Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
Nathan Scott | 7b71876 | 2005-11-02 14:58:39 +1100 | [diff] [blame] | 2 | * Copyright (c) 2000-2005 Silicon Graphics, Inc. |
| 3 | * All Rights Reserved. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4 | * |
Nathan Scott | 7b71876 | 2005-11-02 14:58:39 +1100 | [diff] [blame] | 5 | * This program is free software; you can redistribute it and/or |
| 6 | * modify it under the terms of the GNU General Public License as |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 7 | * published by the Free Software Foundation. |
| 8 | * |
Nathan Scott | 7b71876 | 2005-11-02 14:58:39 +1100 | [diff] [blame] | 9 | * This program is distributed in the hope that it would be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 13 | * |
Nathan Scott | 7b71876 | 2005-11-02 14:58:39 +1100 | [diff] [blame] | 14 | * You should have received a copy of the GNU General Public License |
| 15 | * along with this program; if not, write the Free Software Foundation, |
| 16 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 18 | #include "xfs.h" |
Nathan Scott | a844f45 | 2005-11-02 14:38:42 +1100 | [diff] [blame] | 19 | #include "xfs_fs.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 20 | #include "xfs_types.h" |
Nathan Scott | a844f45 | 2005-11-02 14:38:42 +1100 | [diff] [blame] | 21 | #include "xfs_bit.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 22 | #include "xfs_log.h" |
Nathan Scott | a844f45 | 2005-11-02 14:38:42 +1100 | [diff] [blame] | 23 | #include "xfs_inum.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 24 | #include "xfs_trans.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 25 | #include "xfs_sb.h" |
| 26 | #include "xfs_dir.h" |
| 27 | #include "xfs_dmapi.h" |
| 28 | #include "xfs_mount.h" |
Nathan Scott | a844f45 | 2005-11-02 14:38:42 +1100 | [diff] [blame] | 29 | #include "xfs_buf_item.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 30 | #include "xfs_trans_priv.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 31 | #include "xfs_error.h" |
| 32 | |
| 33 | |
| 34 | kmem_zone_t *xfs_buf_item_zone; |
| 35 | |
| 36 | #ifdef XFS_TRANS_DEBUG |
| 37 | /* |
| 38 | * This function uses an alternate strategy for tracking the bytes |
| 39 | * that the user requests to be logged. This can then be used |
| 40 | * in conjunction with the bli_orig array in the buf log item to |
| 41 | * catch bugs in our callers' code. |
| 42 | * |
| 43 | * We also double check the bits set in xfs_buf_item_log using a |
| 44 | * simple algorithm to check that every byte is accounted for. |
| 45 | */ |
| 46 | STATIC void |
| 47 | xfs_buf_item_log_debug( |
| 48 | xfs_buf_log_item_t *bip, |
| 49 | uint first, |
| 50 | uint last) |
| 51 | { |
| 52 | uint x; |
| 53 | uint byte; |
| 54 | uint nbytes; |
| 55 | uint chunk_num; |
| 56 | uint word_num; |
| 57 | uint bit_num; |
| 58 | uint bit_set; |
| 59 | uint *wordp; |
| 60 | |
| 61 | ASSERT(bip->bli_logged != NULL); |
| 62 | byte = first; |
| 63 | nbytes = last - first + 1; |
| 64 | bfset(bip->bli_logged, first, nbytes); |
| 65 | for (x = 0; x < nbytes; x++) { |
| 66 | chunk_num = byte >> XFS_BLI_SHIFT; |
| 67 | word_num = chunk_num >> BIT_TO_WORD_SHIFT; |
| 68 | bit_num = chunk_num & (NBWORD - 1); |
| 69 | wordp = &(bip->bli_format.blf_data_map[word_num]); |
| 70 | bit_set = *wordp & (1 << bit_num); |
| 71 | ASSERT(bit_set); |
| 72 | byte++; |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | /* |
| 77 | * This function is called when we flush something into a buffer without |
| 78 | * logging it. This happens for things like inodes which are logged |
| 79 | * separately from the buffer. |
| 80 | */ |
| 81 | void |
| 82 | xfs_buf_item_flush_log_debug( |
| 83 | xfs_buf_t *bp, |
| 84 | uint first, |
| 85 | uint last) |
| 86 | { |
| 87 | xfs_buf_log_item_t *bip; |
| 88 | uint nbytes; |
| 89 | |
| 90 | bip = XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t*); |
| 91 | if ((bip == NULL) || (bip->bli_item.li_type != XFS_LI_BUF)) { |
| 92 | return; |
| 93 | } |
| 94 | |
| 95 | ASSERT(bip->bli_logged != NULL); |
| 96 | nbytes = last - first + 1; |
| 97 | bfset(bip->bli_logged, first, nbytes); |
| 98 | } |
| 99 | |
| 100 | /* |
Nathan Scott | c41564b | 2006-03-29 08:55:14 +1000 | [diff] [blame] | 101 | * This function is called to verify that our callers have logged |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 102 | * all the bytes that they changed. |
| 103 | * |
| 104 | * It does this by comparing the original copy of the buffer stored in |
| 105 | * the buf log item's bli_orig array to the current copy of the buffer |
Nathan Scott | c41564b | 2006-03-29 08:55:14 +1000 | [diff] [blame] | 106 | * and ensuring that all bytes which mismatch are set in the bli_logged |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 107 | * array of the buf log item. |
| 108 | */ |
| 109 | STATIC void |
| 110 | xfs_buf_item_log_check( |
| 111 | xfs_buf_log_item_t *bip) |
| 112 | { |
| 113 | char *orig; |
| 114 | char *buffer; |
| 115 | int x; |
| 116 | xfs_buf_t *bp; |
| 117 | |
| 118 | ASSERT(bip->bli_orig != NULL); |
| 119 | ASSERT(bip->bli_logged != NULL); |
| 120 | |
| 121 | bp = bip->bli_buf; |
| 122 | ASSERT(XFS_BUF_COUNT(bp) > 0); |
| 123 | ASSERT(XFS_BUF_PTR(bp) != NULL); |
| 124 | orig = bip->bli_orig; |
| 125 | buffer = XFS_BUF_PTR(bp); |
| 126 | for (x = 0; x < XFS_BUF_COUNT(bp); x++) { |
| 127 | if (orig[x] != buffer[x] && !btst(bip->bli_logged, x)) |
| 128 | cmn_err(CE_PANIC, |
| 129 | "xfs_buf_item_log_check bip %x buffer %x orig %x index %d", |
| 130 | bip, bp, orig, x); |
| 131 | } |
| 132 | } |
| 133 | #else |
| 134 | #define xfs_buf_item_log_debug(x,y,z) |
| 135 | #define xfs_buf_item_log_check(x) |
| 136 | #endif |
| 137 | |
| 138 | STATIC void xfs_buf_error_relse(xfs_buf_t *bp); |
