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 +2304,90 @@ 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))
81 + while ((cur < end) && (*cur == ' '))
83 + value = simple_strtol(cur, &cur, 0);
92 + new_table = kmalloc(num * sizeof(*new_table), GFP_KERNEL);
93 + if (new_table == NULL)
95 + kfree(sbi->s_mb_prealloc_table);
96 + memset(new_table, 0, num * sizeof(*new_table));
97 + sbi->s_mb_prealloc_table = new_table;
98 + sbi->s_mb_prealloc_table_size = num;
101 + while (cur < end && i < num) {
102 + while (cur < end && *cur == ' ')
104 + value = simple_strtol(cur, &cur, 0);
105 + if (ext4_mb_prealloc_table_add(sbi, value) == 0)
109 + sbi->s_mb_prealloc_table_size = i;
114 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
116 + struct ext4_sb_info *sbi = EXT4_SB(m->private);
119 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++)
120 + seq_printf(m, "%ld ", sbi->s_mb_prealloc_table[i]);
121 + seq_printf(m, "\n");
126 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
128 + return single_open(file, mb_prealloc_table_seq_show, PDE_DATA(inode));
131 +static const struct file_operations ext4_mb_prealloc_seq_fops = {
132 + .owner = THIS_MODULE,
133 + .open = mb_prealloc_table_seq_open,
135 + .llseek = seq_lseek,
136 + .release = single_release,
137 + .write = ext4_mb_prealloc_table_proc_write,
140 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
142 struct super_block *sb = PDE_DATA(inode);
143 @@ -2550,7 +2657,7 @@ static int ext4_groupinfo_create_slab(si
144 int ext4_mb_init(struct super_block *sb)
146 struct ext4_sb_info *sbi = EXT4_SB(sb);
148 + unsigned i, j, k, l;
149 unsigned offset, offset_incr;
152 @@ -2595,7 +2702,6 @@ int ext4_mb_init(struct super_block *sb)
153 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
154 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
155 sbi->s_mb_stats = MB_DEFAULT_STATS;
156 - sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
157 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
159 * The default group preallocation is 512, which for 4k block
160 @@ -2619,9 +2725,47 @@ int ext4_mb_init(struct super_block *sb)
161 * RAID stripe size so that preallocations don't fragment
164 - if (sbi->s_stripe > 1) {
165 - sbi->s_mb_group_prealloc = roundup(
166 - sbi->s_mb_group_prealloc, sbi->s_stripe);
168 + if (sbi->s_stripe == 0) {
169 + sbi->s_mb_prealloc_table_size = 10;
170 + i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
171 + sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
172 + if (sbi->s_mb_prealloc_table == NULL) {
176 + memset(sbi->s_mb_prealloc_table, 0, i);
178 + for (k = 0, l = 4; k <= 9; ++k, l *= 2) {
179 + if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
180 + sbi->s_mb_prealloc_table_size = k;
185 + sbi->s_mb_small_req = 256;
186 + sbi->s_mb_large_req = 1024;
187 + sbi->s_mb_group_prealloc = 512;
189 + sbi->s_mb_prealloc_table_size = 3;
190 + i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
191 + sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
192 + if (sbi->s_mb_prealloc_table == NULL) {
196 + memset(sbi->s_mb_prealloc_table, 0, i);
198 + for (k = 0, l = sbi->s_stripe; k <= 2; ++k, l *= 2) {
199 + if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
200 + sbi->s_mb_prealloc_table_size = k;
205 + sbi->s_mb_small_req = sbi->s_stripe;
206 + sbi->s_mb_large_req = sbi->s_stripe * 8;
207 + sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
210 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
211 @@ -2627,9 +2769,13 @@ int ext4_mb_init(struct super_block *sb)
213 goto out_free_locality_groups;
217 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
218 &ext4_mb_seq_groups_fops, sb);
219 + proc_create_data(EXT4_MB_PREALLOC_TABLE, S_IFREG | S_IRUGO |
220 + S_IWUSR, sbi->s_proc,
221 + &ext4_mb_prealloc_seq_fops, sb);
226 @@ -2639,6 +2785,7 @@ out_free_locality_groups:
227 out_free_groupinfo_slab:
228 ext4_groupinfo_destroy_slabs();
230 + kfree(sbi->s_mb_prealloc_table);
231 kfree(sbi->s_mb_offsets);
232 sbi->s_mb_offsets = NULL;
233 kfree(sbi->s_mb_maxs);
234 @@ -2673,8 +2820,10 @@ int ext4_mb_release(struct super_block *
235 struct ext4_sb_info *sbi = EXT4_SB(sb);
236 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
240 remove_proc_entry("mb_groups", sbi->s_proc);
241 + remove_proc_entry(EXT4_MB_PREALLOC_TABLE, sbi->s_proc);
244 if (sbi->s_group_info) {
245 for (i = 0; i < ngroups; i++) {
246 @@ -2985,9 +3134,9 @@ ext4_mb_normalize_request(struct ext4_al
247 struct ext4_allocation_request *ar)
249 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
251 + int bsbits, i, wind;
253 - loff_t size, start_off;
255 loff_t orig_size __maybe_unused;
257 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
258 @@ -3020,50 +3169,34 @@ ext4_mb_normalize_request(struct ext4_al
