Whamcloud - gitweb
LU-17504 build: fix gcc-13 [-Werror=stringop-overread] error
[fs/lustre-release.git] / ldiskfs / kernel_patches / patches / linux-5.8 / ext4-prealloc.patch
1 commit d8d8fd9192a54c7b8caef8cca9b7a1eb5e5e3298
2 Author: Alex Zhuravlev <alex.zhuravlev@sun.com>
3 AuthorDate: Thu Oct 23 10:02:19 2008 +0000
4
5 Subject: ext4: support for tunable preallocation window
6 Add support for tunable preallocation window and new tunables
7 for large/small requests.
8
9 Bugzilla-ID: b=12800
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>
13 ---
14  fs/ext4/ext4.h    |    7 +
15  fs/ext4/inode.c   |    3 
16  fs/ext4/mballoc.c |  215 +++++++++++++++++++++++++++++++++++++++++-------------
17  fs/ext4/sysfs.c   |    8 +-
18  4 files changed, 180 insertions(+), 53 deletions(-)
19
20 --- a/fs/ext4/ext4.h
21 +++ b/fs/ext4/ext4.h
22 @@ -1241,6 +1241,8 @@ extern void ext4_set_bits(void *bm, int
23  /* Metadata checksum algorithm codes */
24  #define EXT4_CRC32C_CHKSUM             1
25  
26 +#define EXT4_MAX_PREALLOC_TABLE        64
27 +
28  /*
29   * Structure of the super block
30   */
31 @@ -1497,11 +1499,13 @@ struct ext4_sb_info {
32  
33         /* tunables */
34         unsigned long s_stripe;
35 -       unsigned int s_mb_stream_request;
36 +       unsigned long s_mb_small_req;
37 +       unsigned long s_mb_large_req;
38         unsigned int s_mb_max_to_scan;
39         unsigned int s_mb_min_to_scan;
40         unsigned int s_mb_stats;
41         unsigned int s_mb_order2_reqs;
42 +       unsigned long *s_mb_prealloc_table;
43         unsigned int s_mb_group_prealloc;
44         unsigned int s_max_dir_size_kb;
45         /* where last allocation was done - for stream allocation */
46 @@ -2645,6 +2649,7 @@ extern int ext4_init_inode_table(struct
47  extern void ext4_end_bitmap_read(struct buffer_head *bh, int uptodate);
48  
49  /* mballoc.c */
50 +extern const struct proc_ops ext4_seq_prealloc_table_fops;
51  extern const struct seq_operations ext4_mb_seq_groups_ops;
52  extern long ext4_mb_stats;
53  extern long ext4_mb_max_to_scan;
54 --- a/fs/ext4/inode.c
55 +++ b/fs/ext4/inode.c
56 @@ -2698,6 +2698,9 @@ static int ext4_writepages(struct addres
57                                                 PAGE_SIZE >> inode->i_blkbits);
58         }
59  
60 +       if (wbc->nr_to_write < sbi->s_mb_small_req)
61 +               wbc->nr_to_write = sbi->s_mb_small_req;
62 +
63         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
64                 range_whole = 1;
65  
66 --- a/fs/ext4/mballoc.c
67 +++ b/fs/ext4/mballoc.c
68 @@ -2455,6 +2455,99 @@ const struct seq_operations ext4_mb_seq_
69         .show   = ext4_mb_seq_groups_show,
70  };
71  
72 +static int ext4_mb_check_and_update_prealloc(struct ext4_sb_info *sbi,
73 +                                                char *str, size_t cnt,
74 +                                                int update)
75 +{
76 +       unsigned long value;
77 +       unsigned long prev = 0;
78 +       char *cur;
79 +       char *next;
80 +       char *end;
81 +       int num = 0;
82 +
83 +       cur = str;
84 +       end = str + cnt;
85 +       while (cur < end) {
86 +               while ((cur < end) && (*cur == ' ')) cur++;
87 +               value = simple_strtol(cur, &next, 0);
88 +               if (value == 0)
89 +                       break;
90 +               if (cur == next)
91 +                       return -EINVAL;
92 +
93 +               cur = next;
94 +
95 +               if (value > (sbi->s_blocks_per_group - 1 - 1 - sbi->s_itb_per_group))
96 +                       return -EINVAL;
97 +
98 +               /* they should add values in order */
99 +               if (value <= prev)
100 +                       return -EINVAL;
101 +
102 +               if (update)
103 +                       sbi->s_mb_prealloc_table[num] = value;
104 +
105 +               prev = value;
106 +               num++;
107 +       }
108 +
109 +       if (num > EXT4_MAX_PREALLOC_TABLE - 1)
110 +               return -EOVERFLOW;
111 +
112 +       if (update)
113 +               sbi->s_mb_prealloc_table[num] = 0;
114 +
115 +       return 0;
116 +}
117 +
118 +static ssize_t ext4_mb_prealloc_table_proc_write(struct file *file,
119 +                                            const char __user *buf,
120 +                                            size_t cnt, loff_t *pos)
121 +{
122 +       struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
123 +       char str[128];
124 +       int rc;
125 +
126 +       if (cnt >= sizeof(str))
127 +               return -EINVAL;
128 +       if (copy_from_user(str, buf, cnt))
129 +               return -EFAULT;
130 +
131 +       rc = ext4_mb_check_and_update_prealloc(sbi, str, cnt, 0);
132 +       if (rc)
133 +               return rc;
134 +
135 +       rc = ext4_mb_check_and_update_prealloc(sbi, str, cnt, 1);
136 +       return rc ? rc : cnt;
137 +}
138 +
139 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
140 +{
141 +       struct ext4_sb_info *sbi = EXT4_SB(m->private);
142 +       int i;
143 +
144 +       for (i = 0; i < EXT4_MAX_PREALLOC_TABLE &&
