Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of the GNU General Public License as |
| 6 | * published by the Free Software Foundation. |
| 7 | * |
| 8 | * This program is distributed in the hope that it would be useful, |
| 9 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 11 | * GNU General Public License for more details. |
| 12 | * |
| 13 | * You should have received a copy of the GNU General Public License |
| 14 | * along with this program; if not, write the Free Software Foundation, |
| 15 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
| 16 | */ |
| 17 | |
| 18 | #include "xfs.h" |
| 19 | #include "xfs_fs.h" |
| 20 | #include "xfs_types.h" |
| 21 | #include "xfs_bit.h" |
| 22 | #include "xfs_log.h" |
| 23 | #include "xfs_inum.h" |
| 24 | #include "xfs_trans.h" |
| 25 | #include "xfs_trans_priv.h" |
| 26 | #include "xfs_log_priv.h" |
| 27 | #include "xfs_sb.h" |
| 28 | #include "xfs_ag.h" |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 29 | #include "xfs_mount.h" |
| 30 | #include "xfs_error.h" |
| 31 | #include "xfs_alloc.h" |
Christoph Hellwig | e84661a | 2011-05-20 13:45:32 +0000 | [diff] [blame] | 32 | #include "xfs_discard.h" |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 33 | |
| 34 | /* |
Christoph Hellwig | 93b8a58 | 2011-12-06 21:58:07 +0000 | [diff] [blame^] | 35 | * Perform initial CIL structure initialisation. |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 36 | */ |
| 37 | int |
| 38 | xlog_cil_init( |
| 39 | struct log *log) |
| 40 | { |
| 41 | struct xfs_cil *cil; |
| 42 | struct xfs_cil_ctx *ctx; |
| 43 | |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 44 | cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL); |
| 45 | if (!cil) |
| 46 | return ENOMEM; |
| 47 | |
| 48 | ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL); |
| 49 | if (!ctx) { |
| 50 | kmem_free(cil); |
| 51 | return ENOMEM; |
| 52 | } |
| 53 | |
| 54 | INIT_LIST_HEAD(&cil->xc_cil); |
| 55 | INIT_LIST_HEAD(&cil->xc_committing); |
| 56 | spin_lock_init(&cil->xc_cil_lock); |
| 57 | init_rwsem(&cil->xc_ctx_lock); |
Dave Chinner | eb40a87 | 2010-12-21 12:09:01 +1100 | [diff] [blame] | 58 | init_waitqueue_head(&cil->xc_commit_wait); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 59 | |
| 60 | INIT_LIST_HEAD(&ctx->committing); |
| 61 | INIT_LIST_HEAD(&ctx->busy_extents); |
| 62 | ctx->sequence = 1; |
| 63 | ctx->cil = cil; |
| 64 | cil->xc_ctx = ctx; |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 65 | cil->xc_current_sequence = ctx->sequence; |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 66 | |
| 67 | cil->xc_log = log; |
| 68 | log->l_cilp = cil; |
| 69 | return 0; |
| 70 | } |
| 71 | |
| 72 | void |
| 73 | xlog_cil_destroy( |
| 74 | struct log *log) |
| 75 | { |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 76 | if (log->l_cilp->xc_ctx) { |
| 77 | if (log->l_cilp->xc_ctx->ticket) |
| 78 | xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket); |
| 79 | kmem_free(log->l_cilp->xc_ctx); |
| 80 | } |
| 81 | |
| 82 | ASSERT(list_empty(&log->l_cilp->xc_cil)); |
| 83 | kmem_free(log->l_cilp); |
| 84 | } |
| 85 | |
| 86 | /* |
| 87 | * Allocate a new ticket. Failing to get a new ticket makes it really hard to |
| 88 | * recover, so we don't allow failure here. Also, we allocate in a context that |
| 89 | * we don't want to be issuing transactions from, so we need to tell the |
| 90 | * allocation code this as well. |
| 91 | * |
| 92 | * We don't reserve any space for the ticket - we are going to steal whatever |
| 93 | * space we require from transactions as they commit. To ensure we reserve all |
| 94 | * the space required, we need to set the current reservation of the ticket to |
| 95 | * zero so that we know to steal the initial transaction overhead from the |
| 96 | * first transaction commit. |
| 97 | */ |
| 98 | static struct xlog_ticket * |
| 99 | xlog_cil_ticket_alloc( |
| 100 | struct log *log) |
| 101 | { |
| 102 | struct xlog_ticket *tic; |
| 103 | |
| 104 | tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0, |
| 105 | KM_SLEEP|KM_NOFS); |
| 106 | tic->t_trans_type = XFS_TRANS_CHECKPOINT; |
| 107 | |
| 108 | /* |
| 109 | * set the current reservation to zero so we know to steal the basic |
| 110 | * transaction overhead reservation from the first transaction commit. |
| 111 | */ |
| 112 | tic->t_curr_res = 0; |
| 113 | return tic; |
| 114 | } |
| 115 | |
| 116 | /* |
| 117 | * After the first stage of log recovery is done, we know where the head and |
| 118 | * tail of the log are. We need this log initialisation done before we can |
| 119 | * initialise the first CIL checkpoint context. |
