| 1 |
/* |
| 2 |
Copyright (c) 2013 Red Hat, Inc. <http://www.redhat.com> |
| 3 |
This file is part of GlusterFS. |
| 4 |
|
| 5 |
This file is licensed to you under your choice of the GNU Lesser |
| 6 |
General Public License, version 3 or any later version (LGPLv3 or |
| 7 |
later), or the GNU General Public License, version 2 (GPLv2), in all |
| 8 |
cases as published by the Free Software Foundation. |
| 9 |
*/ |
| 10 |
|
| 11 |
#include "afr.h" |
| 12 |
#include "afr-self-heal.h" |
| 13 |
#include "protocol-common.h" |
| 14 |
#include "afr-messages.h" |
| 15 |
#include <glusterfs/events.h> |
| 16 |
#include <openssl/md5.h> |
| 17 |
|
| 18 |
#define HAS_HOLES(i) ((i->ia_blocks * 512) < (i->ia_size)) |
| 19 |
static int |
| 20 |
__checksum_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int op_ret, |
| 21 |
int op_errno, uint32_t weak, uint8_t *strong, dict_t *xdata) |
| 22 |
{ |
| 23 |
afr_local_t *local = NULL; |
| 24 |
struct afr_reply *replies = NULL; |
| 25 |
int i = (long)cookie; |
| 26 |
|
| 27 |
local = frame->local; |
| 28 |
replies = local->replies; |
| 29 |
|
| 30 |
replies[i].valid = 1; |
| 31 |
replies[i].op_ret = op_ret; |
| 32 |
replies[i].op_errno = op_errno; |
| 33 |
if (xdata) { |
| 34 |
replies[i].buf_has_zeroes = dict_get_str_boolean( |
| 35 |
xdata, "buf-has-zeroes", _gf_false); |
| 36 |
replies[i].fips_mode_rchecksum = dict_get_str_boolean( |
| 37 |
xdata, "fips-mode-rchecksum", _gf_false); |
| 38 |
} |
| 39 |
if (strong) { |
| 40 |
if (replies[i].fips_mode_rchecksum) { |
| 41 |
memcpy(local->replies[i].checksum, strong, SHA256_DIGEST_LENGTH); |
| 42 |
} else { |
| 43 |
memcpy(local->replies[i].checksum, strong, MD5_DIGEST_LENGTH); |
| 44 |
} |
| 45 |
} |
| 46 |
|
| 47 |
syncbarrier_wake(&local->barrier); |
| 48 |
return 0; |
| 49 |
} |
| 50 |
|
| 51 |
static gf_boolean_t |
| 52 |
__afr_can_skip_data_block_heal(call_frame_t *frame, xlator_t *this, fd_t *fd, |
| 53 |
int source, unsigned char *healed_sinks, |
| 54 |
off_t offset, size_t size, struct iatt *poststat) |
| 55 |
{ |
| 56 |
afr_private_t *priv = NULL; |
| 57 |
afr_local_t *local = NULL; |
| 58 |
unsigned char *wind_subvols = NULL; |
| 59 |
gf_boolean_t checksum_match = _gf_true; |
| 60 |
struct afr_reply *replies = NULL; |
| 61 |
dict_t *xdata = NULL; |
| 62 |
int i = 0; |
| 63 |
|
| 64 |
priv = this->private; |
| 65 |
local = frame->local; |
| 66 |
replies = local->replies; |
| 67 |
|
| 68 |
xdata = dict_new(); |
| 69 |
if (!xdata) |
| 70 |
goto out; |
| 71 |
if (dict_set_int32_sizen(xdata, "check-zero-filled", 1)) { |
| 72 |
dict_unref(xdata); |
| 73 |
goto out; |
| 74 |
} |
| 75 |
|
| 76 |
wind_subvols = alloca0(priv->child_count); |
| 77 |
for (i = 0; i < priv->child_count; i++) { |
| 78 |
if (i == source || healed_sinks[i]) |
| 79 |
wind_subvols[i] = 1; |
| 80 |
} |
| 81 |
|
| 82 |
AFR_ONLIST(wind_subvols, frame, __checksum_cbk, rchecksum, fd, offset, size, |
| 83 |
xdata); |
| 84 |
if (xdata) |
| 85 |
dict_unref(xdata); |
| 86 |
|
| 87 |
if (!replies[source].valid || replies[source].op_ret != 0) |
| 88 |
return _gf_false; |
| 89 |
|
| 90 |
for (i = 0; i < priv->child_count; i++) { |
| 91 |
if (i == source) |
| 92 |
continue; |
| 93 |
if (replies[i].valid) { |
| 94 |
if (memcmp(replies[source].checksum, replies[i].checksum, |
| 95 |
replies[source].fips_mode_rchecksum |
| 96 |
? SHA256_DIGEST_LENGTH |
| 97 |
: MD5_DIGEST_LENGTH)) { |
| 98 |
checksum_match = _gf_false; |
| 99 |
break; |
| 100 |
} |
| 101 |
} |
| 102 |
} |
| 103 |
|
| 104 |
if (checksum_match) { |
| 105 |
if (HAS_HOLES(poststat)) |
| 106 |
return _gf_true; |
| 107 |
|
| 108 |
/* For non-sparse files, we might be better off writing the |
| 109 |
* zeroes to sinks to avoid mismatch of disk-usage in bricks. */ |
| 110 |