| 139 | STATIC void xfs_buf_do_callbacks(xfs_buf_t *bp, xfs_log_item_t *lip); |
| 140 | |
| 141 | /* |
| 142 | * This returns the number of log iovecs needed to log the |
| 143 | * given buf log item. |
| 144 | * |
| 145 | * It calculates this as 1 iovec for the buf log format structure |
| 146 | * and 1 for each stretch of non-contiguous chunks to be logged. |
| 147 | * Contiguous chunks are logged in a single iovec. |
| 148 | * |
| 149 | * If the XFS_BLI_STALE flag has been set, then log nothing. |
| 150 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 151 | STATIC uint |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 152 | xfs_buf_item_size( |
| 153 | xfs_buf_log_item_t *bip) |
| 154 | { |
| 155 | uint nvecs; |
| 156 | int next_bit; |
| 157 | int last_bit; |
| 158 | xfs_buf_t *bp; |
| 159 | |
| 160 | ASSERT(atomic_read(&bip->bli_refcount) > 0); |
| 161 | if (bip->bli_flags & XFS_BLI_STALE) { |
| 162 | /* |
| 163 | * The buffer is stale, so all we need to log |
| 164 | * is the buf log format structure with the |
| 165 | * cancel flag in it. |
| 166 | */ |
| 167 | xfs_buf_item_trace("SIZE STALE", bip); |
| 168 | ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL); |
| 169 | return 1; |
| 170 | } |
| 171 | |
| 172 | bp = bip->bli_buf; |
| 173 | ASSERT(bip->bli_flags & XFS_BLI_LOGGED); |
| 174 | nvecs = 1; |
| 175 | last_bit = xfs_next_bit(bip->bli_format.blf_data_map, |
| 176 | bip->bli_format.blf_map_size, 0); |
| 177 | ASSERT(last_bit != -1); |
| 178 | nvecs++; |
| 179 | while (last_bit != -1) { |
| 180 | /* |
| 181 | * This takes the bit number to start looking from and |
| 182 | * returns the next set bit from there. It returns -1 |
| 183 | * if there are no more bits set or the start bit is |
| 184 | * beyond the end of the bitmap. |
| 185 | */ |
| 186 | next_bit = xfs_next_bit(bip->bli_format.blf_data_map, |
| 187 | bip->bli_format.blf_map_size, |
| 188 | last_bit + 1); |
| 189 | /* |
| 190 | * If we run out of bits, leave the loop, |
| 191 | * else if we find a new set of bits bump the number of vecs, |
| 192 | * else keep scanning the current set of bits. |
| 193 | */ |
| 194 | if (next_bit == -1) { |
| 195 | last_bit = -1; |
| 196 | } else if (next_bit != last_bit + 1) { |
| 197 | last_bit = next_bit; |
| 198 | nvecs++; |
| 199 | } else if (xfs_buf_offset(bp, next_bit * XFS_BLI_CHUNK) != |
| 200 | (xfs_buf_offset(bp, last_bit * XFS_BLI_CHUNK) + |
| 201 | XFS_BLI_CHUNK)) { |
| 202 | last_bit = next_bit; |
| 203 | nvecs++; |
| 204 | } else { |
| 205 | last_bit++; |
| 206 | } |
| 207 | } |
| 208 | |
| 209 | xfs_buf_item_trace("SIZE NORM", bip); |
| 210 | return nvecs; |
| 211 | } |
| 212 | |
| 213 | /* |
| 214 | * This is called to fill in the vector of log iovecs for the |
| 215 | * given log buf item. It fills the first entry with a buf log |
| 216 | * format structure, and the rest point to contiguous chunks |
| 217 | * within the buffer. |
| 218 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 219 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 220 | xfs_buf_item_format( |
| 221 | xfs_buf_log_item_t *bip, |
| 222 | xfs_log_iovec_t *log_vector) |
| 223 | { |
| 224 | uint base_size; |
| 225 | uint nvecs; |
| 226 | xfs_log_iovec_t *vecp; |
| 227 | xfs_buf_t *bp; |
| 228 | int first_bit; |
| 229 | int last_bit; |
| 230 | int next_bit; |
| 231 | uint nbits; |
| 232 | uint buffer_offset; |
| 233 | |
| 234 | ASSERT(atomic_read(&bip->bli_refcount) > 0); |
| 235 | ASSERT((bip->bli_flags & XFS_BLI_LOGGED) || |
| 236 | (bip->bli_flags & XFS_BLI_STALE)); |
| 237 | bp = bip->bli_buf; |
| 238 | ASSERT(XFS_BUF_BP_ISMAPPED(bp)); |
| 239 | vecp = log_vector; |
| 240 | |
| 241 | /* |
| 242 | * The size of the base structure is the size of the |
| 243 | * declared structure plus the space for the extra words |
| 244 | * of the bitmap. We subtract one from the map size, because |
| 245 | * the first element of the bitmap is accounted for in the |
| 246 | * size of the base structure. |
| 247 | */ |
| 248 | base_size = |
| 249 | (uint)(sizeof(xfs_buf_log_format_t) + |
| 250 | ((bip->bli_format.blf_map_size - 1) * sizeof(uint))); |
| 251 | vecp->i_addr = (xfs_caddr_t)&bip->bli_format; |
| 252 | vecp->i_len = base_size; |
Tim Shimmin | 7e9c639 | 2005-09-02 16:42:05 +1000 | [diff] [blame] | 253 | XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BFORMAT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 254 | vecp++; |
| 255 | nvecs = 1; |
| 256 | |
| 257 | if (bip->bli_flags & XFS_BLI_STALE) { |
| 258 | /* |
| 259 | * The buffer is stale, so all we need to log |
| 260 | * is the buf log format structure with the |
| 261 | * cancel flag in it. |
| 262 | */ |
| 263 | xfs_buf_item_trace("FORMAT STALE", bip); |
| 264 | ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL); |
| 265 | bip->bli_format.blf_size = nvecs; |
| 266 | return; |
| 267 | } |
| 268 | |
| 269 | /* |
| 270 | * Fill in an iovec for each set of contiguous chunks. |
| 271 | */ |
| 272 | first_bit = xfs_next_bit(bip->bli_format.blf_data_map, |
| 273 | bip->bli_format.blf_map_size, 0); |
| 274 | ASSERT(first_bit != -1); |
| 275 | last_bit = first_bit; |
| 276 | nbits = 1; |
| 277 | for (;;) { |
| 278 | /* |
| 279 | * This takes the bit number to start looking from and |
| 280 | * returns the next set bit from there. It returns -1 |
| 281 | * if there are no more bits set or the start bit is |
| 282 | * beyond the end of the bitmap. |
| 283 | */ |
| 284 | next_bit = xfs_next_bit(bip->bli_format.blf_data_map, |
| 285 | bip->bli_format.blf_map_size, |
| 286 | (uint)last_bit + 1); |
| 287 | /* |
| 288 | * If we run out of bits fill in the last iovec and get |
| 289 | * out of the loop. |
| 290 | * Else if we start a new set of bits then fill in the |
| 291 | * iovec for the series we were looking at and start |