| 120 | * |
| 121 | * Here we allocate a log ticket to track space usage during a CIL push. This |
| 122 | * ticket is passed to xlog_write() directly so that we don't slowly leak log |
| 123 | * space by failing to account for space used by log headers and additional |
| 124 | * region headers for split regions. |
| 125 | */ |
| 126 | void |
| 127 | xlog_cil_init_post_recovery( |
| 128 | struct log *log) |
| 129 | { |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 130 | log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log); |
| 131 | log->l_cilp->xc_ctx->sequence = 1; |
| 132 | log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle, |
| 133 | log->l_curr_block); |
| 134 | } |
| 135 | |
| 136 | /* |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 137 | * Format log item into a flat buffers |
| 138 | * |
| 139 | * For delayed logging, we need to hold a formatted buffer containing all the |
| 140 | * changes on the log item. This enables us to relog the item in memory and |
| 141 | * write it out asynchronously without needing to relock the object that was |
| 142 | * modified at the time it gets written into the iclog. |
| 143 | * |
| 144 | * This function builds a vector for the changes in each log item in the |
| 145 | * transaction. It then works out the length of the buffer needed for each log |
| 146 | * item, allocates them and formats the vector for the item into the buffer. |
| 147 | * The buffer is then attached to the log item are then inserted into the |
| 148 | * Committed Item List for tracking until the next checkpoint is written out. |
| 149 | * |
| 150 | * We don't set up region headers during this process; we simply copy the |
| 151 | * regions into the flat buffer. We can do this because we still have to do a |
| 152 | * formatting step to write the regions into the iclog buffer. Writing the |
| 153 | * ophdrs during the iclog write means that we can support splitting large |
| 154 | * regions across iclog boundares without needing a change in the format of the |
| 155 | * item/region encapsulation. |
| 156 | * |
| 157 | * Hence what we need to do now is change the rewrite the vector array to point |
| 158 | * to the copied region inside the buffer we just allocated. This allows us to |
| 159 | * format the regions into the iclog as though they are being formatted |
| 160 | * directly out of the objects themselves. |
| 161 | */ |
| 162 | static void |
| 163 | xlog_cil_format_items( |
| 164 | struct log *log, |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 165 | struct xfs_log_vec *log_vector) |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 166 | { |
| 167 | struct xfs_log_vec *lv; |
| 168 | |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 169 | ASSERT(log_vector); |
| 170 | for (lv = log_vector; lv; lv = lv->lv_next) { |
| 171 | void *ptr; |
| 172 | int index; |
| 173 | int len = 0; |
| 174 | |
| 175 | /* build the vector array and calculate it's length */ |
| 176 | IOP_FORMAT(lv->lv_item, lv->lv_iovecp); |
| 177 | for (index = 0; index < lv->lv_niovecs; index++) |
| 178 | len += lv->lv_iovecp[index].i_len; |
| 179 | |
| 180 | lv->lv_buf_len = len; |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 181 | lv->lv_buf = kmem_alloc(lv->lv_buf_len, KM_SLEEP|KM_NOFS); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 182 | ptr = lv->lv_buf; |
| 183 | |
| 184 | for (index = 0; index < lv->lv_niovecs; index++) { |
| 185 | struct xfs_log_iovec *vec = &lv->lv_iovecp[index]; |
| 186 | |
| 187 | memcpy(ptr, vec->i_addr, vec->i_len); |
| 188 | vec->i_addr = ptr; |
| 189 | ptr += vec->i_len; |
| 190 | } |
| 191 | ASSERT(ptr == lv->lv_buf + lv->lv_buf_len); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 192 | } |
| 193 | } |
| 194 | |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 195 | /* |
| 196 | * Prepare the log item for insertion into the CIL. Calculate the difference in |
| 197 | * log space and vectors it will consume, and if it is a new item pin it as |
| 198 | * well. |
| 199 | */ |
| 200 | STATIC void |
| 201 | xfs_cil_prepare_item( |
| 202 | struct log *log, |
| 203 | struct xfs_log_vec *lv, |
| 204 | int *len, |
| 205 | int *diff_iovecs) |
| 206 | { |
| 207 | struct xfs_log_vec *old = lv->lv_item->li_lv; |
| 208 | |
| 209 | if (old) { |
| 210 | /* existing lv on log item, space used is a delta */ |
| 211 | ASSERT(!list_empty(&lv->lv_item->li_cil)); |
| 212 | ASSERT(old->lv_buf && old->lv_buf_len && old->lv_niovecs); |
| 213 | |
| 214 | *len += lv->lv_buf_len - old->lv_buf_len; |
| 215 | *diff_iovecs += lv->lv_niovecs - old->lv_niovecs; |
| 216 | kmem_free(old->lv_buf); |
