1 commit d8d8fd9192a54c7b8caef8cca9b7a1eb5e5e3298
2 Author: Alex Zhuravlev <alex.zhuravlev@sun.com>
3 AuthorDate: Thu Oct 23 10:02:19 2008 +0000
5 Subject: ext4: support for tunable preallocation window
6 Add support for tunable preallocation window and new tunables
7 for large/small requests.
10 Signed-off-by: Alex Zhuravlev <alex.zhuravlev@sun.com>
11 Reviewed-by: Kalpak Shah <kalpak@clusterfs.com>
12 Reviewed-by: Andreas Dilger <andreas.dilger@sun.com>
16 fs/ext4/mballoc.c | 219 +++++++++++++++++++++++++++++++++++-----------
17 fs/ext4/sysfs.c | 8 +-
18 4 files changed, 182 insertions(+), 55 deletions(-)
20 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
21 index b03dbd7..78d3176 100644
24 @@ -1291,6 +1291,8 @@ extern void mb_set_bits(void *bm, int cur, int len);
25 #define EXT4_DFL_MAX_MNT_COUNT 20 /* Allow 20 mounts */
26 #define EXT4_DFL_CHECKINTERVAL 0 /* Don't use interval check */
28 +#define EXT4_MAX_PREALLOC_TABLE 64
31 * Behaviour when detecting errors
33 @@ -1597,11 +1599,13 @@ struct ext4_sb_info {
35 unsigned long s_stripe;
36 unsigned int s_mb_max_linear_groups;
37 - unsigned int s_mb_stream_request;
38 + unsigned long s_mb_small_req;
39 + unsigned long s_mb_large_req;
40 unsigned int s_mb_max_to_scan;
41 unsigned int s_mb_min_to_scan;
42 unsigned int s_mb_stats;
43 unsigned int s_mb_order2_reqs;
44 + unsigned long *s_mb_prealloc_table;
45 unsigned int s_mb_group_prealloc;
46 unsigned int s_mb_max_inode_prealloc;
47 unsigned int s_max_dir_size_kb;
48 @@ -2940,6 +2944,7 @@ int ext4_fc_record_regions(struct super_block *sb, int ino,
52 +extern const struct proc_ops ext4_seq_prealloc_table_fops;
53 extern const struct seq_operations ext4_mb_seq_groups_ops;
54 extern const struct seq_operations ext4_mb_seq_structs_summary_ops;
55 extern long ext4_mb_stats;
56 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
57 index 6723622..2ed00ae 100644
60 @@ -2720,6 +2720,9 @@ static int ext4_writepages(struct address_space *mapping,
61 PAGE_SIZE >> inode->i_blkbits);
64 + if (wbc->nr_to_write < sbi->s_mb_small_req)
65 + wbc->nr_to_write = sbi->s_mb_small_req;
67 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
70 diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
71 index 8164480..94a299b 100644
72 --- a/fs/ext4/mballoc.c
73 +++ b/fs/ext4/mballoc.c
74 @@ -3070,6 +3070,99 @@ const struct seq_operations ext4_mb_seq_structs_summary_ops = {
75 .show = ext4_mb_seq_structs_summary_show,
78 +static int ext4_mb_check_and_update_prealloc(struct ext4_sb_info *sbi,
79 + char *str, size_t cnt,
82 + unsigned long value;
83 + unsigned long prev = 0;
92 + while ((cur < end) && (*cur == ' ')) cur++;
93 + value = simple_strtol(cur, &next, 0);
101 + if (value > (sbi->s_blocks_per_group - 1 - 1 - sbi->s_itb_per_group))
104 + /* they should add values in order */
109 + sbi->s_mb_prealloc_table[num] = value;
115 + if (num > EXT4_MAX_PREALLOC_TABLE - 1)
119 + sbi->s_mb_prealloc_table[num] = 0;
124 +static ssize_t ext4_mb_prealloc_table_proc_write(struct file *file,
125 + const char __user *buf,
126 + size_t cnt, loff_t *pos)
128 + struct ext4_sb_info *sbi = EXT4_SB(pde_data(file_inode(file)));
132 + if (cnt >= sizeof(str))
134 + if (copy_from_user(str, buf, cnt))
137 + rc = ext4_mb_check_and_update_prealloc(sbi, str, cnt, 0);
141 + rc = ext4_mb_check_and_update_prealloc(sbi, str, cnt, 1);
142 + return rc ? rc : cnt;
145 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
147 + struct ext4_sb_info *sbi = EXT4_SB(m->private);
150 + for (i = 0; i < EXT4_MAX_PREALLOC_TABLE &&
151 + sbi->s_mb_prealloc_table[i] != 0; i++)
152 + seq_printf(m, "%ld ", sbi->s_mb_prealloc_table[i]);
153 + seq_printf(m, "\n");
158 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
160 + return single_open(file, mb_prealloc_table_seq_show, pde_data(inode));
163 +const struct proc_ops ext4_seq_prealloc_table_fops = {