| 292 | * counting the bits in the new one. |
| 293 | * Else we're still in the same set of bits so just |
| 294 | * keep counting and scanning. |
| 295 | */ |
| 296 | if (next_bit == -1) { |
| 297 | buffer_offset = first_bit * XFS_BLI_CHUNK; |
| 298 | vecp->i_addr = xfs_buf_offset(bp, buffer_offset); |
| 299 | vecp->i_len = nbits * XFS_BLI_CHUNK; |
Tim Shimmin | 7e9c639 | 2005-09-02 16:42:05 +1000 | [diff] [blame] | 300 | XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BCHUNK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 301 | nvecs++; |
| 302 | break; |
| 303 | } else if (next_bit != last_bit + 1) { |
| 304 | buffer_offset = first_bit * XFS_BLI_CHUNK; |
| 305 | vecp->i_addr = xfs_buf_offset(bp, buffer_offset); |
| 306 | vecp->i_len = nbits * XFS_BLI_CHUNK; |
Tim Shimmin | 7e9c639 | 2005-09-02 16:42:05 +1000 | [diff] [blame] | 307 | XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BCHUNK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 308 | nvecs++; |
| 309 | vecp++; |
| 310 | first_bit = next_bit; |
| 311 | last_bit = next_bit; |
| 312 | nbits = 1; |
| 313 | } else if (xfs_buf_offset(bp, next_bit << XFS_BLI_SHIFT) != |
| 314 | (xfs_buf_offset(bp, last_bit << XFS_BLI_SHIFT) + |
| 315 | XFS_BLI_CHUNK)) { |
| 316 | buffer_offset = first_bit * XFS_BLI_CHUNK; |
| 317 | vecp->i_addr = xfs_buf_offset(bp, buffer_offset); |
| 318 | vecp->i_len = nbits * XFS_BLI_CHUNK; |
Tim Shimmin | 7e9c639 | 2005-09-02 16:42:05 +1000 | [diff] [blame] | 319 | XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_BCHUNK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 320 | /* You would think we need to bump the nvecs here too, but we do not |
| 321 | * this number is used by recovery, and it gets confused by the boundary |
| 322 | * split here |
| 323 | * nvecs++; |
| 324 | */ |
| 325 | vecp++; |
| 326 | first_bit = next_bit; |
| 327 | last_bit = next_bit; |
| 328 | nbits = 1; |
| 329 | } else { |
| 330 | last_bit++; |
| 331 | nbits++; |
| 332 | } |
| 333 | } |
| 334 | bip->bli_format.blf_size = nvecs; |
| 335 | |
| 336 | /* |
| 337 | * Check to make sure everything is consistent. |
| 338 | */ |
| 339 | xfs_buf_item_trace("FORMAT NORM", bip); |
| 340 | xfs_buf_item_log_check(bip); |
| 341 | } |
| 342 | |
| 343 | /* |
| 344 | * This is called to pin the buffer associated with the buf log |
| 345 | * item in memory so it cannot be written out. Simply call bpin() |
| 346 | * on the buffer to do this. |
| 347 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 348 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 349 | xfs_buf_item_pin( |
| 350 | xfs_buf_log_item_t *bip) |
| 351 | { |
| 352 | xfs_buf_t *bp; |
| 353 | |
| 354 | bp = bip->bli_buf; |
| 355 | ASSERT(XFS_BUF_ISBUSY(bp)); |
| 356 | ASSERT(atomic_read(&bip->bli_refcount) > 0); |
| 357 | ASSERT((bip->bli_flags & XFS_BLI_LOGGED) || |
| 358 | (bip->bli_flags & XFS_BLI_STALE)); |
| 359 | xfs_buf_item_trace("PIN", bip); |
| 360 | xfs_buftrace("XFS_PIN", bp); |
| 361 | xfs_bpin(bp); |
| 362 | } |
| 363 | |
| 364 | |
| 365 | /* |
| 366 | * This is called to unpin the buffer associated with the buf log |
| 367 | * item which was previously pinned with a call to xfs_buf_item_pin(). |
| 368 | * Just call bunpin() on the buffer to do this. |
| 369 | * |
| 370 | * Also drop the reference to the buf item for the current transaction. |
| 371 | * If the XFS_BLI_STALE flag is set and we are the last reference, |
| 372 | * then free up the buf log item and unlock the buffer. |
| 373 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 374 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 375 | xfs_buf_item_unpin( |
| 376 | xfs_buf_log_item_t *bip, |
| 377 | int stale) |
| 378 | { |
| 379 | xfs_mount_t *mp; |
| 380 | xfs_buf_t *bp; |
| 381 | int freed; |
| 382 | SPLDECL(s); |
| 383 | |
| 384 | bp = bip->bli_buf; |
| 385 | ASSERT(bp != NULL); |
| 386 | ASSERT(XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t *) == bip); |
| 387 | ASSERT(atomic_read(&bip->bli_refcount) > 0); |
| 388 | xfs_buf_item_trace("UNPIN", bip); |
| 389 | xfs_buftrace("XFS_UNPIN", bp); |
| 390 | |
| 391 | freed = atomic_dec_and_test(&bip->bli_refcount); |
| 392 | mp = bip->bli_item.li_mountp; |
| 393 | xfs_bunpin(bp); |
| 394 | if (freed && stale) { |
| 395 | ASSERT(bip->bli_flags & XFS_BLI_STALE); |
| 396 | ASSERT(XFS_BUF_VALUSEMA(bp) <= 0); |
| 397 | ASSERT(!(XFS_BUF_ISDELAYWRITE(bp))); |
| 398 | ASSERT(XFS_BUF_ISSTALE(bp)); |
| 399 | ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL); |
| 400 | xfs_buf_item_trace("UNPIN STALE", bip); |
| 401 | xfs_buftrace("XFS_UNPIN STALE", bp); |
| 402 | /* |
| 403 | * If we get called here because of an IO error, we may |
| 404 | * or may not have the item on the AIL. xfs_trans_delete_ail() |
| 405 | * will take care of that situation. |
| 406 | * xfs_trans_delete_ail() drops the AIL lock. |
| 407 | */ |
| 408 | if (bip->bli_flags & XFS_BLI_STALE_INODE) { |
| 409 | xfs_buf_do_callbacks(bp, (xfs_log_item_t *)bip); |
| 410 | XFS_BUF_SET_FSPRIVATE(bp, NULL); |
| 411 | XFS_BUF_CLR_IODONE_FUNC(bp); |
| 412 | } else { |
| 413 | AIL_LOCK(mp,s); |
| 414 | xfs_trans_delete_ail(mp, (xfs_log_item_t *)bip, s); |
| 415 | xfs_buf_item_relse(bp); |
| 416 | ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL); |
| 417 | } |
| 418 | xfs_buf_relse(bp); |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | /* |
| 423 | * this is called from uncommit in the forced-shutdown path. |
| 424 | * we need to check to see if the reference count on the log item |
| 425 | * is going to drop to zero. If so, unpin will free the log item |
| 426 | * so we need to free the item's descriptor (that points to the item) |
| 427 | * in the transaction. |
| 428 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 429 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 430 | xfs_buf_item_unpin_remove( |