145 +                       sbi->s_mb_prealloc_table[i] != 0; i++)
146 +               seq_printf(m, "%ld ", sbi->s_mb_prealloc_table[i]);
147 +       seq_printf(m, "\n");
148 +
149 +       return 0;
150 +}
151 +
152 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
153 +{
154 +       return single_open(file, mb_prealloc_table_seq_show, PDE_DATA(inode));
155 +}
156 +
157 +const struct proc_ops ext4_seq_prealloc_table_fops = {
158 +       .proc_open      = mb_prealloc_table_seq_open,
159 +       .proc_read      = seq_read,
160 +       .proc_lseek     = seq_lseek,
161 +       .proc_release   = single_release,
162 +       .proc_write     = ext4_mb_prealloc_table_proc_write,
163 +};
164 +
165  static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
166  {
167         int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
168 @@ -2685,7 +2778,7 @@ static int ext4_groupinfo_create_slab(si
169  int ext4_mb_init(struct super_block *sb)
170  {
171         struct ext4_sb_info *sbi = EXT4_SB(sb);
172 -       unsigned i, j;
173 +       unsigned i, j, k, l;
174         unsigned offset, offset_incr;
175         unsigned max;
176         int ret;
177 @@ -2734,7 +2827,6 @@ int ext4_mb_init(struct super_block *sb)
178         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
179         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
180         sbi->s_mb_stats = MB_DEFAULT_STATS;
181 -       sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
182         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
183         /*
184          * The default group preallocation is 512, which for 4k block
185 @@ -2758,9 +2850,29 @@ int ext4_mb_init(struct super_block *sb)
186          * RAID stripe size so that preallocations don't fragment
187          * the stripes.
188          */
189 -       if (sbi->s_stripe > 1) {
190 -               sbi->s_mb_group_prealloc = roundup(
191 -                       sbi->s_mb_group_prealloc, sbi->s_stripe);
192 +
193 +       /* Allocate table once */
194 +       sbi->s_mb_prealloc_table = kzalloc(
195 +               EXT4_MAX_PREALLOC_TABLE * sizeof(unsigned long), GFP_NOFS);
196 +       if (sbi->s_mb_prealloc_table == NULL) {
197 +               ret = -ENOMEM;
198 +               goto out;
199 +       }
200 +
201 +       if (sbi->s_stripe == 0) {
202 +               for (k = 0, l = 4; k <= 9; ++k, l *= 2)
203 +                       sbi->s_mb_prealloc_table[k] = l;
204 +
205 +               sbi->s_mb_small_req = 256;
206 +               sbi->s_mb_large_req = 1024;
207 +               sbi->s_mb_group_prealloc = 512;
208 +       } else {
209 +               for (k = 0, l = sbi->s_stripe; k <= 2; ++k, l *= 2)
210 +                       sbi->s_mb_prealloc_table[k] = l;
211 +
212 +               sbi->s_mb_small_req = sbi->s_stripe;
213 +               sbi->s_mb_large_req = sbi->s_stripe * 8;
214 +               sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
215         }
216  
217         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
218 @@ -2788,6 +2900,7 @@ out_free_locality_groups:
219         free_percpu(sbi->s_locality_groups);
220         sbi->s_locality_groups = NULL;
221  out:
222 +       kfree(sbi->s_mb_prealloc_table);
223         kfree(sbi->s_mb_offsets);
224         sbi->s_mb_offsets = NULL;
225         kfree(sbi->s_mb_maxs);
226 @@ -3057,7 +3170,6 @@ ext4_mb_mark_diskspace_used(struct ext4_
227         int err, len;
228  
229         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
230 -       BUG_ON(ac->ac_b_ex.fe_len <= 0);
231  
232         sb = ac->ac_sb;
233         sbi = EXT4_SB(sb);
234 @@ -3187,13 +3299,14 @@ ext4_mb_normalize_request(struct ext4_al
235                                 struct ext4_allocation_request *ar)
236  {
237         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
238 -       int bsbits, max;
239 +       int bsbits, i, wind;
240         ext4_lblk_t end;
241 -       loff_t size, start_off;
242 +       loff_t size;
243         loff_t orig_size __maybe_unused;
244         ext4_lblk_t start;
245         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
246         struct ext4_prealloc_space *pa;
247 +       unsigned long value, last_non_zero;
248  
249         /* do normalize only data requests, metadata requests
250            do not need preallocation */
251 @@ -3222,51 +3335,46 @@ ext4_mb_normalize_request(struct ext4_al
252         size = size << bsbits;
253         if (size < i_size_read(ac->ac_inode))
254                 size = i_size_read(ac->ac_inode);
255 -       orig_size = size;
256 +       size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
257 +
258 +       start = wind = 0;
259 +       value = last_non_zero = 0;
260  
261 -       /* max size of free chunks */
262 -       max = 2 << bsbits;
263 +       /* let's choose preallocation window depending on file size */
264 +       for (i = 0; i < EXT4_MAX_PREALLOC_TABLE; i++) {
265 +               value = sbi->s_mb_prealloc_table[i];
266 +               if (value == 0)
267 +                       break;
268 +               else