if (local->replies[source].buf_has_zeroes) |
| 111 |
return _gf_false; |
| 112 |
else |
| 113 |
return _gf_true; |
| 114 |
} |
| 115 |
out: |
| 116 |
return _gf_false; |
| 117 |
} |
| 118 |
|
| 119 |
static gf_boolean_t |
| 120 |
__afr_is_sink_zero_filled(xlator_t *this, fd_t *fd, size_t size, off_t offset, |
| 121 |
int sink) |
| 122 |
{ |
| 123 |
afr_private_t *priv = NULL; |
| 124 |
struct iobref *iobref = NULL; |
| 125 |
struct iovec *iovec = NULL; |
| 126 |
int count = 0; |
| 127 |
int ret = 0; |
| 128 |
gf_boolean_t zero_filled = _gf_false; |
| 129 |
|
| 130 |
priv = this->private; |
| 131 |
ret = syncop_readv(priv->children[sink], fd, size, offset, 0, &iovec, |
| 132 |
&count, &iobref, NULL, NULL, NULL); |
| 133 |
if (ret < 0) |
| 134 |
goto out; |
| 135 |
ret = iov_0filled(iovec, count); |
| 136 |
if (!ret) |
| 137 |
zero_filled = _gf_true; |
| 138 |
out: |
| 139 |
if (iovec) |
| 140 |
GF_FREE(iovec); |
| 141 |
if (iobref) |
| 142 |
iobref_unref(iobref); |
| 143 |
return zero_filled; |
| 144 |
} |
| 145 |
|
| 146 |
static int |
| 147 |
__afr_selfheal_data_read_write(call_frame_t *frame, xlator_t *this, fd_t *fd, |
| 148 |
int source, unsigned char *healed_sinks, |
| 149 |
off_t offset, size_t size, |
| 150 |
struct afr_reply *replies, int type) |
| 151 |
{ |
| 152 |
struct iovec *iovec = NULL; |
| 153 |
int count = 0; |
| 154 |
struct iobref *iobref = NULL; |
| 155 |
int ret = 0; |
| 156 |
int i = 0; |
| 157 |
afr_private_t *priv = NULL; |
| 158 |
|
| 159 |
priv = this->private; |
| 160 |
|
| 161 |
ret = syncop_readv(priv->children[source], fd, size, offset, 0, &iovec, |
| 162 |
&count, &iobref, NULL, NULL, NULL); |
| 163 |
if (ret <= 0) |
| 164 |
return ret; |
| 165 |
|
| 166 |
for (i = 0; i < priv->child_count; i++) { |
| 167 |
if (!healed_sinks[i]) |
| 168 |
continue; |
| 169 |
|
| 170 |
/* |
| 171 |
* TODO: Use fiemap() and discard() to heal holes |
| 172 |
* in the future. |
| 173 |
* |
| 174 |
* For now, |
| 175 |
* |
| 176 |
* - if the source had any holes at all, |
| 177 |
* AND |
| 178 |
* - if we are writing past the original file size |
| 179 |
* of the sink |
| 180 |
* AND |
| 181 |
* - is NOT the last block of the source file. if |
| 182 |
* the block contains EOF, it has to be written |
| 183 |
* in order to set the file size even if the |
| 184 |
* last block is 0-filled. |
| 185 |
* AND |
| 186 |
* - if the read buffer is filled with only 0's |
| 187 |
* |
| 188 |
* then, skip writing to this source. We don't depend |
| 189 |
* on the write to happen to update the size as we |
| 190 |
* have performed an ftruncate() upfront anyways. |
| 191 |
*/ |
| 192 |
#define is_last_block(o, b, s) ((s >= o) && (s <= (o + b))) |
| 193 |
if (HAS_HOLES((&replies[source].poststat)) && |
| 194 |
offset >= replies[i].poststat.ia_size && |
| 195 |
!is_last_block(offset, size, replies[source].poststat.ia_size) && |
| 196 |
(iov_0filled(iovec, count) == 0)) |
| 197 |
continue; |
| 198 |
|
| 199 |
/* Avoid filling up sparse regions of the sink with 0-filled |
| 200 |
* writes.*/ |
| 201 |
if (type == AFR_SELFHEAL_DATA_FULL && |
| 202 |
HAS_HOLES((&replies[source].poststat)) && |
| 203 |
((offset + size) <= replies[i].poststat.ia_size) && |
| 204 |
(iov_0filled(iovec, count) == 0) && |
| 205 |
__afr_is_sink_zero_filled(this, fd, size, offset, i)) { |
| 206 |
continue; |
| 207 |
} |
| 208 |
|
| 209 |
ret = syncop_writev(priv->children[i], fd, iovec, count, offset, iobref, |
| 210 |
0, NULL, NULL, NULL, NULL); |
| 211 |
if (ret != iov_length(iovec, count)) { |
| 212 |
/* write() failed on this sink. unset the corresponding |
| 213 |
member in sinks[] (which is healed_sinks[] in the |
| 214 |
caller) so that this server does NOT get considered |
| 215 |
as successfully healed. |
| 216 |
*/ |
