mdev->tconn->req_lock, flags); |
b411b363 PR |
1720 | } |
1721 | } | |
1722 | ||
1723 | ||
1724 | /** | |
1725 | * drbd_md_first_sector() - Returns the first sector number of the meta data area | |
1726 | * @bdev: Meta data block device. | |
1727 | * | |
1728 | * BTW, for internal meta data, this happens to be the maximum capacity | |
1729 | * we could agree upon with our peer node. | |
1730 | */ | |
daeda1cc | 1731 | static inline sector_t _drbd_md_first_sector(int meta_dev_idx, struct drbd_backing_dev *bdev) |
b411b363 | 1732 | { |
daeda1cc | 1733 | switch (meta_dev_idx) { |
b411b363 PR |
1734 | case DRBD_MD_INDEX_INTERNAL: |
1735 | case DRBD_MD_INDEX_FLEX_INT: | |
1736 | return bdev->md.md_offset + bdev->md.bm_offset; | |
1737 | case DRBD_MD_INDEX_FLEX_EXT: | |
1738 | default: | |
1739 | return bdev->md.md_offset; | |
1740 | } | |
1741 | } | |
1742 | ||
daeda1cc PR |
1743 | static inline sector_t drbd_md_first_sector(struct drbd_backing_dev *bdev) |
1744 | { | |
1745 | int meta_dev_idx; | |
1746 | ||
1747 | rcu_read_lock(); | |
1748 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1749 | rcu_read_unlock(); | |
1750 | ||
1751 | return _drbd_md_first_sector(meta_dev_idx, bdev); | |
1752 | } | |
1753 | ||
b411b363 PR |
1754 | /** |
1755 | * drbd_md_last_sector() - Return the last sector number of the meta data area | |
1756 | * @bdev: Meta data block device. | |
1757 | */ | |
1758 | static inline sector_t drbd_md_last_sector(struct drbd_backing_dev *bdev) | |
1759 | { | |
daeda1cc PR |
1760 | int meta_dev_idx; |
1761 | ||
1762 | rcu_read_lock(); | |
1763 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1764 | rcu_read_unlock(); | |
1765 | ||
1766 | switch (meta_dev_idx) { | |
b411b363 PR |
1767 | case DRBD_MD_INDEX_INTERNAL: |
1768 | case DRBD_MD_INDEX_FLEX_INT: | |
1769 | return bdev->md.md_offset + MD_AL_OFFSET - 1; | |
1770 | case DRBD_MD_INDEX_FLEX_EXT: | |
1771 | default: | |
1772 | return bdev->md.md_offset + bdev->md.md_size_sect; | |
1773 | } | |
1774 | } | |
1775 | ||
1776 | /* Returns the number of 512 byte sectors of the device */ | |
1777 | static inline sector_t drbd_get_capacity(struct block_device *bdev) | |
1778 | { | |
1779 | /* return bdev ? get_capacity(bdev->bd_disk) : 0; */ | |
77304d2a | 1780 | return bdev ? i_size_read(bdev->bd_inode) >> 9 : 0; |
b411b363 PR |
1781 | } |
1782 | ||
1783 | /** | |
1784 | * drbd_get_max_capacity() - Returns the capacity we announce to out peer | |
1785 | * @bdev: Meta data block device. | |
1786 | * | |
1787 | * returns the capacity we announce to out peer. we clip ourselves at the | |
1788 | * various MAX_SECTORS, because if we don't, current implementation will | |
1789 | * oops sooner or later | |
1790 | */ | |
1791 | static inline sector_t drbd_get_max_capacity(struct drbd_backing_dev *bdev) | |
1792 | { | |
1793 | sector_t s; | |
daeda1cc PR |
1794 | int meta_dev_idx; |
1795 | ||
1796 | rcu_read_lock(); | |
1797 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1798 | rcu_read_unlock(); | |
1799 | ||
1800 | switch (meta_dev_idx) { | |
b411b363 PR |
1801 | case DRBD_MD_INDEX_INTERNAL: |
1802 | case DRBD_MD_INDEX_FLEX_INT: | |
