Merge git://git.kernel.org/pub/scm/linux/kernel/git/steve/gfs2-3.0-nmw
[~shefty/rdma-dev.git] / fs / gfs2 / trans.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2006 Red Hat, Inc.  All rights reserved.
4  *
5  * This copyrighted material is made available to anyone wishing to use,
6  * modify, copy, or redistribute it subject to the terms and conditions
7  * of the GNU General Public License version 2.
8  */
9
10 #include <linux/sched.h>
11 #include <linux/slab.h>
12 #include <linux/spinlock.h>
13 #include <linux/completion.h>
14 #include <linux/buffer_head.h>
15 #include <linux/kallsyms.h>
16 #include <linux/gfs2_ondisk.h>
17
18 #include "gfs2.h"
19 #include "incore.h"
20 #include "glock.h"
21 #include "inode.h"
22 #include "log.h"
23 #include "lops.h"
24 #include "meta_io.h"
25 #include "trans.h"
26 #include "util.h"
27 #include "trace_gfs2.h"
28
29 int gfs2_trans_begin(struct gfs2_sbd *sdp, unsigned int blocks,
30                      unsigned int revokes)
31 {
32         struct gfs2_trans *tr;
33         int error;
34
35         BUG_ON(current->journal_info);
36         BUG_ON(blocks == 0 && revokes == 0);
37
38         if (!test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))
39                 return -EROFS;
40
41         tr = kzalloc(sizeof(struct gfs2_trans), GFP_NOFS);
42         if (!tr)
43                 return -ENOMEM;
44
45         tr->tr_ip = (unsigned long)__builtin_return_address(0);
46         tr->tr_blocks = blocks;
47         tr->tr_revokes = revokes;
48         tr->tr_reserved = 1;
49         if (blocks)
50                 tr->tr_reserved += 6 + blocks;
51         if (revokes)
52                 tr->tr_reserved += gfs2_struct2blk(sdp, revokes,
53                                                    sizeof(u64));
54         sb_start_intwrite(sdp->sd_vfs);
55         gfs2_holder_init(sdp->sd_trans_gl, LM_ST_SHARED, 0, &tr->tr_t_gh);
56
57         error = gfs2_glock_nq(&tr->tr_t_gh);
58         if (error)
59                 goto fail_holder_uninit;
60
61         error = gfs2_log_reserve(sdp, tr->tr_reserved);
62         if (error)
63                 goto fail_gunlock;
64
65         current->journal_info = tr;
66
67         return 0;
68
69 fail_gunlock:
70         gfs2_glock_dq(&tr->tr_t_gh);
71
72 fail_holder_uninit:
73         sb_end_intwrite(sdp->sd_vfs);
74         gfs2_holder_uninit(&tr->tr_t_gh);
75         kfree(tr);
76
77         return error;
78 }
79
80 /**
81  * gfs2_log_release - Release a given number of log blocks
82  * @sdp: The GFS2 superblock
83  * @blks: The number of blocks
84  *
85  */
86
87 static void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
88 {
89
90         atomic_add(blks, &sdp->sd_log_blks_free);
91         trace_gfs2_log_blocks(sdp, blks);
92         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
93                                   sdp->sd_jdesc->jd_blocks);
94         up_read(&sdp->sd_log_flush_lock);
95 }
96
97 static void gfs2_print_trans(const struct gfs2_trans *tr)
98 {
99         print_symbol(KERN_WARNING "GFS2: Transaction created at: %s\n", tr->tr_ip);
100         printk(KERN_WARNING "GFS2: blocks=%u revokes=%u reserved=%u touched=%d\n",
101                tr->tr_blocks, tr->tr_revokes, tr->tr_reserved, tr->tr_touched);
102         printk(KERN_WARNING "GFS2: Buf %u/%u Databuf %u/%u Revoke %u/%u\n",
103                tr->tr_num_buf_new, tr->tr_num_buf_rm,
104                tr->tr_num_databuf_new, tr->tr_num_databuf_rm,
105                tr->tr_num_revoke, tr->tr_num_revoke_rm);
106 }
107
108 void gfs2_trans_end(struct gfs2_sbd *sdp)
109 {
110         struct gfs2_trans *tr = current->journal_info;
111         s64 nbuf;
112         BUG_ON(!tr);
113         current->journal_info = NULL;
114
