ldiskfs/kernel_patches/patches/suse15/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 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/ext4.h
  15 ===================================================================
  16 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/ext4.h
  17 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/ext4.h
  18 @@ -1243,11 +1243,14 @@ struct ext4_sb_info {
  19
  20         /* tunables */
  21         unsigned long s_stripe;
  22 -       unsigned int s_mb_stream_request;
  23 +       unsigned long s_mb_small_req;
  24 +       unsigned long s_mb_large_req;
  25         unsigned int s_mb_max_to_scan;
  26         unsigned int s_mb_min_to_scan;
  27         unsigned int s_mb_stats;
  28         unsigned int s_mb_order2_reqs;
  29 +       unsigned long *s_mb_prealloc_table;
  30 +       unsigned long s_mb_prealloc_table_size;
  31         unsigned int s_mb_group_prealloc;
  32         unsigned int s_max_writeback_mb_bump;
  33         unsigned int s_max_dir_size_kb;
  34 @@ -2243,6 +1243,7 @@ struct ext4_sb_info {
  35  extern void ext4_end_bitmap_read(struct buffer_head *bh, int uptodate);
  36
  37  /* mballoc.c */
  38 +extern const struct file_operations ext4_seq_prealloc_table_fops;
  39  extern const struct file_operations ext4_seq_mb_groups_fops;
  40  extern long ext4_mb_stats;
  41  extern long ext4_mb_max_to_scan;
  42 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.c
  43 ===================================================================
  44 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/mballoc.c
  45 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.c
  46 @@ -1828,6 +1828,26 @@ int ext4_mb_find_by_goal(struct ext4_all
  47         return 0;
  48  }
  49
  50 +static int ext4_mb_prealloc_table_add(struct ext4_sb_info *sbi, int value)
  51 +{
  52 +       int i;
  53 +
  54 +       if (value > (sbi->s_blocks_per_group - 1 - 1 - sbi->s_itb_per_group))
  55 +               return -1;
  56 +
  57 +       for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
  58 +               if (sbi->s_mb_prealloc_table[i] == 0) {
  59 +                       sbi->s_mb_prealloc_table[i] = value;
  60 +                       return 0;
  61 +               }
  62 +
  63 +               /* they should add values in order */
  64 +               if (value <= sbi->s_mb_prealloc_table[i])
  65 +                       return -1;
  66 +       }
  67 +       return -1;
  68 +}
  69 +
  70  /*
  71   * The routine scans buddy structures (not bitmap!) from given order
  72   * to max order and tries to find big enough chunk to satisfy the req
  73 @@ -2263,6 +2282,88 @@ static const struct seq_operations ext4_
  74         .show   = ext4_mb_seq_groups_show,
  75  };
  76
  77 +static ssize_t ext4_mb_prealloc_table_proc_write(struct file *file,
  78 +                                            const char __user *buf,
  79 +                                            size_t cnt, loff_t *pos)
  80 +{
  81 +       struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
  82 +       unsigned long value;
  83 +       unsigned long prev = 0;
  84 +       char str[128];
  85 +       char *cur;
  86 +       char *end;
  87 +       unsigned long *new_table;
  88 +       int num = 0;
  89 +       int i = 0;
  90 +
  91 +       if (cnt >= sizeof(str))
  92 +               return -EINVAL;
  93 +       if (copy_from_user(str, buf, cnt))
