ldiskfs/kernel_patches/patches/rhel7.4/ext4-prealloc.patch

   1 Index: linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/ext4.h
   2 ===================================================================
   3 --- linux-3.10.0-514.16.1.el7.x86_64.orig/fs/ext4/ext4.h
   4 +++ linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/ext4.h
   5 @@ -1270,11 +1270,14 @@ struct ext4_sb_info {
   6
   7         /* tunables */
   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-3.10.0-514.16.1.el7.x86_64/fs/ext4/mballoc.c
  22 ===================================================================
  23 --- linux-3.10.0-514.16.1.el7.x86_64.orig/fs/ext4/mballoc.c
  24 +++ linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/mballoc.c
  25 @@ -1862,6 +1862,26 @@ int ext4_mb_find_by_goal(struct ext4_all
  26         return 0;
  27  }
  28
  29 +static int ext4_mb_prealloc_table_add(struct ext4_sb_info *sbi, int value)
  30 +{
  31 +       int i;
  32 +
  33 +       if (value > (sbi->s_blocks_per_group - 1 - 1 - sbi->s_itb_per_group))
  34 +               return -1;
  35 +
  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;
  39 +                       return 0;
  40 +               }
  41 +
  42 +               /* they should add values in order */
  43 +               if (value <= sbi->s_mb_prealloc_table[i])
  44 +                       return -1;
  45 +       }
  46 +       return -1;
  47 +}
  48 +
  49  /*
  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 @@ -2301,6 +2321,93 @@ static const struct seq_operations ext4_
  53         .show   = ext4_mb_seq_groups_show,
  54  };
  55
  56 +#define EXT4_MB_PREALLOC_TABLE          "prealloc_table"
  57 +
  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)
  61 +{
  62 +       struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
  63 +       unsigned long value;
  64 +       unsigned long prev = 0;
  65 +       char str[128];
  66 +       char *cur;
  67 +       char *end;
  68 +       unsigned long *new_table;
  69 +       int num = 0;
  70 +       int i = 0;
  71 +
  72 +       if (cnt >= sizeof(str))
  73 +               return -EINVAL;
  74 +       if (copy_from_user(str, buf, cnt))
  75 +               return -EFAULT;
  76 +
  77 +       num = 0;
  78 +       cur = str;
  79 +       end = str + cnt;
  80 +       while (cur < end) {
  81 +               int rc;
  82 +               while ((cur < end) && (*cur == ' '))
  83 +                       cur++;
  84 +               rc = kstrtol(cur, 0, &value);
  85 +               if (rc != 0)
  86 +                       return -EINVAL;
  87 +               if (value == 0)
  88 +                       break;
  89 +               if (value <= prev)
  90 +                       return -EINVAL;
  91 +               prev = value;
  92 +               num++;
  93 +       }
  94 +
  95 +       new_table = kmalloc(num * sizeof(*new_table), GFP_KERNEL);
  96 +       if (new_table == NULL)
  97 +               return -ENOMEM;
  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;
 102 +       cur = str;
 103 +       end = str + cnt;
 104 +       while (cur < end && i < num) {
 105 +               while (cur < end && *cur == ' ')
 106 +                       cur++;
 107 +               value = simple_strtol(cur, &cur, 0);
 108 +               if (ext4_mb_prealloc_table_add(sbi, value) == 0)
 109 +                       ++i;
 110 +       }
 111 +       if (i != num)
 112 +               sbi->s_mb_prealloc_table_size = i;
 113 +
 114 +       return cnt;
 115 +}
 116 +
 117 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
 118 +{
 119 +       struct ext4_sb_info *sbi = EXT4_SB(m->private);
 120 +       int i;
 121 +
 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");
 125 +
 126 +       return 0;
 127 +}
 128 +
 129 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
 130 +{
 131 +       return single_open(file, mb_prealloc_table_seq_show, PDE_DATA(inode));
 132 +}
 133 +
 134 +static const struct file_operations ext4_mb_prealloc_seq_fops = {
 135 +       .owner   = THIS_MODULE,
 136 +       .open    = mb_prealloc_table_seq_open,
 137 +       .read    = seq_read,
 138 +       .llseek  = seq_lseek,
 139 +       .release = single_release,
 140 +       .write   = ext4_mb_prealloc_table_proc_write,
 141 +};
 142 +
 143  static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
 144  {
 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)
 148  {
 149         struct ext4_sb_info *sbi = EXT4_SB(sb);
 150 -       unsigned i, j;
 151 +       unsigned i, j, k, l;
 152         unsigned offset, offset_incr;
 153         unsigned max;
 154         int ret;
 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;
 161         /*
 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
 165          * the stripes.
