1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2013 Jie Liu.
4 * All Rights Reserved.
5 */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_da_format.h"
14 #include "xfs_trans_space.h"
15 #include "xfs_da_btree.h"
16 #include "xfs_bmap_btree.h"
17 #include "xfs_trace.h"
18
19 /*
20 * Shortly after enabling the large extents count feature in 2023, longstanding
21 * bugs were found in the code that computes the minimum log size. Luckily,
22 * the bugs resulted in over-estimates of that size, so there's no impact to
23 * existing users. However, we don't want to reduce the minimum log size
24 * because that can create the situation where a newer mkfs writes a new
25 * filesystem that an older kernel won't mount.
26 *
27 * Several years prior, we also discovered that the transaction reservations
28 * for rmap and reflink operations were unnecessarily large. That was fixed,
29 * but the minimum log size computation was left alone to avoid the
30 * compatibility problems noted above. Fix that too.
31 *
32 * Therefore, we only may correct the computation starting with filesystem
33 * features that didn't exist in 2023. In other words, only turn this on if
34 * the filesystem has parent pointers.
35 *
36 * This function can be called before the XFS_HAS_* flags have been set up,
37 * (e.g. mkfs) so we must check the ondisk superblock.
38 */
39 static inline bool
xfs_want_minlogsize_fixes(struct xfs_sb * sb)40 xfs_want_minlogsize_fixes(
41 struct xfs_sb *sb)
42 {
43 return xfs_sb_is_v5(sb) &&
44 xfs_sb_has_incompat_feature(sb, XFS_SB_FEAT_INCOMPAT_PARENT);
45 }
46
47 /*
48 * Calculate the maximum length in bytes that would be required for a local
49 * attribute value as large attributes out of line are not logged.
50 */
51 STATIC int
xfs_log_calc_max_attrsetm_res(struct xfs_mount * mp)52 xfs_log_calc_max_attrsetm_res(
53 struct xfs_mount *mp)
54 {
55 int size;
56 int nblks;
57
58 size = xfs_attr_leaf_entsize_local_max(mp->m_attr_geo->blksize) -
59 MAXNAMELEN - 1;
60 nblks = XFS_DAENTER_SPACE_RES(mp, XFS_ATTR_FORK);
61 nblks += XFS_B_TO_FSB(mp, size);
62
63 /*
64 * If the feature set is new enough, correct a unit conversion error in
65 * the xattr transaction reservation code that resulted in oversized
66 * minimum log size computations.
67 */
68 if (xfs_want_minlogsize_fixes(&mp->m_sb))
69 size = XFS_B_TO_FSB(mp, size);
70
71 nblks += XFS_NEXTENTADD_SPACE_RES(mp, size, XFS_ATTR_FORK);
72
73 return M_RES(mp)->tr_attrsetm.tr_logres +
74 M_RES(mp)->tr_attrsetrt.tr_logres * nblks;
75 }
76
77 /*
78 * Compute an alternate set of log reservation sizes for use exclusively with
79 * minimum log size calculations.
80 */
81 static void
xfs_log_calc_trans_resv_for_minlogblocks(struct xfs_mount * mp,struct xfs_trans_resv * resv)82 xfs_log_calc_trans_resv_for_minlogblocks(
83 struct xfs_mount *mp,
84 struct xfs_trans_resv *resv)
85 {
86 unsigned int rmap_maxlevels = mp->m_rmap_maxlevels;
87
88 /*
89 * If the feature set is new enough, drop the oversized minimum log
90 * size computation introduced by the original reflink code.
91 */
92 if (xfs_want_minlogsize_fixes(&mp->m_sb)) {
93 xfs_trans_resv_calc(mp, resv);
94 resv->tr_atomic_ioend = M_RES(mp)->tr_atomic_ioend;
95 return;
96 }
97
98 /*
99 * In the early days of rmap+reflink, we always set the rmap maxlevels
100 * to 9 even if the AG was small enough that it would never grow to
101 * that height. Transaction reservation sizes influence the minimum
102 * log size calculation, which influences the size of the log that mkfs
103 * creates. Use the old value here to ensure that newly formatted
104 * small filesystems will mount on older kernels.
105 */
106 if (xfs_has_rmapbt(mp) && xfs_has_reflink(mp))
107 mp->m_rmap_maxlevels = XFS_OLD_REFLINK_RMAP_MAXLEVELS;
108
109 xfs_trans_resv_calc(mp, resv);
110
111 /* Copy the dynamic transaction reservation types from the running fs */
112 resv->tr_atomic_ioend = M_RES(mp)->tr_atomic_ioend;
113
114 if (xfs_has_reflink(mp)) {
115 /*
116 * In the early days of reflink, typical log operation counts
117 * were greatly overestimated.
118 */
119 resv->tr_write.tr_logcount = XFS_WRITE_LOG_COUNT_REFLINK;
120 resv->tr_itruncate.tr_logcount =
121 XFS_ITRUNCATE_LOG_COUNT_REFLINK;
122 resv->tr_qm_dqalloc.tr_logcount = XFS_WRITE_LOG_COUNT_REFLINK;
123 } else if (xfs_has_rmapbt(mp)) {
124 /*
125 * In the early days of non-reflink rmap, the impact of rmapbt
126 * updates on log counts were not taken into account at all.