1803 | s = drbd_get_capacity(bdev->backing_bdev) | |
1804 | ? min_t(sector_t, DRBD_MAX_SECTORS_FLEX, | |
daeda1cc | 1805 | _drbd_md_first_sector(meta_dev_idx, bdev)) |
b411b363 PR |
1806 | : 0; |
1807 | break; | |
1808 | case DRBD_MD_INDEX_FLEX_EXT: | |
1809 | s = min_t(sector_t, DRBD_MAX_SECTORS_FLEX, | |
1810 | drbd_get_capacity(bdev->backing_bdev)); | |
1811 | /* clip at maximum size the meta device can support */ | |
1812 | s = min_t(sector_t, s, | |
1813 | BM_EXT_TO_SECT(bdev->md.md_size_sect | |
1814 | - bdev->md.bm_offset)); | |
1815 | break; | |
1816 | default: | |
1817 | s = min_t(sector_t, DRBD_MAX_SECTORS, | |
1818 | drbd_get_capacity(bdev->backing_bdev)); | |
1819 | } | |
1820 | return s; | |
1821 | } | |
1822 | ||
1823 | /** | |
1824 | * drbd_md_ss__() - Return the sector number of our meta data super block | |
1825 | * @mdev: DRBD device. | |
1826 | * @bdev: Meta data block device. | |
1827 | */ | |
1828 | static inline sector_t drbd_md_ss__(struct drbd_conf *mdev, | |
1829 | struct drbd_backing_dev *bdev) | |
1830 | { | |
daeda1cc PR |
1831 | int meta_dev_idx; |
1832 | ||
1833 | rcu_read_lock(); | |
1834 | meta_dev_idx = rcu_dereference(bdev->disk_conf)->meta_dev_idx; | |
1835 | rcu_read_unlock(); | |
1836 | ||
1837 | switch (meta_dev_idx) { | |
b411b363 | 1838 | default: /* external, some index */ |
daeda1cc | 1839 | return MD_RESERVED_SECT * meta_dev_idx; |
b411b363 PR |
1840 | case DRBD_MD_INDEX_INTERNAL: |
1841 | /* with drbd08, internal meta data is always "flexible" */ | |
1842 | case DRBD_MD_INDEX_FLEX_INT: | |
1843 | /* sizeof(struct md_on_disk_07) == 4k | |
1844 | * position: last 4k aligned block of 4k size */ | |
1845 | if (!bdev->backing_bdev) { | |
1846 | if (__ratelimit(&drbd_ratelimit_state)) { | |
1847 | dev_err(DEV, "bdev->backing_bdev==NULL\n"); | |
1848 | dump_stack(); | |
1849 | } | |
1850 | return 0; | |
1851 | } | |
1852 | return (drbd_get_capacity(bdev->backing_bdev) & ~7ULL) | |
1853 | - MD_AL_OFFSET; | |
1854 | case DRBD_MD_INDEX_FLEX_EXT: | |
1855 | return 0; | |
1856 | } | |
1857 | } | |
1858 | ||
b411b363 PR |
1859 | static inline void |
1860 | drbd_queue_work_front(struct drbd_work_queue *q, struct drbd_work *w) | |
1861 | { | |
1862 | unsigned long flags; | |
1863 | spin_lock_irqsave(&q->q_lock, flags); | |
1864 | list_add(&w->list, &q->q); | |
b411b363 | 1865 | spin_unlock_irqrestore(&q->q_lock, flags); |
8c0785a5 | 1866 | wake_up(&q->q_wait); |
b411b363 PR |
1867 | } |
1868 | ||
1869 | static inline void | |
1870 | drbd_queue_work(struct drbd_work_queue *q, struct drbd_work *w) | |
1871 | { | |
1872 | unsigned long flags; | |
1873 | spin_lock_irqsave(&q->q_lock, flags); | |
1874 | list_add_tail(&w->list, &q->q); | |
b411b363 | 1875 | spin_unlock_irqrestore(&q->q_lock, flags); |
8c0785a5 | 1876 | wake_up(&q->q_wait); |
b411b363 PR |
1877 | } |
1878 | ||
0625ac19 | 1879 | static inline void wake_asender(struct drbd_tconn *tconn) |
b411b363 | 1880 | { |
0625ac19 PR |
1881 | if (test_bit(SIGNAL_ASENDER, &tconn->flags)) |
1882 | force_sig(DRBD_SIG, tconn->asender.task); | |
b411b363 PR |
1883 | } |
1884 | ||