 166          */
 167 -       if (sbi->s_stripe > 1) {
 168 -               sbi->s_mb_group_prealloc = roundup(
 169 -                       sbi->s_mb_group_prealloc, sbi->s_stripe);
 170 +
 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) {
 176 +                       ret = -ENOMEM;
 177 +                       goto out;
 178 +               }
 179 +               memset(sbi->s_mb_prealloc_table, 0, i);
 180 +
 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;
 184 +                               break;
 185 +                       }
 186 +               }
 187 +
 188 +               sbi->s_mb_small_req = 256;
 189 +               sbi->s_mb_large_req = 1024;
 190 +               sbi->s_mb_group_prealloc = 512;
 191 +       } else {
 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) {
 196 +                       ret = -ENOMEM;
 197 +                       goto out;
 198 +               }
 199 +               memset(sbi->s_mb_prealloc_table, 0, i);
 200 +
 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;
 204 +                               break;
 205 +                       }
 206 +               }
 207 +
 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;
 211         }
 212
 213         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
 214 @@ -2643,9 +2787,13 @@ int ext4_mb_init(struct super_block *sb)
 215         if (ret != 0)
 216                 goto out_free_locality_groups;
 217
 218 -       if (sbi->s_proc)
 219 +       if (sbi->s_proc) {
 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);
 225 +       }
 226
 227         return 0;
 228
 229 @@ -2653,6 +2801,7 @@ out_free_locality_groups:
 230         free_percpu(sbi->s_locality_groups);
 231         sbi->s_locality_groups = NULL;
 232  out:
 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 @@ -2687,8 +2836,10 @@ 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);
 240
 241 -       if (sbi->s_proc)
 242 +       if (sbi->s_proc) {
 243                 remove_proc_entry("mb_groups", sbi->s_proc);
 244 +               remove_proc_entry(EXT4_MB_PREALLOC_TABLE, sbi->s_proc);
 245 +       }
 246
 247         if (sbi->s_group_info) {
 248                 for (i = 0; i < ngroups; i++) {
 249 @@ -3000,9 +3151,9 @@ ext4_mb_normalize_request(struct ext4_al
 250                                 struct ext4_allocation_request *ar)
 251  {
 252         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
 253 -       int bsbits, max;
 254 +       int bsbits, i, wind;
 255         ext4_lblk_t end;
 256 -       loff_t size, start_off;
 257 +       loff_t size;
 258         loff_t orig_size __maybe_unused;
 259         ext4_lblk_t start;
 260         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
 261 @@ -3035,51 +3186,34 @@ 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);
 265 -       orig_size = size;
 266 +       size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
 267
 268 -       /* max size of free chunks */
 269 -       max = 2 << bsbits;
 270 +       start = wind = 0;
 271
 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];
 278 +                       break;
 279 +               }
 280 +       }
 281 +       size = wind;
 282
 283 -       /* first, try to predict filesize */
 284 -       /* XXX: should this table be tunable? */
 285 -       start_off = 0;
 286 -       if (size <= 16 * 1024) {
 287 -               size = 16 * 1024;
 288 -       } else if (size <= 32 * 1024) {
 289 -               size = 32 * 1024;
 290 -       } else if (size <= 64 * 1024) {
 291 -               size = 64 * 1024;
 292 -       } else if (size <= 128 * 1024) {
 293 -               size = 128 * 1024;
 294 -       } else if (size <= 256 * 1024) {
 295 -               size = 256 * 1024;
 296 -       } else if (size <= 512 * 1024) {
 297 -               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;
 313 -       } else {
 314 -               start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
 315 -               size      = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
 316 -                                             ac->ac_o_ex.fe_len) << bsbits;