| 217 |
healed_sinks[i] = 0; |
| 218 |
} |
| 219 |
} |
| 220 |
if (iovec) |
| 221 |
GF_FREE(iovec); |
| 222 |
if (iobref) |
| 223 |
iobref_unref(iobref); |
| 224 |
|
| 225 |
return ret; |
| 226 |
} |
| 227 |
|
| 228 |
static gf_boolean_t |
| 229 |
afr_source_sinks_locked(xlator_t *this, unsigned char *locked_on, int source, |
| 230 |
unsigned char *healed_sinks) |
| 231 |
{ |
| 232 |
afr_private_t *priv = this->private; |
| 233 |
int i = 0; |
| 234 |
|
| 235 |
if (!locked_on[source]) |
| 236 |
return _gf_false; |
| 237 |
|
| 238 |
for (i = 0; i < priv->child_count; i++) { |
| 239 |
if (healed_sinks[i] && locked_on[i]) |
| 240 |
return _gf_true; |
| 241 |
} |
| 242 |
|
| 243 |
return _gf_false; |
| 244 |
} |
| 245 |
|
| 246 |
static int |
| 247 |
afr_selfheal_data_block(call_frame_t *frame, xlator_t *this, fd_t *fd, |
| 248 |
int source, unsigned char *healed_sinks, off_t offset, |
| 249 |
size_t size, int type, struct afr_reply *replies) |
| 250 |
{ |
| 251 |
int ret = -1; |
| 252 |
afr_private_t *priv = NULL; |
| 253 |
unsigned char *data_lock = NULL; |
| 254 |
|
| 255 |
priv = this->private; |
| 256 |
data_lock = alloca0(priv->child_count); |
| 257 |
|
| 258 |
gf_msg_debug(this->name, 0, "gfid:%s, offset=%jd, size=%zu", |
| 259 |
uuid_utoa(fd->inode->gfid), offset, size); |
| 260 |
|
| 261 |
ret = afr_selfheal_inodelk(frame, this, fd->inode, this->name, offset, size, |
| 262 |
data_lock); |
| 263 |
{ |
| 264 |
if (!afr_source_sinks_locked(this, data_lock, source, healed_sinks)) { |
| 265 |
ret = -ENOTCONN; |
| 266 |
goto unlock; |
| 267 |
} |
| 268 |
|
| 269 |
if (type == AFR_SELFHEAL_DATA_DIFF && |
| 270 |
__afr_can_skip_data_block_heal(frame, this, fd, source, |
| 271 |
healed_sinks, offset, size, |
| 272 |
&replies[source].poststat)) { |
| 273 |
ret = 0; |
| 274 |
goto unlock; |
| 275 |
} |
| 276 |
|
| 277 |
ret = __afr_selfheal_data_read_write( |
| 278 |
frame, this, fd, source, healed_sinks, offset, size, replies, type); |
| 279 |
} |
| 280 |
unlock: |
| 281 |
afr_selfheal_uninodelk(frame, this, fd->inode, this->name, offset, size, |
| 282 |
data_lock); |
| 283 |
return ret; |
| 284 |
} |
| 285 |
|
| 286 |
static int |
| 287 |
afr_selfheal_data_fsync(call_frame_t *frame, xlator_t *this, fd_t *fd, |
| 288 |
unsigned char *healed_sinks) |
| 289 |
{ |
| 290 |
afr_local_t *local = NULL; |
| 291 |
afr_private_t *priv = NULL; |
| 292 |
int i = 0; |
| 293 |
|
| 294 |
local = frame->local; |
| 295 |
priv = this->private; |
| 296 |
|
| 297 |
if (!priv->ensure_durability) |
| 298 |
return 0; |
| 299 |
|
| 300 |
AFR_ONLIST(healed_sinks, frame, afr_sh_generic_fop_cbk, fsync, fd, 0, NULL); |
| 301 |
|
| 302 |
for (i = 0; i < priv->child_count; i++) |
| 303 |
if (healed_sinks[i] && local->replies[i].op_ret != 0) |
| 304 |
/* fsync() failed. Do NOT consider this server |
| 305 |
as successfully healed. Mark it so. |
| 306 |
*/ |
| 307 |
healed_sinks[i] = 0; |
| 308 |
return 0; |
| 309 |
} |
| 310 |
|
| 311 |
static int |
| 312 |
afr_data_self_heal_type_get(afr_private_t *priv, unsigned char *healed_sinks, |
| 313 |
int source, struct afr_reply *replies) |
| 314 |
{ |
| 315 |
int type = AFR_SELFHEAL_DATA_FULL; |
| 316 |
int i = 0; |
| 317 |
|
| 318 |
if (priv->data_self_heal_algorithm == AFR_SELFHEAL_DATA_DYNAMIC) { |
| 319 |
type = AFR_SELFHEAL_DATA_FULL; |
| 320 |
for (i = 0; i < priv->child_count; i++) { |
| 321 |
if (!healed_sinks[i] && i != source) |
| 322 |
continue; |
| 323 |
if (replies[i].poststat.ia_size) { |
| 324 |
type = AFR_SELFHEAL_DATA_DIFF; |
| 325 |
break; |
| 326 |
} |
| 327 |
} |
| 328 |
} else { |
| 329 |
type = priv->data_self_heal_algorithm; |
| 330 |
} |
| 331 |
return type; |
| 332 |
} |
| 333 |
|
| 334 |
static int |
| 335 |
afr_selfheal_data_do(call_frame_t *frame, xlator_t *this, fd_t *fd, int source, |