| 217 | kmem_free(old); |
| 218 | } else { |
| 219 | /* new lv, must pin the log item */ |
| 220 | ASSERT(!lv->lv_item->li_lv); |
| 221 | ASSERT(list_empty(&lv->lv_item->li_cil)); |
| 222 | |
| 223 | *len += lv->lv_buf_len; |
| 224 | *diff_iovecs += lv->lv_niovecs; |
| 225 | IOP_PIN(lv->lv_item); |
| 226 | |
| 227 | } |
| 228 | |
| 229 | /* attach new log vector to log item */ |
| 230 | lv->lv_item->li_lv = lv; |
| 231 | |
| 232 | /* |
| 233 | * If this is the first time the item is being committed to the |
| 234 | * CIL, store the sequence number on the log item so we can |
| 235 | * tell in future commits whether this is the first checkpoint |
| 236 | * the item is being committed into. |
| 237 | */ |
| 238 | if (!lv->lv_item->li_seq) |
| 239 | lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence; |
| 240 | } |
| 241 | |
| 242 | /* |
| 243 | * Insert the log items into the CIL and calculate the difference in space |
| 244 | * consumed by the item. Add the space to the checkpoint ticket and calculate |
| 245 | * if the change requires additional log metadata. If it does, take that space |
| 246 | * as well. Remove the amount of space we addded to the checkpoint ticket from |
| 247 | * the current transaction ticket so that the accounting works out correctly. |
| 248 | */ |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 249 | static void |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 250 | xlog_cil_insert_items( |
| 251 | struct log *log, |
| 252 | struct xfs_log_vec *log_vector, |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 253 | struct xlog_ticket *ticket) |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 254 | { |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 255 | struct xfs_cil *cil = log->l_cilp; |
| 256 | struct xfs_cil_ctx *ctx = cil->xc_ctx; |
| 257 | struct xfs_log_vec *lv; |
| 258 | int len = 0; |
| 259 | int diff_iovecs = 0; |
| 260 | int iclog_space; |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 261 | |
| 262 | ASSERT(log_vector); |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 263 | |
| 264 | /* |
| 265 | * Do all the accounting aggregation and switching of log vectors |
| 266 | * around in a separate loop to the insertion of items into the CIL. |
| 267 | * Then we can do a separate loop to update the CIL within a single |
| 268 | * lock/unlock pair. This reduces the number of round trips on the CIL |
| 269 | * lock from O(nr_logvectors) to O(1) and greatly reduces the overall |
| 270 | * hold time for the transaction commit. |
| 271 | * |
| 272 | * If this is the first time the item is being placed into the CIL in |
| 273 | * this context, pin it so it can't be written to disk until the CIL is |
| 274 | * flushed to the iclog and the iclog written to disk. |
| 275 | * |
| 276 | * We can do this safely because the context can't checkpoint until we |
| 277 | * are done so it doesn't matter exactly how we update the CIL. |
| 278 | */ |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 279 | for (lv = log_vector; lv; lv = lv->lv_next) |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 280 | xfs_cil_prepare_item(log, lv, &len, &diff_iovecs); |
| 281 | |
| 282 | /* account for space used by new iovec headers */ |
| 283 | len += diff_iovecs * sizeof(xlog_op_header_t); |
| 284 | |
| 285 | spin_lock(&cil->xc_cil_lock); |
| 286 | |
| 287 | /* move the items to the tail of the CIL */ |
| 288 | for (lv = log_vector; lv; lv = lv->lv_next) |
| 289 | list_move_tail(&lv->lv_item->li_cil, &cil->xc_cil); |
| 290 | |
| 291 | ctx->nvecs += diff_iovecs; |
| 292 | |
| 293 | /* |
| 294 | * Now transfer enough transaction reservation to the context ticket |
| 295 | * for the checkpoint. The context ticket is special - the unit |
| 296 | * reservation has to grow as well as the current reservation as we |
| 297 | * steal from tickets so we can correctly determine the space used |
| 298 | * during the transaction commit. |
| 299 | */ |
| 300 | if (ctx->ticket->t_curr_res == 0) { |
| 301 | /* first commit in checkpoint, steal the header reservation */ |
| 302 | ASSERT(ticket->t_curr_res >= ctx->ticket->t_unit_res + len); |
| 303 | ctx->ticket->t_curr_res = ctx->ticket->t_unit_res; |
| 304 | ticket->t_curr_res -= ctx->ticket->t_unit_res; |
| 305 | } |
| 306 | |
| 307 | /* do we need space for more log record headers? */ |
| 308 | iclog_space = log->l_iclog_size - log->l_iclog_hsize; |
| 309 | if (len > 0 && (ctx->space_used / iclog_space != |
| 310 | (ctx->space_used + len) / iclog_space)) { |
| 311 | int hdrs; |
| 312 | |
| 313 | hdrs = (len + iclog_space - 1) / iclog_space; |
