ldiskfs/kernel_patches/patches/sles12sp2/ext4-prealloc.patch

   1 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/ext4.h
   2 ===================================================================
   3 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/ext4.h
   4 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/ext4.h
   5 @@ -1243,11 +1243,14 @@ 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_writeback_mb_bump;
  20         unsigned int s_max_dir_size_kb;
  21 @@ -2243,6 +1243,7 @@ struct ext4_sb_info {
  22  extern void ext4_end_bitmap_read(struct buffer_head *bh, int uptodate);
  23
  24  /* mballoc.c */
  25 +extern const struct file_operations ext4_seq_prealloc_table_fops;
  26  extern const struct file_operations ext4_seq_mb_groups_fops;
  27  extern long ext4_mb_stats;
  28  extern long ext4_mb_max_to_scan;
  29 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.c
  30 ===================================================================
  31 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/mballoc.c
  32 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.c
  33 @@ -1828,6 +1828,26 @@ int ext4_mb_find_by_goal(struct ext4_all
  34         return 0;
  35  }
  36
  37 +static int ext4_mb_prealloc_table_add(struct ext4_sb_info *sbi, int value)
  38 +{
  39 +       int i;
  40 +
  41 +       if (value > (sbi->s_blocks_per_group - 1 - 1 - sbi->s_itb_per_group))
  42 +               return -1;
  43 +
  44 +       for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
  45 +               if (sbi->s_mb_prealloc_table[i] == 0) {
  46 +                       sbi->s_mb_prealloc_table[i] = value;
  47 +                       return 0;
  48 +               }
  49 +
  50 +               /* they should add values in order */
  51 +               if (value <= sbi->s_mb_prealloc_table[i])
  52 +                       return -1;
  53 +       }
  54 +       return -1;
  55 +}
  56 +
  57  /*
  58   * The routine scans buddy structures (not bitmap!) from given order
  59   * to max order and tries to find big enough chunk to satisfy the req
  60 @@ -2263,6 +2282,88 @@ static const struct seq_operations ext4_
  61         .show   = ext4_mb_seq_groups_show,
  62  };
  63
  64 +static ssize_t ext4_mb_prealloc_table_proc_write(struct file *file,
  65 +                                            const char __user *buf,
  66 +                                            size_t cnt, loff_t *pos)
  67 +{
  68 +       struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
  69 +       unsigned long value;
  70 +       unsigned long prev = 0;
  71 +       char str[128];
  72 +       char *cur;
  73 +       char *end;
  74 +       unsigned long *new_table;
  75 +       int num = 0;
  76 +       int i = 0;
  77 +
  78 +       if (cnt >= sizeof(str))
  79 +               return -EINVAL;
  80 +       if (copy_from_user(str, buf, cnt))
  81 +               return -EFAULT;
  82 +
  83 +       num = 0;
  84 +       cur = str;
  85 +       end = str + cnt;
  86 +       while (cur < end) {
  87 +               while ((cur < end) && (*cur == ' '))
  88 +                       cur++;
  89 +               value = simple_strtol(cur, &cur, 0);
  90 +               if (value == 0)
  91 +                       break;
  92 +               if (value <= prev)
  93 +                       return -EINVAL;
  94 +               prev = value;
  95 +               num++;
  96 +       }
  97 +
  98 +       new_table = kmalloc(num * sizeof(*new_table), GFP_KERNEL);
  99 +       if (new_table == NULL)
 100 +               return -ENOMEM;
 101 +       kfree(sbi->s_mb_prealloc_table);
 102 +       memset(new_table, 0, num * sizeof(*new_table));
 103 +       sbi->s_mb_prealloc_table = new_table;
 104 +       sbi->s_mb_prealloc_table_size = num;
 105 +       cur = str;
 106 +       end = str + cnt;
 107 +       while (cur < end && i < num) {
 108 +               while (cur < end && *cur == ' ')
 109 +                       cur++;
 110 +               value = simple_strtol(cur, &cur, 0);
 111 +               if (ext4_mb_prealloc_table_add(sbi, value) == 0)
 112 +                       ++i;
 113 +       }
 114 +       if (i != num)
 115 +               sbi->s_mb_prealloc_table_size = i;
 116 +
 117 +       return cnt;
 118 +}
 119 +
 120 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
 121 +{
 122 +       struct ext4_sb_info *sbi = EXT4_SB(m->private);
 123 +       int i;
 124 +
 125 +       for (i = 0; i < sbi->s_mb_prealloc_table_size; i++)
 126 +               seq_printf(m, "%ld ", sbi->s_mb_prealloc_table[i]);
 127 +       seq_printf(m, "\n");
 128 +
 129 +       return 0;
 130 +}
 131 +
 132 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
 133 +{
 134 +       return single_open(file, mb_prealloc_table_seq_show, PDE_DATA(inode));
 135 +}
 136 +
 137 +const struct file_operations ext4_seq_prealloc_table_fops = {
 138 +       .owner   = THIS_MODULE,
 139 +       .open    = mb_prealloc_table_seq_open,
 140 +       .read    = seq_read,
 141 +       .llseek  = seq_lseek,
 142 +       .release = single_release,
 143 +       .write   = ext4_mb_prealloc_table_proc_write,
 144 +};
 145 +
 146  static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
 147  {
 148         struct super_block *sb = PDE_DATA(inode);
 149 @@ -2550,7 +2657,7 @@ static int ext4_groupinfo_create_slab(si
 150  int ext4_mb_init(struct super_block *sb)
 151  {
 152         struct ext4_sb_info *sbi = EXT4_SB(sb);
 153 -       unsigned i, j;
 154 +       unsigned i, j, k, l;
 155         unsigned offset, offset_incr;
 156         unsigned max;
 157         int ret;
 158 @@ -2595,7 +2702,6 @@ int ext4_mb_init(struct super_block *sb)
 159         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
 160         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
 161         sbi->s_mb_stats = MB_DEFAULT_STATS;
 162 -       sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
 163         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
 164         /*
 165          * The default group preallocation is 512, which for 4k block
 166 @@ -2619,9 +2725,47 @@ int ext4_mb_init(struct super_block *sb)
 167          * RAID stripe size so that preallocations don't fragment
 168          * the stripes.