0625ac19 | 1885 | static inline void request_ping(struct drbd_tconn *tconn) |
b411b363 | 1886 | { |
0625ac19 PR |
1887 | set_bit(SEND_PING, &tconn->flags); |
1888 | wake_asender(tconn); | |
b411b363 PR |
1889 | } |
1890 | ||
dba58587 AG |
1891 | extern void *conn_prepare_command(struct drbd_tconn *, struct drbd_socket *); |
1892 | extern void *drbd_prepare_command(struct drbd_conf *, struct drbd_socket *); | |
1893 | extern int conn_send_command(struct drbd_tconn *, struct drbd_socket *, | |
1894 | enum drbd_packet, unsigned int, void *, | |
1895 | unsigned int); | |
1896 | extern int drbd_send_command(struct drbd_conf *, struct drbd_socket *, | |
1897 | enum drbd_packet, unsigned int, void *, | |
1898 | unsigned int); | |
1899 | ||
e307f352 AG |
1900 | extern int drbd_send_ping(struct drbd_tconn *tconn); |
1901 | extern int drbd_send_ping_ack(struct drbd_tconn *tconn); | |
9f5bdc33 AG |
1902 | extern int drbd_send_state_req(struct drbd_conf *, union drbd_state, union drbd_state); |
1903 | extern int conn_send_state_req(struct drbd_tconn *, union drbd_state, union drbd_state); | |
cf29c9d8 | 1904 | |
b411b363 PR |
1905 | static inline void drbd_thread_stop(struct drbd_thread *thi) |
1906 | { | |
81e84650 | 1907 | _drbd_thread_stop(thi, false, true); |
b411b363 PR |
1908 | } |
1909 | ||
1910 | static inline void drbd_thread_stop_nowait(struct drbd_thread *thi) | |
1911 | { | |
81e84650 | 1912 | _drbd_thread_stop(thi, false, false); |
b411b363 PR |
1913 | } |
1914 | ||
1915 | static inline void drbd_thread_restart_nowait(struct drbd_thread *thi) | |
1916 | { | |
81e84650 | 1917 | _drbd_thread_stop(thi, true, false); |
b411b363 PR |
1918 | } |
1919 | ||
1920 | /* counts how many answer packets packets we expect from our peer, | |
1921 | * for either explicit application requests, | |
1922 | * or implicit barrier packets as necessary. | |
1923 | * increased: | |
1924 | * w_send_barrier | |
8554df1c | 1925 | * _req_mod(req, QUEUE_FOR_NET_WRITE or QUEUE_FOR_NET_READ); |
b411b363 PR |
1926 | * it is much easier and equally valid to count what we queue for the |
1927 | * worker, even before it actually was queued or send. | |
1928 | * (drbd_make_request_common; recovery path on read io-error) | |
1929 | * decreased: | |
1930 | * got_BarrierAck (respective tl_clear, tl_clear_barrier) | |
8554df1c | 1931 | * _req_mod(req, DATA_RECEIVED) |
b411b363 | 1932 | * [from receive_DataReply] |
8554df1c | 1933 | * _req_mod(req, WRITE_ACKED_BY_PEER or RECV_ACKED_BY_PEER or NEG_ACKED) |
b411b363 PR |
1934 | * [from got_BlockAck (P_WRITE_ACK, P_RECV_ACK)] |
1935 | * for some reason it is NOT decreased in got_NegAck, | |
1936 | * but in the resulting cleanup code from report_params. | |
1937 | * we should try to remember the reason for that... | |
8554df1c AG |
1938 | * _req_mod(req, SEND_FAILED or SEND_CANCELED) |
1939 | * _req_mod(req, CONNECTION_LOST_WHILE_PENDING) | |
b411b363 PR |
1940 | * [from tl_clear_barrier] |
1941 | */ | |
1942 | static inline void inc_ap_pending(struct drbd_conf *mdev) | |
1943 | { | |
1944 | atomic_inc(&mdev->ap_pending_cnt); | |
1945 | } | |
1946 | ||
49559d87 PR |
1947 | #define ERR_IF_CNT_IS_NEGATIVE(which, func, line) \ |