| 431 | xfs_buf_log_item_t *bip, |
| 432 | xfs_trans_t *tp) |
| 433 | { |
| 434 | xfs_buf_t *bp; |
| 435 | xfs_log_item_desc_t *lidp; |
| 436 | int stale = 0; |
| 437 | |
| 438 | bp = bip->bli_buf; |
| 439 | /* |
| 440 | * will xfs_buf_item_unpin() call xfs_buf_item_relse()? |
| 441 | */ |
| 442 | if ((atomic_read(&bip->bli_refcount) == 1) && |
| 443 | (bip->bli_flags & XFS_BLI_STALE)) { |
| 444 | ASSERT(XFS_BUF_VALUSEMA(bip->bli_buf) <= 0); |
| 445 | xfs_buf_item_trace("UNPIN REMOVE", bip); |
| 446 | xfs_buftrace("XFS_UNPIN_REMOVE", bp); |
| 447 | /* |
| 448 | * yes -- clear the xaction descriptor in-use flag |
| 449 | * and free the chunk if required. We can safely |
| 450 | * do some work here and then call buf_item_unpin |
| 451 | * to do the rest because if the if is true, then |
| 452 | * we are holding the buffer locked so no one else |
| 453 | * will be able to bump up the refcount. |
| 454 | */ |
| 455 | lidp = xfs_trans_find_item(tp, (xfs_log_item_t *) bip); |
| 456 | stale = lidp->lid_flags & XFS_LID_BUF_STALE; |
| 457 | xfs_trans_free_item(tp, lidp); |
| 458 | /* |
| 459 | * Since the transaction no longer refers to the buffer, |
| 460 | * the buffer should no longer refer to the transaction. |
| 461 | */ |
| 462 | XFS_BUF_SET_FSPRIVATE2(bp, NULL); |
| 463 | } |
| 464 | |
| 465 | xfs_buf_item_unpin(bip, stale); |
| 466 | |
| 467 | return; |
| 468 | } |
| 469 | |
| 470 | /* |
| 471 | * This is called to attempt to lock the buffer associated with this |
| 472 | * buf log item. Don't sleep on the buffer lock. If we can't get |
| 473 | * the lock right away, return 0. If we can get the lock, pull the |
| 474 | * buffer from the free list, mark it busy, and return 1. |
| 475 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 476 | STATIC uint |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 477 | xfs_buf_item_trylock( |
| 478 | xfs_buf_log_item_t *bip) |
| 479 | { |
| 480 | xfs_buf_t *bp; |
| 481 | |
| 482 | bp = bip->bli_buf; |
| 483 | |
| 484 | if (XFS_BUF_ISPINNED(bp)) { |
| 485 | return XFS_ITEM_PINNED; |
| 486 | } |
| 487 | |
| 488 | if (!XFS_BUF_CPSEMA(bp)) { |
| 489 | return XFS_ITEM_LOCKED; |
| 490 | } |
| 491 | |
| 492 | /* |
| 493 | * Remove the buffer from the free list. Only do this |
| 494 | * if it's on the free list. Private buffers like the |
| 495 | * superblock buffer are not. |
| 496 | */ |
| 497 | XFS_BUF_HOLD(bp); |
| 498 | |
| 499 | ASSERT(!(bip->bli_flags & XFS_BLI_STALE)); |
| 500 | xfs_buf_item_trace("TRYLOCK SUCCESS", bip); |
| 501 | return XFS_ITEM_SUCCESS; |
| 502 | } |
| 503 | |
| 504 | /* |
| 505 | * Release the buffer associated with the buf log item. |
| 506 | * If there is no dirty logged data associated with the |
| 507 | * buffer recorded in the buf log item, then free the |
| 508 | * buf log item and remove the reference to it in the |
| 509 | * buffer. |
| 510 | * |
| 511 | * This call ignores the recursion count. It is only called |
| 512 | * when the buffer should REALLY be unlocked, regardless |
| 513 | * of the recursion count. |
| 514 | * |
| 515 | * If the XFS_BLI_HOLD flag is set in the buf log item, then |
| 516 | * free the log item if necessary but do not unlock the buffer. |
| 517 | * This is for support of xfs_trans_bhold(). Make sure the |
| 518 | * XFS_BLI_HOLD field is cleared if we don't free the item. |
| 519 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 520 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 521 | xfs_buf_item_unlock( |
| 522 | xfs_buf_log_item_t *bip) |
| 523 | { |
| 524 | int aborted; |
| 525 | xfs_buf_t *bp; |
| 526 | uint hold; |
| 527 | |
| 528 | bp = bip->bli_buf; |
| 529 | xfs_buftrace("XFS_UNLOCK", bp); |
| 530 | |
| 531 | /* |
| 532 | * Clear the buffer's association with this transaction. |
| 533 | */ |
| 534 | XFS_BUF_SET_FSPRIVATE2(bp, NULL); |
| 535 | |
| 536 | /* |
| 537 | * If this is a transaction abort, don't return early. |
| 538 | * Instead, allow the brelse to happen. |
| 539 | * Normally it would be done for stale (cancelled) buffers |
| 540 | * at unpin time, but we'll never go through the pin/unpin |
| 541 | * cycle if we abort inside commit. |
| 542 | */ |
| 543 | aborted = (bip->bli_item.li_flags & XFS_LI_ABORTED) != 0; |
| 544 | |
| 545 | /* |
| 546 | * If the buf item is marked stale, then don't do anything. |
| 547 | * We'll unlock the buffer and free the buf item when the |
| 548 | * buffer is unpinned for the last time. |
| 549 | */ |
| 550 | if (bip->bli_flags & XFS_BLI_STALE) { |
| 551 | bip->bli_flags &= ~XFS_BLI_LOGGED; |
| 552 | xfs_buf_item_trace("UNLOCK STALE", bip); |
| 553 | ASSERT(bip->bli_format.blf_flags & XFS_BLI_CANCEL); |
| 554 | if (!aborted) |
| 555 | return; |
| 556 | } |
| 557 | |
| 558 | /* |
| 559 | * Drop the transaction's reference to the log item if |
| 560 | * it was not logged as part of the transaction. Otherwise |
| 561 | * we'll drop the reference in xfs_buf_item_unpin() when |
| 562 | * the transaction is really through with the buffer. |
| 563 | */ |
| 564 | if (!(bip->bli_flags & XFS_BLI_LOGGED)) { |
| 565 | atomic_dec(&bip->bli_refcount); |
| 566 | } else { |
| 567 | /* |
| 568 | * Clear the logged flag since this is per |
| 569 | * transaction state. |
| 570 | */ |
| 571 | bip->bli_flags &= ~XFS_BLI_LOGGED; |
| 572 | } |
| 573 | |
| 574 | /* |
| 575 | * Before possibly freeing the buf item, determine if we should |
| 576 | * release the buffer at the end of this routine. |
| 577 | */ |
| 578 | hold = bip->bli_flags & XFS_BLI_HOLD; |
| 579 | xfs_buf_item_trace("UNLOCK", bip); |
| 580 | |
| 581 | /* |
| 582 | * If the buf item isn't tracking any data, free it. |
| 583 | * Otherwise, if XFS_BLI_HOLD is set clear it. |
| 584 | */ |
| 585 | if (xfs_count_bits(bip->bli_format.blf_data_map, |
| 586 | bip->bli_format.blf_map_size, 0) == 0) { |