 169          */
 170 -       if (sbi->s_stripe > 1) {
 171 -               sbi->s_mb_group_prealloc = roundup(
 172 -                       sbi->s_mb_group_prealloc, sbi->s_stripe);
 173 +
 174 +       if (sbi->s_stripe == 0) {
 175 +               sbi->s_mb_prealloc_table_size = 10;
 176 +               i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
 177 +               sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
 178 +               if (sbi->s_mb_prealloc_table == NULL) {
 179 +                       ret = -ENOMEM;
 180 +                       goto out;
 181 +               }
 182 +               memset(sbi->s_mb_prealloc_table, 0, i);
 183 +
 184 +               for (k = 0, l = 4; k <= 9; ++k, l *= 2) {
 185 +                       if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
 186 +                               sbi->s_mb_prealloc_table_size = k;
 187 +                               break;
 188 +                       }
 189 +               }
 190 +
 191 +               sbi->s_mb_small_req = 256;
 192 +               sbi->s_mb_large_req = 1024;
 193 +               sbi->s_mb_group_prealloc = 512;
 194 +       } else {
 195 +               sbi->s_mb_prealloc_table_size = 3;
 196 +               i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
 197 +               sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
 198 +               if (sbi->s_mb_prealloc_table == NULL) {
 199 +                       ret = -ENOMEM;
 200 +                       goto out;
 201 +               }
 202 +               memset(sbi->s_mb_prealloc_table, 0, i);
 203 +
 204 +               for (k = 0, l = sbi->s_stripe; k <= 2; ++k, l *= 2) {
 205 +                       if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
 206 +                               sbi->s_mb_prealloc_table_size = k;
 207 +                               break;
 208 +                       }
 209 +               }
 210 +
 211 +               sbi->s_mb_small_req = sbi->s_stripe;
 212 +               sbi->s_mb_large_req = sbi->s_stripe * 8;
 213 +               sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
 214         }
 215
 216         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
 217 @@ -2963,9 +3107,9 @@ ext4_mb_normalize_request(struct ext4_al
 218                                 struct ext4_allocation_request *ar)
 219  {
 220         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
 221 -       int bsbits, max;
 222 +       int bsbits, i, wind;
 223         ext4_lblk_t end;
 224 -       loff_t size, start_off;
 225 +       loff_t size;
 226         loff_t orig_size __maybe_unused;
 227         ext4_lblk_t start;
 228         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
 229 @@ -2998,51 +3142,34 @@ ext4_mb_normalize_request(struct ext4_al
 230         size = size << bsbits;
 231         if (size < i_size_read(ac->ac_inode))
 232                 size = i_size_read(ac->ac_inode);
 233 -       orig_size = size;
 234 +       size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
 235
 236 -       /* max size of free chunks */
 237 -       max = 2 << bsbits;
 238 +       start = wind = 0;
 239
 240 -#define NRL_CHECK_SIZE(req, size, max, chunk_size)     \
 241 -               (req <= (size) || max <= (chunk_size))
 242 +       /* let's choose preallocation window depending on file size */
 243 +       for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
 244 +               if (size <= sbi->s_mb_prealloc_table[i]) {
 245 +                       wind = sbi->s_mb_prealloc_table[i];
 246 +                       break;
 247 +               }
 248 +       }
 249 +       size = wind;
 250
 251 -       /* first, try to predict filesize */
 252 -       /* XXX: should this table be tunable? */
 253 -       start_off = 0;
 254 -       if (size <= 16 * 1024) {
 255 -               size = 16 * 1024;
 256 -       } else if (size <= 32 * 1024) {
 257 -               size = 32 * 1024;
 258 -       } else if (size <= 64 * 1024) {
 259 -               size = 64 * 1024;
 260 -       } else if (size <= 128 * 1024) {
 261 -               size = 128 * 1024;
 262 -       } else if (size <= 256 * 1024) {
 263 -               size = 256 * 1024;
 264 -       } else if (size <= 512 * 1024) {
 265 -               size = 512 * 1024;
 266 -       } else if (size <= 1024 * 1024) {
 267 -               size = 1024 * 1024;
 268 -       } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
 269 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
 270 -                                               (21 - bsbits)) << 21;
 271 -               size = 2 * 1024 * 1024;
 272 -       } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
 273 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
 274 -                                                       (22 - bsbits)) << 22;
 275 -               size = 4 * 1024 * 1024;
 276 -       } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
 277 -                                       (8<<20)>>bsbits, max, 8 * 1024)) {
