1 Index: linux-3.12.39-47.1/fs/ext4/ext4.h
2 ===================================================================
3 --- linux-3.12.39-47.1.orig/fs/ext4/ext4.h
4 +++ linux-3.12.39-47.1/fs/ext4/ext4.h
5 @@ -1251,11 +1251,14 @@ struct ext4_sb_info {
8 unsigned long s_stripe;
9 - unsigned int s_mb_stream_request;
10 + unsigned long s_mb_small_req;
11 + unsigned long s_mb_large_req;
12 unsigned int s_mb_max_to_scan;
13 unsigned int s_mb_min_to_scan;
14 unsigned int s_mb_stats;
15 unsigned int s_mb_order2_reqs;
16 + unsigned long *s_mb_prealloc_table;
17 + unsigned long s_mb_prealloc_table_size;
18 unsigned int s_mb_group_prealloc;
19 unsigned int s_max_dir_size_kb;
20 /* where last allocation was done - for stream allocation */
21 Index: linux-3.12.39-47.1/fs/ext4/mballoc.c
22 ===================================================================
23 --- linux-3.12.39-47.1.orig/fs/ext4/mballoc.c
24 +++ linux-3.12.39-47.1/fs/ext4/mballoc.c
25 @@ -1847,6 +1847,26 @@ int ext4_mb_find_by_goal(struct ext4_all
29 +static int ext4_mb_prealloc_table_add(struct ext4_sb_info *sbi, int value)
33 + if (value > (sbi->s_clusters_per_group - 1 - 1 - sbi->s_itb_per_group))
36 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
37 + if (sbi->s_mb_prealloc_table[i] == 0) {
38 + sbi->s_mb_prealloc_table[i] = value;
42 + /* they should add values in order */
43 + if (value <= sbi->s_mb_prealloc_table[i])
50 * The routine scans buddy structures (not bitmap!) from given order
51 * to max order and tries to find big enough chunk to satisfy the req
52 @@ -2285,6 +2304,93 @@ static const struct seq_operations ext4_
53 .show = ext4_mb_seq_groups_show,
56 +#define EXT4_MB_PREALLOC_TABLE "prealloc_table"
58 +static ssize_t ext4_mb_prealloc_table_proc_write(struct file *file,
59 + const char __user *buf,
60 + size_t cnt, loff_t *pos)
62 + struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
63 + unsigned long value;
64 + unsigned long prev = 0;
68 + unsigned long *new_table;
72 + if (cnt >= sizeof(str))
74 + if (copy_from_user(str, buf, cnt))
82 + while ((cur < end) && (*cur == ' '))
84 + rc = kstrtol(cur, 0, &value);
95 + new_table = kmalloc(num * sizeof(*new_table), GFP_KERNEL);
96 + if (new_table == NULL)
98 + kfree(sbi->s_mb_prealloc_table);
99 + memset(new_table, 0, num * sizeof(*new_table));
100 + sbi->s_mb_prealloc_table = new_table;
101 + sbi->s_mb_prealloc_table_size = num;
104 + while (cur < end && i < num) {
105 + while (cur < end && *cur == ' ')
107 + value = simple_strtol(cur, &cur, 0);
108 + if (ext4_mb_prealloc_table_add(sbi, value) == 0)
112 + sbi->s_mb_prealloc_table_size = i;
117 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
119 + struct ext4_sb_info *sbi = EXT4_SB(m->private);
122 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++)
123 + seq_printf(m, "%ld ", sbi->s_mb_prealloc_table[i]);
124 + seq_printf(m, "\n");
129 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
131 + return single_open(file, mb_prealloc_table_seq_show, PDE_DATA(inode));
134 +static const struct file_operations ext4_mb_prealloc_seq_fops = {
135 + .owner = THIS_MODULE,
136 + .open = mb_prealloc_table_seq_open,
138 + .llseek = seq_lseek,
139 + .release = single_release,
140 + .write = ext4_mb_prealloc_table_proc_write,
143 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
145 struct super_block *sb = PDE_DATA(inode);
146 @@ -2550,7 +2657,7 @@ static int ext4_groupinfo_create_slab(si
147 int ext4_mb_init(struct super_block *sb)
149 struct ext4_sb_info *sbi = EXT4_SB(sb);
151 + unsigned i, j, k, l;
152 unsigned offset, offset_incr;
155 @@ -2595,7 +2702,6 @@ int ext4_mb_init(struct super_block *sb)
156 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
157 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
158 sbi->s_mb_stats = MB_DEFAULT_STATS;