| 587 | xfs_buf_item_relse(bp); |
| 588 | } else if (hold) { |
| 589 | bip->bli_flags &= ~XFS_BLI_HOLD; |
| 590 | } |
| 591 | |
| 592 | /* |
| 593 | * Release the buffer if XFS_BLI_HOLD was not set. |
| 594 | */ |
| 595 | if (!hold) { |
| 596 | xfs_buf_relse(bp); |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | /* |
| 601 | * This is called to find out where the oldest active copy of the |
| 602 | * buf log item in the on disk log resides now that the last log |
| 603 | * write of it completed at the given lsn. |
| 604 | * We always re-log all the dirty data in a buffer, so usually the |
| 605 | * latest copy in the on disk log is the only one that matters. For |
| 606 | * those cases we simply return the given lsn. |
| 607 | * |
| 608 | * The one exception to this is for buffers full of newly allocated |
| 609 | * inodes. These buffers are only relogged with the XFS_BLI_INODE_BUF |
| 610 | * flag set, indicating that only the di_next_unlinked fields from the |
| 611 | * inodes in the buffers will be replayed during recovery. If the |
| 612 | * original newly allocated inode images have not yet been flushed |
| 613 | * when the buffer is so relogged, then we need to make sure that we |
| 614 | * keep the old images in the 'active' portion of the log. We do this |
| 615 | * by returning the original lsn of that transaction here rather than |
| 616 | * the current one. |
| 617 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 618 | STATIC xfs_lsn_t |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 619 | xfs_buf_item_committed( |
| 620 | xfs_buf_log_item_t *bip, |
| 621 | xfs_lsn_t lsn) |
| 622 | { |
| 623 | xfs_buf_item_trace("COMMITTED", bip); |
| 624 | if ((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) && |
| 625 | (bip->bli_item.li_lsn != 0)) { |
| 626 | return bip->bli_item.li_lsn; |
| 627 | } |
| 628 | return (lsn); |
| 629 | } |
| 630 | |
| 631 | /* |
| 632 | * This is called when the transaction holding the buffer is aborted. |
| 633 | * Just behave as if the transaction had been cancelled. If we're shutting down |
| 634 | * and have aborted this transaction, we'll trap this buffer when it tries to |
| 635 | * get written out. |
| 636 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 637 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 638 | xfs_buf_item_abort( |
| 639 | xfs_buf_log_item_t *bip) |
| 640 | { |
| 641 | xfs_buf_t *bp; |
| 642 | |
| 643 | bp = bip->bli_buf; |
| 644 | xfs_buftrace("XFS_ABORT", bp); |
| 645 | XFS_BUF_SUPER_STALE(bp); |
| 646 | xfs_buf_item_unlock(bip); |
| 647 | return; |
| 648 | } |
| 649 | |
| 650 | /* |
| 651 | * This is called to asynchronously write the buffer associated with this |
| 652 | * buf log item out to disk. The buffer will already have been locked by |
| 653 | * a successful call to xfs_buf_item_trylock(). If the buffer still has |
| 654 | * B_DELWRI set, then get it going out to disk with a call to bawrite(). |
| 655 | * If not, then just release the buffer. |
| 656 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 657 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 658 | xfs_buf_item_push( |
| 659 | xfs_buf_log_item_t *bip) |
| 660 | { |
| 661 | xfs_buf_t *bp; |
| 662 | |
| 663 | ASSERT(!(bip->bli_flags & XFS_BLI_STALE)); |
| 664 | xfs_buf_item_trace("PUSH", bip); |
| 665 | |
| 666 | bp = bip->bli_buf; |
| 667 | |
| 668 | if (XFS_BUF_ISDELAYWRITE(bp)) { |
| 669 | xfs_bawrite(bip->bli_item.li_mountp, bp); |
| 670 | } else { |
| 671 | xfs_buf_relse(bp); |
| 672 | } |
| 673 | } |
| 674 | |
| 675 | /* ARGSUSED */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 676 | STATIC void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 677 | xfs_buf_item_committing(xfs_buf_log_item_t *bip, xfs_lsn_t commit_lsn) |
| 678 | { |
| 679 | } |
| 680 | |
| 681 | /* |
| 682 | * This is the ops vector shared by all buf log items. |
| 683 | */ |
Christoph Hellwig | ba0f32d | 2005-06-21 15:36:52 +1000 | [diff] [blame] | 684 | STATIC struct xfs_item_ops xfs_buf_item_ops = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 685 | .iop_size = (uint(*)(xfs_log_item_t*))xfs_buf_item_size, |
| 686 | .iop_format = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*)) |
| 687 | xfs_buf_item_format, |
| 688 | .iop_pin = (void(*)(xfs_log_item_t*))xfs_buf_item_pin, |
| 689 | .iop_unpin = (void(*)(xfs_log_item_t*, int))xfs_buf_item_unpin, |
| 690 | .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t *)) |
| 691 | xfs_buf_item_unpin_remove, |
| 692 | .iop_trylock = (uint(*)(xfs_log_item_t*))xfs_buf_item_trylock, |
| 693 | .iop_unlock = (void(*)(xfs_log_item_t*))xfs_buf_item_unlock, |
| 694 | .iop_committed = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t)) |
| 695 | xfs_buf_item_committed, |
| 696 | .iop_push = (void(*)(xfs_log_item_t*))xfs_buf_item_push, |
| 697 | .iop_abort = (void(*)(xfs_log_item_t*))xfs_buf_item_abort, |
| 698 | .iop_pushbuf = NULL, |
| 699 | .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t)) |
| 700 | xfs_buf_item_committing |
| 701 | }; |
| 702 | |
| 703 | |
| 704 | /* |
| 705 | * Allocate a new buf log item to go with the given buffer. |
| 706 | * Set the buffer's b_fsprivate field to point to the new |
| 707 | * buf log item. If there are other item's attached to the |
| 708 | * buffer (see xfs_buf_attach_iodone() below), then put the |
| 709 | * buf log item at the front. |
| 710 | */ |
| 711 | void |
| 712 | xfs_buf_item_init( |
| 713 | xfs_buf_t *bp, |
| 714 | xfs_mount_t *mp) |
| 715 | { |
| 716 | xfs_log_item_t *lip; |
| 717 | xfs_buf_log_item_t *bip; |
| 718 | int chunks; |
| 719 | int map_size; |
| 720 | |
| 721 | /* |
| 722 | * Check to see if there is already a buf log item for |
| 723 | * this buffer. If there is, it is guaranteed to be |
| 724 | * the first. If we do already have one, there is |