| 336 |
unsigned char *healed_sinks, struct afr_reply *replies) |
| 337 |
{ |
| 338 |
afr_private_t *priv = NULL; |
| 339 |
off_t off = 0; |
| 340 |
size_t block = 0; |
| 341 |
int type = AFR_SELFHEAL_DATA_FULL; |
| 342 |
int ret = -1; |
| 343 |
call_frame_t *iter_frame = NULL; |
| 344 |
unsigned char arbiter_sink_status = 0; |
| 345 |
|
| 346 |
gf_msg(this->name, GF_LOG_INFO, 0, AFR_MSG_SELF_HEAL_INFO, |
| 347 |
"performing data selfheal on %s", uuid_utoa(fd->inode->gfid)); |
| 348 |
|
| 349 |
priv = this->private; |
| 350 |
if (priv->arbiter_count) { |
| 351 |
arbiter_sink_status = healed_sinks[ARBITER_BRICK_INDEX]; |
| 352 |
healed_sinks[ARBITER_BRICK_INDEX] = 0; |
| 353 |
} |
| 354 |
|
| 355 |
block = 128 * 1024 * priv->data_self_heal_window_size; |
| 356 |
if (HAS_HOLES((&replies[source].poststat))) { |
| 357 |
/*Reduce the possibility of data-block allocations in case of files |
| 358 |
* with holes. Correct way to fix it would be to use seek fop while |
| 359 |
* healing data*/ |
| 360 |
block = 128 * 1024; |
| 361 |
} |
| 362 |
|
| 363 |
type = afr_data_self_heal_type_get(priv, healed_sinks, source, replies); |
| 364 |
|
| 365 |
iter_frame = afr_copy_frame(frame); |
| 366 |
if (!iter_frame) { |
| 367 |
ret = -ENOMEM; |
| 368 |
goto out; |
| 369 |
} |
| 370 |
|
| 371 |
for (off = 0; off < replies[source].poststat.ia_size; off += block) { |
| 372 |
if (AFR_COUNT(healed_sinks, priv->child_count) == 0) { |
| 373 |
ret = -ENOTCONN; |
| 374 |
goto out; |
| 375 |
} |
| 376 |
|
| 377 |
ret = afr_selfheal_data_block(iter_frame, this, fd, source, |
| 378 |
healed_sinks, off, block, type, replies); |
| 379 |
if (ret < 0) |
| 380 |
goto out; |
| 381 |
|
| 382 |
AFR_STACK_RESET(iter_frame); |
| 383 |
if (iter_frame->local == NULL) { |
| 384 |
ret = -ENOTCONN; |
| 385 |
goto out; |
| 386 |
} |
| 387 |
} |
| 388 |
|
| 389 |
ret = afr_selfheal_data_fsync(frame, this, fd, healed_sinks); |
| 390 |
|
| 391 |
out: |
| 392 |
if (arbiter_sink_status) |
| 393 |
healed_sinks[ARBITER_BRICK_INDEX] = arbiter_sink_status; |
| 394 |
|
| 395 |
if (iter_frame) |
| 396 |
AFR_STACK_DESTROY(iter_frame); |
| 397 |
return ret; |
| 398 |
} |
| 399 |
|
| 400 |
static int |
| 401 |
__afr_selfheal_truncate_sinks(call_frame_t *frame, xlator_t *this, fd_t *fd, |
| 402 |
unsigned char *healed_sinks, uint64_t size) |
| 403 |
{ |
| 404 |
afr_local_t *local = NULL; |
| 405 |
afr_private_t *priv = NULL; |
| 406 |
int i = 0; |
| 407 |
|
| 408 |
local = frame->local; |
| 409 |
priv = this->private; |
| 410 |
|
| 411 |
/* This will send truncate on the arbiter brick as well if it is marked as |
| 412 |
* sink. If changelog is enabled on the volume it captures truncate as a |
| 413 |
* data transactions on the arbiter brick. This will help geo-rep to |
| 414 |
* properly sync the data from primary to secondary if arbiter is the ACTIVE |
| 415 |
* brick during syncing and which had got some entries healed for data as |
| 416 |
* part of self heal. |
| 417 |
*/ |
| 418 |
AFR_ONLIST(healed_sinks, frame, afr_sh_generic_fop_cbk, ftruncate, fd, size, |
| 419 |
NULL); |
| 420 |
|
| 421 |
for (i = 0; i < priv->child_count; i++) |
| 422 |
if (healed_sinks[i] && local->replies[i].op_ret == -1) |
| 423 |
/* truncate() failed. Do NOT consider this server |
| 424 |
as successfully healed. Mark it so. |
| 425 |
*/ |
| 426 |
healed_sinks[i] = 0; |
| 427 |
|
| 428 |
return 0; |
| 429 |
} |
| 430 |
|
| 431 |
static gf_boolean_t |
| 432 |
afr_has_source_witnesses(xlator_t *this, unsigned char *sources, |
| 433 |
uint64_t *witness) |
| 434 |
{ |
| 435 |
int i = 0; |
| 436 |
afr_private_t *priv = NULL; |
| 437 |
|
| 438 |
priv = this->private; |
| 439 |
|
| 440 |
for (i = 0; i < priv->child_count; i++) { |