159 - sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
160 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
162 * The default group preallocation is 512, which for 4k block
163 @@ -2619,9 +2725,47 @@ int ext4_mb_init(struct super_block *sb)
164 * RAID stripe size so that preallocations don't fragment
167 - if (sbi->s_stripe > 1) {
168 - sbi->s_mb_group_prealloc = roundup(
169 - sbi->s_mb_group_prealloc, sbi->s_stripe);
171 + if (sbi->s_stripe == 0) {
172 + sbi->s_mb_prealloc_table_size = 10;
173 + i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
174 + sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
175 + if (sbi->s_mb_prealloc_table == NULL) {
179 + memset(sbi->s_mb_prealloc_table, 0, i);
181 + for (k = 0, l = 4; k <= 9; ++k, l *= 2) {
182 + if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
183 + sbi->s_mb_prealloc_table_size = k;
188 + sbi->s_mb_small_req = 256;
189 + sbi->s_mb_large_req = 1024;
190 + sbi->s_mb_group_prealloc = 512;
192 + sbi->s_mb_prealloc_table_size = 3;
193 + i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
194 + sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
195 + if (sbi->s_mb_prealloc_table == NULL) {
199 + memset(sbi->s_mb_prealloc_table, 0, i);
201 + for (k = 0, l = sbi->s_stripe; k <= 2; ++k, l *= 2) {
202 + if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
203 + sbi->s_mb_prealloc_table_size = k;
208 + sbi->s_mb_small_req = sbi->s_stripe;
209 + sbi->s_mb_large_req = sbi->s_stripe * 8;
210 + sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
213 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
214 @@ -2627,9 +2769,13 @@ int ext4_mb_init(struct super_block *sb)
216 goto out_free_locality_groups;
220 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
221 &ext4_mb_seq_groups_fops, sb);
222 + proc_create_data(EXT4_MB_PREALLOC_TABLE, S_IFREG | S_IRUGO |
223 + S_IWUSR, sbi->s_proc,
224 + &ext4_mb_prealloc_seq_fops, sb);
229 @@ -2639,6 +2785,7 @@ out_free_locality_groups:
230 out_free_groupinfo_slab:
231 ext4_groupinfo_destroy_slabs();
233 + kfree(sbi->s_mb_prealloc_table);
234 kfree(sbi->s_mb_offsets);
235 sbi->s_mb_offsets = NULL;
236 kfree(sbi->s_mb_maxs);
237 @@ -2673,8 +2820,10 @@ int ext4_mb_release(struct super_block *
238 struct ext4_sb_info *sbi = EXT4_SB(sb);
239 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
243 remove_proc_entry("mb_groups", sbi->s_proc);
244 + remove_proc_entry(EXT4_MB_PREALLOC_TABLE, sbi->s_proc);
247 if (sbi->s_group_info) {
248 for (i = 0; i < ngroups; i++) {
249 @@ -2985,9 +3134,9 @@ ext4_mb_normalize_request(struct ext4_al
250 struct ext4_allocation_request *ar)
252 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
254 + int bsbits, i, wind;
256 - loff_t size, start_off;
258 loff_t orig_size __maybe_unused;
260 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
261 @@ -3020,50 +3169,34 @@ ext4_mb_normalize_request(struct ext4_al
262 size = size << bsbits;
263 if (size < i_size_read(ac->ac_inode))
264 size = i_size_read(ac->ac_inode);
266 + size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
268 - /* max size of free chunks */
272 -#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
273 - (req <= (size) || max <= (chunk_size))
274 + /* let's choose preallocation window depending on file size */
275 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
276 + if (size <= sbi->s_mb_prealloc_table[i]) {
277 + wind = sbi->s_mb_prealloc_table[i];
283 - /* first, try to predict filesize */
284 - /* XXX: should this table be tunable? */
286 - if (size <= 16 * 1024) {
288 - } else if (size <= 32 * 1024) {
290 - } else if (size <= 64 * 1024) {
292 - } else if (size <= 128 * 1024) {
294 - } else if (size <= 256 * 1024) {
296 - } else if (size <= 512 * 1024) {
298 - } else if (size <= 1024 * 1024) {
299 - size = 1024 * 1024;