 317 +       if (wind == 0) {
 318 +               __u64 tstart, tend;
 319 +               /* file is quite large, we now preallocate with
 320 +                * the biggest configured window with regart to
 321 +                * logical offset */
 322 +               wind = sbi->s_mb_prealloc_table[i - 1];
 323 +               tstart = ac->ac_o_ex.fe_logical;
 324 +               do_div(tstart, wind);
 325 +               start = tstart * wind;
 326 +               tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
 327 +               do_div(tend, wind);
 328 +               tend = tend * wind + wind;
 329 +               size = tend - start;
 330         }
 331 -       size = size >> bsbits;
 332 -       start = start_off >> bsbits;
 333 +       orig_size = size;
 334
 335         /* don't cover already allocated blocks in selected range */
 336         if (ar->pleft && start <= ar->lleft) {
 337 @@ -3154,7 +3288,6 @@ ext4_mb_normalize_request(struct ext4_al
 338                          (unsigned long) ac->ac_o_ex.fe_logical);
 339                 BUG();
 340         }
 341 -       BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
 342
 343         /* now prepare goal request */
 344
 345 @@ -4119,11 +4252,19 @@ static void ext4_mb_group_or_file(struct
 346
 347         /* don't use group allocation for large files */
 348         size = max(size, isize);
 349 -       if (size > sbi->s_mb_stream_request) {
 350 +       if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
 351 +           (size >= sbi->s_mb_large_req)) {
 352                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
 353                 return;
 354         }
 355
 356 +       /*
 357 +        * request is so large that we don't care about
 358 +        * streaming - it overweights any possible seek
 359 +        */
 360 +       if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
 361 +               return;
 362 +
 363         BUG_ON(ac->ac_lg != NULL);
 364         /*
 365          * locality group prealloc space are per cpu. The reason for having
 366 Index: linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/super.c
 367 ===================================================================
 368 --- linux-3.10.0-514.16.1.el7.x86_64.orig/fs/ext4/super.c
 369 +++ linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/super.c
 370 @@ -2672,7 +2672,8 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats
 371  EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
 372  EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
 373  EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
 374 -EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
 375 +EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
 376 +EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
 377  EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
 378  EXT4_DEPRECATED_ATTR(max_writeback_mb_bump, 128);
 379  EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
 380 @@ -2698,7 +2699,8 @@ static struct attribute *ext4_attrs[] =
 381         ATTR_LIST(mb_max_to_scan),
 382         ATTR_LIST(mb_min_to_scan),
 383         ATTR_LIST(mb_order2_req),
 384 -       ATTR_LIST(mb_stream_req),
 385 +       ATTR_LIST(mb_small_req),
 386 +       ATTR_LIST(mb_large_req),
 387         ATTR_LIST(mb_group_prealloc),
 388         ATTR_LIST(max_writeback_mb_bump),
 389         ATTR_LIST(extent_max_zeroout_kb),
 390 Index: linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/inode.c
 391 ===================================================================
 392 --- linux-3.10.0-514.16.1.el7.x86_64.orig/fs/ext4/inode.c
 393 +++ linux-3.10.0-514.16.1.el7.x86_64/fs/ext4/inode.c
 394 @@ -2399,6 +2399,9 @@ static int ext4_writepages(struct addres
 395                 ext4_journal_stop(handle);
 396         }
 397
 398 +       if (wbc->nr_to_write < sbi->s_mb_small_req)
 399 +               wbc->nr_to_write = sbi->s_mb_small_req;
 400 +
 401         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 402                 range_whole = 1;
 403