115         if (!tr->tr_touched) {
116                 gfs2_log_release(sdp, tr->tr_reserved);
117                 if (tr->tr_t_gh.gh_gl) {
118                         gfs2_glock_dq(&tr->tr_t_gh);
119                         gfs2_holder_uninit(&tr->tr_t_gh);
120                         kfree(tr);
121                 }
122                 sb_end_intwrite(sdp->sd_vfs);
123                 return;
124         }
125
126         nbuf = tr->tr_num_buf_new + tr->tr_num_databuf_new;
127         nbuf -= tr->tr_num_buf_rm;
128         nbuf -= tr->tr_num_databuf_rm;
129
130         if (gfs2_assert_withdraw(sdp, (nbuf <= tr->tr_blocks) &&
131                                        (tr->tr_num_revoke <= tr->tr_revokes)))
132                 gfs2_print_trans(tr);
133
134         gfs2_log_commit(sdp, tr);
135         if (tr->tr_t_gh.gh_gl) {
136                 gfs2_glock_dq(&tr->tr_t_gh);
137                 gfs2_holder_uninit(&tr->tr_t_gh);
138                 kfree(tr);
139         }
140
141         if (sdp->sd_vfs->s_flags & MS_SYNCHRONOUS)
142                 gfs2_log_flush(sdp, NULL);
143         sb_end_intwrite(sdp->sd_vfs);
144 }
145
146 static struct gfs2_bufdata *gfs2_alloc_bufdata(struct gfs2_glock *gl,
147                                                struct buffer_head *bh,
148                                                const struct gfs2_log_operations *lops)
149 {
150         struct gfs2_bufdata *bd;
151
152         bd = kmem_cache_zalloc(gfs2_bufdata_cachep, GFP_NOFS | __GFP_NOFAIL);
153         bd->bd_bh = bh;
154         bd->bd_gl = gl;
155         bd->bd_ops = lops;
156         INIT_LIST_HEAD(&bd->bd_list);
157         bh->b_private = bd;
158         return bd;
159 }
160
161 /**
162  * gfs2_trans_add_data - Add a databuf to the transaction.
163  * @gl: The inode glock associated with the buffer
164  * @bh: The buffer to add
165  *
166  * This is used in two distinct cases:
167  * i) In ordered write mode
168  *    We put the data buffer on a list so that we can ensure that its
169  *    synced to disk at the right time
170  * ii) In journaled data mode
171  *    We need to journal the data block in the same way as metadata in
172  *    the functions above. The difference is that here we have a tag
173  *    which is two __be64's being the block number (as per meta data)
174  *    and a flag which says whether the data block needs escaping or
175  *    not. This means we need a new log entry for each 251 or so data
176  *    blocks, which isn't an enormous overhead but twice as much as
177  *    for normal metadata blocks.
178  */
179 void gfs2_trans_add_data(struct gfs2_glock *gl, struct buffer_head *bh)
180 {
181         struct gfs2_trans *tr = current->journal_info;
182         struct gfs2_sbd *sdp = gl->gl_sbd;
183         struct address_space *mapping = bh->b_page->mapping;
184         struct gfs2_inode *ip = GFS2_I(mapping->host);
185         struct gfs2_bufdata *bd;
186
187         if (!gfs2_is_jdata(ip)) {
188                 gfs2_ordered_add_inode(ip);
189                 return;
190         }
191
192         lock_buffer(bh);
193         gfs2_log_lock(sdp);
194         bd = bh->b_private;
195         if (bd == NULL) {
196                 gfs2_log_unlock(sdp);
197                 unlock_buffer(bh);
198                 if (bh->b_private == NULL)
199                         bd = gfs2_alloc_bufdata(gl, bh, &gfs2_databuf_lops);
200                 lock_buffer(bh);
201                 gfs2_log_lock(sdp);
202         }
203         gfs2_assert(sdp, bd->bd_gl == gl);
204         tr->tr_touched = 1;
205         if (list_empty(&bd->bd_list)) {
206                 set_bit(GLF_LFLUSH, &bd->bd_gl->gl_flags);
207                 set_bit(GLF_DIRTY, &bd->bd_gl->gl_flags);
208                 gfs2_pin(sdp, bd->bd_bh);