| 314 | /* need to take into account split region headers, too */ |
| 315 | hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header); |
| 316 | ctx->ticket->t_unit_res += hdrs; |
| 317 | ctx->ticket->t_curr_res += hdrs; |
| 318 | ticket->t_curr_res -= hdrs; |
| 319 | ASSERT(ticket->t_curr_res >= len); |
| 320 | } |
| 321 | ticket->t_curr_res -= len; |
| 322 | ctx->space_used += len; |
| 323 | |
| 324 | spin_unlock(&cil->xc_cil_lock); |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 325 | } |
| 326 | |
| 327 | static void |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 328 | xlog_cil_free_logvec( |
| 329 | struct xfs_log_vec *log_vector) |
| 330 | { |
| 331 | struct xfs_log_vec *lv; |
| 332 | |
| 333 | for (lv = log_vector; lv; ) { |
| 334 | struct xfs_log_vec *next = lv->lv_next; |
| 335 | kmem_free(lv->lv_buf); |
| 336 | kmem_free(lv); |
| 337 | lv = next; |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | /* |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 342 | * Mark all items committed and clear busy extents. We free the log vector |
| 343 | * chains in a separate pass so that we unpin the log items as quickly as |
| 344 | * possible. |
| 345 | */ |
| 346 | static void |
| 347 | xlog_cil_committed( |
| 348 | void *args, |
| 349 | int abort) |
| 350 | { |
| 351 | struct xfs_cil_ctx *ctx = args; |
Christoph Hellwig | e84661a | 2011-05-20 13:45:32 +0000 | [diff] [blame] | 352 | struct xfs_mount *mp = ctx->cil->xc_log->l_mp; |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 353 | |
Dave Chinner | 0e57f6a | 2010-12-20 12:02:19 +1100 | [diff] [blame] | 354 | xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain, |
| 355 | ctx->start_lsn, abort); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 356 | |
Christoph Hellwig | 8a072a4 | 2011-04-24 19:06:17 +0000 | [diff] [blame] | 357 | xfs_alloc_busy_sort(&ctx->busy_extents); |
Christoph Hellwig | e84661a | 2011-05-20 13:45:32 +0000 | [diff] [blame] | 358 | xfs_alloc_busy_clear(mp, &ctx->busy_extents, |
| 359 | (mp->m_flags & XFS_MOUNT_DISCARD) && !abort); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 360 | |
| 361 | spin_lock(&ctx->cil->xc_cil_lock); |
| 362 | list_del(&ctx->committing); |
| 363 | spin_unlock(&ctx->cil->xc_cil_lock); |
| 364 | |
| 365 | xlog_cil_free_logvec(ctx->lv_chain); |
Christoph Hellwig | e84661a | 2011-05-20 13:45:32 +0000 | [diff] [blame] | 366 | |
| 367 | if (!list_empty(&ctx->busy_extents)) { |
| 368 | ASSERT(mp->m_flags & XFS_MOUNT_DISCARD); |
| 369 | |
| 370 | xfs_discard_extents(mp, &ctx->busy_extents); |
| 371 | xfs_alloc_busy_clear(mp, &ctx->busy_extents, false); |
| 372 | } |
| 373 | |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 374 | kmem_free(ctx); |
| 375 | } |
| 376 | |
| 377 | /* |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 378 | * Push the Committed Item List to the log. If @push_seq flag is zero, then it |
| 379 | * is a background flush and so we can chose to ignore it. Otherwise, if the |
| 380 | * current sequence is the same as @push_seq we need to do a flush. If |
| 381 | * @push_seq is less than the current sequence, then it has already been |
| 382 | * flushed and we don't need to do anything - the caller will wait for it to |
| 383 | * complete if necessary. |
| 384 | * |
| 385 | * @push_seq is a value rather than a flag because that allows us to do an |
| 386 | * unlocked check of the sequence number for a match. Hence we can allows log |
| 387 | * forces to run racily and not issue pushes for the same sequence twice. If we |
| 388 | * get a race between multiple pushes for the same sequence they will block on |
| 389 | * the first one and then abort, hence avoiding needless pushes. |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 390 | */ |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 391 | STATIC int |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 392 | xlog_cil_push( |
| 393 | struct log *log, |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 394 | xfs_lsn_t push_seq) |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 395 | { |
| 396 | struct xfs_cil *cil = log->l_cilp; |
| 397 | struct xfs_log_vec *lv; |
| 398 | struct xfs_cil_ctx *ctx; |
| 399 | struct xfs_cil_ctx *new_ctx; |
| 400 | struct xlog_in_core *commit_iclog; |
| 401 | struct xlog_ticket *tic; |
| 402 | int num_lv; |
| 403 | int num_iovecs; |
| 404 | int len; |
| 405 | int error = 0; |
| 406 | struct xfs_trans_header thdr; |
| 407 | struct xfs_log_iovec lhdr; |
| 408 | struct xfs_log_vec lvhdr = { NULL }; |
| 409 | xfs_lsn_t commit_lsn; |
| 410 | |
| 411 | if (!cil) |
| 412 | return 0; |
