2 * Copyright (c) 2010 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * HAMMER redo - REDO record support for the UNDO/REDO FIFO.
38 * See also hammer_undo.c
43 RB_GENERATE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
44 hammer_redo_rb_compare, hammer_off_t, redo_fifo_start);
47 * HAMMER version 4+ REDO support.
49 * REDO records are used to improve fsync() performance. Instead of having
50 * to go through a complete double-flush cycle involving at least two disk
51 * synchronizations the fsync need only flush UNDO/REDO FIFO buffers through
52 * the related REDO records, which is a single synchronization requiring
53 * no track seeking. If a recovery becomes necessary the recovery code
54 * will generate logical data writes based on the REDO records encountered.
55 * That is, the recovery code will UNDO any partial meta-data/data writes
56 * at the raw disk block level and then REDO the data writes at the logical
60 hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
61 hammer_off_t file_off, uint32_t flags,
65 hammer_volume_t root_volume;
66 hammer_blockmap_t undomap;
67 hammer_buffer_t buffer = NULL;
68 hammer_fifo_redo_t redo;
69 hammer_fifo_tail_t tail;
70 hammer_off_t next_offset;
80 root_volume = trans->rootvol;
81 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
84 * No undo recursion when modifying the root volume
86 hammer_modify_volume_noundo(NULL, root_volume);
87 hammer_lock_ex(&hmp->undo_lock);
89 /* undo had better not roll over (loose test) */
90 if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
91 hpanic("insufficient UNDO/REDO FIFO space for redo!");
94 * Loop until the undo for the entire range has been laid down.
95 * Loop at least once (len might be 0 as a degenerate case).
99 * Fetch the layout offset in the UNDO FIFO, wrap it as
102 if (undomap->next_offset == undomap->alloc_offset)
103 undomap->next_offset = HAMMER_ENCODE_UNDO(0);
104 next_offset = undomap->next_offset;
107 * This is a tail-chasing FIFO, when we hit the start of a new
108 * buffer we don't have to read it in.
110 if ((next_offset & HAMMER_BUFMASK) == 0) {
111 redo = hammer_bnew(hmp, next_offset, &error, &buffer);
112 hammer_format_undo(redo, hmp->undo_seqno ^ 0x40000000);
114 redo = hammer_bread(hmp, next_offset, &error, &buffer);
118 hammer_modify_buffer_noundo(NULL, buffer);
121 * Calculate how big a media structure fits up to the next
122 * alignment point and how large a data payload we can
125 * If n calculates to 0 or negative there is no room for
126 * anything but a PAD.
128 bytes = HAMMER_UNDO_ALIGN -
129 ((int)next_offset & HAMMER_UNDO_MASK);
131 (int)sizeof(struct hammer_fifo_redo) -
132 (int)sizeof(struct hammer_fifo_tail);
135 * If available space is insufficient for any payload
136 * we have to lay down a PAD.
138 * The minimum PAD is 8 bytes and the head and tail will
139 * overlap each other in that case. PADs do not have
140 * sequence numbers or CRCs.
142 * A PAD may not start on a boundary. That is, every
143 * 512-byte block in the UNDO/REDO FIFO must begin with
144 * a record containing a sequence number.
147 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
148 KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
149 tail = (void *)((char *)redo + bytes - sizeof(*tail));
150 if ((void *)redo != (void *)tail) {
151 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
152 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
153 tail->tail_size = bytes;
155 redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
156 redo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
157 redo->head.hdr_size = bytes;
158 /* NO CRC OR SEQ NO */
159 undomap->next_offset += bytes;
160 hammer_modify_buffer_done(buffer);
161 hammer_stats_redo += bytes;
166 * When generating an inode-related REDO record we track
167 * the point in the UNDO/REDO FIFO containing the inode's
168 * earliest REDO record. See hammer_generate_redo_sync().
170 * redo_fifo_next is cleared when an inode is staged to
171 * the backend and then used to determine how to reassign
172 * redo_fifo_start after the inode flush completes.
175 redo->redo_objid = ip->obj_id;
176 redo->redo_localization = ip->obj_localization;
177 if ((ip->flags & HAMMER_INODE_RDIRTY) == 0) {
178 ip->redo_fifo_start = next_offset;
179 if (RB_INSERT(hammer_redo_rb_tree,
180 &hmp->rb_redo_root, ip)) {
181 hpanic("cannot insert inode %p on "
184 ip->flags |= HAMMER_INODE_RDIRTY;
186 if (ip->redo_fifo_next == 0)
187 ip->redo_fifo_next = next_offset;
189 redo->redo_objid = 0;
190 redo->redo_localization = 0;
194 * Calculate the actual payload and recalculate the size
195 * of the media structure as necessary. If no data buffer
196 * is supplied there is no payload.