164 + .proc_open = mb_prealloc_table_seq_open,
165 + .proc_read = seq_read,
166 + .proc_lseek = seq_lseek,
167 + .proc_release = single_release,
168 + .proc_write = ext4_mb_prealloc_table_proc_write,
171 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
173 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
174 @@ -3382,7 +3475,7 @@ static void ext4_discard_work(struct work_struct *work)
175 int ext4_mb_init(struct super_block *sb)
177 struct ext4_sb_info *sbi = EXT4_SB(sb);
179 + unsigned i, j, k, l;
180 unsigned offset, offset_incr;
183 @@ -3454,7 +3547,6 @@ int ext4_mb_init(struct super_block *sb)
184 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
185 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
186 sbi->s_mb_stats = MB_DEFAULT_STATS;
187 - sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
188 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
189 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
191 @@ -3479,9 +3571,29 @@ int ext4_mb_init(struct super_block *sb)
192 * RAID stripe size so that preallocations don't fragment
195 - if (sbi->s_stripe > 1) {
196 - sbi->s_mb_group_prealloc = roundup(
197 - sbi->s_mb_group_prealloc, sbi->s_stripe);
199 + /* Allocate table once */
200 + sbi->s_mb_prealloc_table = kzalloc(
201 + EXT4_MAX_PREALLOC_TABLE * sizeof(unsigned long), GFP_NOFS);
202 + if (sbi->s_mb_prealloc_table == NULL) {
207 + if (sbi->s_stripe == 0) {
208 + for (k = 0, l = 4; k <= 9; ++k, l *= 2)
209 + sbi->s_mb_prealloc_table[k] = l;
211 + sbi->s_mb_small_req = 256;
212 + sbi->s_mb_large_req = 1024;
213 + sbi->s_mb_group_prealloc = 512;
215 + for (k = 0, l = sbi->s_stripe; k <= 2; ++k, l *= 2)
216 + sbi->s_mb_prealloc_table[k] = l;
218 + sbi->s_mb_small_req = sbi->s_stripe;
219 + sbi->s_mb_large_req = sbi->s_stripe * 8;
220 + sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
223 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
224 @@ -3515,6 +3627,7 @@ out_free_locality_groups:
226 kfree(sbi->s_mb_largest_free_orders);
227 kfree(sbi->s_mb_largest_free_orders_locks);
228 + kfree(sbi->s_mb_prealloc_table);
229 kfree(sbi->s_mb_offsets);
230 sbi->s_mb_offsets = NULL;
231 kfree(sbi->s_mb_maxs);
232 @@ -3783,7 +3896,6 @@ ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
235 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
236 - BUG_ON(ac->ac_b_ex.fe_len <= 0);
240 @@ -4026,13 +4138,14 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
241 struct ext4_allocation_request *ar)
243 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
245 + int bsbits, i, wind;
247 - loff_t size, start_off;
249 loff_t orig_size __maybe_unused;
251 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
252 struct ext4_prealloc_space *pa;
253 + unsigned long value, last_non_zero;
255 /* do normalize only data requests, metadata requests
256 do not need preallocation */
257 @@ -4061,51 +4174,46 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
258 size = size << bsbits;
259 if (size < i_size_read(ac->ac_inode))
260 size = i_size_read(ac->ac_inode);
262 + size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
265 + value = last_non_zero = 0;
267 + /* let's choose preallocation window depending on file size */
268 + for (i = 0; i < EXT4_MAX_PREALLOC_TABLE; i++) {
269 + value = sbi->s_mb_prealloc_table[i];
273 + last_non_zero = value;
275 - /* max size of free chunks */
278 -#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
279 - (req <= (size) || max <= (chunk_size))
281 - /* first, try to predict filesize */
282 - /* XXX: should this table be tunable? */
284 - if (size <= 16 * 1024) {
286 - } else if (size <= 32 * 1024) {
288 - } else if (size <= 64 * 1024) {
290 - } else if (size <= 128 * 1024) {
292 - } else if (size <= 256 * 1024) {
294 - } else if (size <= 512 * 1024) {
296 - } else if (size <= 1024 * 1024) {