| 413 | |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 414 | ASSERT(!push_seq || push_seq <= cil->xc_ctx->sequence); |
| 415 | |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 416 | new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS); |
| 417 | new_ctx->ticket = xlog_cil_ticket_alloc(log); |
| 418 | |
Dave Chinner | 8016867 | 2010-09-24 18:13:44 +1000 | [diff] [blame] | 419 | /* |
| 420 | * Lock out transaction commit, but don't block for background pushes |
| 421 | * unless we are well over the CIL space limit. See the definition of |
| 422 | * XLOG_CIL_HARD_SPACE_LIMIT() for the full explanation of the logic |
| 423 | * used here. |
| 424 | */ |
Dave Chinner | df80615 | 2010-05-17 15:52:13 +1000 | [diff] [blame] | 425 | if (!down_write_trylock(&cil->xc_ctx_lock)) { |
Dave Chinner | 8016867 | 2010-09-24 18:13:44 +1000 | [diff] [blame] | 426 | if (!push_seq && |
| 427 | cil->xc_ctx->space_used < XLOG_CIL_HARD_SPACE_LIMIT(log)) |
Dave Chinner | df80615 | 2010-05-17 15:52:13 +1000 | [diff] [blame] | 428 | goto out_free_ticket; |
| 429 | down_write(&cil->xc_ctx_lock); |
| 430 | } |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 431 | ctx = cil->xc_ctx; |
| 432 | |
| 433 | /* check if we've anything to push */ |
| 434 | if (list_empty(&cil->xc_cil)) |
| 435 | goto out_skip; |
| 436 | |
Dave Chinner | df80615 | 2010-05-17 15:52:13 +1000 | [diff] [blame] | 437 | /* check for spurious background flush */ |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 438 | if (!push_seq && cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log)) |
| 439 | goto out_skip; |
| 440 | |
| 441 | /* check for a previously pushed seqeunce */ |
Dave Chinner | 8016867 | 2010-09-24 18:13:44 +1000 | [diff] [blame] | 442 | if (push_seq && push_seq < cil->xc_ctx->sequence) |
Dave Chinner | df80615 | 2010-05-17 15:52:13 +1000 | [diff] [blame] | 443 | goto out_skip; |
| 444 | |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 445 | /* |
| 446 | * pull all the log vectors off the items in the CIL, and |
| 447 | * remove the items from the CIL. We don't need the CIL lock |
| 448 | * here because it's only needed on the transaction commit |
| 449 | * side which is currently locked out by the flush lock. |
| 450 | */ |
| 451 | lv = NULL; |
| 452 | num_lv = 0; |
| 453 | num_iovecs = 0; |
| 454 | len = 0; |
| 455 | while (!list_empty(&cil->xc_cil)) { |
| 456 | struct xfs_log_item *item; |
| 457 | int i; |
| 458 | |
| 459 | item = list_first_entry(&cil->xc_cil, |
| 460 | struct xfs_log_item, li_cil); |
| 461 | list_del_init(&item->li_cil); |
| 462 | if (!ctx->lv_chain) |
| 463 | ctx->lv_chain = item->li_lv; |
| 464 | else |
| 465 | lv->lv_next = item->li_lv; |
| 466 | lv = item->li_lv; |
| 467 | item->li_lv = NULL; |
| 468 | |
| 469 | num_lv++; |
| 470 | num_iovecs += lv->lv_niovecs; |
| 471 | for (i = 0; i < lv->lv_niovecs; i++) |
| 472 | len += lv->lv_iovecp[i].i_len; |
| 473 | } |
| 474 | |
| 475 | /* |
| 476 | * initialise the new context and attach it to the CIL. Then attach |
| 477 | * the current context to the CIL committing lsit so it can be found |
| 478 | * during log forces to extract the commit lsn of the sequence that |
| 479 | * needs to be forced. |
| 480 | */ |
| 481 | INIT_LIST_HEAD(&new_ctx->committing); |
| 482 | INIT_LIST_HEAD(&new_ctx->busy_extents); |
| 483 | new_ctx->sequence = ctx->sequence + 1; |
| 484 | new_ctx->cil = cil; |
| 485 | cil->xc_ctx = new_ctx; |
| 486 | |
| 487 | /* |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 488 | * mirror the new sequence into the cil structure so that we can do |
| 489 | * unlocked checks against the current sequence in log forces without |
| 490 | * risking deferencing a freed context pointer. |
| 491 | */ |
| 492 | cil->xc_current_sequence = new_ctx->sequence; |
| 493 | |
| 494 | /* |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 495 | * The switch is now done, so we can drop the context lock and move out |
| 496 | * of a shared context. We can't just go straight to the commit record, |
| 497 | * though - we need to synchronise with previous and future commits so |
| 498 | * that the commit records are correctly ordered in the log to ensure |
| 499 | * that we process items during log IO completion in the correct order. |
| 500 | * |
| 501 | * For example, if we get an EFI in one checkpoint and the EFD in the |
| 502 | * next (e.g. due to log forces), we do not want the checkpoint with |
| 503 | * the EFD to be committed before the checkpoint with the EFI. Hence |
| 504 | * we must strictly order the commit records of the checkpoints so |
| 505 | * that: a) the checkpoint callbacks are attached to the iclogs in the |