1948 | if (atomic_read(&mdev->which) < 0) \ | |
b411b363 | 1949 | dev_err(DEV, "in %s:%d: " #which " = %d < 0 !\n", \ |
49559d87 PR |
1950 | func, line, \ |
1951 | atomic_read(&mdev->which)) | |
b411b363 | 1952 | |
49559d87 PR |
1953 | #define dec_ap_pending(mdev) _dec_ap_pending(mdev, __FUNCTION__, __LINE__) |
1954 | static inline void _dec_ap_pending(struct drbd_conf *mdev, const char *func, int line) | |
1955 | { | |
1956 | if (atomic_dec_and_test(&mdev->ap_pending_cnt)) | |
1957 | wake_up(&mdev->misc_wait); | |
1958 | ERR_IF_CNT_IS_NEGATIVE(ap_pending_cnt, func, line); | |
1959 | } | |
b411b363 PR |
1960 | |
1961 | /* counts how many resync-related answers we still expect from the peer | |
1962 | * increase decrease | |
1963 | * C_SYNC_TARGET sends P_RS_DATA_REQUEST (and expects P_RS_DATA_REPLY) | |
25985edc | 1964 | * C_SYNC_SOURCE sends P_RS_DATA_REPLY (and expects P_WRITE_ACK with ID_SYNCER) |
b411b363 PR |
1965 | * (or P_NEG_ACK with ID_SYNCER) |
1966 | */ | |
1967 | static inline void inc_rs_pending(struct drbd_conf *mdev) | |
1968 | { | |
1969 | atomic_inc(&mdev->rs_pending_cnt); | |
1970 | } | |
1971 | ||
49559d87 PR |
1972 | #define dec_rs_pending(mdev) _dec_rs_pending(mdev, __FUNCTION__, __LINE__) |
1973 | static inline void _dec_rs_pending(struct drbd_conf *mdev, const char *func, int line) | |
1974 | { | |
1975 | atomic_dec(&mdev->rs_pending_cnt); | |
1976 | ERR_IF_CNT_IS_NEGATIVE(rs_pending_cnt, func, line); | |
1977 | } | |
b411b363 PR |
1978 | |
1979 | /* counts how many answers we still need to send to the peer. | |
1980 | * increased on | |
1981 | * receive_Data unless protocol A; | |
1982 | * we need to send a P_RECV_ACK (proto B) | |
1983 | * or P_WRITE_ACK (proto C) | |
1984 | * receive_RSDataReply (recv_resync_read) we need to send a P_WRITE_ACK | |
1985 | * receive_DataRequest (receive_RSDataRequest) we need to send back P_DATA | |
1986 | * receive_Barrier_* we need to send a P_BARRIER_ACK | |
1987 | */ | |
1988 | static inline void inc_unacked(struct drbd_conf *mdev) | |
1989 | { | |
1990 | atomic_inc(&mdev->unacked_cnt); | |
1991 | } | |
1992 | ||
49559d87 PR |
1993 | #define dec_unacked(mdev) _dec_unacked(mdev, __FUNCTION__, __LINE__) |
1994 | static inline void _dec_unacked(struct drbd_conf *mdev, const char *func, int line) | |
1995 | { | |
1996 | atomic_dec(&mdev->unacked_cnt); | |
1997 | ERR_IF_CNT_IS_NEGATIVE(unacked_cnt, func, line); | |
1998 | } | |
b411b363 | 1999 | |
49559d87 PR |
2000 | #define sub_unacked(mdev, n) _sub_unacked(mdev, n, __FUNCTION__, __LINE__) |
2001 | static inline void _sub_unacked(struct drbd_conf *mdev, int n, const char *func, int line) | |
2002 | { | |
2003 | atomic_sub(n, &mdev->unacked_cnt); | |
2004 | ERR_IF_CNT_IS_NEGATIVE(unacked_cnt, func, line); | |
2005 | } | |
b411b363 | 2006 | |
b411b363 PR |
2007 | /** |
2008 | * get_ldev() - Increase the ref count on mdev->ldev. Returns 0 if there is no ldev | |
2009 | * @M: DRBD device. | |
2010 | * | |
2011 | * You have to call put_ldev() when finished working with mdev->ldev. | |
2012 | */ | |
2013 | #define get_ldev(M) __cond_lock(local, _get_ldev_if_state(M,D_INCONSISTENT)) | |