| 725 | * nothing to do here so return. |
| 726 | */ |
| 727 | if (XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *) != mp) |
| 728 | XFS_BUF_SET_FSPRIVATE3(bp, mp); |
| 729 | XFS_BUF_SET_BDSTRAT_FUNC(bp, xfs_bdstrat_cb); |
| 730 | if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) { |
| 731 | lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *); |
| 732 | if (lip->li_type == XFS_LI_BUF) { |
| 733 | return; |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | /* |
| 738 | * chunks is the number of XFS_BLI_CHUNK size pieces |
| 739 | * the buffer can be divided into. Make sure not to |
| 740 | * truncate any pieces. map_size is the size of the |
| 741 | * bitmap needed to describe the chunks of the buffer. |
| 742 | */ |
| 743 | chunks = (int)((XFS_BUF_COUNT(bp) + (XFS_BLI_CHUNK - 1)) >> XFS_BLI_SHIFT); |
| 744 | map_size = (int)((chunks + NBWORD) >> BIT_TO_WORD_SHIFT); |
| 745 | |
| 746 | bip = (xfs_buf_log_item_t*)kmem_zone_zalloc(xfs_buf_item_zone, |
| 747 | KM_SLEEP); |
| 748 | bip->bli_item.li_type = XFS_LI_BUF; |
| 749 | bip->bli_item.li_ops = &xfs_buf_item_ops; |
| 750 | bip->bli_item.li_mountp = mp; |
| 751 | bip->bli_buf = bp; |
| 752 | bip->bli_format.blf_type = XFS_LI_BUF; |
| 753 | bip->bli_format.blf_blkno = (__int64_t)XFS_BUF_ADDR(bp); |
| 754 | bip->bli_format.blf_len = (ushort)BTOBB(XFS_BUF_COUNT(bp)); |
| 755 | bip->bli_format.blf_map_size = map_size; |
| 756 | #ifdef XFS_BLI_TRACE |
| 757 | bip->bli_trace = ktrace_alloc(XFS_BLI_TRACE_SIZE, KM_SLEEP); |
| 758 | #endif |
| 759 | |
| 760 | #ifdef XFS_TRANS_DEBUG |
| 761 | /* |
| 762 | * Allocate the arrays for tracking what needs to be logged |
| 763 | * and what our callers request to be logged. bli_orig |
| 764 | * holds a copy of the original, clean buffer for comparison |
| 765 | * against, and bli_logged keeps a 1 bit flag per byte in |
| 766 | * the buffer to indicate which bytes the callers have asked |
| 767 | * to have logged. |
| 768 | */ |
| 769 | bip->bli_orig = (char *)kmem_alloc(XFS_BUF_COUNT(bp), KM_SLEEP); |
| 770 | memcpy(bip->bli_orig, XFS_BUF_PTR(bp), XFS_BUF_COUNT(bp)); |
| 771 | bip->bli_logged = (char *)kmem_zalloc(XFS_BUF_COUNT(bp) / NBBY, KM_SLEEP); |
| 772 | #endif |
| 773 | |
| 774 | /* |
| 775 | * Put the buf item into the list of items attached to the |
| 776 | * buffer at the front. |
| 777 | */ |
| 778 | if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) { |
| 779 | bip->bli_item.li_bio_list = |
| 780 | XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *); |
| 781 | } |
| 782 | XFS_BUF_SET_FSPRIVATE(bp, bip); |
| 783 | } |
| 784 | |
| 785 | |
| 786 | /* |
| 787 | * Mark bytes first through last inclusive as dirty in the buf |
| 788 | * item's bitmap. |
| 789 | */ |
| 790 | void |
| 791 | xfs_buf_item_log( |
| 792 | xfs_buf_log_item_t *bip, |
| 793 | uint first, |
| 794 | uint last) |
| 795 | { |
| 796 | uint first_bit; |
| 797 | uint last_bit; |
| 798 | uint bits_to_set; |
| 799 | uint bits_set; |
| 800 | uint word_num; |
| 801 | uint *wordp; |
| 802 | uint bit; |
| 803 | uint end_bit; |
| 804 | uint mask; |
| 805 | |
| 806 | /* |
| 807 | * Mark the item as having some dirty data for |
| 808 | * quick reference in xfs_buf_item_dirty. |
| 809 | */ |
| 810 | bip->bli_flags |= XFS_BLI_DIRTY; |
| 811 | |
| 812 | /* |
| 813 | * Convert byte offsets to bit numbers. |
| 814 | */ |
| 815 | first_bit = first >> XFS_BLI_SHIFT; |
| 816 | last_bit = last >> XFS_BLI_SHIFT; |
| 817 | |
| 818 | /* |
| 819 | * Calculate the total number of bits to be set. |
| 820 | */ |
| 821 | bits_to_set = last_bit - first_bit + 1; |
| 822 | |
| 823 | /* |
| 824 | * Get a pointer to the first word in the bitmap |
| 825 | * to set a bit in. |
| 826 | */ |
| 827 | word_num = first_bit >> BIT_TO_WORD_SHIFT; |
| 828 | wordp = &(bip->bli_format.blf_data_map[word_num]); |
| 829 | |
| 830 | /* |
| 831 | * Calculate the starting bit in the first word. |
| 832 | */ |
| 833 | bit = first_bit & (uint)(NBWORD - 1); |
| 834 | |
| 835 | /* |
| 836 | * First set any bits in the first word of our range. |
| 837 | * If it starts at bit 0 of the word, it will be |
| 838 | * set below rather than here. That is what the variable |
| 839 | * bit tells us. The variable bits_set tracks the number |
| 840 | * of bits that have been set so far. End_bit is the number |
| 841 | * of the last bit to be set in this word plus one. |
| 842 | */ |
| 843 | if (bit) { |
| 844 | end_bit = MIN(bit + bits_to_set, (uint)NBWORD); |
| 845 | mask = ((1 << (end_bit - bit)) - 1) << bit; |
| 846 | *wordp |= mask; |
| 847 | wordp++; |
| 848 | bits_set = end_bit - bit; |
| 849 | } else { |
| 850 | bits_set = 0; |
| 851 | } |
| 852 | |
| 853 | /* |
| 854 | * Now set bits a whole word at a time that are between |
| 855 | * first_bit and last_bit. |
| 856 | */ |
| 857 | while ((bits_to_set - bits_set) >= NBWORD) { |
| 858 | *wordp |= 0xffffffff; |
| 859 | bits_set += NBWORD; |
| 860 | wordp++; |
| 861 | } |
| 862 | |
| 863 | /* |
| 864 | * Finally, set any bits left to be set in one last partial word. |
| 865 | */ |
| 866 | end_bit = bits_to_set - bits_set; |
| 867 | if (end_bit) { |
| 868 | mask = (1 << end_bit) - 1; |
| 869 | *wordp |= mask; |
| 870 | } |
| 871 | |
| 872 | xfs_buf_item_log_debug(bip, first, last); |
| 873 | } |
| 874 | |
| 875 | |
| 876 | /* |
| 877 | * Return 1 if the buffer has some data that has been logged (at any |
| 878 | * point, not just the current transaction) and 0 if not. |
| 879 | */ |
| 880 | uint |
| 881 | xfs_buf_item_dirty( |
| 882 | xfs_buf_log_item_t *bip) |
| 883 | { |
| 884 | return (bip->bli_flags & XFS_BLI_DIRTY); |
| 885 | } |
| 886 | |
| 887 | /* |
| 888 | * This is called when the buf log item is no longer needed. It should |
| 889 | * free the buf log item associated with the given buffer and clear |