259 size = size << bsbits;
260 if (size < i_size_read(ac->ac_inode))
261 size = i_size_read(ac->ac_inode);
263 + size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
265 - /* max size of free chunks */
269 -#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
270 - (req <= (size) || max <= (chunk_size))
271 + /* let's choose preallocation window depending on file size */
272 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
273 + if (size <= sbi->s_mb_prealloc_table[i]) {
274 + wind = sbi->s_mb_prealloc_table[i];
280 - /* first, try to predict filesize */
281 - /* XXX: should this table be tunable? */
283 - if (size <= 16 * 1024) {
285 - } else if (size <= 32 * 1024) {
287 - } else if (size <= 64 * 1024) {
289 - } else if (size <= 128 * 1024) {
291 - } else if (size <= 256 * 1024) {
293 - } else if (size <= 512 * 1024) {
295 - } else if (size <= 1024 * 1024) {
296 - size = 1024 * 1024;
297 - } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
298 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
299 - (21 - bsbits)) << 21;
300 - size = 2 * 1024 * 1024;
301 - } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
302 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
303 - (22 - bsbits)) << 22;
304 - size = 4 * 1024 * 1024;
305 - } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
306 - (8<<20)>>bsbits, max, 8 * 1024)) {
307 - start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
308 - (23 - bsbits)) << 23;
309 - size = 8 * 1024 * 1024;
311 - start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
312 - size = ac->ac_o_ex.fe_len << bsbits;
314 + __u64 tstart, tend;
315 + /* file is quite large, we now preallocate with
316 + * the biggest configured window with regart to
317 + * logical offset */
318 + wind = sbi->s_mb_prealloc_table[i - 1];
319 + tstart = ac->ac_o_ex.fe_logical;
320 + do_div(tstart, wind);
321 + start = tstart * wind;
322 + tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
323 + do_div(tend, wind);
324 + tend = tend * wind + wind;
325 + size = tend - start;
327 - size = size >> bsbits;
328 - start = start_off >> bsbits;
331 /* don't cover already allocated blocks in selected range */
332 if (ar->pleft && start <= ar->lleft) {
333 @@ -3139,7 +3272,6 @@ ext4_mb_normalize_request(struct ext4_al
335 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
336 start > ac->ac_o_ex.fe_logical);
337 - BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
339 /* now prepare goal request */
341 @@ -4105,11 +4237,19 @@ static void ext4_mb_group_or_file(struct
343 /* don't use group allocation for large files */
344 size = max(size, isize);
345 - if (size > sbi->s_mb_stream_request) {
346 + if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
347 + (size >= sbi->s_mb_large_req)) {
348 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
353 + * request is so large that we don't care about
354 + * streaming - it overweights any possible seek
356 + if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
359 BUG_ON(ac->ac_lg != NULL);
361 * locality group prealloc space are per cpu. The reason for having
362 Index: linux-3.12.39-47.1/fs/ext4/super.c
363 ===================================================================
364 --- linux-3.12.39-47.1.orig/fs/ext4/super.c
365 +++ linux-3.12.39-47.1/fs/ext4/super.c
366 @@ -2592,7 +2592,8 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats
367 EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
368 EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
369 EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
370 -EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
371 +EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
372 +EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
373 EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
374 EXT4_DEPRECATED_ATTR(max_writeback_mb_bump, 128);
375 EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
376 @@ -2609,7 +2610,8 @@ static struct attribute *ext4_attrs[] =
377 ATTR_LIST(mb_max_to_scan),
378 ATTR_LIST(mb_min_to_scan),
379 ATTR_LIST(mb_order2_req),
380 - ATTR_LIST(mb_stream_req),
381 + ATTR_LIST(mb_small_req),
382 + ATTR_LIST(mb_large_req),
383 ATTR_LIST(mb_group_prealloc),
384 ATTR_LIST(max_writeback_mb_bump),
385 ATTR_LIST(extent_max_zeroout_kb),
386 Index: linux-3.12.39-47.1/fs/ext4/inode.c
387 ===================================================================
388 --- linux-3.12.39-47.1.orig/fs/ext4/inode.c
389 +++ linux-3.12.39-47.1/fs/ext4/inode.c
390 @@ -2457,6 +2457,9 @@ static int ext4_writepages(struct addres
391 ext4_journal_stop(handle);
394 + if (wbc->nr_to_write < sbi->s_mb_small_req)
395 + wbc->nr_to_write = sbi->s_mb_small_req;
397 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)