| 506 | * correct order; and b) the checkpoints are replayed in correct order |
| 507 | * in log recovery. |
| 508 | * |
| 509 | * Hence we need to add this context to the committing context list so |
| 510 | * that higher sequences will wait for us to write out a commit record |
| 511 | * before they do. |
| 512 | */ |
| 513 | spin_lock(&cil->xc_cil_lock); |
| 514 | list_add(&ctx->committing, &cil->xc_committing); |
| 515 | spin_unlock(&cil->xc_cil_lock); |
| 516 | up_write(&cil->xc_ctx_lock); |
| 517 | |
| 518 | /* |
| 519 | * Build a checkpoint transaction header and write it to the log to |
| 520 | * begin the transaction. We need to account for the space used by the |
| 521 | * transaction header here as it is not accounted for in xlog_write(). |
| 522 | * |
| 523 | * The LSN we need to pass to the log items on transaction commit is |
| 524 | * the LSN reported by the first log vector write. If we use the commit |
| 525 | * record lsn then we can move the tail beyond the grant write head. |
| 526 | */ |
| 527 | tic = ctx->ticket; |
| 528 | thdr.th_magic = XFS_TRANS_HEADER_MAGIC; |
| 529 | thdr.th_type = XFS_TRANS_CHECKPOINT; |
| 530 | thdr.th_tid = tic->t_tid; |
| 531 | thdr.th_num_items = num_iovecs; |
Christoph Hellwig | 4e0d5f9 | 2010-06-23 18:11:15 +1000 | [diff] [blame] | 532 | lhdr.i_addr = &thdr; |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 533 | lhdr.i_len = sizeof(xfs_trans_header_t); |
| 534 | lhdr.i_type = XLOG_REG_TYPE_TRANSHDR; |
| 535 | tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t); |
| 536 | |
| 537 | lvhdr.lv_niovecs = 1; |
| 538 | lvhdr.lv_iovecp = &lhdr; |
| 539 | lvhdr.lv_next = ctx->lv_chain; |
| 540 | |
| 541 | error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0); |
| 542 | if (error) |
Dave Chinner | 7db37c5 | 2011-01-27 12:02:00 +1100 | [diff] [blame] | 543 | goto out_abort_free_ticket; |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 544 | |
| 545 | /* |
| 546 | * now that we've written the checkpoint into the log, strictly |
| 547 | * order the commit records so replay will get them in the right order. |
| 548 | */ |
| 549 | restart: |
| 550 | spin_lock(&cil->xc_cil_lock); |
| 551 | list_for_each_entry(new_ctx, &cil->xc_committing, committing) { |
| 552 | /* |
| 553 | * Higher sequences will wait for this one so skip them. |
| 554 | * Don't wait for own own sequence, either. |
| 555 | */ |
| 556 | if (new_ctx->sequence >= ctx->sequence) |
| 557 | continue; |
| 558 | if (!new_ctx->commit_lsn) { |
| 559 | /* |
| 560 | * It is still being pushed! Wait for the push to |
| 561 | * complete, then start again from the beginning. |
| 562 | */ |
Dave Chinner | eb40a87 | 2010-12-21 12:09:01 +1100 | [diff] [blame] | 563 | xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 564 | goto restart; |
| 565 | } |
| 566 | } |
| 567 | spin_unlock(&cil->xc_cil_lock); |
| 568 | |
Dave Chinner | 7db37c5 | 2011-01-27 12:02:00 +1100 | [diff] [blame] | 569 | /* xfs_log_done always frees the ticket on error. */ |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 570 | commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0); |
Dave Chinner | 7db37c5 | 2011-01-27 12:02:00 +1100 | [diff] [blame] | 571 | if (commit_lsn == -1) |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 572 | goto out_abort; |
| 573 | |
| 574 | /* attach all the transactions w/ busy extents to iclog */ |
| 575 | ctx->log_cb.cb_func = xlog_cil_committed; |
| 576 | ctx->log_cb.cb_arg = ctx; |
| 577 | error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb); |
| 578 | if (error) |
| 579 | goto out_abort; |
| 580 | |
| 581 | /* |
| 582 | * now the checkpoint commit is complete and we've attached the |
| 583 | * callbacks to the iclog we can assign the commit LSN to the context |
| 584 | * and wake up anyone who is waiting for the commit to complete. |
| 585 | */ |
| 586 | spin_lock(&cil->xc_cil_lock); |
| 587 | ctx->commit_lsn = commit_lsn; |
Dave Chinner | eb40a87 | 2010-12-21 12:09:01 +1100 | [diff] [blame] | 588 | wake_up_all(&cil->xc_commit_wait); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 589 | spin_unlock(&cil->xc_cil_lock); |
| 590 | |
| 591 | /* release the hounds! */ |
| 592 | return xfs_log_release_iclog(log->l_mp, commit_iclog); |
| 593 | |
| 594 | out_skip: |
| 595 | up_write(&cil->xc_ctx_lock); |
Dave Chinner | df80615 | 2010-05-17 15:52:13 +1000 | [diff] [blame] | 596 | out_free_ticket: |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 597 | xfs_log_ticket_put(new_ctx->ticket); |
| 598 | kmem_free(new_ctx); |
| 599 | return 0; |
| 600 | |
Dave Chinner | 7db37c5 | 2011-01-27 12:02:00 +1100 | [diff] [blame] | 601 | out_abort_free_ticket: |