| 890 | * the buffer's pointer to the buf log item. If there are no more |
| 891 | * items in the list, clear the b_iodone field of the buffer (see |
| 892 | * xfs_buf_attach_iodone() below). |
| 893 | */ |
| 894 | void |
| 895 | xfs_buf_item_relse( |
| 896 | xfs_buf_t *bp) |
| 897 | { |
| 898 | xfs_buf_log_item_t *bip; |
| 899 | |
| 900 | xfs_buftrace("XFS_RELSE", bp); |
| 901 | bip = XFS_BUF_FSPRIVATE(bp, xfs_buf_log_item_t*); |
| 902 | XFS_BUF_SET_FSPRIVATE(bp, bip->bli_item.li_bio_list); |
| 903 | if ((XFS_BUF_FSPRIVATE(bp, void *) == NULL) && |
| 904 | (XFS_BUF_IODONE_FUNC(bp) != NULL)) { |
| 905 | ASSERT((XFS_BUF_ISUNINITIAL(bp)) == 0); |
| 906 | XFS_BUF_CLR_IODONE_FUNC(bp); |
| 907 | } |
| 908 | |
| 909 | #ifdef XFS_TRANS_DEBUG |
| 910 | kmem_free(bip->bli_orig, XFS_BUF_COUNT(bp)); |
| 911 | bip->bli_orig = NULL; |
| 912 | kmem_free(bip->bli_logged, XFS_BUF_COUNT(bp) / NBBY); |
| 913 | bip->bli_logged = NULL; |
| 914 | #endif /* XFS_TRANS_DEBUG */ |
| 915 | |
| 916 | #ifdef XFS_BLI_TRACE |
| 917 | ktrace_free(bip->bli_trace); |
| 918 | #endif |
| 919 | kmem_zone_free(xfs_buf_item_zone, bip); |
| 920 | } |
| 921 | |
| 922 | |
| 923 | /* |
| 924 | * Add the given log item with its callback to the list of callbacks |
| 925 | * to be called when the buffer's I/O completes. If it is not set |
| 926 | * already, set the buffer's b_iodone() routine to be |
| 927 | * xfs_buf_iodone_callbacks() and link the log item into the list of |
| 928 | * items rooted at b_fsprivate. Items are always added as the second |
| 929 | * entry in the list if there is a first, because the buf item code |
| 930 | * assumes that the buf log item is first. |
| 931 | */ |
| 932 | void |
| 933 | xfs_buf_attach_iodone( |
| 934 | xfs_buf_t *bp, |
| 935 | void (*cb)(xfs_buf_t *, xfs_log_item_t *), |
| 936 | xfs_log_item_t *lip) |
| 937 | { |
| 938 | xfs_log_item_t *head_lip; |
| 939 | |
| 940 | ASSERT(XFS_BUF_ISBUSY(bp)); |
| 941 | ASSERT(XFS_BUF_VALUSEMA(bp) <= 0); |
| 942 | |
| 943 | lip->li_cb = cb; |
| 944 | if (XFS_BUF_FSPRIVATE(bp, void *) != NULL) { |
| 945 | head_lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *); |
| 946 | lip->li_bio_list = head_lip->li_bio_list; |
| 947 | head_lip->li_bio_list = lip; |
| 948 | } else { |
| 949 | XFS_BUF_SET_FSPRIVATE(bp, lip); |
| 950 | } |
| 951 | |
| 952 | ASSERT((XFS_BUF_IODONE_FUNC(bp) == xfs_buf_iodone_callbacks) || |
| 953 | (XFS_BUF_IODONE_FUNC(bp) == NULL)); |
| 954 | XFS_BUF_SET_IODONE_FUNC(bp, xfs_buf_iodone_callbacks); |
| 955 | } |
| 956 | |
| 957 | STATIC void |
| 958 | xfs_buf_do_callbacks( |
| 959 | xfs_buf_t *bp, |
| 960 | xfs_log_item_t *lip) |
| 961 | { |
| 962 | xfs_log_item_t *nlip; |
| 963 | |
| 964 | while (lip != NULL) { |
| 965 | nlip = lip->li_bio_list; |
| 966 | ASSERT(lip->li_cb != NULL); |
| 967 | /* |
| 968 | * Clear the next pointer so we don't have any |
| 969 | * confusion if the item is added to another buf. |
| 970 | * Don't touch the log item after calling its |
| 971 | * callback, because it could have freed itself. |
| 972 | */ |
| 973 | lip->li_bio_list = NULL; |
| 974 | lip->li_cb(bp, lip); |
| 975 | lip = nlip; |
| 976 | } |
| 977 | } |
| 978 | |
| 979 | /* |
| 980 | * This is the iodone() function for buffers which have had callbacks |
| 981 | * attached to them by xfs_buf_attach_iodone(). It should remove each |
| 982 | * log item from the buffer's list and call the callback of each in turn. |
| 983 | * When done, the buffer's fsprivate field is set to NULL and the buffer |
| 984 | * is unlocked with a call to iodone(). |
| 985 | */ |
| 986 | void |
| 987 | xfs_buf_iodone_callbacks( |
| 988 | xfs_buf_t *bp) |
| 989 | { |
| 990 | xfs_log_item_t *lip; |
| 991 | static ulong lasttime; |
| 992 | static xfs_buftarg_t *lasttarg; |
| 993 | xfs_mount_t *mp; |
| 994 | |
| 995 | ASSERT(XFS_BUF_FSPRIVATE(bp, void *) != NULL); |
| 996 | lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *); |
| 997 | |
| 998 | if (XFS_BUF_GETERROR(bp) != 0) { |
| 999 | /* |
| 1000 | * If we've already decided to shutdown the filesystem |
| 1001 | * because of IO errors, there's no point in giving this |
| 1002 | * a retry. |
| 1003 | */ |
| 1004 | mp = lip->li_mountp; |
| 1005 | if (XFS_FORCED_SHUTDOWN(mp)) { |
| 1006 | ASSERT(XFS_BUF_TARGET(bp) == mp->m_ddev_targp); |
| 1007 | XFS_BUF_SUPER_STALE(bp); |
| 1008 | xfs_buftrace("BUF_IODONE_CB", bp); |
| 1009 | xfs_buf_do_callbacks(bp, lip); |
| 1010 | XFS_BUF_SET_FSPRIVATE(bp, NULL); |
| 1011 | XFS_BUF_CLR_IODONE_FUNC(bp); |
| 1012 | |
| 1013 | /* |
| 1014 | * XFS_SHUT flag gets set when we go thru the |
| 1015 | * entire buffer cache and deliberately start |
| 1016 | * throwing away delayed write buffers. |
| 1017 | * Since there's no biowait done on those, |
| 1018 | * we should just brelse them. |
| 1019 | */ |
| 1020 | if (XFS_BUF_ISSHUT(bp)) { |
| 1021 | XFS_BUF_UNSHUT(bp); |
| 1022 | xfs_buf_relse(bp); |
| 1023 | } else { |
| 1024 | xfs_biodone(bp); |
| 1025 | } |
| 1026 | |
| 1027 | return; |
| 1028 | } |
| 1029 | |
| 1030 | if ((XFS_BUF_TARGET(bp) != lasttarg) || |
| 1031 | (time_after(jiffies, (lasttime + 5*HZ)))) { |
| 1032 | lasttime = jiffies; |
Nathan Scott | b657452 | 2006-06-09 15:29:40 +1000 | [diff] [blame] | 1033 | cmn_err(CE_ALERT, "Device %s, XFS metadata write error" |
| 1034 | " block 0x%llx in %s", |
| 1035 | XFS_BUFTARG_NAME(XFS_BUF_TARGET(bp)), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1036 | (__uint64_t)XFS_BUF_ADDR(bp), mp->m_fsname); |
| 1037 | } |
| 1038 | lasttarg = XFS_BUF_TARGET(bp); |
| 1039 | |
| 1040 | if (XFS_BUF_ISASYNC(bp)) { |
| 1041 | /* |
| 1042 | * If the write was asynchronous then noone will be |
| 1043 | * looking for the error. Clear the error state |
| 1044 | * and write the buffer out again delayed write. |
| 1045 | * |
| 1046 | * XXXsup This is OK, so long as we catch these |