200 } else if (n > len) {
203 bytes = HAMMER_HEAD_DOALIGN(n) +
204 (int)sizeof(struct hammer_fifo_redo) +
205 (int)sizeof(struct hammer_fifo_tail);
206 if (hammer_debug_general & 0x0080) {
207 hdkprintf("redo %016jx %d %d\n",
208 (intmax_t)next_offset, bytes, n);
211 redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
212 redo->head.hdr_type = HAMMER_HEAD_TYPE_REDO;
213 redo->head.hdr_size = bytes;
214 redo->head.hdr_seq = hmp->undo_seqno++;
215 redo->head.hdr_crc = 0;
216 redo->redo_offset = file_off;
217 redo->redo_flags = flags;
220 * Incremental payload. If no payload we throw the entire
221 * len into redo_data_bytes and will not loop.
224 redo->redo_data_bytes = n;
225 bcopy(base, redo + 1, n);
227 base = (char *)base + n;
230 redo->redo_data_bytes = len;
235 tail = (void *)((char *)redo + bytes - sizeof(*tail));
236 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
237 tail->tail_type = HAMMER_HEAD_TYPE_REDO;
238 tail->tail_size = bytes;
240 KKASSERT(bytes >= sizeof(redo->head));
241 hammer_crc_set_fifo_head(&redo->head, bytes);
242 undomap->next_offset += bytes;
243 hammer_stats_redo += bytes;
246 * Before we finish off the buffer we have to deal with any
247 * junk between the end of the media structure we just laid
248 * down and the UNDO alignment boundary. We do this by laying
249 * down a dummy PAD. Even though we will probably overwrite
250 * it almost immediately we have to do this so recovery runs
251 * can iterate the UNDO space without having to depend on
252 * the indices in the volume header.
254 * This dummy PAD will be overwritten on the next undo so
255 * we do not adjust undomap->next_offset.
257 bytes = HAMMER_UNDO_ALIGN -
258 ((int)undomap->next_offset & HAMMER_UNDO_MASK);
259 if (bytes != HAMMER_UNDO_ALIGN) {
260 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
261 redo = (void *)(tail + 1);
262 tail = (void *)((char *)redo + bytes - sizeof(*tail));
263 if ((void *)redo != (void *)tail) {
264 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
265 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
266 tail->tail_size = bytes;
268 redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
269 redo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
270 redo->head.hdr_size = bytes;
271 /* NO CRC OR SEQ NO */
273 hammer_modify_buffer_done(buffer);
277 hammer_modify_volume_done(root_volume);
278 hammer_unlock(&hmp->undo_lock);
281 hammer_rel_buffer(buffer, 0);
284 * Make sure the nominal undo span contains at least one REDO_SYNC,
285 * otherwise the REDO recovery will not be triggered.
287 if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
288 flags != HAMMER_REDO_SYNC) {
289 hammer_generate_redo_sync(trans);
296 * Generate a REDO SYNC record. At least one such record must be generated
297 * in the nominal recovery span for the recovery code to be able to run
298 * REDOs outside of the span.
300 * The SYNC record contains the aggregate earliest UNDO/REDO FIFO offset
301 * for all inodes with active REDOs. This changes dynamically as inodes
304 * During recovery stage2 any new flush cycles must specify the original
305 * redo sync offset. That way a crash will re-run the REDOs, at least
306 * up to the point where the UNDO FIFO does not overwrite the area.
309 hammer_generate_redo_sync(hammer_transaction_t trans)
311 hammer_mount_t hmp = trans->hmp;
313 hammer_off_t redo_fifo_start;
315 if (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) {
317 redo_fifo_start = hmp->recover_stage2_offset;
319 ip = RB_FIRST(hammer_redo_rb_tree, &hmp->rb_redo_root);
321 redo_fifo_start = ip->redo_fifo_start;
325 if (redo_fifo_start) {
326 if (hammer_debug_io & 0x0004) {
327 hdkprintf("SYNC IP %p %016jx\n",
328 ip, (intmax_t)redo_fifo_start);
330 hammer_generate_redo(trans, NULL, redo_fifo_start,
331 HAMMER_REDO_SYNC, NULL, 0);
332 trans->hmp->flags |= HAMMER_MOUNT_REDO_SYNC;
337 * This is called when an inode is queued to the backend.
340 hammer_redo_fifo_start_flush(hammer_inode_t ip)
342 ip->redo_fifo_next = 0;
346 * This is called when an inode backend flush is finished. We have to make
347 * sure that RDIRTY is not set unless dirty bufs are present. Dirty bufs
348 * can get destroyed through operations such as truncations and leave
349 * us with a stale redo_fifo_next.
352 hammer_redo_fifo_end_flush(hammer_inode_t ip)
354 hammer_mount_t hmp = ip->hmp;
356 if (ip->flags & HAMMER_INODE_RDIRTY) {
357 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
358 ip->flags &= ~HAMMER_INODE_RDIRTY;
360 if ((ip->flags & HAMMER_INODE_BUFS) == 0)
361 ip->redo_fifo_next = 0;
362 if (ip->redo_fifo_next) {
363 ip->redo_fifo_start = ip->redo_fifo_next;
364 if (RB_INSERT(hammer_redo_rb_tree, &hmp->rb_redo_root, ip)) {
365 hpanic("cannot reinsert inode %p on redo FIFO", ip);
367 ip->flags |= HAMMER_INODE_RDIRTY;