| 441 |
if (sources[i] && witness[i]) |
| 442 |
return _gf_true; |
| 443 |
} |
| 444 |
return _gf_false; |
| 445 |
} |
| 446 |
|
| 447 |
static gf_boolean_t |
| 448 |
afr_does_size_mismatch(xlator_t *this, unsigned char *sources, |
| 449 |
struct afr_reply *replies) |
| 450 |
{ |
| 451 |
int i = 0; |
| 452 |
afr_private_t *priv = NULL; |
| 453 |
struct iatt *min = NULL; |
| 454 |
struct iatt *max = NULL; |
| 455 |
|
| 456 |
priv = this->private; |
| 457 |
|
| 458 |
for (i = 0; i < priv->child_count; i++) { |
| 459 |
if (!replies[i].valid) |
| 460 |
continue; |
| 461 |
|
| 462 |
if (replies[i].op_ret < 0) |
| 463 |
continue; |
| 464 |
|
| 465 |
if (!sources[i]) |
| 466 |
continue; |
| 467 |
|
| 468 |
if (AFR_IS_ARBITER_BRICK(priv, i) && (replies[i].poststat.ia_size == 0)) |
| 469 |
continue; |
| 470 |
|
| 471 |
if (!min) |
| 472 |
min = &replies[i].poststat; |
| 473 |
|
| 474 |
if (!max) |
| 475 |
max = &replies[i].poststat; |
| 476 |
|
| 477 |
if (min->ia_size > replies[i].poststat.ia_size) |
| 478 |
min = &replies[i].poststat; |
| 479 |
|
| 480 |
if (max->ia_size < replies[i].poststat.ia_size) |
| 481 |
max = &replies[i].poststat; |
| 482 |
} |
| 483 |
|
| 484 |
if (min && max) { |
| 485 |
if (min->ia_size != max->ia_size) |
| 486 |
return _gf_true; |
| 487 |
} |
| 488 |
|
| 489 |
return _gf_false; |
| 490 |
} |
| 491 |
|
| 492 |
static void |
| 493 |
afr_mark_biggest_witness_as_source(xlator_t *this, unsigned char *sources, |
| 494 |
uint64_t *witness) |
| 495 |
{ |
| 496 |
int i = 0; |
| 497 |
afr_private_t *priv = NULL; |
| 498 |
uint64_t biggest_witness = 0; |
| 499 |
|
| 500 |
priv = this->private; |
| 501 |
/* Find source with biggest witness count */ |
| 502 |
for (i = 0; i < priv->child_count; i++) { |
| 503 |
if (!sources[i]) |
| 504 |
continue; |
| 505 |
if (biggest_witness < witness[i]) |
| 506 |
biggest_witness = witness[i]; |
| 507 |
} |
| 508 |
|
| 509 |
/* Mark files with less witness count as not source */ |
| 510 |
for (i = 0; i < priv->child_count; i++) { |
| 511 |
if (!sources[i]) |
| 512 |
continue; |
| 513 |
if (witness[i] < biggest_witness) |
| 514 |
sources[i] = 0; |
| 515 |
} |
| 516 |
|
| 517 |
return; |
| 518 |
} |
| 519 |
|
| 520 |
/* This is a tie breaker function. Only one source be assigned here */ |
| 521 |
static void |
| 522 |
afr_mark_newest_file_as_source(xlator_t *this, unsigned char *sources, |
| 523 |
struct afr_reply *replies) |
| 524 |
{ |
| 525 |
int i = 0; |
| 526 |
afr_private_t *priv = NULL; |
| 527 |
int source = -1; |
| 528 |
uint64_t max_ctime = 0; |
| 529 |
|
| 530 |
priv = this->private; |
| 531 |
/* Find source with latest ctime */ |
| 532 |
for (i = 0; i < priv->child_count; i++) { |
| 533 |
if (!sources[i]) |
| 534 |
continue; |
| 535 |
|
| 536 |
if (max_ctime <= replies[i].poststat.ia_ctime) { |
| 537 |
source = i; |
| 538 |
max_ctime = replies[i].poststat.ia_ctime; |
| 539 |
} |
| 540 |
} |
| 541 |
|
| 542 |
/* Only mark one of the files as source to break ties */ |
| 543 |
memset(sources, 0, sizeof(*sources) * priv->child_count); |
| 544 |
sources[source] = 1; |
| 545 |
} |
| 546 |
|
| 547 |
static int |
| 548 |
__afr_selfheal_data_finalize_source( |
| 549 |
call_frame_t *frame, xlator_t *this, inode_t *inode, unsigned char *sources, |
| 550 |
unsigned char *sinks, unsigned char *healed_sinks, unsigned char *locked_on, |
| 551 |
unsigned char *undid_pending, struct afr_reply *replies, uint64_t *witness) |
| 552 |
{ |
| 553 |
afr_private_t *priv = NULL; |
| 554 |
int source = -1; |
| 555 |
int sources_count = 0; |
| 556 |
priv = this->private; |
| 557 |
|
| 558 |
sources_count = AFR_COUNT(sources, priv->child_count); |
| 559 |
|
| 560 |
if ((AFR_CMP(locked_on, healed_sinks, priv->child_count) == 0) || |
| 561 |
!sources_count) { |