209                 tr->tr_num_databuf_new++;
210                 sdp->sd_log_num_databuf++;
211                 list_add_tail(&bd->bd_list, &sdp->sd_log_le_databuf);
212         }
213         gfs2_log_unlock(sdp);
214         unlock_buffer(bh);
215 }
216
217 static void meta_lo_add(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
218 {
219         struct gfs2_meta_header *mh;
220         struct gfs2_trans *tr;
221
222         tr = current->journal_info;
223         tr->tr_touched = 1;
224         if (!list_empty(&bd->bd_list))
225                 return;
226         set_bit(GLF_LFLUSH, &bd->bd_gl->gl_flags);
227         set_bit(GLF_DIRTY, &bd->bd_gl->gl_flags);
228         mh = (struct gfs2_meta_header *)bd->bd_bh->b_data;
229         if (unlikely(mh->mh_magic != cpu_to_be32(GFS2_MAGIC))) {
230                 printk(KERN_ERR
231                        "Attempting to add uninitialised block to journal (inplace block=%lld)\n",
232                        (unsigned long long)bd->bd_bh->b_blocknr);
233                 BUG();
234         }
235         gfs2_pin(sdp, bd->bd_bh);
236         mh->__pad0 = cpu_to_be64(0);
237         mh->mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
238         sdp->sd_log_num_buf++;
239         list_add(&bd->bd_list, &sdp->sd_log_le_buf);
240         tr->tr_num_buf_new++;
241 }
242
243 void gfs2_trans_add_meta(struct gfs2_glock *gl, struct buffer_head *bh)
244 {
245
246         struct gfs2_sbd *sdp = gl->gl_sbd;
247         struct gfs2_bufdata *bd;
248
249         lock_buffer(bh);
250         gfs2_log_lock(sdp);
251         bd = bh->b_private;
252         if (bd == NULL) {
253                 gfs2_log_unlock(sdp);
254                 unlock_buffer(bh);
255                 lock_page(bh->b_page);
256                 if (bh->b_private == NULL)
257                         bd = gfs2_alloc_bufdata(gl, bh, &gfs2_buf_lops);
258                 unlock_page(bh->b_page);
259                 lock_buffer(bh);
260                 gfs2_log_lock(sdp);
261         }
262         gfs2_assert(sdp, bd->bd_gl == gl);
263         meta_lo_add(sdp, bd);
264         gfs2_log_unlock(sdp);
265         unlock_buffer(bh);
266 }
267
268 void gfs2_trans_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
269 {
270         struct gfs2_glock *gl = bd->bd_gl;
271         struct gfs2_trans *tr = current->journal_info;
272
273         BUG_ON(!list_empty(&bd->bd_list));
274         BUG_ON(!list_empty(&bd->bd_ail_st_list));
275         BUG_ON(!list_empty(&bd->bd_ail_gl_list));
276         bd->bd_ops = &gfs2_revoke_lops;
277         tr->tr_touched = 1;
278         tr->tr_num_revoke++;
279         sdp->sd_log_num_revoke++;
280         atomic_inc(&gl->gl_revokes);
281         set_bit(GLF_LFLUSH, &gl->gl_flags);
282         list_add(&bd->bd_list, &sdp->sd_log_le_revoke);
283 }
284
285 void gfs2_trans_add_unrevoke(struct gfs2_sbd *sdp, u64 blkno, unsigned int len)
286 {
287         struct gfs2_bufdata *bd, *tmp;
288         struct gfs2_trans *tr = current->journal_info;
289         unsigned int n = len;
290
291         gfs2_log_lock(sdp);
292         list_for_each_entry_safe(bd, tmp, &sdp->sd_log_le_revoke, bd_list) {
293                 if ((bd->bd_blkno >= blkno) && (bd->bd_blkno < (blkno + len))) {
294                         list_del_init(&bd->bd_list);
295                         gfs2_assert_withdraw(sdp, sdp->sd_log_num_revoke);
296                         sdp->sd_log_num_revoke--;
297                         kmem_cache_free(gfs2_bufdata_cachep, bd);
298                         tr->tr_num_revoke_rm++;
299                         if (--n == 0)
300                                 break;
301                 }
302         }
303         gfs2_log_unlock(sdp);
304 }
305