2014 | #define get_ldev_if_state(M,MINS) __cond_lock(local, _get_ldev_if_state(M,MINS)) | |
2015 | ||
2016 | static inline void put_ldev(struct drbd_conf *mdev) | |
2017 | { | |
1d7734a0 | 2018 | int i = atomic_dec_return(&mdev->local_cnt); |
9a0d9d03 LE |
2019 | |
2020 | /* This may be called from some endio handler, | |
2021 | * so we must not sleep here. */ | |
2022 | ||
b411b363 | 2023 | __release(local); |
1d7734a0 | 2024 | D_ASSERT(i >= 0); |
e9e6f3ec | 2025 | if (i == 0) { |
82f59cc6 LE |
2026 | if (mdev->state.disk == D_DISKLESS) |
2027 | /* even internal references gone, safe to destroy */ | |
2028 | drbd_ldev_destroy(mdev); | |
e9e6f3ec | 2029 | if (mdev->state.disk == D_FAILED) |
82f59cc6 | 2030 | /* all application IO references gone. */ |
e9e6f3ec | 2031 | drbd_go_diskless(mdev); |
b411b363 | 2032 | wake_up(&mdev->misc_wait); |
e9e6f3ec | 2033 | } |
b411b363 PR |
2034 | } |
2035 | ||
2036 | #ifndef __CHECKER__ | |
2037 | static inline int _get_ldev_if_state(struct drbd_conf *mdev, enum drbd_disk_state mins) | |
2038 | { | |
2039 | int io_allowed; | |
2040 | ||
82f59cc6 LE |
2041 | /* never get a reference while D_DISKLESS */ |
2042 | if (mdev->state.disk == D_DISKLESS) | |
2043 | return 0; | |
2044 | ||
b411b363 PR |
2045 | atomic_inc(&mdev->local_cnt); |
2046 | io_allowed = (mdev->state.disk >= mins); | |
2047 | if (!io_allowed) | |
2048 | put_ldev(mdev); | |
2049 | return io_allowed; | |
2050 | } | |
2051 | #else | |
2052 | extern int _get_ldev_if_state(struct drbd_conf *mdev, enum drbd_disk_state mins); | |
2053 | #endif | |
2054 | ||
2055 | /* you must have an "get_ldev" reference */ | |
2056 | static inline void drbd_get_syncer_progress(struct drbd_conf *mdev, | |
2057 | unsigned long *bits_left, unsigned int *per_mil_done) | |
2058 | { | |
4b0715f0 LE |
2059 | /* this is to break it at compile time when we change that, in case we |
2060 | * want to support more than (1<<32) bits on a 32bit arch. */ | |
b411b363 PR |
2061 | typecheck(unsigned long, mdev->rs_total); |
2062 | ||
2063 | /* note: both rs_total and rs_left are in bits, i.e. in | |
2064 | * units of BM_BLOCK_SIZE. | |
2065 | * for the percentage, we don't care. */ | |
2066 | ||
439d5953 LE |
2067 | if (mdev->state.conn == C_VERIFY_S || mdev->state.conn == C_VERIFY_T) |
2068 | *bits_left = mdev->ov_left; | |
2069 | else | |
2070 | *bits_left = drbd_bm_total_weight(mdev) - mdev->rs_failed; | |
b411b363 PR |
2071 | /* >> 10 to prevent overflow, |
2072 | * +1 to prevent division by zero */ | |
2073 | if (*bits_left > mdev->rs_total) { | |
2074 | /* doh. maybe a logic bug somewhere. | |
2075 | * may also be just a race condition | |
2076 | * between this and a disconnect during sync. | |
2077 | * for now, just prevent in-kernel buffer overflow. | |
2078 | */ | |
2079 | smp_rmb(); | |
2080 | dev_warn(DEV, "cs:%s rs_left=%lu > rs_total=%lu (rs_failed %lu)\n", | |
2081 | drbd_conn_str(mdev->state.conn), | |
2082 | *bits_left, mdev->rs_total, mdev->rs_failed); | |
2083 | *per_mil_done = 0; | |
2084 | } else { | |
4b0715f0 LE |
2085 | /* Make sure the division happens in long context. |