| 562 |
/* split brain */ |
| 563 |
source = afr_mark_split_brain_source_sinks( |
| 564 |
frame, this, inode, sources, sinks, healed_sinks, locked_on, |
| 565 |
replies, AFR_DATA_TRANSACTION); |
| 566 |
if (source < 0) { |
| 567 |
gf_event(EVENT_AFR_SPLIT_BRAIN, |
| 568 |
"client-pid=%d;" |
| 569 |
"subvol=%s;type=data;" |
| 570 |
"file=%s", |
| 571 |
this->ctx->cmd_args.client_pid, this->name, |
| 572 |
uuid_utoa(inode->gfid)); |
| 573 |
return -EIO; |
| 574 |
} |
| 575 |
|
| 576 |
_afr_fav_child_reset_sink_xattrs( |
| 577 |
frame, this, inode, source, healed_sinks, undid_pending, |
| 578 |
AFR_DATA_TRANSACTION, locked_on, replies); |
| 579 |
goto out; |
| 580 |
} |
| 581 |
|
| 582 |
/* No split brain at this point. If we were called from |
| 583 |
* afr_heal_splitbrain_file(), abort.*/ |
| 584 |
if (afr_dict_contains_heal_op(frame)) |
| 585 |
return -EIO; |
| 586 |
|
| 587 |
/* If there are no witnesses/size-mismatches on sources we are done*/ |
| 588 |
if (!afr_does_size_mismatch(this, sources, replies) && |
| 589 |
!afr_has_source_witnesses(this, sources, witness)) |
| 590 |
goto out; |
| 591 |
|
| 592 |
afr_mark_largest_file_as_source(this, sources, replies); |
| 593 |
afr_mark_biggest_witness_as_source(this, sources, witness); |
| 594 |
afr_mark_newest_file_as_source(this, sources, replies); |
| 595 |
if (priv->arbiter_count) |
| 596 |
/* Choose non-arbiter brick as source for empty files. */ |
| 597 |
afr_mark_source_sinks_if_file_empty(this, sources, sinks, healed_sinks, |
| 598 |
locked_on, replies, |
| 599 |
AFR_DATA_TRANSACTION); |
| 600 |
|
| 601 |
out: |
| 602 |
afr_mark_active_sinks(this, sources, locked_on, healed_sinks); |
| 603 |
source = afr_choose_source_by_policy(priv, sources, AFR_DATA_TRANSACTION); |
| 604 |
|
| 605 |
return source; |
| 606 |
} |
| 607 |
|
| 608 |
/* |
| 609 |
* __afr_selfheal_data_prepare: |
| 610 |
* |
| 611 |
* This function inspects the on-disk xattrs and determines which subvols |
| 612 |
* are sources and sinks. |
| 613 |
* |
| 614 |
* The return value is the index of the subvolume to be used as the source |
| 615 |
* for self-healing, or -1 if no healing is necessary/split brain. |
| 616 |
*/ |
| 617 |
int |
| 618 |
__afr_selfheal_data_prepare(call_frame_t *frame, xlator_t *this, inode_t *inode, |
| 619 |
unsigned char *locked_on, unsigned char *sources, |
| 620 |
unsigned char *sinks, unsigned char *healed_sinks, |
| 621 |
unsigned char *undid_pending, |
| 622 |
struct afr_reply *replies, unsigned char *pflag) |
| 623 |
{ |
| 624 |
int ret = -1; |
| 625 |
int source = -1; |
| 626 |
afr_private_t *priv = NULL; |
| 627 |
uint64_t *witness = NULL; |
| 628 |
|
| 629 |
priv = this->private; |
| 630 |
|
| 631 |
ret = afr_selfheal_unlocked_discover(frame, inode, inode->gfid, replies); |
| 632 |
|
| 633 |
if (ret) |
| 634 |
return ret; |
| 635 |
|
| 636 |
witness = alloca0(priv->child_count * sizeof(*witness)); |
| 637 |
ret = afr_selfheal_find_direction(frame, this, replies, |
| 638 |
AFR_DATA_TRANSACTION, locked_on, sources, |
| 639 |
sinks, witness, pflag); |
| 640 |
if (ret) |
| 641 |
return ret; |
| 642 |
|
| 643 |
/* Initialize the healed_sinks[] array optimistically to |
| 644 |
the intersection of to-be-healed (i.e sinks[]) and |
| 645 |
the list of servers which are up (i.e locked_on[]). |
| 646 |
As we encounter failures in the healing process, we |
| 647 |
will unmark the respective servers in the healed_sinks[] |
| 648 |
array. |
| 649 |
*/ |
| 650 |
AFR_INTERSECT(healed_sinks, sinks, locked_on, priv->child_count); |
| 651 |
|
| 652 |
source = __afr_selfheal_data_finalize_source( |
| 653 |
frame, this, inode, sources, sinks, healed_sinks, locked_on, |
| 654 |
undid_pending, replies, witness); |
| 655 |
if (source < 0) |
| 656 |