297 - size = 1024 * 1024;
298 - } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
299 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
300 - (21 - bsbits)) << 21;
301 - size = 2 * 1024 * 1024;
302 - } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
303 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
304 - (22 - bsbits)) << 22;
305 - size = 4 * 1024 * 1024;
306 - } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
307 - (8<<20)>>bsbits, max, 8 * 1024)) {
308 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
309 - (23 - bsbits)) << 23;
310 - size = 8 * 1024 * 1024;
311 + if (size <= value) {
318 + if (last_non_zero != 0) {
319 + __u64 tstart, tend;
320 + /* file is quite large, we now preallocate with
321 + * the biggest configured window with regart to
322 + * logical offset */
323 + wind = last_non_zero;
324 + tstart = ac->ac_o_ex.fe_logical;
325 + do_div(tstart, wind);
326 + start = tstart * wind;
327 + tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
328 + do_div(tend, wind);
329 + tend = tend * wind + wind;
330 + size = tend - start;
333 - start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
334 - size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
335 - ac->ac_o_ex.fe_len) << bsbits;
338 - size = size >> bsbits;
339 - start = start_off >> bsbits;
345 * For tiny groups (smaller than 8MB) the chosen allocation
346 @@ -4196,7 +4304,6 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
347 (unsigned long) ac->ac_o_ex.fe_logical);
350 - BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
352 /* now prepare goal request */
354 @@ -5199,11 +5306,19 @@ static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
356 /* don't use group allocation for large files */
357 size = max(size, isize);
358 - if (size > sbi->s_mb_stream_request) {
359 + if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
360 + (size >= sbi->s_mb_large_req)) {
361 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
366 + * request is so large that we don't care about
367 + * streaming - it overweights any possible seek
369 + if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
372 BUG_ON(ac->ac_lg != NULL);
374 * locality group prealloc space are per cpu. The reason for having
375 diff --git a/fs/ext4/sysfs.c b/fs/ext4/sysfs.c
376 index d233c24..9d2254f 100644
377 --- a/fs/ext4/sysfs.c
378 +++ b/fs/ext4/sysfs.c
379 @@ -212,7 +212,8 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats);
380 EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
381 EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
382 EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
383 -EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
384 +EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
385 +EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
386 EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
387 EXT4_RW_ATTR_SBI_UI(mb_max_inode_prealloc, s_mb_max_inode_prealloc);
388 EXT4_RW_ATTR_SBI_UI(mb_max_linear_groups, s_mb_max_linear_groups);
389 @@ -262,7 +263,8 @@ static struct attribute *ext4_attrs[] = {
390 ATTR_LIST(mb_max_to_scan),
391 ATTR_LIST(mb_min_to_scan),
392 ATTR_LIST(mb_order2_req),
393 - ATTR_LIST(mb_stream_req),
394 + ATTR_LIST(mb_small_req),
395 + ATTR_LIST(mb_large_req),
396 ATTR_LIST(mb_group_prealloc),
397 ATTR_LIST(mb_max_inode_prealloc),
398 ATTR_LIST(mb_max_linear_groups),
399 @@ -543,6 +545,8 @@ int ext4_register_sysfs(struct super_block *sb)
400 ext4_fc_info_show, sb);
401 proc_create_seq_data("mb_groups", S_IRUGO, sbi->s_proc,
402 &ext4_mb_seq_groups_ops, sb);
403 + proc_create_data("prealloc_table", S_IRUGO, sbi->s_proc,
404 + &ext4_seq_prealloc_table_fops, sb);
405 proc_create_single_data("mb_stats", 0444, sbi->s_proc,
406 ext4_seq_mb_stats_show, sb);
407 proc_create_seq_data("mb_structs_summary", 0444, sbi->s_proc,