| 602 | xfs_log_ticket_put(tic); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 603 | out_abort: |
| 604 | xlog_cil_committed(ctx, XFS_LI_ABORTED); |
| 605 | return XFS_ERROR(EIO); |
| 606 | } |
| 607 | |
| 608 | /* |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 609 | * Commit a transaction with the given vector to the Committed Item List. |
| 610 | * |
| 611 | * To do this, we need to format the item, pin it in memory if required and |
| 612 | * account for the space used by the transaction. Once we have done that we |
| 613 | * need to release the unused reservation for the transaction, attach the |
| 614 | * transaction to the checkpoint context so we carry the busy extents through |
| 615 | * to checkpoint completion, and then unlock all the items in the transaction. |
| 616 | * |
| 617 | * For more specific information about the order of operations in |
| 618 | * xfs_log_commit_cil() please refer to the comments in |
| 619 | * xfs_trans_commit_iclog(). |
| 620 | * |
| 621 | * Called with the context lock already held in read mode to lock out |
| 622 | * background commit, returns without it held once background commits are |
| 623 | * allowed again. |
| 624 | */ |
Dave Chinner | c6f990d | 2011-01-27 13:23:28 +1100 | [diff] [blame] | 625 | void |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 626 | xfs_log_commit_cil( |
| 627 | struct xfs_mount *mp, |
| 628 | struct xfs_trans *tp, |
| 629 | struct xfs_log_vec *log_vector, |
| 630 | xfs_lsn_t *commit_lsn, |
| 631 | int flags) |
| 632 | { |
| 633 | struct log *log = mp->m_log; |
| 634 | int log_flags = 0; |
| 635 | int push = 0; |
| 636 | |
| 637 | if (flags & XFS_TRANS_RELEASE_LOG_RES) |
| 638 | log_flags = XFS_LOG_REL_PERM_RESERV; |
| 639 | |
Dave Chinner | 3b93c7a | 2010-08-24 11:45:53 +1000 | [diff] [blame] | 640 | /* |
| 641 | * do all the hard work of formatting items (including memory |
| 642 | * allocation) outside the CIL context lock. This prevents stalling CIL |
| 643 | * pushes when we are low on memory and a transaction commit spends a |
| 644 | * lot of time in memory reclaim. |
| 645 | */ |
| 646 | xlog_cil_format_items(log, log_vector); |
| 647 | |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 648 | /* lock out background commit */ |
| 649 | down_read(&log->l_cilp->xc_ctx_lock); |
Dave Chinner | d1583a3 | 2010-09-24 18:14:13 +1000 | [diff] [blame] | 650 | if (commit_lsn) |
| 651 | *commit_lsn = log->l_cilp->xc_ctx->sequence; |
| 652 | |
| 653 | xlog_cil_insert_items(log, log_vector, tp->t_ticket); |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 654 | |
| 655 | /* check we didn't blow the reservation */ |
| 656 | if (tp->t_ticket->t_curr_res < 0) |
| 657 | xlog_print_tic_res(log->l_mp, tp->t_ticket); |
| 658 | |
| 659 | /* attach the transaction to the CIL if it has any busy extents */ |
| 660 | if (!list_empty(&tp->t_busy)) { |
| 661 | spin_lock(&log->l_cilp->xc_cil_lock); |
| 662 | list_splice_init(&tp->t_busy, |
| 663 | &log->l_cilp->xc_ctx->busy_extents); |
| 664 | spin_unlock(&log->l_cilp->xc_cil_lock); |
| 665 | } |
| 666 | |
| 667 | tp->t_commit_lsn = *commit_lsn; |
| 668 | xfs_log_done(mp, tp->t_ticket, NULL, log_flags); |
| 669 | xfs_trans_unreserve_and_mod_sb(tp); |
| 670 | |
| 671 | /* |
| 672 | * Once all the items of the transaction have been copied to the CIL, |
| 673 | * the items can be unlocked and freed. |
| 674 | * |
| 675 | * This needs to be done before we drop the CIL context lock because we |
| 676 | * have to update state in the log items and unlock them before they go |
| 677 | * to disk. If we don't, then the CIL checkpoint can race with us and |
| 678 | * we can run checkpoint completion before we've updated and unlocked |
| 679 | * the log items. This affects (at least) processing of stale buffers, |
| 680 | * inodes and EFIs. |
| 681 | */ |
| 682 | xfs_trans_free_items(tp, *commit_lsn, 0); |
| 683 | |
| 684 | /* check for background commit before unlock */ |
| 685 | if (log->l_cilp->xc_ctx->space_used > XLOG_CIL_SPACE_LIMIT(log)) |
| 686 | push = 1; |
| 687 | |
| 688 | up_read(&log->l_cilp->xc_ctx_lock); |
| 689 | |
| 690 | /* |
| 691 | * We need to push CIL every so often so we don't cache more than we |
| 692 | * can fit in the log. The limit really is that a checkpoint can't be |
| 693 | * more than half the log (the current checkpoint is not allowed to |
| 694 | * overwrite the previous checkpoint), but commit latency and memory |
| 695 | * usage limit this to a smaller size in most cases. |
| 696 | */ |
| 697 | if (push) |
| 698 | xlog_cil_push(log, 0); |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 699 | } |