return -EIO; |
| 657 |
|
| 658 |
return source; |
| 659 |
} |
| 660 |
|
| 661 |
static int |
| 662 |
__afr_selfheal_data(call_frame_t *frame, xlator_t *this, fd_t *fd, |
| 663 |
unsigned char *locked_on) |
| 664 |
{ |
| 665 |
afr_private_t *priv = NULL; |
| 666 |
int ret = -1; |
| 667 |
unsigned char *sources = NULL; |
| 668 |
unsigned char *sinks = NULL; |
| 669 |
unsigned char *data_lock = NULL; |
| 670 |
unsigned char *healed_sinks = NULL; |
| 671 |
unsigned char *undid_pending = NULL; |
| 672 |
struct afr_reply *locked_replies = NULL; |
| 673 |
int source = -1; |
| 674 |
gf_boolean_t did_sh = _gf_true; |
| 675 |
gf_boolean_t is_arbiter_the_only_sink = _gf_false; |
| 676 |
gf_boolean_t empty_file = _gf_false; |
| 677 |
|
| 678 |
priv = this->private; |
| 679 |
|
| 680 |
sources = alloca0(priv->child_count); |
| 681 |
sinks = alloca0(priv->child_count); |
| 682 |
healed_sinks = alloca0(priv->child_count); |
| 683 |
data_lock = alloca0(priv->child_count); |
| 684 |
undid_pending = alloca0(priv->child_count); |
| 685 |
|
| 686 |
locked_replies = alloca0(sizeof(*locked_replies) * priv->child_count); |
| 687 |
|
| 688 |
ret = afr_selfheal_inodelk(frame, this, fd->inode, this->name, 0, 0, |
| 689 |
data_lock); |
| 690 |
{ |
| 691 |
if (ret < priv->child_count) { |
| 692 |
gf_msg_debug(this->name, 0, |
| 693 |
"%s: Skipping " |
| 694 |
"self-heal as only %d number " |
| 695 |
"of subvolumes " |
| 696 |
"could be locked", |
| 697 |
uuid_utoa(fd->inode->gfid), ret); |
| 698 |
ret = -ENOTCONN; |
| 699 |
goto unlock; |
| 700 |
} |
| 701 |
|
| 702 |
ret = __afr_selfheal_data_prepare(frame, this, fd->inode, data_lock, |
| 703 |
sources, sinks, healed_sinks, |
| 704 |
undid_pending, locked_replies, NULL); |
| 705 |
if (ret < 0) |
| 706 |
goto unlock; |
| 707 |
|
| 708 |
if (AFR_COUNT(healed_sinks, priv->child_count) == 0) { |
| 709 |
did_sh = _gf_false; |
| 710 |
goto unlock; |
| 711 |
} |
| 712 |
|
| 713 |
source = ret; |
| 714 |
|
| 715 |
if (AFR_IS_ARBITER_BRICK(priv, source)) { |
| 716 |
empty_file = afr_is_file_empty_on_all_children(priv, |
| 717 |
locked_replies); |
| 718 |
if (empty_file) |
| 719 |
goto restore_time; |
| 720 |
|
| 721 |
did_sh = _gf_false; |
| 722 |
goto unlock; |
| 723 |
} |
| 724 |
|
| 725 |
ret = __afr_selfheal_truncate_sinks( |
| 726 |
frame, this, fd, healed_sinks, |
| 727 |
locked_replies[source].poststat.ia_size); |
| 728 |
if (ret < 0) |
| 729 |
goto unlock; |
| 730 |
|
| 731 |
if (priv->arbiter_count && |
| 732 |
AFR_COUNT(healed_sinks, priv->child_count) == 1 && |
| 733 |
healed_sinks[ARBITER_BRICK_INDEX]) { |
| 734 |
is_arbiter_the_only_sink = _gf_true; |
| 735 |
goto restore_time; |
| 736 |
} |
| 737 |
ret = 0; |
| 738 |
} |
| 739 |
unlock: |
| 740 |
afr_selfheal_uninodelk(frame, this, fd->inode, this->name, 0, 0, data_lock); |
| 741 |
if (ret < 0) |
| 742 |
goto out; |
| 743 |
|
| 744 |
if (!did_sh) |
| 745 |
goto out; |
| 746 |
|
| 747 |
ret = afr_selfheal_data_do(frame, this, fd, source, healed_sinks, |
| 748 |
locked_replies); |
| 749 |
if (ret) |
| 750 |
goto out; |
| 751 |
restore_time: |
| 752 |
afr_selfheal_restore_time(frame, this, fd->inode, source, healed_sinks, |
| 753 |
locked_replies); |
| 754 |
|
| 755 |
if (!is_arbiter_the_only_sink && !empty_file) { |
| 756 |
ret = afr_selfheal_inodelk(frame, this, fd->inode, this->name, 0, 0, |
| 757 |
data_lock); |
| 758 |
if (ret < priv->child_count) { |
| 759 |
ret = -ENOTCONN; |
| 760 |
did_sh = _gf_false; |
| 761 |
goto skip_undo_pending; |
| 762 |
} |
| 763 |
} |
| 764 |
ret = afr_selfheal_undo_pending( |
| 765 |
frame, this, fd->inode, sources, sinks, healed_sinks, undid_pending, |
| 766 |
AFR_DATA_TRANSACTION, locked_replies, data_lock); |
| 767 |
skip_undo_pending: |
| 768 |