2086 | * We allow up to one petabyte storage right now, | |
2087 | * at a granularity of 4k per bit that is 2**38 bits. | |
2088 | * After shift right and multiplication by 1000, | |
2089 | * this should still fit easily into a 32bit long, | |
2090 | * so we don't need a 64bit division on 32bit arch. | |
2091 | * Note: currently we don't support such large bitmaps on 32bit | |
2092 | * arch anyways, but no harm done to be prepared for it here. | |
2093 | */ | |
3b7cd457 | 2094 | unsigned int shift = mdev->rs_total > UINT_MAX ? 16 : 10; |
4b0715f0 LE |
2095 | unsigned long left = *bits_left >> shift; |
2096 | unsigned long total = 1UL + (mdev->rs_total >> shift); | |
2097 | unsigned long tmp = 1000UL - left * 1000UL/total; | |
b411b363 PR |
2098 | *per_mil_done = tmp; |
2099 | } | |
2100 | } | |
2101 | ||
2102 | ||
2103 | /* this throttles on-the-fly application requests | |
2104 | * according to max_buffers settings; | |
2105 | * maybe re-implement using semaphores? */ | |
2106 | static inline int drbd_get_max_buffers(struct drbd_conf *mdev) | |
2107 | { | |
44ed167d PR |
2108 | struct net_conf *nc; |
2109 | int mxb; | |
2110 | ||
2111 | rcu_read_lock(); | |
2112 | nc = rcu_dereference(mdev->tconn->net_conf); | |
2113 | mxb = nc ? nc->max_buffers : 1000000; /* arbitrary limit on open requests */ | |
2114 | rcu_read_unlock(); | |
2115 | ||
b411b363 PR |
2116 | return mxb; |
2117 | } | |
2118 | ||
3719094e | 2119 | static inline int drbd_state_is_stable(struct drbd_conf *mdev) |
b411b363 | 2120 | { |
da9fbc27 | 2121 | union drbd_dev_state s = mdev->state; |
b411b363 PR |
2122 | |
2123 | /* DO NOT add a default clause, we want the compiler to warn us | |
2124 | * for any newly introduced state we may have forgotten to add here */ | |
2125 | ||
2126 | switch ((enum drbd_conns)s.conn) { | |
2127 | /* new io only accepted when there is no connection, ... */ | |
2128 | case C_STANDALONE: | |
2129 | case C_WF_CONNECTION: | |
2130 | /* ... or there is a well established connection. */ | |
2131 | case C_CONNECTED: | |
2132 | case C_SYNC_SOURCE: | |
2133 | case C_SYNC_TARGET: | |
2134 | case C_VERIFY_S: | |
2135 | case C_VERIFY_T: | |
2136 | case C_PAUSED_SYNC_S: | |
2137 | case C_PAUSED_SYNC_T: | |
67531718 PR |
2138 | case C_AHEAD: |
2139 | case C_BEHIND: | |
3719094e | 2140 | /* transitional states, IO allowed */ |
b411b363 PR |
2141 | case C_DISCONNECTING: |
2142 | case C_UNCONNECTED: | |
2143 | case C_TIMEOUT: | |
2144 | case C_BROKEN_PIPE: | |
2145 | case C_NETWORK_FAILURE: | |
2146 | case C_PROTOCOL_ERROR: | |
2147 | case C_TEAR_DOWN: | |
2148 | case C_WF_REPORT_PARAMS: | |
2149 | case C_STARTING_SYNC_S: | |
2150 | case C_STARTING_SYNC_T: | |
3719094e PR |
2151 | break; |
2152 | ||
2153 | /* Allow IO in BM exchange states with new protocols */ | |
b411b363 | 2154 | case C_WF_BITMAP_S: |
31890f4a | 2155 | if (mdev->tconn->agreed_pro_version < 96) |
3719094e PR |
2156 | return 0; |
2157 | break; | |
2158 | ||
2159 | /* no new io accepted in these states */ | |
b411b363 PR |
2160 | case C_WF_BITMAP_T: |
2161 | case C_WF_SYNC_UUID: | |
2162 | case C_MASK: | |
2163 | /* not "stable" */ | |