  94 +               return -EFAULT;
  95 +
  96 +       num = 0;
  97 +       cur = str;
  98 +       end = str + cnt;
  99 +       while (cur < end) {
 100 +               while ((cur < end) && (*cur == ' '))
 101 +                       cur++;
 102 +               value = simple_strtol(cur, &cur, 0);
 103 +               if (value == 0)
 104 +                       break;
 105 +               if (value <= prev)
 106 +                       return -EINVAL;
 107 +               prev = value;
 108 +               num++;
 109 +       }
 110 +
 111 +       new_table = kmalloc(num * sizeof(*new_table), GFP_KERNEL);
 112 +       if (new_table == NULL)
 113 +               return -ENOMEM;
 114 +       kfree(sbi->s_mb_prealloc_table);
 115 +       memset(new_table, 0, num * sizeof(*new_table));
 116 +       sbi->s_mb_prealloc_table = new_table;
 117 +       sbi->s_mb_prealloc_table_size = num;
 118 +       cur = str;
 119 +       end = str + cnt;
 120 +       while (cur < end && i < num) {
 121 +               while (cur < end && *cur == ' ')
 122 +                       cur++;
 123 +               value = simple_strtol(cur, &cur, 0);
 124 +               if (ext4_mb_prealloc_table_add(sbi, value) == 0)
 125 +                       ++i;
 126 +       }
 127 +       if (i != num)
 128 +               sbi->s_mb_prealloc_table_size = i;
 129 +
 130 +       return cnt;
 131 +}
 132 +
 133 +static int mb_prealloc_table_seq_show(struct seq_file *m, void *v)
 134 +{
 135 +       struct ext4_sb_info *sbi = EXT4_SB(m->private);
 136 +       int i;
 137 +
 138 +       for (i = 0; i < sbi->s_mb_prealloc_table_size; i++)
 139 +               seq_printf(m, "%ld ", sbi->s_mb_prealloc_table[i]);
 140 +       seq_printf(m, "\n");
 141 +
 142 +       return 0;
 143 +}
 144 +
 145 +static int mb_prealloc_table_seq_open(struct inode *inode, struct file *file)
 146 +{
 147 +       return single_open(file, mb_prealloc_table_seq_show, PDE_DATA(inode));
 148 +}
 149 +
 150 +const struct file_operations ext4_seq_prealloc_table_fops = {
 151 +       .owner   = THIS_MODULE,
 152 +       .open    = mb_prealloc_table_seq_open,
 153 +       .read    = seq_read,
 154 +       .llseek  = seq_lseek,
 155 +       .release = single_release,
 156 +       .write   = ext4_mb_prealloc_table_proc_write,
 157 +};
 158 +
 159  static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
 160  {
 161         struct super_block *sb = PDE_DATA(inode);
 162 @@ -2550,7 +2657,7 @@ static int ext4_groupinfo_create_slab(si
 163  int ext4_mb_init(struct super_block *sb)
 164  {
 165         struct ext4_sb_info *sbi = EXT4_SB(sb);
 166 -       unsigned i, j;
 167 +       unsigned i, j, k, l;
 168         unsigned offset, offset_incr;
 169         unsigned max;
 170         int ret;
 171 @@ -2595,7 +2702,6 @@ int ext4_mb_init(struct super_block *sb)
 172         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
 173         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
 174         sbi->s_mb_stats = MB_DEFAULT_STATS;
 175 -       sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
 176         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
 177         /*
 178          * The default group preallocation is 512, which for 4k block
 179 @@ -2619,9 +2725,47 @@ int ext4_mb_init(struct super_block *sb)
 180          * RAID stripe size so that preallocations don't fragment
 181          * the stripes.