| 1047 | * before we start the umount; we don't want these |
| 1048 | * DELWRI metadata bufs to be hanging around. |
| 1049 | */ |
| 1050 | XFS_BUF_ERROR(bp,0); /* errno of 0 unsets the flag */ |
| 1051 | |
| 1052 | if (!(XFS_BUF_ISSTALE(bp))) { |
| 1053 | XFS_BUF_DELAYWRITE(bp); |
| 1054 | XFS_BUF_DONE(bp); |
| 1055 | XFS_BUF_SET_START(bp); |
| 1056 | } |
| 1057 | ASSERT(XFS_BUF_IODONE_FUNC(bp)); |
| 1058 | xfs_buftrace("BUF_IODONE ASYNC", bp); |
| 1059 | xfs_buf_relse(bp); |
| 1060 | } else { |
| 1061 | /* |
| 1062 | * If the write of the buffer was not asynchronous, |
| 1063 | * then we want to make sure to return the error |
| 1064 | * to the caller of bwrite(). Because of this we |
| 1065 | * cannot clear the B_ERROR state at this point. |
| 1066 | * Instead we install a callback function that |
| 1067 | * will be called when the buffer is released, and |
| 1068 | * that routine will clear the error state and |
| 1069 | * set the buffer to be written out again after |
| 1070 | * some delay. |
| 1071 | */ |
| 1072 | /* We actually overwrite the existing b-relse |
| 1073 | function at times, but we're gonna be shutting down |
| 1074 | anyway. */ |
| 1075 | XFS_BUF_SET_BRELSE_FUNC(bp,xfs_buf_error_relse); |
| 1076 | XFS_BUF_DONE(bp); |
| 1077 | XFS_BUF_V_IODONESEMA(bp); |
| 1078 | } |
| 1079 | return; |
| 1080 | } |
| 1081 | #ifdef XFSERRORDEBUG |
| 1082 | xfs_buftrace("XFS BUFCB NOERR", bp); |
| 1083 | #endif |
| 1084 | xfs_buf_do_callbacks(bp, lip); |
| 1085 | XFS_BUF_SET_FSPRIVATE(bp, NULL); |
| 1086 | XFS_BUF_CLR_IODONE_FUNC(bp); |
| 1087 | xfs_biodone(bp); |
| 1088 | } |
| 1089 | |
| 1090 | /* |
| 1091 | * This is a callback routine attached to a buffer which gets an error |
| 1092 | * when being written out synchronously. |
| 1093 | */ |
| 1094 | STATIC void |
| 1095 | xfs_buf_error_relse( |
| 1096 | xfs_buf_t *bp) |
| 1097 | { |
| 1098 | xfs_log_item_t *lip; |
| 1099 | xfs_mount_t *mp; |
| 1100 | |
| 1101 | lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *); |
| 1102 | mp = (xfs_mount_t *)lip->li_mountp; |
| 1103 | ASSERT(XFS_BUF_TARGET(bp) == mp->m_ddev_targp); |
| 1104 | |
| 1105 | XFS_BUF_STALE(bp); |
| 1106 | XFS_BUF_DONE(bp); |
| 1107 | XFS_BUF_UNDELAYWRITE(bp); |
| 1108 | XFS_BUF_ERROR(bp,0); |
| 1109 | xfs_buftrace("BUF_ERROR_RELSE", bp); |
| 1110 | if (! XFS_FORCED_SHUTDOWN(mp)) |
Nathan Scott | 7d04a33 | 2006-06-09 14:58:38 +1000 | [diff] [blame] | 1111 | xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1112 | /* |
| 1113 | * We have to unpin the pinned buffers so do the |
| 1114 | * callbacks. |
| 1115 | */ |
| 1116 | xfs_buf_do_callbacks(bp, lip); |
| 1117 | XFS_BUF_SET_FSPRIVATE(bp, NULL); |
| 1118 | XFS_BUF_CLR_IODONE_FUNC(bp); |
| 1119 | XFS_BUF_SET_BRELSE_FUNC(bp,NULL); |
| 1120 | xfs_buf_relse(bp); |
| 1121 | } |
| 1122 | |
| 1123 | |
| 1124 | /* |
| 1125 | * This is the iodone() function for buffers which have been |
| 1126 | * logged. It is called when they are eventually flushed out. |
| 1127 | * It should remove the buf item from the AIL, and free the buf item. |
| 1128 | * It is called by xfs_buf_iodone_callbacks() above which will take |
| 1129 | * care of cleaning up the buffer itself. |
| 1130 | */ |
| 1131 | /* ARGSUSED */ |
| 1132 | void |
| 1133 | xfs_buf_iodone( |
| 1134 | xfs_buf_t *bp, |
| 1135 | xfs_buf_log_item_t *bip) |
| 1136 | { |
| 1137 | struct xfs_mount *mp; |
| 1138 | SPLDECL(s); |
| 1139 | |
| 1140 | ASSERT(bip->bli_buf == bp); |
| 1141 | |
| 1142 | mp = bip->bli_item.li_mountp; |
| 1143 | |
| 1144 | /* |
| 1145 | * If we are forcibly shutting down, this may well be |
| 1146 | * off the AIL already. That's because we simulate the |
| 1147 | * log-committed callbacks to unpin these buffers. Or we may never |
| 1148 | * have put this item on AIL because of the transaction was |
| 1149 | * aborted forcibly. xfs_trans_delete_ail() takes care of these. |
| 1150 | * |
| 1151 | * Either way, AIL is useless if we're forcing a shutdown. |
| 1152 | */ |
| 1153 | AIL_LOCK(mp,s); |
| 1154 | /* |
| 1155 | * xfs_trans_delete_ail() drops the AIL lock. |
| 1156 | */ |
| 1157 | xfs_trans_delete_ail(mp, (xfs_log_item_t *)bip, s); |
| 1158 | |
| 1159 | #ifdef XFS_TRANS_DEBUG |
| 1160 | kmem_free(bip->bli_orig, XFS_BUF_COUNT(bp)); |
| 1161 | bip->bli_orig = NULL; |
| 1162 | kmem_free(bip->bli_logged, XFS_BUF_COUNT(bp) / NBBY); |
| 1163 | bip->bli_logged = NULL; |
| 1164 | #endif /* XFS_TRANS_DEBUG */ |
| 1165 | |
| 1166 | #ifdef XFS_BLI_TRACE |
| 1167 | ktrace_free(bip->bli_trace); |
| 1168 | #endif |
| 1169 | kmem_zone_free(xfs_buf_item_zone, bip); |
| 1170 | } |
| 1171 | |
| 1172 | #if defined(XFS_BLI_TRACE) |
| 1173 | void |
| 1174 | xfs_buf_item_trace( |
| 1175 | char *id, |
| 1176 | xfs_buf_log_item_t *bip) |
| 1177 | { |
| 1178 | xfs_buf_t *bp; |
| 1179 | ASSERT(bip->bli_trace != NULL); |
| 1180 | |
| 1181 | bp = bip->bli_buf; |
| 1182 | ktrace_enter(bip->bli_trace, |
| 1183 | (void *)id, |
| 1184 | (void *)bip->bli_buf, |
| 1185 | (void *)((unsigned long)bip->bli_flags), |
| 1186 | (void *)((unsigned long)bip->bli_recur), |
| 1187 | (void *)((unsigned long)atomic_read(&bip->bli_refcount)), |
| 1188 | (void *)((unsigned long) |
| 1189 | (0xFFFFFFFF & XFS_BUF_ADDR(bp) >> 32)), |
| 1190 | (void *)((unsigned long)(0xFFFFFFFF & XFS_BUF_ADDR(bp))), |
| 1191 | (void *)((unsigned long)XFS_BUF_COUNT(bp)), |
| 1192 | (void *)((unsigned long)XFS_BUF_BFLAGS(bp)), |
| 1193 | XFS_BUF_FSPRIVATE(bp, void *), |
| 1194 | XFS_BUF_FSPRIVATE2(bp, void *), |
| 1195 | (void *)(unsigned long)XFS_BUF_ISPINNED(bp), |
| 1196 | (void *)XFS_BUF_IODONE_FUNC(bp), |
| 1197 | (void *)((unsigned long)(XFS_BUF_VALUSEMA(bp))), |
| 1198 | (void *)bip->bli_item.li_desc, |
| 1199 | (void *)((unsigned long)bip->bli_item.li_flags)); |
| 1200 | } |
| 1201 | #endif /* XFS_BLI_TRACE */ |