2164 | return 0; | |
2165 | } | |
2166 | ||
2167 | switch ((enum drbd_disk_state)s.disk) { | |
2168 | case D_DISKLESS: | |
2169 | case D_INCONSISTENT: | |
2170 | case D_OUTDATED: | |
2171 | case D_CONSISTENT: | |
2172 | case D_UP_TO_DATE: | |
cdfda633 | 2173 | case D_FAILED: |
b411b363 PR |
2174 | /* disk state is stable as well. */ |
2175 | break; | |
2176 | ||
d942ae44 | 2177 | /* no new io accepted during transitional states */ |
b411b363 | 2178 | case D_ATTACHING: |
b411b363 PR |
2179 | case D_NEGOTIATING: |
2180 | case D_UNKNOWN: | |
2181 | case D_MASK: | |
2182 | /* not "stable" */ | |
2183 | return 0; | |
2184 | } | |
2185 | ||
2186 | return 1; | |
2187 | } | |
2188 | ||
2aebfabb | 2189 | static inline int drbd_suspended(struct drbd_conf *mdev) |
fb22c402 | 2190 | { |
8e0af25f PR |
2191 | struct drbd_tconn *tconn = mdev->tconn; |
2192 | ||
2193 | return tconn->susp || tconn->susp_fen || tconn->susp_nod; | |
fb22c402 PR |
2194 | } |
2195 | ||
1b881ef7 | 2196 | static inline bool may_inc_ap_bio(struct drbd_conf *mdev) |
b411b363 PR |
2197 | { |
2198 | int mxb = drbd_get_max_buffers(mdev); | |
2199 | ||
2aebfabb | 2200 | if (drbd_suspended(mdev)) |
1b881ef7 | 2201 | return false; |
b411b363 | 2202 | if (test_bit(SUSPEND_IO, &mdev->flags)) |
1b881ef7 | 2203 | return false; |
b411b363 PR |
2204 | |
2205 | /* to avoid potential deadlock or bitmap corruption, | |
2206 | * in various places, we only allow new application io | |
2207 | * to start during "stable" states. */ | |
2208 | ||
2209 | /* no new io accepted when attaching or detaching the disk */ | |
3719094e | 2210 | if (!drbd_state_is_stable(mdev)) |
1b881ef7 | 2211 | return false; |
b411b363 PR |
2212 | |
2213 | /* since some older kernels don't have atomic_add_unless, | |
2214 | * and we are within the spinlock anyways, we have this workaround. */ | |
2215 | if (atomic_read(&mdev->ap_bio_cnt) > mxb) | |
1b881ef7 | 2216 | return false; |
b411b363 | 2217 | if (test_bit(BITMAP_IO, &mdev->flags)) |
1b881ef7 AG |
2218 | return false; |
2219 | return true; | |
b411b363 PR |
2220 | } |
2221 | ||
23361cf3 | 2222 | static inline bool inc_ap_bio_cond(struct drbd_conf *mdev) |
b411b363 | 2223 | { |
1b881ef7 | 2224 | bool rv = false; |
8869d683 | 2225 | |
87eeee41 | 2226 | spin_lock_irq(&mdev->tconn->req_lock); |
1b881ef7 | 2227 | rv = may_inc_ap_bio(mdev); |
8869d683 | 2228 | if (rv) |
23361cf3 | 2229 | atomic_inc(&mdev->ap_bio_cnt); |
87eeee41 | 2230 | spin_unlock_irq(&mdev->tconn->req_lock); |
8869d683 PR |
2231 | |
2232 | return rv; | |
2233 | } | |
b411b363 | 2234 | |
23361cf3 | 2235 | static inline void inc_ap_bio(struct drbd_conf *mdev) |
8869d683 | 2236 | { |
b411b363 PR |
2237 | /* we wait here |
2238 | * as long as the device is suspended | |
2239 | * until the bitmap is no longer on the fly during connection | |
d942ae44 | 2240 | * handshake as long as we would exceed the max_buffer limit. |
b411b363 PR |
2241 | * |
2242 | * to avoid races with the reconnect code, | |
2243 | * we need to atomic_inc within the spinlock. */ | |
2244 | ||
23361cf3 | 2245 | wait_event(mdev->misc_wait, inc_ap_bio_cond(mdev)); |
b411b363 PR |
2246 | } |