127 */
128 resv->tr_write.tr_logcount = XFS_WRITE_LOG_COUNT;
129 resv->tr_itruncate.tr_logcount = XFS_ITRUNCATE_LOG_COUNT;
130 resv->tr_qm_dqalloc.tr_logcount = XFS_WRITE_LOG_COUNT;
131 }
132
133 /*
134 * In the early days of reflink, we did not use deferred refcount
135 * update log items, so log reservations must be recomputed using the
136 * old calculations.
137 */
138 resv->tr_write.tr_logres =
139 xfs_calc_write_reservation_minlogsize(mp);
140 resv->tr_itruncate.tr_logres =
141 xfs_calc_itruncate_reservation_minlogsize(mp);
142 resv->tr_qm_dqalloc.tr_logres =
143 xfs_calc_qm_dqalloc_reservation_minlogsize(mp);
144
145 /* Put everything back the way it was. This goes at the end. */
146 mp->m_rmap_maxlevels = rmap_maxlevels;
147 }
148
149 /*
150 * Iterate over the log space reservation table to figure out and return
151 * the maximum one in terms of the pre-calculated values which were done
152 * at mount time.
153 */
154 void
xfs_log_get_max_trans_res(struct xfs_mount * mp,struct xfs_trans_res * max_resp)155 xfs_log_get_max_trans_res(
156 struct xfs_mount *mp,
157 struct xfs_trans_res *max_resp)
158 {
159 struct xfs_trans_resv resv = {};
160 struct xfs_trans_res *resp;
161 struct xfs_trans_res *end_resp;
162 unsigned int i;
163 int log_space = 0;
164 int attr_space;
165
166 attr_space = xfs_log_calc_max_attrsetm_res(mp);
167
168 xfs_log_calc_trans_resv_for_minlogblocks(mp, &resv);
169
170 resp = (struct xfs_trans_res *)&resv;
171 end_resp = (struct xfs_trans_res *)(&resv + 1);
172 for (i = 0; resp < end_resp; i++, resp++) {
173 int tmp = resp->tr_logcount > 1 ?
174 resp->tr_logres * resp->tr_logcount :
175 resp->tr_logres;
176
177 trace_xfs_trans_resv_calc_minlogsize(mp, i, resp);
178 if (log_space < tmp) {
179 log_space = tmp;
180 *max_resp = *resp; /* struct copy */
181 }
182 }
183
184 if (attr_space > log_space) {
185 *max_resp = resv.tr_attrsetm; /* struct copy */
186 max_resp->tr_logres = attr_space;
187 }
188 trace_xfs_log_get_max_trans_res(mp, max_resp);
189 }
190
191 /*
192 * Calculate the minimum valid log size for the given superblock configuration.
193 * Used to calculate the minimum log size at mkfs time, and to determine if
194 * the log is large enough or not at mount time. Returns the minimum size in
195 * filesystem block size units.
196 */
197 int
xfs_log_calc_minimum_size(struct xfs_mount * mp)198 xfs_log_calc_minimum_size(
199 struct xfs_mount *mp)
200 {
201 struct xfs_trans_res tres = {0};
202 int max_logres;
203 int min_logblks = 0;
204 int lsunit = 0;
205
206 xfs_log_get_max_trans_res(mp, &tres);
207
208 max_logres = xfs_log_calc_unit_res(mp, tres.tr_logres);
209 if (tres.tr_logcount > 1)
210 max_logres *= tres.tr_logcount;
211
212 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
213 lsunit = BTOBB(mp->m_sb.sb_logsunit);
214
215 /*
216 * Two factors should be taken into account for calculating the minimum
217 * log space.
218 * 1) The fundamental limitation is that no single transaction can be
219 * larger than half size of the log.
220 *
221 * From mkfs.xfs, this is considered by the XFS_MIN_LOG_FACTOR
222 * define, which is set to 3. That means we can definitely fit
223 * maximally sized 2 transactions in the log. We'll use this same
224 * value here.
225 *
226 * 2) If the lsunit option is specified, a transaction requires 2 LSU
227 * for the reservation because there are two log writes that can
228 * require padding - the transaction data and the commit record which
229 * are written separately and both can require padding to the LSU.
230 * Consider that we can have an active CIL reservation holding 2*LSU,
231 * but the CIL is not over a push threshold, in this case, if we
232 * don't have enough log space for at one new transaction, which
233 * includes another 2*LSU in the reservation, we will run into dead
234 * loop situation in log space grant procedure. i.e.
235 * xlog_grant_head_wait().
236 *
237 * Hence the log size needs to be able to contain two maximally sized
238 * and padded transactions, which is (2 * (2 * LSU + maxlres)).
239 *
240 * Also, the log size should be a multiple of the log stripe unit, round
241 * it up to lsunit boundary if lsunit is specified.
242 */
243 if (lsunit) {
244 min_logblks = roundup_64(BTOBB(max_logres), lsunit) +
245 2 * lsunit;
246 } else
247 min_logblks = BTOBB(max_logres) + 2 * BBSIZE;
248 min_logblks *= XFS_MIN_LOG_FACTOR;
249
250 return XFS_BB_TO_FSB(mp, min_logblks);
251 }
252