269 +                       last_non_zero = value;
270  
271 -#define NRL_CHECK_SIZE(req, size, max, chunk_size)     \
272 -               (req <= (size) || max <= (chunk_size))
273 +               if (size <= value) {
274 +                       wind = value;
275 +                       break;
276 +               }
277 +       }
278  
279 -       /* first, try to predict filesize */
280 -       /* XXX: should this table be tunable? */
281 -       start_off = 0;
282 -       if (size <= 16 * 1024) {
283 -               size = 16 * 1024;
284 -       } else if (size <= 32 * 1024) {
285 -               size = 32 * 1024;
286 -       } else if (size <= 64 * 1024) {
287 -               size = 64 * 1024;
288 -       } else if (size <= 128 * 1024) {
289 -               size = 128 * 1024;
290 -       } else if (size <= 256 * 1024) {
291 -               size = 256 * 1024;
292 -       } else if (size <= 512 * 1024) {
293 -               size = 512 * 1024;
294 -       } else if (size <= 1024 * 1024) {
295 -               size = 1024 * 1024;
296 -       } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
297 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
298 -                                               (21 - bsbits)) << 21;
299 -               size = 2 * 1024 * 1024;
300 -       } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
301 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
302 -                                                       (22 - bsbits)) << 22;
303 -               size = 4 * 1024 * 1024;
304 -       } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
305 -                                       (8<<20)>>bsbits, max, 8 * 1024)) {
306 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
307 -                                                       (23 - bsbits)) << 23;
308 -               size = 8 * 1024 * 1024;
309 +       if (wind == 0) {
310 +               if (last_non_zero != 0) {
311 +                       __u64 tstart, tend;
312 +                       /* file is quite large, we now preallocate with
313 +                       * the biggest configured window with regart to
314 +                       * logical offset */
315 +                       wind = last_non_zero;
316 +                       tstart = ac->ac_o_ex.fe_logical;
317 +                       do_div(tstart, wind);
318 +                       start = tstart * wind;
319 +                       tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
320 +                       do_div(tend, wind);
321 +                       tend = tend * wind + wind;
322 +                       size = tend - start;
323 +               }
324         } else {
325 -               start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
326 -               size      = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
327 -                                             ac->ac_o_ex.fe_len) << bsbits;
328 +               size = wind;
329         }
330 -       size = size >> bsbits;
331 -       start = start_off >> bsbits;
332 +
333 +
334 +       orig_size = size;
335  
336         /* don't cover already allocated blocks in selected range */
337         if (ar->pleft && start <= ar->lleft) {
338 @@ -3348,7 +3456,6 @@ ext4_mb_normalize_request(struct ext4_al
339                          (unsigned long) ac->ac_o_ex.fe_logical);
340                 BUG();
341         }
342 -       BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
343  
344         /* now prepare goal request */
345  
346 @@ -4341,11 +4448,19 @@ static void ext4_mb_group_or_file(struct
347  
348         /* don't use group allocation for large files */
349         size = max(size, isize);
350 -       if (size > sbi->s_mb_stream_request) {
351 +       if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
352 +           (size >= sbi->s_mb_large_req)) {
353                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
354                 return;
355         }
356  
357 +       /*
358 +        * request is so large that we don't care about
359 +        * streaming - it overweights any possible seek
360 +        */
361 +       if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
362 +               return;
363 +
364         BUG_ON(ac->ac_lg != NULL);
365         /*
366          * locality group prealloc space are per cpu. The reason for having
367 --- a/fs/ext4/sysfs.c
368 +++ b/fs/ext4/sysfs.c
369 @@ -213,7 +213,8 @@ 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_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
378  EXT4_ATTR(trigger_fs_error, 0200, trigger_test_error);
379 @@ -255,7 +256,8 @@ 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 @@ -510,6 +512,8 @@ int ext4_register_sysfs(struct super_blo
390                                 sb);
391                 proc_create_seq_data("mb_groups", S_IRUGO, sbi->s_proc,
392                                 &ext4_mb_seq_groups_ops, sb);
393 +               proc_create_data("prealloc_table", S_IRUGO, sbi->s_proc,
394 +                               &ext4_seq_prealloc_table_fops, sb);
395         }
396         return 0;
397  }