2247 | ||
2248 | static inline void dec_ap_bio(struct drbd_conf *mdev) | |
2249 | { | |
2250 | int mxb = drbd_get_max_buffers(mdev); | |
2251 | int ap_bio = atomic_dec_return(&mdev->ap_bio_cnt); | |
2252 | ||
2253 | D_ASSERT(ap_bio >= 0); | |
970fbde1 LE |
2254 | |
2255 | if (ap_bio == 0 && test_bit(BITMAP_IO, &mdev->flags)) { | |
2256 | if (!test_and_set_bit(BITMAP_IO_QUEUED, &mdev->flags)) | |
2257 | drbd_queue_work(&mdev->tconn->sender_work, &mdev->bm_io_work.w); | |
2258 | } | |
2259 | ||
b411b363 PR |
2260 | /* this currently does wake_up for every dec_ap_bio! |
2261 | * maybe rather introduce some type of hysteresis? | |
2262 | * e.g. (ap_bio == mxb/2 || ap_bio == 0) ? */ | |
2263 | if (ap_bio < mxb) | |
2264 | wake_up(&mdev->misc_wait); | |
b411b363 PR |
2265 | } |
2266 | ||
58ffa580 LE |
2267 | static inline bool verify_can_do_stop_sector(struct drbd_conf *mdev) |
2268 | { | |
2269 | return mdev->tconn->agreed_pro_version >= 97 && | |
2270 | mdev->tconn->agreed_pro_version != 100; | |
2271 | } | |
2272 | ||
62b0da3a | 2273 | static inline int drbd_set_ed_uuid(struct drbd_conf *mdev, u64 val) |
b411b363 | 2274 | { |
62b0da3a | 2275 | int changed = mdev->ed_uuid != val; |
b411b363 | 2276 | mdev->ed_uuid = val; |
62b0da3a | 2277 | return changed; |
b411b363 PR |
2278 | } |
2279 | ||
b411b363 PR |
2280 | static inline int drbd_queue_order_type(struct drbd_conf *mdev) |
2281 | { | |
2282 | /* sorry, we currently have no working implementation | |
2283 | * of distributed TCQ stuff */ | |
2284 | #ifndef QUEUE_ORDERED_NONE | |
2285 | #define QUEUE_ORDERED_NONE 0 | |
2286 | #endif | |
2287 | return QUEUE_ORDERED_NONE; | |
2288 | } | |
2289 | ||
b411b363 PR |
2290 | static inline void drbd_md_flush(struct drbd_conf *mdev) |
2291 | { | |
2292 | int r; | |
2293 | ||
a8a4e51e | 2294 | if (test_bit(MD_NO_FUA, &mdev->flags)) |
b411b363 PR |
2295 | return; |
2296 | ||
8747d30a | 2297 | r = blkdev_issue_flush(mdev->ldev->md_bdev, GFP_NOIO, NULL); |
b411b363 | 2298 | if (r) { |
a8a4e51e | 2299 | set_bit(MD_NO_FUA, &mdev->flags); |
b411b363 PR |
2300 | dev_err(DEV, "meta data flush failed with status %d, disabling md-flushes\n", r); |
2301 | } | |
2302 | } | |
2303 | ||
2304 | #endif | |
c1fd29a1 PR |
2305 | |
2306 | /* This is defined in drivers/md/md.h as well. Should go into wait.h */ | |
2307 | #define __wait_event_lock_irq(wq, condition, lock, cmd) \ | |
2308 | do { \ | |
2309 | wait_queue_t __wait; \ | |
2310 | init_waitqueue_entry(&__wait, current); \ | |
2311 | \ | |
2312 | add_wait_queue(&wq, &__wait); \ | |
2313 | for (;;) { \ | |
2314 | set_current_state(TASK_UNINTERRUPTIBLE); \ | |
2315 | if (condition) \ | |
2316 | break; \ | |
2317 | spin_unlock_irq(&lock); \ | |
2318 | cmd; \ | |
2319 | schedule(); \ | |
2320 | spin_lock_irq(&lock); \ | |
2321 | } \ | |
2322 | current->state = TASK_RUNNING; \ | |
2323 | remove_wait_queue(&wq, &__wait); \ | |
2324 | } while (0) | |
2325 | ||
2326 | #define wait_event_lock_irq(wq, condition, lock, cmd) \ | |
2327 | do { \ | |
2328 | if (condition) \ | |
2329 | break; \ | |
2330 | __wait_event_lock_irq(wq, condition, lock, cmd); \ | |
2331 | } while (0) |