HAMMER 40F/Many: UNDO cleanup & stabilization.
[dragonfly.git] / sys / vfs / hammer / hammer_undo.c
1 /*
2  * Copyright (c) 2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
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
16  *    distribution.
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.
20  * 
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
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_undo.c,v 1.11 2008/05/04 09:06:45 dillon Exp $
35  */
36
37 /*
38  * HAMMER undo - undo buffer/FIFO management.
39  */
40
41 #include "hammer.h"
42
43 static int hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2);
44
45 RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
46              hammer_und_rb_compare, hammer_off_t, offset);
47
48 /*
49  * Convert a zone-3 undo offset into a zone-2 buffer offset.
50  */
51 hammer_off_t
52 hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone3_off, int *errorp)
53 {
54         hammer_volume_t root_volume;
55         hammer_blockmap_t undomap;
56         struct hammer_blockmap_layer2 *layer2;
57         hammer_off_t result_offset;
58         int i;
59
60         KKASSERT((zone3_off & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_UNDO);
61         root_volume = hammer_get_root_volume(hmp, errorp);
62         if (*errorp)
63                 return(0);
64         undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
65         KKASSERT(HAMMER_ZONE_DECODE(undomap->alloc_offset) == HAMMER_ZONE_UNDO_INDEX);
66         KKASSERT (zone3_off < undomap->alloc_offset);
67
68         i = (zone3_off & HAMMER_OFF_SHORT_MASK) / HAMMER_LARGEBLOCK_SIZE;
69         layer2 = &root_volume->ondisk->vol0_undo_array[i];
70         result_offset = layer2->u.phys_offset +
71                         (zone3_off & HAMMER_LARGEBLOCK_MASK64);
72
73         hammer_rel_volume(root_volume, 0);
74         return(result_offset);
75 }
76
77 /*
78  * Generate an UNDO record for the block of data at the specified zone1
79  * or zone2 offset.
80  *
81  * The recovery code will execute UNDOs in reverse order, allowing overlaps.
82  * All the UNDOs are executed together so if we already laid one down we
83  * do not have to lay another one down for the same range.
84  */
85 int
86 hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
87                      hammer_off_t zone_off, void *base, int len)
88 {
89         hammer_volume_t root_volume;
90         hammer_volume_ondisk_t ondisk;
91         hammer_blockmap_t undomap;
92         hammer_buffer_t buffer = NULL;
93         hammer_fifo_undo_t undo;
94         hammer_fifo_tail_t tail;
95         hammer_off_t next_offset;
96         int error;
97         int bytes;
98
99         /*
100          * Enter the offset into our undo history.  If there is an existing
101          * undo we do not have to generate a new one.
102          */
103         if (hammer_enter_undo_history(trans->hmp, zone_off, len) == EALREADY)
104                 return(0);
105
106         root_volume = trans->rootvol;
107         ondisk = root_volume->ondisk;
108         undomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
109
110         /* no undo recursion */
111         hammer_modify_volume(NULL, root_volume, NULL, 0);
112
113 again:
114         /*
115          * Allocate space in the FIFO
116          */
117         bytes = ((len + HAMMER_HEAD_ALIGN_MASK) & ~HAMMER_HEAD_ALIGN_MASK) +
118                 sizeof(struct hammer_fifo_undo) +
119                 sizeof(struct hammer_fifo_tail);
120         if (hammer_undo_space(trans->hmp) < bytes + HAMMER_BUFSIZE*2)
121                 panic("hammer: insufficient undo FIFO space!");
122
123         next_offset = undomap->next_offset;
124
125         /*
126          * Wrap next_offset
127          */
128         if (undomap->next_offset == undomap->alloc_offset) {
129                 next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
130                 undomap->next_offset = next_offset;
131                 kprintf("undo zone's next_offset wrapped\n");
132         }
133
134         undo = hammer_bread(trans->hmp, next_offset, &error, &buffer);
135
136         /*
137          * We raced another thread, try again.
138          */
139         if (undomap->next_offset != next_offset)
140                 goto again;
141
142         hammer_modify_buffer(NULL, buffer, NULL, 0);
143
144         /* XXX eventually goto again again, but for now catch it */
145         KKASSERT(undomap->next_offset == next_offset);
146
147         /*
148          * The FIFO entry would cross a buffer boundary, PAD to the end
149          * of the buffer and try again.  Due to our data alignment, the
150          * worst case (smallest) PAD record is 8 bytes.  PAD records only
151          * populate the first 8 bytes of hammer_fifo_head and the tail may
152          * be at the same offset as the head.