 278 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
 279 -                                                       (23 - bsbits)) << 23;
 280 -               size = 8 * 1024 * 1024;
 281 -       } else {
 282 -               start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
 283 -               size      = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
 284 -                                             ac->ac_o_ex.fe_len) << bsbits;
 285 +       if (wind == 0) {
 286 +               __u64 tstart, tend;
 287 +               /* file is quite large, we now preallocate with
 288 +                * the biggest configured window with regart to
 289 +                * logical offset */
 290 +               wind = sbi->s_mb_prealloc_table[i - 1];
 291 +               tstart = ac->ac_o_ex.fe_logical;
 292 +               do_div(tstart, wind);
 293 +               start = tstart * wind;
 294 +               tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
 295 +               do_div(tend, wind);
 296 +               tend = tend * wind + wind;
 297 +               size = tend - start;
 298         }
 299 -       size = size >> bsbits;
 300 -       start = start_off >> bsbits;
 301 +       orig_size = size;
 302
 303         /* don't cover already allocated blocks in selected range */
 304         if (ar->pleft && start <= ar->lleft) {
 305 @@ -3117,7 +3245,6 @@ ext4_mb_normalize_request(struct ext4_al
 306         BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
 307                         start > ac->ac_o_ex.fe_logical);
 308         }
 309 -       BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
 310
 311         /* now prepare goal request */
 312
 313 @@ -4056,11 +4183,19 @@ static void ext4_mb_group_or_file(struct
 314
 315         /* don't use group allocation for large files */
 316         size = max(size, isize);
 317 -       if (size > sbi->s_mb_stream_request) {
 318 +       if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
 319 +           (size >= sbi->s_mb_large_req)) {
 320                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
 321                 return;
 322         }
 323
 324 +       /*
 325 +        * request is so large that we don't care about
 326 +        * streaming - it overweights any possible seek
 327 +        */
 328 +       if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
 329 +               return;
 330 +
 331         BUG_ON(ac->ac_lg != NULL);
 332         /*
 333          * locality group prealloc space are per cpu. The reason for having
 334 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 335 ===================================================================
 336 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/sysfs.c
 337 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 338 @@ -2555,7 +2555,8 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats
 339  EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
 340  EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
 341  EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
 342 -EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
 343 +EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
 344 +EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
 345  EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
 346  EXT4_RW_ATTR_SBI_UI(max_writeback_mb_bump, s_max_writeback_mb_bump);
 347  EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
 348 @@ -2578,7 +2579,8 @@ static struct attribute *ext4_attrs[] =
 349         ATTR_LIST(mb_max_to_scan),
 350         ATTR_LIST(mb_min_to_scan),
 351         ATTR_LIST(mb_order2_req),
 352 -       ATTR_LIST(mb_stream_req),
 353 +       ATTR_LIST(mb_small_req),
 354 +       ATTR_LIST(mb_large_req),
 355         ATTR_LIST(mb_group_prealloc),
 356         ATTR_LIST(max_writeback_mb_bump),
 357         ATTR_LIST(extent_max_zeroout_kb),
 358 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/inode.c
 359 ===================================================================
 360 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/inode.c
 361 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/inode.c
 362 @@ -2476,6 +2476,9 @@ static int ext4_da_writepages(struct add
 363         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
 364                 return -EROFS;
 365
 366 +       if (wbc->nr_to_write < sbi->s_mb_small_req)
 367 +               wbc->nr_to_write = sbi->s_mb_small_req;
 368 +
 369         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 370                 range_whole = 1;
 371
 372 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 373 ===================================================================
 374 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/sysfs.c
 375 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 376 @@ -1243,6 +1243,7 @@ struct ext4_sb_info {
 377         PROC_FILE_LIST(options),
 378         PROC_FILE_LIST(es_shrinker_info),
 379         PROC_FILE_LIST(mb_groups),
 380 +       PROC_FILE_LIST(prealloc_table),
 381         { NULL, NULL },
 382  };
 383