300 - } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
301 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
302 - (21 - bsbits)) << 21;
303 - size = 2 * 1024 * 1024;
304 - } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
305 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
306 - (22 - bsbits)) << 22;
307 - size = 4 * 1024 * 1024;
308 - } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
309 - (8<<20)>>bsbits, max, 8 * 1024)) {
310 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
311 - (23 - bsbits)) << 23;
312 - size = 8 * 1024 * 1024;
314 - start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
315 - size = ac->ac_o_ex.fe_len << bsbits;
317 + __u64 tstart, tend;
318 + /* file is quite large, we now preallocate with
319 + * the biggest configured window with regart to
320 + * logical offset */
321 + wind = sbi->s_mb_prealloc_table[i - 1];
322 + tstart = ac->ac_o_ex.fe_logical;
323 + do_div(tstart, wind);
324 + start = tstart * wind;
325 + tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
326 + do_div(tend, wind);
327 + tend = tend * wind + wind;
328 + size = tend - start;
330 - size = size >> bsbits;
331 - start = start_off >> bsbits;
334 /* don't cover already allocated blocks in selected range */
335 if (ar->pleft && start <= ar->lleft) {
336 @@ -3139,7 +3272,6 @@ ext4_mb_normalize_request(struct ext4_al
338 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
339 start > ac->ac_o_ex.fe_logical);
340 - BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
342 /* now prepare goal request */
344 @@ -4105,11 +4237,19 @@ static void ext4_mb_group_or_file(struct
346 /* don't use group allocation for large files */
347 size = max(size, isize);
348 - if (size > sbi->s_mb_stream_request) {
349 + if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
350 + (size >= sbi->s_mb_large_req)) {
351 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
356 + * request is so large that we don't care about
357 + * streaming - it overweights any possible seek
359 + if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
362 BUG_ON(ac->ac_lg != NULL);
364 * locality group prealloc space are per cpu. The reason for having
365 Index: linux-3.12.39-47.1/fs/ext4/super.c
366 ===================================================================
367 --- linux-3.12.39-47.1.orig/fs/ext4/super.c
368 +++ linux-3.12.39-47.1/fs/ext4/super.c
369 @@ -2592,7 +2592,8 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats
370 EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
371 EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
372 EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
373 -EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
374 +EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
375 +EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
376 EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
377 EXT4_DEPRECATED_ATTR(max_writeback_mb_bump, 128);
378 EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
379 @@ -2609,7 +2610,8 @@ static struct attribute *ext4_attrs[] =
380 ATTR_LIST(mb_max_to_scan),
381 ATTR_LIST(mb_min_to_scan),
382 ATTR_LIST(mb_order2_req),
383 - ATTR_LIST(mb_stream_req),
384 + ATTR_LIST(mb_small_req),
385 + ATTR_LIST(mb_large_req),
386 ATTR_LIST(mb_group_prealloc),
387 ATTR_LIST(max_writeback_mb_bump),
388 ATTR_LIST(extent_max_zeroout_kb),
389 Index: linux-3.12.39-47.1/fs/ext4/inode.c
390 ===================================================================
391 --- linux-3.12.39-47.1.orig/fs/ext4/inode.c
392 +++ linux-3.12.39-47.1/fs/ext4/inode.c
393 @@ -2457,6 +2457,9 @@ static int ext4_writepages(struct addres
394 ext4_journal_stop(handle);
397 + if (wbc->nr_to_write < sbi->s_mb_small_req)
398 + wbc->nr_to_write = sbi->s_mb_small_req;
400 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)