 182          */
 183 -       if (sbi->s_stripe > 1) {
 184 -               sbi->s_mb_group_prealloc = roundup(
 185 -                       sbi->s_mb_group_prealloc, sbi->s_stripe);
 186 +
 187 +       if (sbi->s_stripe == 0) {
 188 +               sbi->s_mb_prealloc_table_size = 10;
 189 +               i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
 190 +               sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
 191 +               if (sbi->s_mb_prealloc_table == NULL) {
 192 +                       ret = -ENOMEM;
 193 +                       goto out;
 194 +               }
 195 +               memset(sbi->s_mb_prealloc_table, 0, i);
 196 +
 197 +               for (k = 0, l = 4; k <= 9; ++k, l *= 2) {
 198 +                       if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
 199 +                               sbi->s_mb_prealloc_table_size = k;
 200 +                               break;
 201 +                       }
 202 +               }
 203 +
 204 +               sbi->s_mb_small_req = 256;
 205 +               sbi->s_mb_large_req = 1024;
 206 +               sbi->s_mb_group_prealloc = 512;
 207 +       } else {
 208 +               sbi->s_mb_prealloc_table_size = 3;
 209 +               i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
 210 +               sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
 211 +               if (sbi->s_mb_prealloc_table == NULL) {
 212 +                       ret = -ENOMEM;
 213 +                       goto out;
 214 +               }
 215 +               memset(sbi->s_mb_prealloc_table, 0, i);
 216 +
 217 +               for (k = 0, l = sbi->s_stripe; k <= 2; ++k, l *= 2) {
 218 +                       if (ext4_mb_prealloc_table_add(sbi, l) < 0) {
 219 +                               sbi->s_mb_prealloc_table_size = k;
 220 +                               break;
 221 +                       }
 222 +               }
 223 +
 224 +               sbi->s_mb_small_req = sbi->s_stripe;
 225 +               sbi->s_mb_large_req = sbi->s_stripe * 8;
 226 +               sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
 227         }
 228
 229         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
 230 @@ -2800,6 +2840,7 @@ int ext4_mb_release(struct super_block *sb)
 231                 kvfree(group_info);
 232                 rcu_read_unlock();
 233         }
 234 +       kfree(sbi->s_mb_prealloc_table);
 235         kfree(sbi->s_mb_offsets);
 236         kfree(sbi->s_mb_maxs);
 237         iput(sbi->s_buddy_cache);
 238 @@ -2963,9 +3107,9 @@ ext4_mb_normalize_request(struct ext4_al
 239                                 struct ext4_allocation_request *ar)
 240  {
 241         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
 242 -       int bsbits, max;
 243 +       int bsbits, i, wind;
 244         ext4_lblk_t end;
 245 -       loff_t size, start_off;
 246 +       loff_t size;
 247         loff_t orig_size __maybe_unused;
 248         ext4_lblk_t start;
 249         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
 250 @@ -2998,51 +3142,34 @@ ext4_mb_normalize_request(struct ext4_al
 251         size = size << bsbits;
 252         if (size < i_size_read(ac->ac_inode))
 253                 size = i_size_read(ac->ac_inode);
 254 -       orig_size = size;
 255 +       size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
 256
 257 -       /* max size of free chunks */
 258 -       max = 2 << bsbits;
 259 +       start = wind = 0;
 260
 261 -#define NRL_CHECK_SIZE(req, size, max, chunk_size)     \
 262 -               (req <= (size) || max <= (chunk_size))
 263 +       /* let's choose preallocation window depending on file size */
 264 +       for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
 265 +               if (size <= sbi->s_mb_prealloc_table[i]) {
 266 +                       wind = sbi->s_mb_prealloc_table[i];
 267 +                       break;
 268 +               }
 269 +       }
 270 +       size = wind;
 271
 272 -       /* first, try to predict filesize */
 273 -       /* XXX: should this table be tunable? */
 274 -       start_off = 0;
 275 -       if (size <= 16 * 1024) {
 276 -               size = 16 * 1024;
 277 -       } else if (size <= 32 * 1024) {
 278 -               size = 32 * 1024;
 279 -       } else if (size <= 64 * 1024) {
 280 -               size = 64 * 1024;
 281 -       } else if (size <= 128 * 1024) {
 282 -               size = 128 * 1024;
 283 -       } else if (size <= 256 * 1024) {
 284 -               size = 256 * 1024;
 285 -       } else if (size <= 512 * 1024) {
 286 -               size = 512 * 1024;
 287 -       } else if (size <= 1024 * 1024) {
 288 -               size = 1024 * 1024;
 289 -       } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
 290 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
 291 -                                               (21 - bsbits)) << 21;
 292 -               size = 2 * 1024 * 1024;
 293 -       } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
 294 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
 295 -                                                       (22 - bsbits)) << 22;
 296 -               size = 4 * 1024 * 1024;