| 700 | |
| 701 | /* |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 702 | * Conditionally push the CIL based on the sequence passed in. |
| 703 | * |
| 704 | * We only need to push if we haven't already pushed the sequence |
| 705 | * number given. Hence the only time we will trigger a push here is |
| 706 | * if the push sequence is the same as the current context. |
| 707 | * |
| 708 | * We return the current commit lsn to allow the callers to determine if a |
| 709 | * iclog flush is necessary following this call. |
| 710 | * |
| 711 | * XXX: Initially, just push the CIL unconditionally and return whatever |
| 712 | * commit lsn is there. It'll be empty, so this is broken for now. |
| 713 | */ |
| 714 | xfs_lsn_t |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 715 | xlog_cil_force_lsn( |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 716 | struct log *log, |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 717 | xfs_lsn_t sequence) |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 718 | { |
| 719 | struct xfs_cil *cil = log->l_cilp; |
| 720 | struct xfs_cil_ctx *ctx; |
| 721 | xfs_lsn_t commit_lsn = NULLCOMMITLSN; |
| 722 | |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 723 | ASSERT(sequence <= cil->xc_current_sequence); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 724 | |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 725 | /* |
| 726 | * check to see if we need to force out the current context. |
| 727 | * xlog_cil_push() handles racing pushes for the same sequence, |
| 728 | * so no need to deal with it here. |
| 729 | */ |
| 730 | if (sequence == cil->xc_current_sequence) |
| 731 | xlog_cil_push(log, sequence); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 732 | |
| 733 | /* |
| 734 | * See if we can find a previous sequence still committing. |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 735 | * We need to wait for all previous sequence commits to complete |
| 736 | * before allowing the force of push_seq to go ahead. Hence block |
| 737 | * on commits for those as well. |
| 738 | */ |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 739 | restart: |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 740 | spin_lock(&cil->xc_cil_lock); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 741 | list_for_each_entry(ctx, &cil->xc_committing, committing) { |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 742 | if (ctx->sequence > sequence) |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 743 | continue; |
| 744 | if (!ctx->commit_lsn) { |
| 745 | /* |
| 746 | * It is still being pushed! Wait for the push to |
| 747 | * complete, then start again from the beginning. |
| 748 | */ |
Dave Chinner | eb40a87 | 2010-12-21 12:09:01 +1100 | [diff] [blame] | 749 | xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock); |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 750 | goto restart; |
| 751 | } |
Dave Chinner | a44f13e | 2010-08-24 11:40:03 +1000 | [diff] [blame] | 752 | if (ctx->sequence != sequence) |
Dave Chinner | 71e330b | 2010-05-21 14:37:18 +1000 | [diff] [blame] | 753 | continue; |
| 754 | /* found it! */ |
| 755 | commit_lsn = ctx->commit_lsn; |
| 756 | } |
| 757 | spin_unlock(&cil->xc_cil_lock); |
| 758 | return commit_lsn; |
| 759 | } |
Dave Chinner | ccf7c23 | 2010-05-20 23:19:42 +1000 | [diff] [blame] | 760 | |
| 761 | /* |
| 762 | * Check if the current log item was first committed in this sequence. |
| 763 | * We can't rely on just the log item being in the CIL, we have to check |
| 764 | * the recorded commit sequence number. |
| 765 | * |
| 766 | * Note: for this to be used in a non-racy manner, it has to be called with |
| 767 | * CIL flushing locked out. As a result, it should only be used during the |
| 768 | * transaction commit process when deciding what to format into the item. |
| 769 | */ |
| 770 | bool |
| 771 | xfs_log_item_in_current_chkpt( |
| 772 | struct xfs_log_item *lip) |
| 773 | { |
| 774 | struct xfs_cil_ctx *ctx; |
| 775 | |
Dave Chinner | ccf7c23 | 2010-05-20 23:19:42 +1000 | [diff] [blame] | 776 | if (list_empty(&lip->li_cil)) |
| 777 | return false; |
| 778 | |
| 779 | ctx = lip->li_mountp->m_log->l_cilp->xc_ctx; |
| 780 | |
| 781 | /* |
| 782 | * li_seq is written on the first commit of a log item to record the |
| 783 | * first checkpoint it is written to. Hence if it is different to the |
| 784 | * current sequence, we're in a new checkpoint. |
| 785 | */ |
| 786 | if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0) |
| 787 | return false; |
| 788 | return true; |
| 789 | } |