afr_selfheal_uninodelk(frame, this, fd->inode, this->name, 0, 0, data_lock); |
| 769 |
out: |
| 770 |
|
| 771 |
if (did_sh) |
| 772 |
afr_log_selfheal(fd->inode->gfid, this, ret, "data", source, sources, |
| 773 |
healed_sinks); |
| 774 |
else |
| 775 |
ret = 1; |
| 776 |
|
| 777 |
if (locked_replies) |
| 778 |
afr_replies_wipe(locked_replies, priv->child_count); |
| 779 |
|
| 780 |
return ret; |
| 781 |
} |
| 782 |
|
| 783 |
static int |
| 784 |
afr_selfheal_data_open_cbk(call_frame_t *frame, void *cookie, xlator_t *this, |
| 785 |
int32_t op_ret, int32_t op_errno, fd_t *fd, |
| 786 |
dict_t *xdata) |
| 787 |
{ |
| 788 |
afr_local_t *local = NULL; |
| 789 |
int i = (long)cookie; |
| 790 |
|
| 791 |
local = frame->local; |
| 792 |
|
| 793 |
local->replies[i].valid = 1; |
| 794 |
local->replies[i].op_ret = op_ret; |
| 795 |
local->replies[i].op_errno = op_errno; |
| 796 |
|
| 797 |
syncbarrier_wake(&local->barrier); |
| 798 |
|
| 799 |
return 0; |
| 800 |
} |
| 801 |
|
| 802 |
int |
| 803 |
afr_selfheal_data_open(xlator_t *this, inode_t *inode, fd_t **fd) |
| 804 |
{ |
| 805 |
int ret = 0; |
| 806 |
fd_t *fd_tmp = NULL; |
| 807 |
loc_t loc = { |
| 808 |
0, |
| 809 |
}; |
| 810 |
call_frame_t *frame = NULL; |
| 811 |
afr_local_t *local = NULL; |
| 812 |
afr_private_t *priv = NULL; |
| 813 |
int i = 0; |
| 814 |
|
| 815 |
priv = this->private; |
| 816 |
|
| 817 |
fd_tmp = fd_create(inode, 0); |
| 818 |
if (!fd_tmp) |
| 819 |
return -ENOMEM; |
| 820 |
|
| 821 |
loc.inode = inode_ref(inode); |
| 822 |
gf_uuid_copy(loc.gfid, inode->gfid); |
| 823 |
|
| 824 |
frame = afr_frame_create(this, &ret); |
| 825 |
if (!frame) { |
| 826 |
ret = -ret; |
| 827 |
fd_unref(fd_tmp); |
| 828 |
goto out; |
| 829 |
} |
| 830 |
local = frame->local; |
| 831 |
|
| 832 |
AFR_ONLIST(local->child_up, frame, afr_selfheal_data_open_cbk, open, &loc, |
| 833 |
O_RDWR | O_LARGEFILE, fd_tmp, NULL); |
| 834 |
|
| 835 |
ret = -ENOTCONN; |
| 836 |
for (i = 0; i < priv->child_count; i++) { |
| 837 |
if (!local->replies[i].valid) |
| 838 |
continue; |
| 839 |
|
| 840 |
if (local->replies[i].op_ret < 0) { |
| 841 |
ret = -local->replies[i].op_errno; |
| 842 |
continue; |
| 843 |
} |
| 844 |
|
| 845 |
ret = 0; |
| 846 |
break; |
| 847 |
} |
| 848 |
|
| 849 |
if (ret < 0) { |
| 850 |
fd_unref(fd_tmp); |
| 851 |
goto out; |
| 852 |
} else { |
| 853 |
fd_bind(fd_tmp); |
| 854 |
} |
| 855 |
|
| 856 |
*fd = fd_tmp; |
| 857 |
out: |
| 858 |
loc_wipe(&loc); |
| 859 |
if (frame) |
| 860 |
AFR_STACK_DESTROY(frame); |
| 861 |
return ret; |
| 862 |
} |
| 863 |
|
| 864 |
int |
| 865 |
afr_selfheal_data(call_frame_t *frame, xlator_t *this, fd_t *fd) |
| 866 |
{ |
| 867 |
afr_private_t *priv = NULL; |
| 868 |
unsigned char *locked_on = NULL; |
| 869 |
int ret = 0; |
| 870 |
inode_t *inode = fd->inode; |
| 871 |
|
| 872 |
priv = this->private; |
| 873 |
|
| 874 |
locked_on = alloca0(priv->child_count); |
| 875 |
|
| 876 |
ret = afr_selfheal_tie_breaker_inodelk(frame, this, inode, priv->sh_domain, |
| 877 |
0, 0, locked_on); |
| 878 |
{ |
| 879 |
if (ret < priv->child_count) { |
| 880 |
gf_msg_debug(this->name, 0, |
| 881 |
"%s: Skipping " |
| 882 |
"self-heal as only %d number of " |
| 883 |
"subvolumes could be locked", |
| 884 |
uuid_utoa(fd->inode->gfid), ret); |
| 885 |
/* Either less than two subvols available, or another |
| 886 |
selfheal (from another server) is in progress. Skip |
| 887 |
for now in any case there isn't anything to do. |
| 888 |
*/ |
| 889 |
ret = -ENOTCONN; |
| 890 |
goto unlock; |
| 891 |
} |
| 892 |
|
| 893 |
ret = __afr_selfheal_data(frame, this, fd, locked_on); |
| 894 |
} |
| 895 |
unlock: |
| 896 |
afr_selfheal_uninodelk(frame, this, inode, priv->sh_domain, 0, 0, |
| 897 |
locked_on); |
| 898 |
|
| 899 |
return ret; |
| 900 |
} |