 297 -       } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
 298 -                                       (8<<20)>>bsbits, max, 8 * 1024)) {
 299 -               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
 300 -                                                       (23 - bsbits)) << 23;
 301 -               size = 8 * 1024 * 1024;
 302 -       } else {
 303 -               start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
 304 -               size      = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
 305 -                                             ac->ac_o_ex.fe_len) << bsbits;
 306 +       if (wind == 0) {
 307 +               __u64 tstart, tend;
 308 +               /* file is quite large, we now preallocate with
 309 +                * the biggest configured window with regart to
 310 +                * logical offset */
 311 +               wind = sbi->s_mb_prealloc_table[i - 1];
 312 +               tstart = ac->ac_o_ex.fe_logical;
 313 +               do_div(tstart, wind);
 314 +               start = tstart * wind;
 315 +               tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
 316 +               do_div(tend, wind);
 317 +               tend = tend * wind + wind;
 318 +               size = tend - start;
 319         }
 320 -       size = size >> bsbits;
 321 -       start = start_off >> bsbits;
 322 +       orig_size = size;
 323
 324         /* don't cover already allocated blocks in selected range */
 325         if (ar->pleft && start <= ar->lleft) {
 326 @@ -3117,7 +3245,6 @@ ext4_mb_normalize_request(struct ext4_al
 327         BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
 328                         start > ac->ac_o_ex.fe_logical);
 329         }
 330 -       BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
 331
 332         /* now prepare goal request */
 333
 334 @@ -4056,11 +4183,19 @@ static void ext4_mb_group_or_file(struct
 335
 336         /* don't use group allocation for large files */
 337         size = max(size, isize);
 338 -       if (size > sbi->s_mb_stream_request) {
 339 +       if ((ac->ac_o_ex.fe_len >= sbi->s_mb_small_req) ||
 340 +           (size >= sbi->s_mb_large_req)) {
 341                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
 342                 return;
 343         }
 344
 345 +       /*
 346 +        * request is so large that we don't care about
 347 +        * streaming - it overweights any possible seek
 348 +        */
 349 +       if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
 350 +               return;
 351 +
 352         BUG_ON(ac->ac_lg != NULL);
 353         /*
 354          * locality group prealloc space are per cpu. The reason for having
 355 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 356 ===================================================================
 357 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/sysfs.c
 358 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 359 @@ -2555,7 +2555,8 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats
 360  EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
 361  EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
 362  EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
 363 -EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
 364 +EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
 365 +EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
 366  EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
 367  EXT4_RW_ATTR_SBI_UI(max_writeback_mb_bump, s_max_writeback_mb_bump);
 368  EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
 369 @@ -2578,7 +2579,8 @@ static struct attribute *ext4_attrs[] =
 370         ATTR_LIST(mb_max_to_scan),
 371         ATTR_LIST(mb_min_to_scan),
 372         ATTR_LIST(mb_order2_req),
 373 -       ATTR_LIST(mb_stream_req),
 374 +       ATTR_LIST(mb_small_req),
 375 +       ATTR_LIST(mb_large_req),
 376         ATTR_LIST(mb_group_prealloc),
 377         ATTR_LIST(max_writeback_mb_bump),
 378         ATTR_LIST(extent_max_zeroout_kb),
 379 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/inode.c
 380 ===================================================================
 381 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/inode.c
 382 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/inode.c
 383 @@ -2476,6 +2476,9 @@ static int ext4_da_writepages(struct add
 384         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
 385                 return -EROFS;
 386
 387 +       if (wbc->nr_to_write < sbi->s_mb_small_req)
 388 +               wbc->nr_to_write = sbi->s_mb_small_req;
 389 +
 390         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 391                 range_whole = 1;
 392
 393 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 394 ===================================================================
 395 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/sysfs.c
 396 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/sysfs.c
 397 @@ -1243,6 +1243,7 @@ struct ext4_sb_info {
 398         PROC_FILE_LIST(options),
 399         PROC_FILE_LIST(es_shrinker_info),
 400         PROC_FILE_LIST(mb_groups),
 401 +       PROC_FILE_LIST(prealloc_table),
 402         { NULL, NULL },
 403  };
 404