153          */
154         if ((next_offset ^ (next_offset + bytes)) & ~HAMMER_BUFMASK64) {
155                 bytes = HAMMER_BUFSIZE - ((int)next_offset & HAMMER_BUFMASK);
156                 tail = (void *)((char *)undo + bytes - sizeof(*tail));
157                 if ((void *)undo != (void *)tail) {
158                         tail->tail_signature = HAMMER_TAIL_SIGNATURE;
159                         tail->tail_type = HAMMER_HEAD_TYPE_PAD;
160                         tail->tail_size = bytes;
161                 }
162                 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
163                 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
164                 undo->head.hdr_size = bytes;
165                 undomap->next_offset += bytes;
166                 hammer_modify_buffer_done(buffer);
167                 goto again;
168         }
169         if (hammer_debug_general & 0x0080)
170                 kprintf("undo %016llx %d %d\n", next_offset, bytes, len);
171
172         /*
173          * We're good, create the entry.
174          */
175         undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
176         undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
177         undo->head.hdr_size = bytes;
178         undo->head.reserved01 = 0;
179         undo->head.hdr_crc = 0;
180         undo->undo_offset = zone_off;
181         undo->undo_data_bytes = len;
182         bcopy(base, undo + 1, len);
183
184         tail = (void *)((char *)undo + bytes - sizeof(*tail));
185         tail->tail_signature = HAMMER_TAIL_SIGNATURE;
186         tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
187         tail->tail_size = bytes;
188
189         undo->head.hdr_crc = crc32(undo, bytes);
190         undomap->next_offset += bytes;
191
192         hammer_modify_buffer_done(buffer);
193         hammer_modify_volume_done(root_volume);
194
195         if (buffer)
196                 hammer_rel_buffer(buffer, 0);
197         return(error);
198 }
199
200 /*
201  * UNDO HISTORY API
202  *
203  * It is not necessary to layout an undo record for the same address space
204  * multiple times.  Maintain a cache of recent undo's.
205  */
206
207 /*
208  * Enter an undo into the history.  Return EALREADY if the request completely
209  * covers a previous request.
210  */
211 int
212 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
213 {
214         hammer_undo_t node;
215         hammer_undo_t onode;
216
217         node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
218         if (node) {
219                 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
220                 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
221                 if (bytes <= node->bytes)
222                         return(EALREADY);
223                 node->bytes = bytes;
224                 return(0);
225         }
226         if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
227                 node = &hmp->undos[hmp->undo_alloc++];
228         } else {
229                 node = TAILQ_FIRST(&hmp->undo_lru_list);
230                 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
231                 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
232         }
233         node->offset = offset;
234         node->bytes = bytes;
235         TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
236         onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
237         KKASSERT(onode == NULL);
238         return(0);
239 }
240
241 void
242 hammer_clear_undo_history(hammer_mount_t hmp)
243 {
244         RB_INIT(&hmp->rb_undo_root);
245         TAILQ_INIT(&hmp->undo_lru_list);
246         hmp->undo_alloc = 0;
247 }
248
249 /*
250  * Misc helper routines.  Return available space and total space.
251  */
252 int64_t
253 hammer_undo_used(hammer_mount_t hmp)
254 {
255         hammer_blockmap_t rootmap;
256         int64_t max_bytes;
257         int64_t bytes;
258
259         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
260
261         if (rootmap->first_offset <= rootmap->next_offset) {
262                 bytes = rootmap->next_offset - rootmap->first_offset;
263         } else {
264                 bytes = rootmap->alloc_offset - rootmap->first_offset +
265                         (rootmap->next_offset & HAMMER_OFF_LONG_MASK);
266         }
267         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
268         KKASSERT(bytes <= max_bytes);
269         return(bytes);
270 }
271
272 int64_t
273 hammer_undo_space(hammer_mount_t hmp)
274 {
275         hammer_blockmap_t rootmap;
276         int64_t max_bytes;
277
278         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
279         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
280         return(max_bytes - hammer_undo_used(hmp));
281 }
282
283 int64_t
284 hammer_undo_max(hammer_mount_t hmp)
285 {
286         hammer_blockmap_t rootmap;
287         int64_t max_bytes;
288
289         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
290         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
291
292         return(max_bytes);
293 }
294
295 static int
296 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
297 {
298         if (node1->offset < node2->offset)
299                 return(-1);
300         if (node1->offset > node2->offset)
301                 return(1);
302         return(0);
303 }
304